text
stringlengths 0
316k
| year
stringclasses 50
values | No
stringclasses 911
values |
|---|---|---|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3298–3309 August 1–6, 2021. ©2021 Association for Computational Linguistics 3298 Evidence-based Factual Error Correction James Thorne Department of Computer Science University of Cambridge [email protected] Andreas Vlachos Department of Computer Science University of Cambridge [email protected] Abstract This paper introduces the task of factual error correction: performing edits to a claim so that the generated rewrite is better supported by evidence. This extends the well-studied task of fact verification by providing a mechanism to correct written texts that are refuted or only partially supported by evidence. We demonstrate that it is feasible to train factual error correction systems from existing fact checking datasets which only contain labeled claims accompanied by evidence, but not the correction. We achieve this by employing a two-stage distant supervision approach that incorporates evidence into masked claims when generating corrections. Our approach, based on the T5 transformer and using retrieved evidence, achieved better results than existing work which used a pointer copy network and gold evidence, producing accurate factual error corrections for 5x more instances in human evaluation and a .125 increase in SARI score. The evaluation is conducted on a dataset of 65,000 instances based on a recent fact verification shared task and we release it to enable further work on the task.1 1 Introduction Fact verification is the task of predicting whether claims are true or false using evidence. With the availability of a number of resources (Wang, 2017; Karadzhov et al., 2017; Thorne et al., 2018; Augenstein et al., 2019; Wadden et al., 2020), the task has attracted significant attention and spawned the development of new models, architectures and approaches. With potentially sensitive applications, recent works have focused on building explainable variants of fact checking (Atanasova et al., 2020; Stammbach and Ash, 2020; Kotonya and Toni, 2020). Exposing the evidence source and 1https://github.com/j6mes/ 2021-acl-factual-error-correction System Outputs Brown recluse spiders do not bite The brown recluse spider's bite sometimes requires medical attention. Input Claim Similar to other recluse spider bites, their bite sometimes requires medical attention. Retrieved Evidence Fact Verification Wikipedia REFUTED Error Correction Information Retrieval Figure 1: Factual Error Correction uses evidence to make corrections to claims, in contrast to fact verification, which instead classifies the veracity of the claim. decision making process may help the reader uncover subtle issues that cause automated systems to fail. Additionally, using such evidence to continuously update news articles as facts change forms part of the vision outlined by Cohen et al. (2011) for automated newsrooms. In this paper, we propose Factual Error Correction, as an explainable alternative for fact verification. Rather than merely assigning a truth label, possibly accompanied by evidence, our goal is to rewrite claims so that they are better supported by the retrieved evidence. For example, in Figure 1, a claim that would be REFUTED by the evidence using a fact verification system is rewritten so that it becomes supported by evidence retrieved from Wikipedia. This work extends fact guided sentence modification (Shah et al., 2020), which uses short factoid claims to introduce changes to Wikipedia passages. However, they assume that the claim and 3299 Wikipedia text are always incongruous and require a meaning-altering change, our proposal makes no assumptions over the veracity, and is applicable to claims both supported and refuted by evidence. Additionally, we incorporate a retrieval component to select evidence for a given claim from a corpus (in our case, Wikipedia) rather than requiring gold standard evidence to be explicitly provided. A challenge for factual error correction is the lack of datasets consisting of claims paired with their corrections. However, with recent developments in fact checking, there is an abundance of new datasets consisting of claims paired with evidence. To address this data scarcity, we make use of distant supervision to incorporate retrieved evidence into generating the corrections. We release a dataset of 65,000 claims, containing the intermediate annotations from FEVER (Thorne et al., 2018). These consist of factoid sentences that were used to construct the supported and refuted claims in the dataset, and use these as reference targets for automated evaluation.We further verify the findings through a final round of annotation using human raters. Our evaluation finds high correlation between manual scores and the SARI metric (Xu et al., 2016) and our best performing distantlysupervised system generated corrected claims for 24% of instances when using retrieved evidence, with a SARI Final score of .419. A fully-supervised system with gold evidence generated corrections for 69% of instances, indicating plenty of opportunities for future work to extend our contributions. 2 Related Work A number of related works offer methods to make corrections to sentences. However, their use of external information differs. This can be placed on a continuum from only using the knowledge captured during language model pre-training, to conditioning generation based on a context sentence. We briefly outline key methods and approaches below. Grammatical Error Correction (GEC) (Knight and Chander, 1994; Han et al., 2010; Ng et al., 2014) is the task of making meaning-preserving changes to sentences such that grammatical errors made by language learners are removed. No external information is required as the sentence is undergoing a surface-level transformation where the (intended) semantic content of the sentence should remain unchanged. In contrast, the semantic content of sentences undergoing factual error correction will be altered, if needed, to better align the meaning with ground truth evidence. Shah et al. (2020) make meaningaltering updates to sentences in Wikipedia in a two step process that does not require reference corrections in training: salient tokens are masked and a corrector conditionally replaces the masks with ground truth evidence. In this approach, token salience is predicted by querying a model that is trained to perform fact verification for a claim against evidence. Cao et al. (2020) generate corrections as a post-editing step for outputs from abstractive summarization so that they are consistent with the source text. Their approach uses a sequence-tosequence model trained to restore artificially generated corruptions of a reference summary. One potential way to introduce knowledge is to use information stored in the parameters of largescale pre-trained language models (Petroni et al., 2019). The language model can be used recover tokens responsible for causing factual errors that are masked out as a variant of cloze-style evaluation (Taylor, 1953). While such approaches have been employed for fact verification (Lee et al., 2020), these approaches share the following limitations. Without explicit control (Nie et al., 2019), the most likely token when decoded may not be factually accurate, or supported by the retrieved evidence, commonly referred to as a hallucination (Rohrbach et al., 2018; Zhou et al., 2020). Furthermore, even if the information stored within language model parameters could be reliably retrieved for factual error correction, facts change over time and the need to obtain information from up-to-date sources becomes greater as the state of the world diverges from the information captured within the model parameters. Recent language models augmented with a retrieval component such as REALM (Guu et al., 2020) and RAG (Lewis et al., 2020) could be applied, however, task-specific fine-tuning would still be required to condition the generation based on the factual error to mitigate hallucination. 3 Task Definition Training Let a claim c be the input sentence undergoing correction to yield c′. The correction requires incorporating knowledge from retrieved evidence E(c) such that c′ is supported by this evidence, E(c) ⊨ c′. Factual error correction is subject to the following 3 requirements: 3300 John Goodman had the lead role in The Babe. John Goodman had the lead role in # #. John Goodman had the lead role in The Babe. Claim Masked Claim Supervision Target Correction Wiki Page John Goodman. Context His other film performances include lead roles in The Babe (1992) and The Flintstones (1992) Masker Corrector Evidence Training John Goodman acted in Star Wars John Goodman acted in # # John Goodman acted in The Babe Claim Masked Claim Correction Wiki Page John Goodman Context His other film performances include lead roles in The Babe Masker Corrector Evidence Testing Page Star Wars Context Star Wars is an American epic space opera media franchise... Figure 2: The corrector is trained to reconstruct masked claims, conditioned on retrieved evidence, indicated by the dashed arrow. At test time, the corrector is able to incorporate new facts from the evidence to generate corrections. R1 - Intelligible Similar to other language generation tasks, our first requirement is that generated outputs are fluent and intelligible. They must be free of grammatical mistakes and the meaning must be understandable without the aid of additional context or evidence so that their factual correctness can be assessed. R2 - Supported by Evidence The generated correction must be supported by the retrieved evidence. This property follows from previous work (Thorne et al., 2018) and also requires models to condition generation on the retrieved evidence – penalizing models that hallucinate (Holtzman et al., 2020). R3 - Error correction Specific to our task, the corrections should be targeted to the errors present in the inputted claim. While this, in part, can be assessed by R2 we need to compare the correction to the inputted claim to ensure the output is not introducing new unrelated information. For example, an erroneous claim: France is in South America could be supported by evidence if it were rewritten as France is a republic. However, the desired correction should instead state France is in Europe. 4 Task Decomposition The choice of supervision for the error correction system influences the task decomposition. For example, with full supervision, the system can be constructed with an information retrieval module and a sequence-to-sequence module that conditionally generates a correction given the claim and evidence. However, large datasets of claims paired with corrections are not available. The absence of full supervision requires that we distantly-supervise our systems using fact verification datasets, which are an abundant resource. Fact verification datasets contain claims labeled with evidence but do not contain corrections. With this resource, we propose a task decomposition that generated corrections by training models to reconstruct claims with masked tokens using retrieved evidence. 4.1 Distantly-supervised corrections Test time Corrections are generated by a twostage process, illustrated in Figure 2. Tokens from the claim, c, are first masked, yielding ˜c, and then input to the corrector c′ = Corr(˜c, E(c)). The masker, ˜c = Mask(c, E(c)), replaces a subset of tokens in the claim with a blank placeholder, conditioned on E(c). Its purpose is to remove tokens that are salient to the claim being supported or refuted by the evidence. Using the masked claim, ˜c, the corrector replaces the blank placeholders with tokens conditionally generated using retrieved evidence. To correct errors, evidence refuting a claim (E(c) ⊭c) conditions generation of a correction supported by it E(c) ⊨c′. This extends the pro3301 tocol Shah et al. (2020) by conditioning both the masker and corrector with multiple retrieved evidence sentences, rather than a single gold factoid. Training the corrector Similar to masked language modeling, the training objective is to generate the input claim c′ = c conditioned on the masked claim ˜c and evidence E(c). By training the model to generate the input claim, we expect the model to generate the input claim only if it was in complete agreement with the evidence (assuming the masking and the evidence are correct). Otherwise, the generated correction will contain evidence pertinent to the correcting the masked claim, which enables us to generate corrections satisfying requirements R2 and R3. Masker When applied to factual error correction, masking the tokens from the claim acts as a proxy to which tokens need to be removed to correct an error. Parallels can be drawn between masking and generating token-level explanations. We briefly summarize common approaches to generating explanations in Section 5.2. 5 Model 5.1 Evidence retrieval We use GENRE (Cao et al., 2021) and Dense Passage Retrieval (Karpukhin et al., 2020) together to retrieve evidence for claims E(c). Both have shown success for a number of language understanding tasks over Wikipedia (Petroni et al., 2020). GENRE is a pre-trained seq2seq model, trained to predict a Wikipedia page name for a claim. DPR encodes fixed length passages from Wikipedia into vectors using a BERT encoder to build a static index. At test-time, the claim is encoded and the most-similar passages are returned using an innerproduct search. We return the top-k passages returned by DPR from pages predicted by GENRE. 5.2 Token-level explanations as masks At test time, the purpose of the masker is to selectively remove tokens that contribute to the factual errors within a claim. We study how the choice of masker influences the quality of corrections. This considers varying levels of access to model information and different run-time complexity. Both the black- and white-box methods, outlined below, require querying a model trained to classify the veracity of claims given evidence whereas the the language model masker and baselines do not. Black-box masker We evaluate perturbing the input to a classifier that is trained to predict the veracity of a claim given evidence. We use LIME (Ribeiro et al., 2016), a diagnostic that trains a locally linear model to score the importance of input features (in our case, tokens in the claim) with respect to the predicted labels. The model under test is a BERT classifier where evidence and the claim are concatenated in the input. This is referred to as black-box because the model does not undergo modification and no information about internal values or states is exposed. White-box masker In contrast, to obtain whitebox model explanations, the model has undergone modification to expose internal information. We use the Neutrality Masker from (Shah et al., 2020) to predict which tokens, when masked, are likely to cause a label flip from supports or refuted to not enough information. This masker exposes encoded input of an ESIM classifier (Chen et al., 2017), and adds a linear classifier over the hidden states to predict per-token masking probability. At test time, masks can be generated through a single query to the model (unlike LIME in the black-box masker which requires multiple queries to the model), however this requires an additional step to train, using predictions from the classifier as signal. Language model masker We evaluate whether it is possible to generate masks without the need for a fact verification model. We use a BERT pretrained language model (Devlin et al., 2019) to measure the surprisal of tokens in the claim. Our intuition is to identify tokens which introduce misinformation under the hypothesis that the world knowledge (Petroni et al., 2019) captured in retraining would assign lower probabilities to tokens contradictory to the world state. This language model has no additional task-specific fine-tuning. We independently predict the cross-entropy for each token under a masked language modelling objective using BERT and return the top-k tokens. Baselines We additionally consider two simple baseline maskers: random masking of a subset of tokens and also a heuristic method of masking tokens which are not in common between the claim and the retrieved evidence. 5.3 Corrections We train an encoder-decoder transformer model to generate corrections from masked claims and 3302 evidence. Our model uses a pre-trained T5 transformer (Raffel et al., 2020) which we fine-tune with the distant supervision protocol described in Section 4.1. This model jointly encodes the masked claim and evidence by concatenating these two inputs in the input. We also compare against a baseline model from a related task of fact guided sentence modification (Shah et al., 2020) which uses a pointer generator network (See et al., 2017). Unlike our model, which captures long-range dependencies between claim and evidence through the transformer selfattention (Vaswani et al., 2017), the baseline independently encodes the evidence and masked claim using LSTMs (Hochreiter and Schmidhuber, 1997) before decoding using a pointer-copy mechanism. In order to evaluate the impact of conditioning on evidence, we decode tokens from masked claims using a language model without fine-tuning or conditioning, similar to the Language Models as Knowledge Bases hypothesis introduced by Petroni et al. (2019). This would consider correcting claims using the implicit knowledge stored within the model parameters rather than using external evidence. 6 Data We make use of FEVER (Thorne et al., 2018), a commonly used fact verification dataset, as the basis for our experiments. FEVER is one of the largest resources consisting of claims paired with evidence from Wikipedia. There are 185k instances with corresponding evidence sentences and a label as to whether the claim is SUPPORTED or REFUTED by it. Claims where no information could be found are labeled as NOTENOUGHINFO. To comprehensively evaluate the corrections generated manual evaluation is required. However, this is expensive and not suitable for system development and hyper-parameter optimization. To automate system evaluation or to train a seq2seq model with full supervision, a reference “gold standard” correction is also required. For this, we release annotations from the FEVER shared task as follows. The claims in FEVER were generated in a two-stage process: annotators extracted facts from Wikipedia and then performed meaning altering perturbations called mutations over these extracted facts. Each claim was independently labeled using retrieved evidence. Our reference corrections are the unmodified facts extracted from Wikipedia. The class balance and size of the dataset is reported in Table 1. The training and test splits are disjoint by entity. The additional hidden shared task test set was not used. The claims labelled as NOTENOUGHINFO. are used for training fact verification classifiers, but they will not be used for training the error correction systems in this paper as there is no labeled evidence to make corrections from. For completeness, we also release these unused NOTENOUGHINFO instances, as they have claims paired unmodified extracted facts (21934 training, 1870 development and 2037 test). Label Instance Count Train Validation Test Supports 37961 1477 1593 Refutes 20075 2091 2289 Total Training 58036 3568 3891 Table 1: Instance counts by class and dataset partitions 7 Evaluation While it’s convenient to use an automatic metric during development, these metrics compute token overlap against a single reference sentence and cannot capture the nuances required to assess the veracity of the generated corrections against evidence. Thus, our primary evaluation will use human raters to label whether the model predictions meet the task requirements stated in Section 3. Human raters are asked three questions about system outputs to assess whether the corrections meet the requirements of intelligibility, supported by evidence, and error correction introduced in Section 3. For the first 2 requirements, the question has a binary answer. For the third requirement of error correction, the question has 3 answer choices: (1) the information content w.r.t. the evidence improved, (2) information unrelated to the claim was added (i.e. the claim was ignored), (3) no correction was needed (i.e. the claim was already supported by evidence). The raters were shown each question in this sequence without knowledge of which system generated the correction. Negative answers to a question automatically assigned negative answers to subsequent ones (prescribing that an unintelligible sentence could not contain a fact supported by evidence or introduce a correction). 20% of the tasks are assigned to two raters to measure inter-annotator agreement. We used 4 expert participants from our lab (none of them co-authors of the paper) who were familiar with fact verifica3303 tion, but not with error correction. Responses were calibrated using a pilot study on the validation set. For automated evaluation, we use SARI (Xu et al., 2016) which is a metric used for sentence simplification. SARI considers ngrams retained from the source as well added or deleted ngrams through comparison against a reference sentence. We additionally report BLEU (Papineni et al., 2002) and ROUGE (Lin, 2004) to indicate precision and recall of the correction. In Section 9, we report correlation of automated metrics against our manual evaluation. 8 Implementation T5 Masker-Corrector We fine-tuned the T5base pre-trained models released by HuggingFace (Wolf et al., 2020). The number of training epochs and learning rate was selected through optimizing the overall SARI score. The search space for learning rate was {10−5, 5 · 10−5, 104, 5 · 10−4}. We used 5 · 10−5 for all experiments. We found diminishing returns in SARI after 4 epochs and stopped training. Fully Supervised Ceiling We use this model to estimate the ceiling performance of a factual error correction system (assuming a reasonable amount of training data is available) that other methods can be compared against. We fine-tune a T5-base model with supervision of the correction (see Section 6), using the same hyper-parameter choices as the T5 Masker-Corrector. Automated Scoring A single reference sentence from the FEVER dataset is used for automated scoring. We consider BLEU, ROUGE, and SARI. SARI considers the F1 of added tokens, F1 of kept tokens, precision of deletions, and the mean of these 3 scores (denoted final). We use code made available by Xu et al. (2016). Evidence Retrieval We use the Facebook implementation of DPR (Karpukhin et al., 2020) without fine-tuning and constructed an index over the Wikipedia version released with FEVER (Thorne et al., 2018), chunked into passages of 50 tokens. For GENRE, the original authors’ implementation was used. We selected the top matching 2 passages. This resulted in the highest scores on the downstream corrections; SARI was lower when using 1 or 3 passages. Maskers For the white-box masker, we use the implementation provided by Shah et al. (2020) applied to our dataset retaining original hyperparameters trained on FEVER. For the black-box masker, we use the LIME implementation from (Ribeiro et al., 2016) to probe a BERT classifier (Devlin et al., 2019) fine-tuned on FEVER. For the LM and random baseline maskers, where the number of masks was tunable, we masked 50% of the tokens, which was similar to the number of tokens masked by the black- and white-box maskers. Language Model as Correctors? We greedily decode masked tokens using a BERT-base-cased language model using the HuggingFace implementation (Wolf et al., 2020) without fine-tuning. Comparison to Previous Work For comparison to previous work, we use the dual-encoder pointer network implementation from (Shah et al., 2020), retaining the original hyper-parameter choices. 9 Results We first report results from a manual evaluation, assessing the requirements that corrections are intelligible, supported by evidence, and improve the factuality of the claim, as listed in Section 3. Our evaluation considers a sample of 200 instances per system. We report the results in Table 2. For interannotator agreement control, 20% of instances were annotated by two annotators: the Cohen’s κ scores for the 3 questions are 0.92 for intelligible, 0.92 for supported, and 0.86 for corrected. When using retrieved evidence, the white-box masker generated no masks for 41% of instances. Without masked tokens, the T5 corrector copied the input claim to the output. This fits the assumption that, if the claim is already supported well by evidence, no correction is required. The fully supervised models had the highest rate of satisfactory corrections that improved the factuality of the claim (requirement 3), indicating a performance ceiling for the distantly-supervised models. Incorporating retrieved evidence in these supervised models (rather than gold) reduced the number of corrections supported by evidence from 88.9% to 64.7% and the number of satisfactory corrections from 68.9% to 48.9% showing the challenges of incorporating (possibly noisy) retrieved evidence when generating the corrections. When using the masker and corrector distant supervision strategy, different maskers could be used 3304 System Evidence Training Masks Test Masks Aggregated Score (%) Intelligible Supported Corrected T5 Fully Supervised Gold 98.9 88.9 68.9 T5 Fully Supervised Retrieved 97.7 64.7 48.9 T5 Masker + Corrector Retrieved Random Heuristic 89.3 57.9 40.0 T5 Masker + Corrector Retrieved Heuristic Heuristic 90.0 38.0 20.0 T5 Masker + Corrector Retrieved Random Black-box 93.1 42.2 24.0 T5 Masker + Corrector Retrieved Black-box Black-box 91.4 37.0 19.8 T5 Masker + Corrector Retrieved White-box White-box 90.6 41.7 23.9 BERT Language Model Heuristic 48.0 20.7 15.0 BERT Language Model Black-box 30.1 4.9 3.4 Shah et al. (2020) M+C Gold White-box White-box 32.2 10.7 5.0 Table 2: Aggregated scores from human evaluation considering intelligibility, whether generated instances were supported by evidence and errors corrected. to train the corrector to the masker used at test time. We observed that training the corrector with random masks yielded both a higher rate of satisfactory corrections and corrections supported by evidence when using either the black-box or heuristic masker at test time. We further evaluate other maskers with automated metrics in Section 9.2. Using a heuristic masker at test time, which removed tokens from the claim not present in the evidence, generated more claims meeting the supported and corrected requirements than masks generated by querying a fact verification model (both black-box and white-box). An analysis of the masker’s influence on the corrections is provided in Section 9.1. The two baseline systems, Dual Encoder M+C, based on Shah et al. (2020), and a pre-trained BERT language model, generated corrections that were intelligible or supported by evidence at a lower rate than the aforementioned models, further discussed in Sections 9.3 and 9.4. We report the correlation between automated scoring metrics and our manual evaluation in Table 3. The KEEP component of SARI, which measures the F1 of n-grams from the claim retained in the output, had the highest correlation with all three requirements. Overly aggressive maskers which remove too much content from the claim can result in unintelligible outputs, or corrections unrelated to the claim. ROUGE2, which measures the recall of bigrams in the correction w.r.t. the reference, exhibited reasonable correlation to the manual evaluation against the supported and corrected requirements, however does not correlate as well with intelligibility. The ADD and DELETE components of SARI provide further information but do not correlate as strongly with the human judgements. Having only one reference correction reduces the utility of precision-oriented metrics, like BLEU, as valid corrections can differ from the reference. Metric Correlation (Pearson r) Intelligible Supported Corrected SARI Keep .87 .95 .93 SARI Final .78 .92 .91 SARI Delete .72 .82 .91 SARI Add .52 .84 .79 ROUGE2 .75 .90 .91 ROUGE1 .71 .87 .88 BLEU2 −.05 .32 .45 BLEU1 −.46 −.10 .05 Table 3: Both SARI and ROUGE automated scoring metrics have high correlation to manual evaluation. 9.1 Choice of masker When training the corrector with the same masker that is used at test time, both the heuristic and blackbox maskers yielded comparable scores under human evaluation. Inspection of SARI breakdown in Table 4 indicates that more tokens were kept when using the heuristic masker (Keep=.651) whereas the black box model was more aggressive in masking, resulting in less information from the claim being retained (Keep=.594). This correlated well with human judgements as more information retained gives a richer context for generating the correction and prevents erasure of claims already (partially) supported by the evidence. 3305 Both the black-box (LIME) and white-box (the masker from Shah et al. (2020)) methods require querying a veracity classifier to generate the masks. Using retrieved evidence for the veracity classifier, which was used to generate the masks in conjunction with these two methods, had a negative impact on most components of the SARI score. For the black-box masker, using retrieved evidence reduced the number of masked tokens from an average of 4.7 per claim to 3.9. Whereas the number of masked tokens by the white-box masker remained unchanged at 4.7 (approximately 50% of number of tokens in the claim). Most notably, the white-box method of mask generation (row 4 in Table 4) did not to generate masks for 41% of instances when using retrieved evidence, whereas all instances had at least one mask when using gold evidence – an artefact of the noise introduced by retrieval. Masker SARI Score Keep Delete Add Final Black-box (Gold) .630 .582 .088 .433 White-box (Gold) .652 .559 .128 .447 Black-box (IR) .594 .526 .090 .412 White-box (IR) .628 .535 .107 .426 Heuristic (IR) .651 .574 .041 .422 Masked LM .538 .509 .062 .370 Random .619 .475 .087 .390 Table 4: Extrinsic evaluation of maskers, varying the use of evidence when generating the masks, evaluated using the T5 Masker+Corrector model. 9.2 Corrector trained with random masks Generating large quantities of masked training data through querying a model, such as with the blackbox model explanation techniques, can be computationally expensive. In contrast, random masks can be generated without querying a model. Using a corrector trained on random masks resulted in higher quality outputs at test time when paired the black-box and heuristic maskers. Training with random masks promotes good exploration of the task. In contrast, while the black-box and heuristic approaches worked well during testing, correctors trained on these maskers generated worse outputs due to the limited exploration of the task space. Additionally, generating training data using the blackand white-box methods requires making predictions using the model’s training data which may result in different outcomes to making predictions on unseen test data. Masker SARI Score Keep Delete Add Final Black-box (Gold) .618 .622 .102 .447 White-box (Gold) .640 .570 .114 .441 Black-box (IR) .611 .543 .194 .419 White-box (IR) .618 .590 .144 .452 Heuristic (IR) .652 .627 .155 .478 Masked LM .561 .529 .078 .389 Table 5: Using random masks at training resulted in higher scores when testing with different maskers 9.3 Comparison to previous work Previous work uses a dual encoder pointer network (Shah et al., 2020) to make corrections, reported in Table 6. The corrector tended to copy portions of claim rather than correct it, resulting in a SARI KEEP score of .452 which is lower than the T5 model using the same white-box masker (Table 4). Human evaluation considered these corrections mostly unintelligible, even when using gold evidence (Table 2). This was especially the case for rarer entities. Hyper-parameter tuning of the corrector’s coverage ratio, as suggested by the authors, did not yield improvements. System SARI Score Keep Delete Add Final Dual Enc Ptr (Gold) .452 .569 .039 .353 Dual Enc Ptr (IR) .345 .481 .017 .281 Table 6: Results using a dual encoder pointer network (Shah et al., 2020) were low, despite the strong masker. 9.4 Language Models as Correctors? With the exception of the heuristic masker, using a pre-trained language model, without fine-tuning, to correct claims resulted in low SARI scores (Table 7). Without conditioning on the evidence, the correction is not related to the claim or supported by evidence to verify the claim, which is indicated by the low SARI Add scores which consider the precision of the added tokens. As these maskers deleted most tokens, retaining only stop-words, decoding most likely tokens without a prompt or context tokens resulted in unintelligible outputs. For the heuristic masker, more content words were retained yielding more intelligible outputs. However, these were not always supported by evidence, indicated in the human evaluation in Table 2. 3306 Masker SARI Score Keep Delete Add Final Masked LM .360 .472 .019 .289 Heuristic (IR) .629 .651 .034 .438 White-box (IR) .232 .446 .005 .228 Black-box (IR) .364 .003 .001 .122 Table 7: Correcting claims using a language model does not condition the generation on evidence. 10 Qualitative Error Analysis In this section we discuss the following issues which were present in all master-corrector systems: Over-erasure In some instances, the masker removed most or all of the non-stopword tokens from the claim. This resulted in the original meaning of the claim being erased. Without this information the corrector could not reconstruct the claim, resulting in corrections that were unrelated to the input claim. This issue was most prevalent for the blackbox masker, where 15% of instances had more than 5 consecutive tokens masked and 32% of instances had 4 consecutive tokens masked. In contrast, the heuristic masker, which identifies the tokens not present in the retrieved evidence had 5 consecutive tokens masked for 3% of instances and 4 consecutive tokens masked for 9% of instances. While, in some cases, appropriate corrections could be made despite the aggressive masking (e.g. the claim “Exit the King is by man[sic].” was fully masked, but corrected to include the author’s name), others were re-written focusing on a different fact, e.g. a claim about the length of reign of Maria Theresa was rewritten to be about her date of birth. Incorrect masking When the erroneous tokens in a claim were not masked, the corrector would generate outputs not supported by evidence. For example the following claim, which has an incorrect year, was masked but retaining the error: “Ghost, the film was released in 1994” as “[MASK] , [MASK] [MASK] [MASK] [MASK] [MASK] in 1994”. Even with suitable retrieved evidence, indicating the release year is 1990, no appropriate correction could be made. Inadequate evidence retrieval Where the evidence retrieved was related, but not specifically supporting or refuting the claim, the generated corrections were vague: the claim “Poldark aired on HBO” was corrected to “Poldark premiered on TV” as the evidence lacked the name of the correct TV station. Similarly, where incorrect masks were made, additional retrieval retrieval may be required to prevent the corrector from hallucinating information to cover the knowledge missing from the evidence. For example, the name of the TV show was masked in the claim “Two and a half men starred Jamie Fox[sic]”, but as no mention of Jamie Fox was present in the evidence, the model hallucinated a different TV show name. 11 Conclusions and Future Work Going beyond simply identifying errors, factual error correction presents a number of challenges for information retrieval, fact verification and abstractive summarization communities alike. In this paper, we demonstrated that the task can be performed with distant supervision in the form of claims labeled by evidence supporting or refuting them. However, there are a number of outstanding challenges that must be addressed. The data we used from the FEVER task was re-purposed to evaluate whether systems can undo mutations introduced by human annotators and may not be representative of the range of factual errors that would be present in real-world documents. While some automated metrics correlated well with human judgements, future work should consider how automated scoring can be better used to discriminate the adequacy of the generated corrections going beyond similarity to the reference sentence. From a modelling perspective, the masks strongly influenced the corrector and further work is required to generate masks that result in better corrections. We observed where masks mismatched the evidence, the correction was vague, hallucinated or did not correct the factual errors in the claim. This could be addressed through joint training of both components to enable them to avoid error propagation from masking to correction. Acknowledgements The authors wish to thank: Tal Schuster for his helpful comments and feedback; Nicola De Cao for providing the GENRE predictions for FEVER; Amrith Krishna, Guy Aglionby, Rami Aly and Zhijiang Guo for manual evaluation of the model predictions. This research was supported by donation of compute resources from Google Cloud. James Thorne is supported by an Amazon Alexa Graduate Research Fellowship. Andreas Vlachos is supported by the ERC grant AVeriTeC (GA 865958). 3307 Broader Impact Statement Our experiments were performed on publicly available data about common facts from Wikipedia. These data are released under a creative-commons license. The expert raters from our lab who manually reviewed the generated instances were volunteers and were compensated through quid-pro-quo help on their own projects. The intended use of this project is to help explain reasoning using evidence, going beyond singlelabel classification. This adds an additional safeguard, making the decision process more transparent as poor predictions by our model expose limitations that would be hidden by classification. Our data is synthetic in nature and is biased towards synthetic facts from popular entities. Application to political or scientific domains would require additional work. Misinformation about populations that are under-represented in our data may not be accurately identified or corrected without further mitigation. One positive finding in our paper was that some of biases perpetuated in the hallucinations of language models were mitigated when conditioning the generation on retrieved evidence. Model fine-tuning took approximately 2 hours per experiment on a single P100 GPU. Generating LIME explanations of the training dataset took approximately one day – motivating our experiments that used models trained on random or heuristic maskers which required fewer resources by several orders of magnitude. References Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020. Generating fact checking explanations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7352–7364. Association for Computational Linguistics. Isabelle Augenstein, Christina Lioma, Dongsheng Wang, Lucas Chaves Lima, Casper Hansen, Christian Hansen, and Jakob Grue Simonsen. 2019. MultiFC: A real-world multi-domain dataset for evidence-based fact checking of claims. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4685–4697, Hong Kong, China. Association for Computational Linguistics. Meng Cao, Yue Dong, Jiapeng Wu, and Jackie Chi Kit Cheung. 2020. Factual Error Correction for Abstractive Summarization Models. In Empirical Methods in Natural Language Processing, pages 6251–6258. Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021. Autoregressive entity retrieval. In International Conference on Learning Representations. Qian Chen, Xiaodan Zhu, Zhen-Hua Ling, Si Wei, Hui Jiang, and Diana Inkpen. 2017. Enhanced LSTM for natural language inference. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1657–1668, Vancouver, Canada. Association for Computational Linguistics. Sarah Cohen, Chengkai Li, Jun Yang, and Cong Yu. 2011. Computational Journalism: a call to arms to database researchers. Proceedings of the 5th Biennial Conference on Innovative Data Systems Research (CIDR 2011) Asilomar, California, USA., (January):148–151. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-wei Chang. 2020. REALM: Retrieval-Augmented Language Model PreTraining. Na-Rae Han, Joel Tetreault, Soo-Hwa Lee, and JinYoung Ha. 2010. Using an error-annotated learner corpus to develop an ESL/EFL error correction system. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Valletta, Malta. European Language Resources Association (ELRA). Sepp Hochreiter and Jurgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation, 9(8):1735–1780. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In International Conference on Learning Representations. Georgi Karadzhov, Preslav Nakov, Lluís Màrquez, Alberto Barrón-Cedeño, and Ivan Koychev. 2017. Fully automated fact checking using external sources. In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, pages 344–353. INCOMA Ltd. Vladimir Karpukhin, Barlas O˘guz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open-Domain Question Answering. 3308 Kevin Knight and Ishwar Chander. 1994. Automated postediting of documents. In Proceedings of the National Conference on Artificial Intelligence, volume 1, pages 779–784. Neema Kotonya and Francesca Toni. 2020. Explainable Automated Fact-Checking for Public Health Claims. In The 2020 Conference on Empirical Methods in Natural Language Processing. Nayeon Lee, Belinda Li, Sinong Wang, Wen-tau Yih, Hao Ma, and Madian Khabsa. 2020. Language models as fact checkers? In Proceedings of the Third Workshop on Fact Extraction and VERification (FEVER), volume 2, pages 36–41. Association for Computational Linguistics. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for KnowledgeIntensive NLP Tasks. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. 2014. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, July, pages 1– 14. Association for Computational Linguistics. Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, and Chin-Yew Lin. 2019. A simple recipe towards reducing hallucination in neural surface realisation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2673–2679, Florence, Italy. Association for Computational Linguistics. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July, pages 311–318. Association for Computational Linguistics. Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vassilis Plachouras, Tim Rocktäschel, and Sebastian Riedel. 2020. KILT: a Benchmark for Knowledge Intensive Language Tasks. Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 2463–2473, Hong Kong, China. Association for Computational Linguistics. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits of Transfer Learning with a Unified Textto-Text Transformer. Journal of Machine Learning Research, 21:1–67. Marco Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. “why should I trust you?”: Explaining the predictions of any classifier. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations, volume 39, pages 97–101. Association for Computational Linguistics. Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035–4045, Brussels, Belgium. Association for Computational Linguistics. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 1073–1083. Association for Computational Linguistics. Darsh J Shah, Tal Schuster, and Regina Barzilay. 2020. Automatic Fact-guided Sentence Modification. In Proceedings of the AAAI Conference on Artificial Intelligence. Dominik Stammbach and Elliott Ash. 2020. e-FEVER: Explanations and Summaries for Automated Fact Checking. In Proceedings of the 2020 Truth and Trust Online Conference (TTO 2020), page 32. Hacks Hackers. Wilson L. Taylor. 1953. “Cloze Procedure”: A New Tool for Measuring Readability. Journalism Quarterly, 30(4):415–433. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Lilon Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In 31st Conference on Neural Information 3309 Processing Systems (NIPS 2017), Long Beach, CA, USA. David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi. 2020. Fact or Fiction: Verifying Scientific Claims. William Yang Wang. 2017. “liar, liar pants on fire”: A new benchmark dataset for fake news detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 422–426. Association for Computational Linguistics. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. 2016. Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4:401–415. Chunting Zhou, Jiatao Gu, Mona Diab, Paco Guzman, Luke Zettlemoyer, and Marjan Ghazvininejad. 2020. Detecting Hallucinated Content in Conditional Neural Sequence Generation. pages 1–21.
|
2021
|
256
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3310–3321 August 1–6, 2021. ©2021 Association for Computational Linguistics 3310 Probabilistic, Structure-Aware Algorithms for Improved Variety, Accuracy, and Coverage of AMR Alignments Austin Blodgett Nathan Schneider Georgetown University {ajb341, nathan.schneider}@georgetown.edu Abstract We present algorithms for aligning components of Abstract Meaning Representation (AMR) graphs to spans in English sentences. We leverage unsupervised learning in combination with heuristics, taking the best of both worlds from previous AMR aligners. Our unsupervised models, however, are more sensitive to graph substructures, without requiring a separate syntactic parse. Our approach covers a wider variety of AMR substructures than previously considered, achieves higher coverage of nodes and edges, and does so with higher accuracy. We will release our LEAMR datasets and aligner for use in research on AMR parsing, generation, and evaluation. 1 Introduction Research with the Abstract Meaning Representation (AMR; Banarescu et al., 2013), a broadcoverage semantic annotation framework in which sentences are paired with directed acyclic graphs, must contend with the lack of gold-standard alignments between words and semantic units in the English data. A variety of rule-based and statistical algorithms have sought to fill this void, with improvements in alignment accuracy often translating into improvements in AMR parsing accuracy (Pourdamghani et al., 2014; Naseem et al., 2019; Liu et al., 2018). Yet current alignment algorithms still suffer from limited coverage and less-than-ideal accuracy, constraining the design and accuracy of parsing algorithms. Where parsers use latent alignments (e.g., Lyu and Titov, 2018; Cai and Lam, 2020), explicit alignments can still facilitate evaluation and error analysis. Moreover, AMR-to-text generation research and applications using AMR stand to benefit from accurate, human-interpretable alignments. We present Linguistically Enriched AMR (LEAMR) alignment, which achieves full graph coverage via four distinct types of aligned structures: subgraphs, relations, reentrancies, and duplicate subgraphs arising from ellipsis. This formulation lends itself to unsupervised learning of alignment models. Advantages of our algorithm and released alignments include: (1) much improved coverage over previous datasets, (2) increased variety of the substructures aligned, including alignments for all relations, and alignments for diagnosing reentrancies, (3) alignments are made between spans and connected substructures of an AMR, (4) broader identification of spans including named entities and verbal and prepositional multiword expressions. Contributions are as follows: • A novel all-inclusive formulation of AMR alignment in terms of mappings between spans and connected subgraphs, including spans aligned to multiple subgraphs; mappings between spans and inter-subgraph edges; and characterization of reentrancies. Together these alignments fully cover the nodes and edges of the AMR graph (§3). • An algorithm combining rules and EM to align English sentences to AMRs without supervision (§5), achieving higher coverage and quality than existing AMR aligners (§7). • A corpus with automatic alignments for LDC2020 and Little Prince data as well as a few hundred manually annotated sentences for tuning and evaluation (§4). We release this dataset of alignments for over 60,000 sentences along with our aligner code to facilitate more accurate models and greater interpretability in future AMR research. 2 Related Work The main difficulty presented by AMR alignment is that it is a many-to-many mapping problem, with gold alignments often mapping multiple tokens to 3311 multiple nodes while preserving AMR structure. Previous systems use various strategies for aligning. They also have differing approaches to what types of substructures of AMR are aligned—whether they are nodes, subgraphs, or relations—and what they are aligned to—whether individual tokens, token spans, or syntactic parses. Two main alignment strategies remain dominant, though they may be combined or extended in various ways: rule-based strategies as in Flanigan et al. (2014), Flanigan et al. (2016), Liu et al. (2018), and Szubert et al. (2018), and statistical strategies using ExpectationMaximization as in Pourdamghani et al. (2014). JAMR. The JAMR system (Flanigan et al., 2014, 2016) aligns token spans to subgraphs using iterative application of an ordered list of 14 rules which include exact and fuzzy matching. JAMR alignments form a connected subgraph of the AMR by the nature of the rules being applied. A disadvantage of JAMR is that it lacks a method for resolving ambiguities, such as repeated tokens, or of learning novel alignment patterns. ISI. The ISI system (Pourdamghani et al., 2014) produces alignments between tokens and nodes and between tokens and relations via an ExpectationMaximization (EM) algorithm in the style of IBM Model 2 (Brown et al., 1988). First, the AMR is linearized; then EM is applied using a symmetrized scoring function of the form P(a ∣t) + P(t ∣a), where a is any node or edge in the linearized AMR and t is any token in the sentence. Graph connectedness is not enforced for the elements aligning to a given token. Compared to JAMR, ISI produces more novel alignment patterns, but also struggles with rare strings such as dates and names, where a rule-based approach is more appropriate. Extensions and Combinations. TAMR (Tuned Abstract Meaning Representation; Liu et al., 2018) uses the JAMR alignment rules, along with two others, to produce a set of candidate alignments for the sentence. Then, the alignments are “tuned” with a parser oracle to select the candidates that correspond to the oracle parse that is most similar to the gold AMR. Some AMR parsers (Naseem et al., 2019; Fernandez Astudillo et al., 2020) use alignments which are a union of alignments produced by the JAMR and ISI systems. The unioned alignments achieve greater coverage, improving parser performance. Syntax-based. Several alignment systems attempt to incorporate syntax into AMR alignments. nodes edges reentrancies JAMR 91.1 ✗ ✗ ISI 78.7 9.8 ✗ TAMR∗ 94.9 ✗ ✗ Table 1: Coverage and types of previous alignment systems. Scores are evaluated on 200 gold test sentences. ∗TAMR is evaluated on a subset of 91 sentences. Chen and Palmer (2017) perform unsupervised EM alignment between AMR nodes and tokens, taking advantage of a Universal Dependencies (UD) syntactic parse as well as named entity and semantic role features. Szubert et al. (2018) and Chu and Kurohashi (2016) both produce hierachical (nested) alignments between AMR and a syntactic parse. Szubert et al. use a rule-based algorithm to align AMR subgraphs with UD subtrees. Chu and Kurohashi use a supervised algorithm to align AMR subgraphs with constituency parse subtrees. Word Embeddings. Additionally, Anchiêta and Pardo (2020) use an alignment method designed to work well in low-resource settings using pretrained word embeddings for tokens and nodes. Graph Distance. Wang and Xue (2017) use an HMM-based aligner to align tokens and nodes. They include in their aligner a calculation of graph distance as a locality constraint on predicted alignments. This is similar to our use of projection distance as described in §5. Drawbacks of Current Alignments. Alignment methods vary in terms of components of the AMR that are candidates for alignment. Most systems either align nodes (e.g., ISI) or connected subgraphs (e.g., JAMR), with incomplete coverage. Most current systems do not align relations to tokens or spans, and those that do (such as ISI) do so with low coverage and performance. None of the current systems align reentrancies, although Szubert et al. (2020) developed a rule-based set of heuristics for identifying reentrancy types. Table 1 summarizes the coverage and variety of prominent alignment systems. 3 An All-Inclusive Formulation of AMR Alignment Aligning AMRs to English sentences is a vexing problem not only because the English training data lacks gold alignments, but also because AMRs— unlike many semantic representations—are not designed with a derivational process of form–function subunits in mind. Rather, each AMR graph represents the full-sentence meaning, and AMR anno3312 (w / want-01 :ARG0 (p / person :ARG0-of (s / study-01) :ARG1-of (i / include-91 :ARG2 (p2 / person :ARG0-of (s2 / study-01)) :ARG3 (m / most))) :ARG1 (v / visit-01 :ARG0 p :ARG1 (c / city :name (n / name :op1 "New" :op2 "York")) :time (g / graduate-01 :ARG0 p))) Subgraph Alignments Relation Alignments Most →m, of →s :ARG1-of i, of →i, i :ARG2 p2, the →∅, i :ARG3 m; students →(p :ARG0-of s), want →w :ARG0 p, want →w, w :ARG1 v; to →∅, visit →v :ARG0 p, visit →v, v :ARG1 c; New York →(c :name graduate →g :ARG0 p; (n :op1 "New" :op2 "York")), when →v :time g when →∅, Reentrancy Alignments they →∅, want →w :ARG0 p (PRIMARY), graduate →g v :ARG0 p (CONTROL); Duplicate Subgraphs they →g :ARG0 p (COREF) students →(p2 :ARG0-of s2) Figure 1: AMR and alignments for the sentence “Most of the students want to visit New York when they graduate.” Alignments are differentiated by colors: blue (subgraphs), green (duplicate subgraphs), and orange (relations). Relations that also participate in reentrancy alignments are bolded. tation conventions can be opaque with respect to the words or surface structure of the sentence, e.g., by unifying coreferent mentions and making explicit certain elided or pragmatically inferable concepts and relations. Previous efforts toward general tools for AMR alignment have considered mapping tokens, spans, or syntactic units to nodes, edges, or subgraphs (§2). Other approaches to AMR alignment have targeted specific compositional formalisms (Groschwitz et al., 2018; Beschke, 2019; Blodgett and Schneider, 2019). We advocate here for a definition of alignment that is principled—achieving full coverage of the graph structure—while being framework-neutral and easy-to-understand, by aligning graph substructures to shallow token spans on the form side, rather than using syntactic parses. We do use structural considerations to constrain alignments on the meaning side, but by using spans on the form side, we ensure the definition of the alignment search space is not at the mercy of error-prone parsers. Definitions. Given a tokenized sentence w and its corresponding AMR graph G, a complete alignment assumes a segmentation of w into spans s, each containing one or more contiguous tokens; and puts each of the nodes and edges of G in correspondence with some span in s. A span may be aligned to one or more parts of the AMR, or else is null-aligned. Individual alignments for a sentence are grouped into four layers: subgraph alignments, duplicate subgraph alignments, relation alignments, and reentrancy alignments. These are given for an example in figure 1. All alignments are between a single span and a substructure of the AMR. A span may be aligned in multiple layers which are designed to capture different information. Within the subgraph layer, alignments are mutually exclusive with respect to both spans and AMR components. The same holds true within the relation layer. Every node will be aligned exactly once between the subgraph and duplicate subgraph layers. Every edge will be aligned exactly once between the subgraph and relation layers, and may additionally have a secondary alignment in the reentrancy layer. 3.1 Subgraph Layer Alignments in this layer generally reflect the lexical semantic content of words in terms of connected,1 directed acyclic subgraphs of the corresponding AMR. Alignments are mutually exclusive (disjoint) on both the form and meaning sides. 3.2 Duplicate Subgraph Layer A span may be aligned to multiple subgraphs if one is a duplicate of the others, with a matching concept. This is often necessary when dealing with ellipsis constructions, where there is more semantic content in the AMR than is pronounced in the sentence and thus several identical parts of the AMR must be aligned to the same span. In this case, a single subgraph is chosen as the primary alignment (whichever is first based on depth-first order) and is aligned in the subgraph alignment layer, and any others are represented in the duplicates alignment 1Nodes aligned to a span must form a connected subgraph with two exceptions: (1) duplicate alignments are allowed and are separated into subgraph and duplicate layers; (2) a span may be aligned to two terminal nodes that have the same parent. For example, never aligns to :polarity - :time ever, two nodes and two edges which share the same parent. 3313 layer. For example, verb phrase ellipsis, as in I swim and so do you, would involve duplication of the predicate swim, with distinct ARG0s. Similarly, in figure 1, Most of the students involves a subsetsuperset structure where the subset and superset correspond to separate nodes. Because student is represented in AMR like person who studies, there are two 2-node subgraphs aligned to student, one with the variables p and s, and the duplicate with p2 and s2. The difficulty that duplicate subgraphs pose for parsing and generation makes it convenient to put these alignments in a separate layer. 3.3 Relation Layer This layer includes alignments between a span and a single relation—such as when →:time— and alignments mapping a span to its argument structure—such as give →:ARG0 :ARG1 :ARG2. All edges in an AMR that are not contained in a subgraph fit into one of these two categories. English function words such as prepositions and subordinators typically function as connectives between two semantically related words or phrases, and can often be identified with the semantics of AMR relations. But many of these function words are highly ambiguous. Relation alignments make their contribution explicit. For example, when in figure 1 aligns to a :time relation. For spans that are aligned to a subgraph, incoming or outgoing edges attached to that subgraph may also be aligned to the span in the relation layer. These can include core or non-core roles as long as they are evoked by the token span. For example, figure 1 contains visit →:ARG0 :ARG1. 3.4 Reentrancy Layer A reentrant node is one with multiple incoming edges. In figure 1, for example, p appears three times: once as the ARG0 of w (the wanter), once as the ARG0 of v (the visitor), and once as the ARG0 of g (the graduate). The p node is labeled with the concept person—in the PENMAN notation used by annotators, each variable’s concept is only designated on one occurrence of the variable, the choice of occurrence being, in principle, arbitrary. These three ARG0 relations are aligned to their respective predicates in the relation layer. But there are many different causes of reentrancy, and AMR parsers stand to benefit from additional information about the nature of each reentrant edge, such as the fact that the pronoun they is associated with one of the ARG0 relations. The reentrancy layer “explains” the cause of each reentrancy as follows: for the incoming edges of a reentrant node, one of these edges is designated as PRIMARY—this is usually the first mention of the entity in a local surface syntactic attachment, e.g. the argument of a control predicate like want doubles as an argument of an embedded clause predicate. The remaining incoming edges to a reentrant node are aligned to a reentrancy trigger and labeled with one of 8 reentrancy types: coref, repetition, coordination, control, adjunct control, unmarked adjunct control, comparative control, and pragmatic. These are illustrated in table 2. These types, adapted from Szubert et al.’s (2020) classification, correspond to different linguistic phenomena leading to AMR reentrancies—anaphoric and non-anaphoric coreference, coordination, control, etc. The trigger is the word that most directly signals the reentrancy phenomenon in question. For the example in figure 1, the control verb want is aligned to the embedded predicate–argument relation and typed as CONTROL, while the pronoun they serves as the trigger for the third instance of p in when they graduate. 3.5 Validation To validate the annotation scheme we elicited two gold-standard annotations for 40 of the test sentences described in §4 and measured interannotator agreement.2 Interannotator exact-match F1 scores were 94.54 for subgraphs, 90.73 for relations, 76.92 for reentrancies, and 66.67 for duplicate subgraphs (details in appendix A). 4 Released Data We release a dataset3 of the four alignment layers reflecting correpondences between English text and various linguistic phenomena in gold AMR graphs—subgraphs, relations (including argument structures), reentrancies (including coreference, control, etc.), and duplicate subgraphs. Automatic alignments cover the ≈60,000 sentences of the LDC2020T02 dataset (Knight et al., 2020) and ≈1,500 sentences of The Little Prince. We manually created gold alignments for evaluating our automatic aligner, split into a development set (150 sentences) and a test set (200 sen2Both annotators are Ph.D. students with backgrounds in linguistics. One annotator aligned all development and test sentences; the other aligned a subset of 40 test sentences. 3https://github.com/ablodge/leamr 3314 Type Triggered by Example COREF a pronoun (including possessive or reflexive) (anaphora) I love my house REPETITION a repeated name or non-pronominal phrase (non-anaphoric coreference) The U.S. promotes American goods COORDINATION coordination of two or more phrases sharing an argument They cheered and celebrated CONTROL control verbs, control nouns, or control adjectives I was afraid to speak up ADJUNCT CONTROL control within an adjunct phrase I left to buy some milk; Mary cooked while listening to music UNMARKED ADJUNCT CONTROL control within an adjunct phrase with only a bare verb and no subordinating conjunction Mary did her homework listening to music COMPARATIVE CONTROL a comparative construction Be as objective as possible PRAGMATIC Reentrancies that must be resolved using context John met up with a friend Table 2: Reentrancy types with examples. For each reentrant node, one of its incoming edges is labeled PRIMARY and the others are labeled with one of the above reentrancy types. In the examples, the word aligned to an edge labeled with the specified type is underlined, and the word aligned to the parent of that edge is bolded. (h / have-degree-91 :ARG1 (h2 / house :location (l / left)) :ARG2 (b / big) :ARG3 (m / more) :ARG4 (h3 / house :location (r / right))) Figure 2: AMR for the sentence “The house1 on the left is bigger than the house2 on the right.” tences).4 The test sentences were annotated from scratch; the development sentences were first automatically aligned and then hand-corrected. We stress that no preprocessing apart from tokenization is required to prepare the test sentences and AMRs for human annotation. We also release our annotation guidelines as a part of our data release. 5 LEAMR Aligner We formulate statistical models for the alignment layers described above—subgraphs, duplicate subgraphs, relations, and reentrancies—and use the Expectation-Maximization (EM) algorithm to estimate probability distributions without supervision, with a decoding procedure that constrains aligned units to obey structural requirements. In line with Flanigan et al. (2014, 2016), we use rulebased preprocessing to align some substructures using string-matching, morphological features, etc. Before delving into the models and algorithm, we motivate two important characteristics: Structure-Preserving. Constraints on legal candidates during alignment ensure that at any point 4Our test set consists of sentences from the test set of Szubert et al. (2018) but with AMRs updated to the latest release version. This test set contains a mix of English sentences drawn from the LDC data and The Little Prince—some sampled randomly, others hand-selected—as well as several sentences constructed to illustrate particular phenomena. only connected substructures may be aligned to a span. Thus, while our aligner is probabilistic like the ISI aligner, it has the advantage of preserving the AMR graph structure. Projection Distance. The scores calculated for an alignment take into account a distance metric designed to encourage locality—tokens that are close together in a sentence are aligned to subtructures that are close together in the AMR graph. We define the projection distance dist(n1,n2) between two neighboring nodes n1 and n2 to be the signed distance in the corresponding sentence between the span aligned to n1 and the span aligned to n2. This motivates the model to prefer alignments whose spans are close together when aligning nodes which are close together—particularly useful when a word occurs twice with identical subgraphs. Thus, our aligner relies on more information from the AMR graph structure than other aligners (note that the ISI system linearizes the graph). Further details are given in §5.2. 5.1 Overview Algorithm 1 illustrates our base algorithm in pseudocode. The likelihood for a sentence can be expressed as a sum of per-span alignment scores: we write the score of a full set of a sentence’s subgraph alignments A as Score(A ∣G,w) = N ∏ i=1 score(⟨gi,si⟩∣G,w) (1) where s are N aligned spans in the sentence w, and g are sets of subgraphs of the AMR graph G aligned to each span. For relations model and the reentrancies model, each gi consists of relations rather than 3315 subgraphs. Henceforth we assume all alignment scores are conditioned on the sentence and graph and omit w and G for brevity. The score(⋅) component of eq. (1) is calculated differently for each of the three models detailed below. Alignment Pipeline. Alignment proceeds in the following phases, with each phase depending on the output of the previous phase: 1. Preprocessing: Using external tools we extract lemmas, parts of speech, and coreference. 2. Span Segmentation: Tokens are grouped into spans using a rule-based procedure (appendix B). 3. Align Subgraphs & Duplicate Subgraphs: We greedily identify subgraph and duplicate subgraph alignments in the same alignment phase (§5.2). 4. Align Relations: Relations not belonging to a subgraph are greedily aligned in this phase, using POS criteria to identify legal candidates (§5.3). 5. Align Reentrancies: Reentrancies are aligned in this phase, using POS and coreference in criteria for identifying legal candidates (§5.4). The three main alignment phases use different models with different parameters; they also have their own preprocessing rules used to identify some alignments heuristically (appendices C to E).5 In training, parameters for each phase are iteratively learned and used to align the entire training set by running EM to convergence before moving on to the next phase. At test time, the pipeline can be run sentence-by-sentence. Decoding. The three main alignment phases all use essentially the same greedy, substructure-aware search procedure. This searches over node–span candidate pairs based on the scoring function modeling the compatibility between a subgraph (or relation) g and span s, which we denote score(⟨g,s⟩). For each unaligned node (or edge), we identify a set of legal candidate alignments using phase-specific criteria. The incremental score improvement of adding each candidate—either extending a subgraph/set of relations already aligned to the span, or adding a completely new alignment—is calculated as as ∆score = score(⟨g0 ∪{n},s⟩)−score(⟨g0,s⟩), where g0 is the current aligned subgraph, s is the span, and n is an AMR component being considered. Of the candidates for all unaligned nodes, the node–span pair giving the best score improvement is then greedily selected to add to the alignment. 579% of nodes and 89% of edges are aligned by rules. We believe this is why in practice, EM performs well without random restarts. This is repeated until all nodes have been aligned (even if the last ones decrease the score). The procedure is detailed in algorithm 1 for subgraphs; the relations phase and the reentrancies phase use different candidates (respectively: unaligned edges; reentrant edges), different criteria for legal candidates, and different scoring functions. 5.2 Aligning Subgraphs The score assigned to an alignment between a span and subgraph is calculated as score(⟨g,s⟩) = Palign(g ∣s;θ1)⋅∏ di∈D Pdist(di;θ2) 1 ∣D∣⋅IB(g,s) (2) where g is a subgraph, s is a span, di is the projection distance of g with its ith neighboring node, and θ1 and θ2 are model parameters which are updated after each iteration. The subgraph g is represented in the model as a bag of concept labels and (parent concept, relation, child concept) triples. The distributions Palign and Pdist are inspired by IBM Model 2 (Brown et al., 1988), and can be thought of as graph-theoretic extensions of translation (align) and alignment (dist) probabilities. IB stands for inductive bias, explained below. Legal Candidates. For each unaligned node n, the model calculates a score for spans of three possible categories: 1) unaligned spans; 2) spans aligned to a neighboring node (in this case, the aligner considers adding n to an existing subgraph if the resulting subgraph would be connected); 3) spans aligned to a node with the same concept as n (this allows the aligner to identify duplicate subgraphs— candidates in this category receive a score penalty because duplicates are quite rare, so they are generally the option of last resort). Limiting the candidate spans in this way ensures only connected, plausible substructures of the AMR are aligned. To form a multinode subgraph alignment t1 →n1 :rel n2, the aligner could first align n1 to an unaligned span t1, then add n2, which is a legal candidate because t1 is aligned to a neighboring node of n2 (ensuring a connected subgraph). Distance. We model the probability of the projection distance Pdist(d;θ2) using a Skellam distribution, which is the difference of two Poisson distributed random variables D = N1−N2 and can be positive or negative valued. Parameters are updated based on alignments in the previous iteration. For each aligned neighbor ni of a subgraph g, we calculate Pdist(dist(g,ni);θ2) and take the geometric mean of probabilities as Pdist. 3316 Algorithm 1 Procedure for greedily aligning all nodes to spans using a scoring function that decomposes over (span, subgraph) pairs. (Scores are expressed in real space but the implementation is in log space.) 1: function ALIGNSUBGRAPHS(spans, amr) 2: alignments ←dict() ▷map from span to an ordered list of aligned subgraphs 3: unaligned_nodes ←get_unaligned_nodes(amr, alignments) 4: while ∣unaligned_nodes∣> 0 do 5: ∆scores ←[] 6: candidate_s_g_pairs ←[] 7: for n ∈unaligned_nodes do 8: candidate_spans ←get_legal_alignments(n, alignments) 9: for span, i_subgraph ∈candidate_spans do ▷either there is an edge between n and the indicated subgraph already aligned to span, or i_subgraph would be a new subgraph consisting of n 10: current_aligned_nodes ←alignments[span][i_subgraph] ▷∅if this would be a new subgraph 11: new_aligned_nodes ←current_aligned_nodes ∪{n} 12: ∆score ←get_score(span, new_aligned_nodes, alignments) 13: −get_score(span, current_aligned_nodes, alignments) ▷change from adding n into a subgraph aligned to span; get_score queries score(⟨g,s⟩) and multiplies λdup if i_subgraph > 1 14: ∆scores.add(∆score) 15: candidate_s_g_pairs.add((span, new_aligned_nodes, i_subgraph)) 16: span∗, subgraph∗, i_subgraph∗←candidate_s_g_pairs[argmax(∆scores)] ▷update having the best impact on score (equivalently, maximizing sum of scores across individual aligned spans) 17: alignments[span∗][i_subgraph∗] ←subgraph∗ 18: unaligned_nodes ←get_unaligned_nodes(amr, alignments) 19: return alignments Null alignment. The aligner models the possibility of a span being unaligned using a fixed heuristic: Palign(∅∣s) = max{rank(s)−1 2 ,0.01} (3) where rank assigns 1 to the most frequent word, 2 to the 2nd most frequent, etc. Thus, the model expects that very common words are more likely to be null-aligned and rare words should almost always be aligned.6 Factorized Backoff. So that the aligner generalizes to unseen subgraph–span pairs, where Palign(g ∣ s) = 0, we use a backoff factorization into components of the subgraph. In particular, the factors are empirical probabilities of (i) an AMR concept given a span string in the sentence, and (ii) a relation and child node concept given the parent node concept and span string. These cooccurrence probabilities pˆ are estimated directly from the training sentence/AMR pairs (irrespective of latent alignments). The product is scaled by a factor λ. E.g., for a subgraph n1 :rel1 n2 :rel2 n3, where cn is the concept of node n, we have Pfactorized(g ∣s) = λ ⋅pˆ(cn1 ∣s)⋅pˆ(:rel1,cn2 ∣cn1,s) ⋅pˆ(:rel2,cn3 ∣cn1,s) (4) Inductive bias. Lastly, to encourage good initialization, the score function includes an inductive 6We allow several exceptions. For punctuation, words in parentheses, and spans that are coreferent to another span, the probability is 0.5. For repeated spans, the probability is 0.1. bias which does not depend on EM-trained parameters. This inductive bias is based on the empirical probability of a node occurring in the same AMR with a span in the training data. We calculate inductive bias as an average of exponentiated PMIs 1 N ∑iexp(PMI(ni,s)), where N is the number of nodes in g, ni is the ith node contained in the subgraph, and PMI is the PMI of ni and s. Aligning Duplicate Subgraphs. On rare occasion a span should be aligned to multiple subgraphs (§3.2). To encourage the model to align a different span where possible, there is a constant penalty λdup for each additional subgraph aligned to a span beyond the first. Thus the score for a span and its subgraphs is computed as: score(⟨g,s⟩) = λ ∣g∣−1 dup ∏ g∈g score(⟨g,s⟩) (5) 5.3 Aligning Relations For a given relation alignment between a span and a collection of edges, we calculate a score as follows: score(⟨a,s⟩) = Palign(a ∣s;θ3)⋅∏ di∈D1 Pdist(di;θ4) 1 ∣D1∣ ⋅∏ d j∈D2 Pdist(dj;θ5) 1 ∣D2∣ (6) where a is the argument structure (the collection of aligned edges), s is a span, D1 is the projection distances of each edge and its parent, and D2 is 3317 Exact Align Partial Align Spans Coverage P R F1 P R F1 F1 Subgraph Alignments (N = 1707) Our system 93.91 94.02 93.97 95.69 95.81 95.75 96.05 100.0 JAMR 87.21 83.06 85.09 90.29 85.99 88.09 92.38 91.1 ISI 71.56 68.24 69.86 78.03 74.54 76.24 86.59 78.7 TAMR (91 sentences) 85.68 83.38 84.51 88.62 86.24 87.41 93.64 94.9 Relation Alignments (N = 1263) Our system 85.67 85.37 85.52 88.74 88.44 88.59 95.41 100.0 ISI 59.28 8.51 14.89 66.32 9.52 16.65 83.09 9.8 Reentrancy Alignments (N = 293) Ours (labeled) 55.75 54.61 55.17 100.0 Ours (unlabeled) 62.72 61.43 62.07 100.0 Duplicate Subgraph Alignments (N = 17) Our system 66.67 58.82 62.50 70.00 61.76 65.62 100.0 Table 3: Main results on the test set. N represents the denominator of exact alignment recall. There are 2860 gold spans in total, 41% of which are null-aligned and 0.6% of which are aligned to multiple subgraphs. 95% of the spans consist of a single token, and 49% of spans are aligned to a single subgraph consisting of a single node. the projection distances of each edge and its child. The collection of edges a is given a normalized label which represents the relations contained in the alignment (distinguishing incoming versus outgoing relations, and normalizing inverse edges). Legal Candidates. There are two kinds of candidate spans for relation alignment. First, previously unaligned spans7 (with no relation or subgraph alignments), e.g. prepositions and subordinating conjunctions such as in →:location or when → :time. Second, any spans aligned to the relation’s parent or child in the subgraph layer: this facilitates alignment of argument structures such as give → :ARG0 :ARG1 :ARG2. Additionally, we constrain certain types of edges to only align with the parent and others to only align with the child. Distance. For relations there are potentially two distances of interest—the projected distance of the relation from its parent and the projected distance of the relation from its child. We model these separately as parent distance and child distance with distinct parameters. To see why this is useful, consider the sentence “Should we meet at the restaurant or at the office?”, where each at token should be aligned to a :location edge. In English, prepositions like at precede an object and follow a governor. Thus parent distance tends to be to the left (negative valued) while child distance tends to be to the right (positive valued). 7We constrain these to particular parts of speech: prepositions (IN), infinitival to (TO), possessives (POS), and possessive pronouns (PRP$). Additionally, only spans that are between the spans aligned to the parent and any descendent of child nodes of the relation (and are not between the child’s aligned span and any of its descendants’ spans) are allowed. This works well in practice for English. 5.4 Aligning Reentrancies The probability of a reentrancy alignment is similar to eq. (6), but with an extra variable for the reentrancy type: score(⟨r,s,type⟩) = Palign(r,type ∣s;θ6)⋅Pdist(d1;θ7)⋅Pdist(d2;θ8) (7) where r is the role label of the reentrant edge. Legal Candidates. There are 8 reentrancy types (§3.4). For each type, a rule-based test determines if a span and edge are permitted to be aligned. The 8 tests use part of speech, the structure of the AMR, and subgraph and relation alignments. A span may be aligned (rarely) to multiple reentrancies, but these alignments are scored separately. 6 Experimental Setup Sentences are preprocessed with the Stanza library (Qi et al., 2020) to obtain lemmas, part-of-speech tags, and named entities. We identify token spans using a combination of named entities and a fixed list of multiword expressions (details are given in appendix B). Coreference information, which is used to identify legal candidates in the reentrancy alignment phase, is obtained using NeuralCoref.8 Lemmas are used in each alignment phase to normalize representation of spans, while parts of speech and coreference are used to restrict legal candidates in the relation and reentrancy alignment phases. We tune hyperparameters, including penalties for duplicate alignments and our factorized backoff probability, on the development set. 8https://github.com/huggingface/neuralcoref 3318 Exact Align P R F1 Relation Alignments Breakdown Our system: all (1163) 85.67 85.37 85.52 . . . single relations (121) 53.49 56.56 54.98 . . . argument structures (1042) 89.67 88.73 89.20 ISI: all (1163) 59.28 8.51 14.89 . . . single relations (121) 82.89 52.07 63.96 . . . argument structures (1042) 39.56 3.45 6.35 Reentrancy Alignments Breakdown Our system: all (293) 62.37 61.09 61.72 . . . primary (128) 79.37 78.12 78.74 . . . coref (41) 57.14 58.54 57.83 . . . control (36) 73.08 52.78 61.29 . . . coordination (29) 57.14 58.54 57.83 . . . pragmatic (25) 20.93 36.00 26.47 . . . adjunct control (15) 100.00 6.67 12.50 . . . repetition (13) 60.00 46.15 52.17 . . . comparative control (5) 0.0 0.0 0.0 . . . unmarked adjunct control (1) 0.0 0.0 0.0 Table 4: Detailed results for relation alignments and reentrancy alignments. 7 Results Table 3 describes our main results on the 200sentence test set (§4), reporting exact-match and partial-match alignment scores as well as span identification F1 and coverage.9 The partial alignment evaluation metric is designed to be more forgiving of arbitrary or slight differences between alignment systems. We argue that this metric is more comparable across alignment systems. It assigns partial credit equal to the product of Jaccard indices ∣N1∩N2∣ ∣N1∪N2∣⋅∣T1∩T2∣ ∣T1∪T2∣for nodes (or edges) and tokens respectively. This partial credit is calculated for each gold alignment and the closest matching predicted alignment with nodes (or edges) N1 and N2 and tokens T1 and T2. Coverage is the percentage of relevant AMR components that are aligned. Our aligner shows improvements over previous aligners in terms of coverage and accuracy even when using a partial credit metric for evaluation. We demonstrate greater coverage, including coverage of phenomena not aligned by previous systems. Table 4 shows detailed results for relation subtypes and reentrancy subtypes. Here, we see room for improvement. In particular, ISI outperforms our system at aligning single relations. Our reentrancy aligner lacks a baseline to compare to, but the breakdown of results by type suggest there are several categories of reentrancies where scores could be improved. Qualitative Analysis. A number of errors from our subgraph aligner resulted from unseen mul9A previous draft of this work reported lower scores on relations before a constraint was added to improve the legal candidates for relation alignment. Ablations Exact Align P R F1 Subgraphs 93.91 94.02 93.97 Subgraphs (−distance) 92.69 92.85 92.77 Subgraphs (−inductive bias) 93.88 93.44 93.66 Relations 85.67 85.37 85.52 Relations (−distance) 85.14 84.77 84.95 Relations (gold subgraphs) 91.21 90.59 90.90 Table 5: Results when the aligner is trained without projection distance probabilities (−distance) and without the subgraph inductive bias (−inductive bias), as well as a relation aligner with access to gold (instead of trained) subgraphs. tiword expressions in our test data that our span preprocessing failed to recognize and our aligner failed to align. For example, the expression “on the one hand” appears in test and should be aligned to contrast-01. The JAMR aligner suffers without a locality bias; we notice several cases where it misaligns words that are repeated in the sentence. The ISI aligner generally does not align very frequent nodes such as person, thing, country, or name, resulting in generally lower coverage. It also frequently aligns disconnected nodes with the same concept to one token instead of separate tokens. While our relation aligner yields significantly higher coverage, we do observe that the model is overeager to align relations to extremely frequent prepositions (such as to and of), resulting in lower precision of single relations in particular. Ablations. Table 5 shows that projection distance is valuable, adding 1.20 points (exact align F1) for subgraph alignment and 0.57 points for relation alignment. Despite showing anecdotal benefits in early experiments, the inductive bias does not aid the model in a statistically significant way. Using gold subgraphs for relation alignment produces an improvement of over 5 points, indicating the scope of error propagation for the relation aligner. 8 Conclusions We demonstrate structure-aware AMR aligners that combine the best parts of rule-based and statistical methods for AMR alignment. We improve on previous systems in terms of accuracy and particularly in terms of alignment coverage and variety of AMR components to be aligned. Acknowledgments We thank reviewers for their thoughtful feedback, Jakob Prange for assisting with annotation, and members of the NERT lab for their support. 3319 References Rafael Anchiêta and Thiago Pardo. 2020. Semantically inspired AMR alignment for the Portuguese language. In Proc. of EMNLP, pages 1595–1600, Online. Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proc. of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178–186, Sofia, Bulgaria. Sebastian Beschke. 2019. Exploring graph-algebraic CCG combinators for syntactic-semantic AMR parsing. In Proc. of RANLP, pages 112–121, Varna, Bulgaria. Austin Blodgett and Nathan Schneider. 2019. An improved approach for semantic graph composition with CCG. In Proc. of the 13th International Conference on Computational Semantics - Long Papers, pages 55–70, Gothenburg, Sweden. P. Brown, J. Cocke, S. Della Pietra, V. Della Pietra, F. Jelinek, R. Mercer, and P. Roossin. 1988. A statistical approach to language translation. In Proc. of COLING, pages 71–76, Budapest, Hungary. Deng Cai and Wai Lam. 2020. AMR parsing via graphsequence iterative inference. In Proc. of ACL, pages 1290–1301, Online. Wei-Te Chen and Martha Palmer. 2017. Unsupervised AMR-dependency parse alignment. In Proc. of EACL, pages 558–567, Valencia, Spain. Chenhui Chu and Sadao Kurohashi. 2016. Supervised syntax-based alignment between English sentences and Abstract Meaning Representation graphs. arXiv:1606.02126 [cs]. Ramón Fernandez Astudillo, Miguel Ballesteros, Tahira Naseem, Austin Blodgett, and Radu Florian. 2020. Transition-based parsing with stackTransformers. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1001–1007, Online. Jeffrey Flanigan, Chris Dyer, Noah A. Smith, and Jaime Carbonell. 2016. CMU at SemEval-2016 Task 8: Graph-based AMR parsing with infinite ramp loss. In Proc. of SemEval, pages 1202–1206, San Diego, California. Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith. 2014. A discriminative graph-based parser for the Abstract Meaning Representation. In Proc. of ACL, pages 1426–1436, Baltimore, Maryland, USA. Jonas Groschwitz, Matthias Lindemann, Meaghan Fowlie, Mark Johnson, and Alexander Koller. 2018. AMR dependency parsing with a typed semantic algebra. In Proc. of ACL, pages 1831–1841, Melbourne, Australia. Kevin Knight, Bianca Badarau, Laura Baranescu, Claire Bonial, Kira Griffitt, Ulf Hermjakob, Daniel Marcu, Tim O’Gorman, Martha Palmer, Nathan Schneider, and Madalina Bardocz. 2020. Abstract Meaning Representation (AMR) Annotation Release 3.0. Technical Report LDC2020T02, Linguistic Data Consortium, Philadelphia, PA. Yijia Liu, Wanxiang Che, Bo Zheng, Bing Qin, and Ting Liu. 2018. An AMR aligner tuned by transition-based parser. In Proc. of EMNLP, pages 2422–2430, Brussels, Belgium. Chunchuan Lyu and Ivan Titov. 2018. AMR parsing as graph prediction with latent alignment. In Proc. of ACL, pages 397–407, Melbourne, Australia. Tahira Naseem, Abhishek Shah, Hui Wan, Radu Florian, Salim Roukos, and Miguel Ballesteros. 2019. Rewarding smatch: transition-based AMR parsing with reinforcement learning. In Proc. of ACL, pages 4586–4592, Florence, Italy. Nima Pourdamghani, Yang Gao, Ulf Hermjakob, and Kevin Knight. 2014. Aligning English strings with Abstract Meaning Representation graphs. In Proc. of EMNLP, pages 425–429, Doha, Qatar. Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A Python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations. Nathan Schneider, Jena D. Hwang, Vivek Srikumar, Jakob Prange, Austin Blodgett, Sarah R. Moeller, Aviram Stern, Adi Bitan, and Omri Abend. 2018. Comprehensive supersense disambiguation of English prepositions and possessives. In Proc. of ACL, pages 185–196, Melbourne, Australia. Nathan Schneider and Noah A. Smith. 2015. A corpus and model integrating multiword expressions and supersenses. In Proc. of NAACL-HLT, pages 1537– 1547, Denver, Colorado. Ida Szubert, Marco Damonte, Shay B. Cohen, and Mark Steedman. 2020. The role of reentrancies in Abstract Meaning Representation parsing. In Proc. of Findings of EMNLP, pages 2198–2207, Online. Ida Szubert, Adam Lopez, and Nathan Schneider. 2018. A structured syntax-semantics interface for EnglishAMR alignment. In Proc. of NAACL-HLT, pages 1169–1180, New Orleans, Louisiana. Chuan Wang and Nianwen Xue. 2017. Getting the most out of AMR Parsing. In Proc. of EMNLP, pages 1257–1268, Copenhagen, Denmark. 3320 A Interannotator Agreement Table 6 illustrates interannotator agreement for each of the four alignment layers. B Identifying Spans As a preprocessing step, sentences have their tokens grouped into spans based on three criteria, outlined in detail below: 1. Named entity spans identified by Stanza. 2. Spans matching multiword expressions from a fixed list of ≈1600 (a) 143 prepositional MWEs from STREUSLE (Schneider and Smith, 2015; Schneider et al., 2018) (b) 348 verbal MWEs from STREUSLE (c) 1095 MWEs taken from gold AMRs in LDC train data (any concept which is a hyphenated compound of multiple words, e.g., alma-mater or whitecollar) and are not present in the above lists. (d) ≈12 hand-added MWEs 3. Any sequence of tokens which is an exact match to a name in the gold AMR (e.g., “United Kingdom” and (n/name :op1 "United" :op2 "Kingdom")) is also treated as a span. C Rule-based Subgraph Alignment Preprocessing C.1 Token matching We use three phases of rule-based alignment which attempt to align particular spans to particular AMR subgraphs: 1. Exact token matching: If there is a unique full string correspondence between a span and a name or number in the AMR, they are aligned. 2. Exact lemma matching: If there is a unique correspondence between an AMR concept and the lemma of a span (which in the case of a multiword span is the sequence of lemmas of the tokens joined by hyphens), they are aligned. 3. Prefix token matching: A span with a prefix match of length 6, 5, or 4 is aligned if it uniquely corresponds to an AMR named entity. 4. Prefix lemma matching: A span with a prefix match of length 6, 5, or 4 of its lemma is aligned if it uniquely corresponds to an concept. 5. English rules: Several hand-written rules for matching English strings to specific subgraphs are used to match constructions such as dates, currency, and some frequent AMR concepts with many different ways of being expressed, such as and and -. • Parsing dates and times • Numbers written out (e.g., one, two, thousand, etc.) • Currencies (e.g., $, C, etc.) • Decades (e.g., twenties, nineties) • and (matching and, additionally, as well, etc.) • multi-sentence (matching punctuation) • :polarity - (matching not, none, never, etc.) • cause-01 (matching thus, since, because, etc.) • amr-unknown (matching ?, who, when, etc.) • person (matching people) • rate-entity-91 (matching daily, weekly, etc.) • "United" "States" (matching US, U.S., American, etc.) • include-91 (matching out of, include, etc.) • instead-of-91 (matching instead, etc.) • have-03 (matching have, ’s, etc.) • mean-01 (matching : and ,) • how (matching :manner thing or :degree so) • as...as (matching equal) C.2 Graph rules We also perform preprocessing to expand a subgraph alignment to include some neighboring nodes. These fall into two main categories: 1. Some AMR concepts are primarily notational rather than linguistic and should be aligned together with a neighboring node. For example named entities (e.g., (country :name (n/name :op1 :United" :op2 "Kingdom"))) are aligned as a unit rather than one node at a time. Likewise, date entities, and subgraphs matching (x/X-quantity :unit X :quant X) or (x/X-entity :value X) are also aligned as a unit. 2. Neighboring nodes which are associated with morphological information of the aligned span (e.g., biggest → (have-degree-91 :ARG1 big :ARG2 most)) are added to the alignment using a series of rules for identifying comparatives, superlatives, polarity, and suffixes such as -er or -able, etc. D Rule-based Relation Alignment Preprocessing Many of the relations are forced to be aligned in a particular way as a matter of convention. We use a similar approach to that of (Groschwitz et al., 3321 IAA Exact Align Partial Align Spans P R F1 P R F1 F1 Subgraphs (366) 94.54 94.54 94.54 95.56 95.56 95.56 94.97 Relations (260) 91.09 90.38 90.73 93.38 92.66 93.02 93.75 Reentrancies (65) 76.92 76.92 76.92 90.00 90.00 90.00 90.77 Duplicates (5) 75.00 60.00 66.67 79.17 63.33 70.37 66.67 Table 6: Interannotator Agreement for subgraph, relation, reentrancy, and duplicate subgraph layers of alignment scored on a sample of 40 sentences of the gold test data. 2018). 1. :ARGX edges are automatically aligned to the same span as the parent (:ARGX-of edges are automatically aligned to the child). 2. :opX edges are automatically aligned with the parent. 3. :sntX edges are automatically aligned with the parent. 4. :domain edges are automatically aligned with the parent. (We don’t align these edges to copula. Instead, a concept with a :domain edge is thought of as a predicate which takes one argument.) 5. :name, :polarity, and :li edges are automatically aligned with the child. D.1 Token matching Some relations take the form :prep-X or :conj-X where X is a preposition or conjunction in the sentence. We use exact match to align these relations as a preprocessing step. The relations :poss and :part may be automatically aligned to ’s or of if the correspondence is unique within a sentence. E Rule-based Reentrancy Alignment Preprocessing Primary edges are identified as a preprocessing step before aligning reentrancies with the following rules: Any relation which is aligned to the same span as its token (any incoming edge which is a part of a span’s argument structure) is automatically made the primary edge. Otherwise, for each edge pointing to a node, we identify the spans aligned to the parent and child nodes in the subgraph layer. Whichever edge has the shortest distance between the span aligned to the parent and the span aligned to the child is identified as the primary edge. In the event of a tie, the edge whose parent is aligned to the leftmost span is identified as the primary edge. Primary reentrancy edges are always aligned to the same span the edge is aligned to in the relation layer of alignments.
|
2021
|
257
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3322–3335 August 1–6, 2021. ©2021 Association for Computational Linguistics 3322 Meta-Learning to Compositionally Generalize Henry Conklin1∗, Bailin Wang1∗, Kenny Smith1 and Ivan Titov1,2 1University of Edinburgh 2University of Amsterdam {henry.conklin, bailin.wang, kenny.smith}@ed.ac.uk, [email protected] Abstract Natural language is compositional; the meaning of a sentence is a function of the meaning of its parts. This property allows humans to create and interpret novel sentences, generalizing robustly outside their prior experience. Neural networks have been shown to struggle with this kind of generalization, in particular performing poorly on tasks designed to assess compositional generalization (i.e. where training and testing distributions differ in ways that would be trivial for a compositional strategy to resolve). Their poor performance on these tasks may in part be due to the nature of supervised learning which assumes training and testing data to be drawn from the same distribution. We implement a meta-learning augmented version of supervised learning whose objective directly optimizes for out-of-distribution generalization. We construct pairs of tasks for meta-learning by sub-sampling existing training data. Each pair of tasks is constructed to contain relevant examples, as determined by a similarity metric, in an effort to inhibit models from memorizing their input. Experimental results on the COGS and SCAN datasets show that our similaritydriven meta-learning can improve generalization performance. 1 Introduction Compositionality is the property of human language that allows for the meaning of a sentence to be constructed from the meaning of its parts and the way in which they are combined (Cann, 1993). By decomposing phrases into known parts we can generalize to novel sentences despite never having encountered them before. In practice this allows us to produce and interpret a functionally limitless number of sentences given finite means (Chomsky, 1965). ∗Equal contribution. Whether or not neural networks can generalize in this way remains unanswered. Prior work asserts that there exist fundamental differences between cognitive and connectionist architectures that makes compositional generalization by the latter unlikely (Fodor and Pylyshyn, 1988). However, recent work has shown these models’ capacity for learning some syntactic properties. Hupkes et al. (2018) show how some architectures can handle hierarchy in an algebraic context and generalize in a limited way to unseen depths and lengths. Work looking at the latent representations learned by deep machine translation systems show how these models seem to extract constituency and syntactic class information from data (Blevins et al., 2018; Belinkov et al., 2018). These results, and the more general fact that neural models perform a variety of NLP tasks with high fidelity (eg. Vaswani et al., 2017; Dong and Lapata, 2016), suggest these models have some sensitivity to syntactic structure and by extension may be able to learn to generalize compositionally. Recently there have been a number of datasets designed to more formally assess connectionist models’ aptitude for compositional generalization (Kim and Linzen, 2020; Lake and Baroni, 2018; Hupkes et al., 2019). These datasets frame the problem of compositional generalization as one of outof-distribution generalization: the model is trained on one distribution and tested on another which differs in ways that would be trivial for a compositional strategy to resolve. A variety of neural network architectures have shown mixed performance across these tasks, failing to show conclusively that connectionist models are reliably capable of generalizing compositionally (Keysers et al., 2020; Lake and Baroni, 2018). Natural language requires a mixture of memorization and generalization (Jiang et al., 2020), memorizing exceptions and atomic concepts with which to generalize. Previous work 3323 looking at compositional generalization has suggested that models may memorize large spans of sentences multiple words in length (Hupkes et al., 2019; Keysers et al., 2020). This practice may not harm in-domain performance, but if at test time the model encounters a sequence of words it has not encountered before it will be unable to interpret it having not learned the atoms (words) that comprise it. Griffiths (2020) looks at the role of limitations in the development of human cognitive mechanisms. Humans’ finite computational ability and limited memory may be central to the emergence of robust generalization strategies like compositionality. A hard upper-bound on the amount we can memorize may be in part what forces us to generalize as we do. Without the same restriction models may prefer a strategy that memorizes large sections of the input potentially inhibiting their ability to compositionally generalize. In a way the difficulty of these models to generalize out of distribution is unsurprising: supervised learning assumes that training and testing data are drawn from the same distribution, and therefore does not necessarily favour strategies that are robust out of distribution. Data necessarily underspecifies for the generalizations that produced it. Accordingly for a given dataset there may be a large number of generalization strategies that are compatible with the data, only some of which will perform well outside of training (D’Amour et al., 2020). It seems connectionist models do not reliably extract the strategies from their training data that generalize well outside of the training distribution. Here we focus on an approach that tries to to introduce a bias during training such that the model arrives at a more robust strategy. To do this we implement a variant of the model agnostic meta-learning algorithm (MAML, Finn et al., 2017a). The approach used here follows Wang et al. (2020a) which implements an objective function that explicitly optimizes for out-ofdistribution generalization in line with Li et al. (2018). Wang et al. (2020a) creates pairs of tasks for each batch (which here we call meta-train and meta-test) by sub-sampling the existing training data. Each meta-train, meta-test task pair is designed to simulate the divergence between training and testing: meta-train is designed to resemble the training distribution, and meta-test to resemble the test distribution. The training objective then requires that update steps taken on meta-train are also beneficial for meta-test. This serves as a kind of regularizer, inhibiting the model from taking update steps that only benefit meta-train. By manipulating the composition of meta-test we can control the nature of the regularization applied. Unlike other meta-learning methods this is not used for few or zero-shot performance. Instead it acts as a kind of meta-augmented supervised learning, that helps the model to generalize robustly outside of its training distribution. The approach taken by Wang et al. (2020a) relies on the knowledge of the test setting. While it does not assume access to the test distribution, it assumes access to the family of test distributions, from which the actual test distribution will be drawn. While substantially less restrictive than the standard iid setting, it still poses a problem if we do not know the test distribution, or if the model is evaluated in a way that does not lend itself to being represented by discrete pairs of tasks (i.e. if test and train differ in a variety of distinct ways). Here we propose a more general approach that aims to generate meta-train, meta-test pairs which are populated with similar (rather than divergent) examples in an effort to inhibit the model from memorizing its input. Similarity is determined by a string or tree kernel so that for each meta-train task a corresponding meta-test task is created from examples deemed similar. By selecting for similar examples we design the meta-test task to include examples with many of the same words as meta-train, but in novel combinations. As our training objective encourages gradient steps that are beneficial for both tasks we expect the model to be less likely to memorize large chunks which are unlikely to occur in both tasks, and therefore generalize more compositionally. This generalizes the approach from Wang et al. (2020a), by using the meta-test task to apply a bias not-strictly related to the test distribution: the design of the meta-test task allows us to design the bias which it applies. It is worth noting that other recent approaches to this problem have leveraged data augmentation to make the training distribution more representative of the test distribution (Andreas, 2020). We believe this line of work is orthogonal to ours as it does not focus on getting a model to generalize compositionally, but rather making the task simple enough that compositional generalization is not needed. Our method is model agnostic, and does not require prior knowledge of 3324 the target distribution. We summarise our contributions as follows: • We approach the problem of compositional generalization with a meta-learning objective that tries to explicitly reduce input memorization using similarity-driven virtual tasks. • We perform experiments on two text-tosemantic compositional datasets: COGS and SCAN. Our new training objectives lead to significant improvements in accuracy over a baseline parser trained with conventional supervised learning. 1 2 Methods We introduce the meta-learning augmented approach to supervised learning from Li et al. (2018); Wang et al. (2020a) that explicitly optimizes for outof-distribution generalization. Central to this approach is the generation of tasks for meta-learning by sub-sampling training data. We introduce three kinds of similarity metrics used to guide the construction of these tasks. 2.1 Problem Definition Compositional Generalization Lake and Baroni (eg. 2018); Kim and Linzen (eg. 2020) introduce datasets designed to assess compositional generalization. These datasets are created by generating synthetic data with different distributions for testing and training. The differences between the distributions are trivially resolved by a compositional strategy. At their core these tasks tend to assess three key components of compositional ability: systematicity, productivity, and primitive application. Systematicity allows for the use of known parts in novel combinations as in (a). Productivity enables generalization to longer sequences than those seen in training as in (b). Primitive application allows for a word only seen in isolation during training to be applied compositionally at test time as in (c). (a) The cat gives the dog a gift →The dog gives the cat a gift (b) The cat gives the dog a gift →The cat gives the dog a gift and the bird a gift (c) made →The cat made the dog a gift 1Our implementations are available at https:// github.com/berlino/tensor2struct-public. Algorithm 1 MAML Training Algorithm Require: Original training set T Require: Learning rate α, Batch size N 1: for step ←1 to T do 2: Sample a random batch from T as a virtual training set Bt 3: Initialize an empty generalization set Bg 4: for i ←1 to N do 5: Sample an example from ˜p(· | Bt[i]) 6: Add it to Bg 7: end for 8: Construct a virtual task τ := (Bt, Bg) 9: Meta-train update: θ′ ←θ −α∇θLBt(θ) 10: Compute meta-test objective: Lτ(θ) = LBt(θ) + LBg(θ′) 11: Final Update: θ ←Update(θ, ∇θLτ(θ)) 12: end for A compositional grammar like the one that generated the data would be able to resolve these three kinds of generalization easily, and therefore performance on these tasks is taken as an indication of a model’s compositional ability. Conventional Supervised Learning The compositional generalization datasets we look at are semantic parsing tasks, mapping between natural language and a formal representation. A usual supervised learning objective for semantic parsing is to minimize the negative log-likelihood of the correct formal representation given a natural language input sentence, i.e. minimising LB(θ) = −1 N N X i=1 log pθ(y|x) (1) where N is the size of batch B, y is a formal representation and x is a natural language sentence. This approach assumes that the training and testing data are independent and identically distributed. Task Distributions Following from Wang et al. (2020a), we utilize a learning algorithm that can enable a parser to benefit from a distribution of virtual tasks, denoted by p(τ), where τ refers to an instance of a virtual compositional generalization task that has its own training and test examples. 2.2 MAML Training Once we have constructed our pairs of virtual tasks we need a training algorithm that encourages 3325 compositional generalization in each. Like Wang et al. (2020a), we turn to optimization-based metalearning algorithms (Finn et al., 2017b; Li et al., 2018) and apply DG-MAML (Domain Generalization with Model-Agnostic Meta-Learning), a variant of MAML (Finn et al., 2017b). Intuitively, DGMAML encourages optimization on meta-training examples to have a positive effect on the meta-test examples as well. During each learning episode of MAML training we randomly sample a task τ which consists of a training batch Bt and a generalization batch Bg and conduct optimization in two steps, namely metatrain and meta-test. Meta-Train The meta-train task is sampled at random from the training data. The model performs one stochastic gradient descent step on this batch θ′ ←θ −α∇θLBt(θ) (2) where α is the meta-train learning rate. Meta-Test The fine-tuned parameters θ′ are evaluated on the accompanying generalization task, meta-test, by computing their loss on it denoted as LBg(θ′). The final objective for a task τ is then to jointly optimize the following: Lτ(θ) = LBt(θ) + LBg(θ′) = LBt(θ) + LBg(θ −α∇θLβ(θ)) (3) The objective now becomes to reduce the joint loss of both the meta-train and meta-test tasks. Optimizing in this way ensures that updates on metatrain are also beneficial to meta-test. The loss on meta-test acts as a constraint on the loss from metatrain. This is unlike traditional supervised learning (Lτ(θ) = LBt(θ) + LBg(θ)) where the loss on one batch does not constrain the loss on another. With a random Bt and Bg, the joint loss function can be seen as a kind of generic regularizer, ensuring that update steps are not overly beneficial to meta-train alone. By constructing Bt and Bg in ways which we expect to be relevant to compositionality, we aim to allow the MAML algorithm to apply specialized regularization during training. Here we design meta-test to be similar to the metatrain task because we believe this highlights the systematicity generalization that is key to compositional ability: selecting for examples comprised of the same atoms but in different arrangements. In constraining each update step with respect to meta-train by performance on similar examples Source Example: The girl changed a sandwich beside the table . Neighbours using Tree Kernel Similarity A sandwich changed . 0.55 The girl changed . 0.55 The block was changed by the girl . 0.39 The girl changed the cake . 0.39 change 0.32 Neighbours using String Kernel The girl rolled a drink beside the table . 0.35 The girl liked a dealer beside the table . 0.35 The girl cleaned a teacher beside the table . 0.35 The girl froze a bear beside the table . 0.35 The girl grew a pencil beside the table . 0.35 Neighbours using LevDistance The girl rolled a drink beside the table . -2.00 The girl liked a dealer beside the table . -2.00 The girl cleaned a teacher beside the table . -2.00 The girl froze a bear beside the table . -2.00 The girl grew a pencil beside the table . -2.00 Table 1: Top scoring examples according to the tree kernel, string kernel and Levenshtein distance for the sentence ‘The girl changed a sandwich beside the table .’ and accompanying scores. in meta-test we expect the model to dis-prefer a strategy that does not also work for meta-test like memorization of whole phrases or large sections of the input. 2.3 Similarity Metrics Ideally, the design of virtual tasks should reflect specific generalization cases for each dataset. However, in practice this requires some prior knowledge of the distribution to which the model will be expected to generalize, which is not always available. Instead we aim to naively structure the virtual tasks to resemble each other. To do this we use a number of similarity measures intended to help select examples which highlight the systematicity of natural language. Inspired by kernel density estimation (Parzen, 1962), we define a relevance distribution for each example: ˜p(x′, y′|x, y) ∝exp k([x, y], [x′, y′]/η (4) where k is the similarity function, [x, y] is a training example, η is a temperature that controls the sharpness of the distribution. Based on our extended interpretation of relevance, a high ˜p implies that [x, y] is systematically relevant to [x′, y′] - containing many of the same atoms but in a novel combination. We look at three similarity metrics to guide subsampling existing training data into meta-test tasks proportional to each example’s ˜p. 3326 Sentence: A rose was helped by Emma . Logical Form: ∃x help′(rose′(x), Emma) Dependency Tree: help rose emma Partial Trees: help rose help emma help rose emma Sentence: A rose was helped by a dog . Logical Form: ∃x,y help′(rose′(x), dog′(y)) Dependency Tree: help rose dog Partial Trees: help rose help dog help rose dog Figure 1: The dependency-tree forms for the logical forms of two sentences. Shown below each tree are its partial trees. As there are three partial trees shared by the examples their un-normalized tree kernel score is 3. Levenshtein Distance First, we consider Levenshtein distance, a kind of edit distance widely used to measure the dissimilarity between strings. We compute the negative Levenshtein distance at the word-level between natural language sentences of two examples: k([x, y], [x′, y′]) = −1 ∗LevDistance(x, x′) (5) where LevDistance returns the number of edit operations required to transform x into x′. See Table 1 for examples. Another family of similarity metrics for discrete structures are convolution kernels (Haussler, 1999). String-Kernel Similarity We use the string subsequence kernel (Lodhi et al., 2002): k([x, y], [x′, y′]) = SSK(x, x′) (6) where SSK computes the number of common subsequences between natural language sentences at the word-level. See Table 1 for examples. 2 2We use the normalized convolution kernels in this work, i.e., k′(x1, x2) = k(x1, x2)/ p k(x1, x1)k(x2, x2) Tree-Kernel Similarity In semantic parsing, the formal representation y usually has a known grammar which can be used to represent it as a tree structure. In light of this we use tree convolution kernels to compute similarity between examples: 3 k([x, y], [x′, y′]) = TreeKernel(y, y′) (7) where the TreeKernel function is a convolution kernel (Collins and Duffy, 2001) applied to trees. Here we consider a particular case where y is represented as a dependency structure, as shown in Figure 1. We use the partial tree kernel (Moschitti, 2006) which is designed for application to dependency trees. For a given dependency tree partial tree kernels generate a series of all possible partial trees: any set of one or more connected nodes. Given two trees the kernel returns the number of partial trees they have in common, interpreted as a similarity score. Compared with string-based similarity, this kernel prefers sentences that share common syntactic sub-structures, some of which are not assigned high scores in string-based similarity metrics, as shown in Table 1. Though tree-structured formal representations are more informative in obtaining relevance, not all logical forms can be represented as tree structures. In SCAN (Lake and Baroni, 2018) y are action sequences without given grammars. As we will show in the experiments, string-based similarity metrics have a broader scope of applications but are less effective than tree kernels in cases where y can be tree-structured. Sampling for Meta-Test Using our kernels we compute the relevance distribution in Eq 4 to construct virtual tasks for MAML training. We show the resulting procedure in Algorithm 1. In order to construct a virtual task τ, a meta-train batch is first sampled at random from the training data (line 2), then the accompanying meta-test batch is created by sampling examples similar to those in meta-train (line 5). We use Lev-MAML, Str-MAML and Tree-MAML to denote the meta-training using Levenshtein distance, string-kernel and tree-kernel similarity, respectively. 3Alternatively, we can use tree edit-distance (Zhang and Shasha, 1989). 3327 3 Experiments 3.1 Datasets and Splits We evaluate our methods on the following semantic parsing benchmarks that target compositional generalization. SCAN contains a set of natural language commands and their corresponding action sequences (Lake and Baroni, 2018). We use the Maximum Compound Divergence (MCD) splits (Keysers et al., 2020), which are created based on the principle of maximizing the divergence between the compound (e.g., patterns of 2 or more action sequences) distributions of the training and test tests. We apply Lev-MAML and Str-MAML to SCAN where similarity measures are applied to the natural language commands. Tree-MAML (which uses a tree kernel) is not applied as the action sequences do not have an underlying dependency tree-structure. COGS contains a diverse set of natural language sentences paired with logical forms based on lambda calculus (Kim and Linzen, 2020). Compared with SCAN, it covers various systematic linguistic abstractions (e.g., passive to active) including examples of lexical and structural generalization, and thus better reflects the compositionality of natural language. In addition to the standard splits of Train/Dev/Test, COGS provides a generalization (Gen) set drawn from a different distribution that specifically assesses compositional generalization. We apply Lev-MAML, Str-MAML and Tree-MAML to COGS; Lev-MAML and StrMAML make use of the natural language sentences while Tree-MAML uses the dependency structures reconstructed from the logical forms. 3.2 Baselines In general, our method is model-agnostic and can be coupled with any semantic parser to improve its compositional generalization. Additionally LevMAML, and Str-MAML are dataset agnostic provided the dataset has a natural language input. In this work, we apply our methods on two widely used sequence-to-sequences models. 4 LSTM-based Seq2Seq has been the backbone of many neural semantic parsers (Dong and Lapata, 2016; Jia and Liang, 2016). It utilizes 4Details of implementations and hyperparameters can be found in the Appendix. LSTM (Hochreiter and Schmidhuber, 1997) and attention (Bahdanau et al., 2014) under an encoderdecoder (Sutskever et al., 2014) framework. Transformer-based Seq2Seq also follows the encoder-decoder framework, but it uses Transformers (Vaswani et al., 2017) to replace the LSTM for encoding and decoding. It has proved successful in many NLP tasks e.g., machine translation. Recently, it has been adapted for semantic parsing (Wang et al., 2020b) with superior performance. We try to see whether our MAML training can improve the compositional generalization of contemporary semantic parsers, compared with standard supervised learning. Moreover, we include a meta-baseline, referred to as Uni-MAML, that constructs meta-train and meta-test splits by uniformly sampling training examples. By comparing with this meta-baseline, we show the effect of similarity-driven construction of meta-learning splits. Note that we do not focus on making comparisons with other methods that feature specialized architectures for SCAN datasets (see Section 5), as these methods do not generalize well to more complex datasets (Furrer et al., 2020). GECA We additionally apply the good enough compositional augmentation (GECA) method laid out in Andreas (2020) to the SCAN MCD splits. Data augmentation of this kind tries to make the training distribution more representative of the test distribution. This approach is distinct from ours which focuses on the training objective, but the two can be combined with better overall performance as we will show. Specifically, we show the results of GECA applied to the MCD splits as well as GECA combined with our Lev-MAML variant. Note that we elect not to apply GECA to COGS, as the time and space complexity 5 of GECA proves very costly for COGS in our preliminary experiments. 3.3 Construction of Virtual Tasks The similarity-driven sampling distribution ˜p in Eq 4 requires computing the similarity between every pair of training examples, which can be very expensive depending on the size of of the dataset. As the sampling distributions are fixed during training, we compute and cache them beforehand. However, they take an excess of disk space to store as essentially we need to store an N × N matrix where N 5See the original paper for details. 3328 Model MCD1 MCD2 MCD3 LSTM 4.7 ±2.2 7.3 ±2.1 1.8 ±0.7 Transformer 0.4 ±0.4 1.8 ±0.4 0.5 ±0.1 T5-base 26.2 ±1.7 7.9 ±1.6 12.1 ±0.1 T5-11B 7.9 2.4 16.8 LSTM 27.4 ±8.2 31.0 ±0.4 9.6 ±3.7 w. Uni-MAML 44.8 ±5.4 31.9 ±3.4 10.0 ±1.4 w. Lev-MAML 47.6 ±2.3 35.2 ±3.9 11.4 ±3.0 w. Str-MAML 42.2 ±2.6 33.6 ±4.3 11.4 ±2.2 Transformer 2.6 ±0.8 3.1 ±1.0 2.3 ±1.3 w. Uni-MAML 2.8 ±0.7 3.2 ±1.0 3.2 ±1.6 w. Lev-MAML 4.7 ±1.8 6.7 ±1.4 6.5 ±1.2 w. Str-MAML 2.8 ±0.6 5.6 ±1.6 6.7 ±1.4 GECA + LSTM 51.5 ±4.4 30.4 ±4.8 12.0 ±6.8 w. Lev-MAML 58.9 ±6.4 34.5 ±2.5 12.3 ±4.9 Table 2: Main results on SCAN MCD splits. We show the mean and variance (95% confidence interval) of 10 runs. Cells with a grey background are results obtained in this paper, whereas cells with a white background are from Furrer et al. (2020). is the number of training examples. To allow efficient storage and sampling, we use the following approximation. First, we found that usually each example only has a small set of neighbours that are relevant to it. 6 Motivated by this observation, we only store the top 1000 relevant neighbours for each example sorted by similarity, and use it to construct the sampling distribution denoted as ˜ptop1000. To allow examples out of top 1000 being sampled, we use a linear interpolation between ˜ptop1000 and a uniform distribution. Specifically, we end up using the following sampling distribution: ˜p(x′, y′|x, y) = λ ˜ptop1000(x′, y′|x, y)+(1−λ) 1 N where ˜ptop1000 assigns 0 probability to out-of top 1000 examples, N is the number of training examples, and λ is a hyperparameter for interpolation. In practice, we set λ to 0.5 in all experiments. To sample from this distribution, we first decide whether the sample is in the top 1000 by sampling from a Bernoulli distribution parameterized by λ. If it is, we use ˜ptop1000 to do the sampling; otherwise, we uniformly sample an example from the training set. 3.4 Development Set Many tasks that assess out-of-distribution (O.O.D.) generalization (e.g. COGS) do not have an O.O.D. 6For example, in COGS, each example only retrieves 3.6% of the whole training set as its neighbours (i.e., have non-zero tree-kernel similarity) on average. Model Gen Dev Test Gen LSTM 99 16 ±8 Transformer 96 35 ±6 LSTM 30.3 ±7.3 99.7 34.5 ±4.5 w. Uni-MAML 36.1 ±6.7 99.7 36.4 ±3.6 w. Lev-MAML 35.6 ±5.3 99.7 36.4 ±5.2 w. Str-MAML 36.3 ±4.2 99.7 36.8 ±3.5 w. Tree-MAML 41.2 ±2.8 99.7 41.0 ±4.9 Transformer 54.7 ±4.0 99.5 58.6 ±3.7 w. Uni-MAML 60.9 ±2.8 99.6 64.4 ±4.0 w. Lev-MAML 62.7 ±3.8 99.7 64.9 ±6.3 w. Str-MAML 62.3 ±3.0 99.6 64.8 ±5.5 w. Tree-MAML 64.1 ±3.2 99.6 66.7 ±4.4 Table 3: Main results on the COGS dataset. We show the mean and variance (standard deviation) of 10 runs. Cells with a grey background are results obtained in this paper, whereas cells with a white background are from Kim and Linzen (2020). Dev set that is representative of the generalization distribution. This is desirable as a parser in principle should never have knowledge of the Gen set during training. In practice though the lack of an O.O.D. Dev set makes model selection extremely difficult and not reproducible. 7 In this work, we propose the following strategy to alleviate this issue: 1) we sample a small subset from the Gen set, denoted as ‘Gen Dev’ for tuning meta-learning hyperparmeters, 2) we use two disjoint sets of random seeds for development and testing respectively, i.e., retraining the selected models from scratch before applying them to the final test set. In this way, we make sure that our tuning is not exploiting the models resulting from specific random seeds: we do not perform random seed tuning. At no point are any of our models trained on the Gen Dev set. 3.5 Main Results On SCAN, as shown in Table 2, Lev-MAML substantially helps both base parsers achieve better performance across three different splits constructed according to the MCD principle. 8 Though our models do not utilize pre-training such as T5 (Raffel et al., 2019), our best model (Lev-MAML + LSTM) still outperforms T5 based models significantly in MCD1 and MCD2. We show that GECA is also effective for MCD splits (especially 7We elaborate on this issue in the Appendix. 8Our base parsers also perform much better than previous methods, likely due to the choice of hyperparameters. 3329 in MCD1). More importantly, augmenting GECA with Lev-MAML further boosts the performance substantially in MCD1 and MCD2, signifying that our MAML training is complementary to GECA to some degree. Table 3 shows our results on COGS. TreeMAML boosts the performance of both LSTM and Transformer base parsers by a large margin: 6.5% and 8.1% respectively in average accuracy. Moreover, Tree-MAML is consistently better than other MAML variants, showing the effectiveness of exploiting tree structures of formal representation to construct virtual tasks. 9 4 Discussion 4.1 SCAN Discussion The application of our string-similarity driven metalearning approaches to the SCAN dataset improved the performance of the LSTM baseline parser. Our results are reported on three splits of the dataset generated according to the maximum compound divergence (MCD) principle. We report results on the only MCD tasks for SCAN as these tasks explicitly focus on the systematicity of language. As such they assess a model’s ability to extract sufficiently atomic concepts from its input, such that it can still recognize those concepts in a new context (i.e. as part of a different compound). To succeed here a model must learn atoms from the training data and apply them compositionally at test time. The improvement in performance our approach achieves on this task suggests that it does disincentivise the model from memorizing large sections - or entire compounds - from its input. GECA applied to the SCAN MCD splits does improve performance of the baseline, however not to the same extent as when applied to other SCAN tasks in Andreas (2020). GECA’s improvement is comparable to our meta-learning method, despite the fact that our method does not leverage any data augmentation. This means that our method achieves high performance by generalizing robustly outside of its training distribution, rather than by making its training data more representative of the test distribution. The application of our LevMAML approach to GECA-augmented data results in further improvements in performance, suggest9The improvement of all of our MAML variants applied to the Transformer are significant (p < 0.03) compared to the baseline, of our methods applied to LSTMs, Tree-MAML is significant (p < 0.01) compared to the baseline. ing that these approaches aid the model in distinct yet complementary ways. 4.2 COGS Discussion All variants of our meta-learning approach improved both the LSTM and Transformer baseline parsers’ performance on the COGS dataset. The Tree-MAML method outperforms the Lev-MAML, Str-MAML, and Uni-MAML versions. The only difference between these methods is the similarity metric used, and so differences in performance must be driven by what each metric selects for. For further analysis of the metrics refer to the appendix. The strong performance of the Uni-MAML variant highlights the usefulness of our approach generally in improving models’ generalization performance. Even without a specially designed metatest task this approach substantially improves on the baseline Transformer model. We see this as evidence that this kind of meta-augmented supervised learning acts as a robust regularizer particularly for tasks requiring out of distribution generalization. Although the Uni-MAML, Lev-MAML, and StrMAML versions perform similarly overall on the COGS dataset they may select for different generalization strategies. The COGS generalization set is comprised of 21 sub-tasks which can be used to better understand the ways in which a model is generalizing (refer to Table 4 for examples of subtask performance). Despite having very similar overall performance Uni-MAML and Str-MAML perform distinctly on individual COGS tasks - with their performance appearing to diverge on a number of of them. This would suggest that the design of the meta-test task may have a substantive impact on the kind of generalization strategy that emerges in the model. For further analysis of COGS sub-task performance see the appendix. Our approaches’ strong results on both of these datasets suggest that it aids compositional generalization generally. However it is worth nothing that both datasets shown here are synthetic, and although COGS endeavours to be similar to natural data, the application of our methods outside of synthetic datasets is important future work. 5 Related Work Compositional Generalization A large body of work on compositional generalization provide models with strong compositional bias, such as specialized neural architectures (Li et al., 2019; Russin 3330 Case Training Generalization Accuracy Distribution Primitive noun →Subject (common noun) shark A shark examined the child. 0.5 1 Baseline Tree-MAML Primitive noun →Subject (proper noun) Paula Paula sketched William. 0.4 0.6 0.8 1 Baseline Tree-MAML Primitive noun →Object (common noun) shark A chief heard the shark. 0 0.2 0.4 Baseline Tree-MAML Primitive noun →Object (proper noun) Paula The child helped Paula. 0 0.5 1 Baseline Tree-MAML Table 4: Accuracy on COGS by generalization case. Each dot represents a single run of the model. et al., 2019; Gordon et al., 2019), or grammar-based models that accommodate alignments between natural language utterances and programs (Shaw et al., 2020; Herzig and Berant, 2020). Another line of work utilizes data augmentation via fixed rules (Andreas, 2020) or a learned network (Akyürek et al., 2020) in an effort to transform the out-of-distribution compositional generalization task into an in-distribution one. Our work follows an orthogonal direction, injecting compositional bias using a specialized training algorithm. A related area of research looks at the emergence of compositional languages, often showing that languages which seem to lack natural-language like compositional structure may still be able to generalize to novel concepts (Kottur et al., 2017; Chaabouni et al., 2020). This may help to explain the ways in which models can generalize robustly on in-distribution data unseen during training while still struggling on tasks specifically targeting compositionality. Meta-Learning for NLP Meta-learning methods (Vinyals et al., 2016; Ravi and Larochelle, 2016; Finn et al., 2017b) that are widely used for few-shot learning, have been adapted for NLP applications like machine translation (Gu et al., 2018) and relation classification (Obamuyide and Vlachos, 2019). In this work, we extend the conventional MAML (Finn et al., 2017b) algorithm, which was initially proposed for few-shot learning, as a tool to inject inductive bias, inspired by Li et al. (2018); Wang et al. (2020a). For compositional generalization, Lake (2019) proposes a meta-learning procedure to train a memory-augmented neural model. However, its meta-learning algorithm is specialized for the SCAN dataset (Lake and Baroni, 2018) and not suitable to more realistic datasets. 6 Conclusion Our work highlights the importance of training objectives that select for robust generalization strategies. The meta-learning augmented approach to supervised learning used here allows for the specification of different constraints on learning through the design of the meta-tasks. Our similarity-driven task design improved on baseline performance on two different compositional generalization datasets, by inhibiting the model’s ability to memorize large sections of its input. Importantly though the overall approach used here is model agnostic, with portions of it (Str-MAML, Lev-MAML, and Uni-MAML) proving dataset agnostic as well requiring only that the input be a natural language sentence. Our methods are simple to implement compared with other approaches to improving compositional generalization, and we look forward to their use in combination with other techniques to further improve models’ compositional ability. Acknowledgements This work was supported in part by the UKRI Centre for Doctoral Training in Natural Language Processing, funded by the UKRI (grant EP/S022481/1) and the University of Edinburgh, School of Informatics and School of Philosophy, Psychology & Language Sciences. We also acknowledge the financial support of the European Research Council (Titov, ERC StG BroadSem 678254) and the Dutch National Science Foundation (Titov, NWO VIDI 639.022.518). References Ekin Akyürek, Afra Feyza Akyürek, and Jacob Andreas. 2020. Learning to recombine and resam3331 ple data for compositional generalization. arXiv preprint arXiv:2010.03706. Jacob Andreas. 2020. Good-enough compositional data augmentation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7556–7566, Online. Association for Computational Linguistics. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. Yonatan Belinkov, Lluís Màrquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2018. Evaluating Layers of Representation in Neural Machine Translation on Part-of-Speech and Semantic Tagging Tasks. arXiv:1801.07772 [cs]. ArXiv: 1801.07772. Terra Blevins, Omer Levy, and Luke Zettlemoyer. 2018. Deep RNNs Encode Soft Hierarchical Syntax. arXiv:1805.04218 [cs]. ArXiv: 1805.04218. Ronnie Cann. 1993. Formal semantics an introduction. Cambridge University Press, Cambridge [etc. OCLC: 1120437841. Rahma Chaabouni, Eugene Kharitonov, Diane Bouchacourt, Emmanuel Dupoux, and Marco Baroni. 2020. Compositionality and generalization in emergent languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4427–4442, Online. Association for Computational Linguistics. Noam Chomsky. 1965. Aspects of the theory of syntax, 50th anniversary edition edition. Number no. 11 in Massachusetts Institute of Technology. Research Laboratory of Electronics. Special technical report. The MIT Press, Cambridge, Massachusetts. Michael Collins and Nigel Duffy. 2001. Convolution kernels for natural language. In Advances in neural information processing systems, pages 625–632. Alexander D’Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D. Hoffman, Farhad Hormozdiari, Neil Houlsby, Shaobo Hou, Ghassen Jerfel, Alan Karthikesalingam, Mario Lucic, Yian Ma, Cory McLean, Diana Mincu, Akinori Mitani, Andrea Montanari, Zachary Nado, Vivek Natarajan, Christopher Nielson, Thomas F. Osborne, Rajiv Raman, Kim Ramasamy, Rory Sayres, Jessica Schrouff, Martin Seneviratne, Shannon Sequeira, Harini Suresh, Victor Veitch, Max Vladymyrov, Xuezhi Wang, Kellie Webster, Steve Yadlowsky, Taedong Yun, Xiaohua Zhai, and D. Sculley. 2020. Underspecification Presents Challenges for Credibility in Modern Machine Learning. arXiv:2011.03395 [cs, stat]. ArXiv: 2011.03395. Li Dong and Mirella Lapata. 2016. Language to logical form with neural attention. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 33–43, Berlin, Germany. Association for Computational Linguistics. Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017a. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. arXiv:1703.03400 [cs]. ArXiv: 1703.03400. Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017b. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1126–1135. JMLR. org. Jerry A. Fodor and Zenon W. Pylyshyn. 1988. Connectionism and cognitive architecture: A critical analysis. Cognition, 28(1-2):3–71. Daniel Furrer, Marc van Zee, Nathan Scales, and Nathanael Schärli. 2020. Compositional generalization in semantic parsing: Pre-training vs. specialized architectures. arXiv preprint arXiv:2007.08970. Jonathan Gordon, David Lopez-Paz, Marco Baroni, and Diane Bouchacourt. 2019. Permutation equivariant models for compositional generalization in language. In International Conference on Learning Representations. Thomas L Griffiths. 2020. Understanding human intelligence through human limitations. Trends in Cognitive Sciences. Jiatao Gu, Yong Wang, Yun Chen, Victor O. K. Li, and Kyunghyun Cho. 2018. Meta-learning for lowresource neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3622–3631, Brussels, Belgium. Association for Computational Linguistics. David Haussler. 1999. Convolution kernels on discrete structures. Technical report, Technical report, Department of Computer Science, University of California .... Jonathan Herzig and Jonathan Berant. 2020. Spanbased semantic parsing for compositional generalization. arXiv preprint arXiv:2009.06040. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Dieuwke Hupkes, Verna Dankers, Mathijs Mul, and Elia Bruni. 2019. The compositionality of neural networks: integrating symbolism and connectionism. arXiv:1908.08351 [cs, stat]. ArXiv: 1908.08351. 3332 Dieuwke Hupkes, Sara Veldhoen, and Willem Zuidema. 2018. Visualisation and’diagnostic classifiers’ reveal how recurrent and recursive neural networks process hierarchical structure. Journal of Artificial Intelligence Research, 61:907–926. Robin Jia and Percy Liang. 2016. Data recombination for neural semantic parsing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12–22, Berlin, Germany. Association for Computational Linguistics. Ziheng Jiang, Chiyuan Zhang, Kunal Talwar, and Michael C. Mozer. 2020. Characterizing Structural Regularities of Labeled Data in Overparameterized Models. arXiv:2002.03206 [cs, stat]. ArXiv: 2002.03206. Daniel Keysers, Nathanael Schärli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, and Olivier Bousquet. 2020. Measuring compositional generalization: A comprehensive method on realistic data. In International Conference on Learning Representations. Najoung Kim and Tal Linzen. 2020. COGS: A Compositional Generalization Challenge Based on Semantic Interpretation. arXiv:2010.05465 [cs]. ArXiv: 2010.05465. Satwik Kottur, José MF Moura, Stefan Lee, and Dhruv Batra. 2017. Natural language does not emerge’naturally’in multi-agent dialog. arXiv preprint arXiv:1706.08502. Brenden Lake and Marco Baroni. 2018. Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks. In International Conference on Machine Learning, pages 2873–2882. PMLR. Brenden M Lake. 2019. Compositional generalization through meta sequence-to-sequence learning. arXiv preprint arXiv:1906.05381. Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. 2018. Learning to generalize: Metalearning for domain generalization. In ThirtySecond AAAI Conference on Artificial Intelligence. Yuanpeng Li, Liang Zhao, Jianyu Wang, and Joel Hestness. 2019. Compositional generalization for primitive substitutions. arXiv preprint arXiv:1910.02612. Huma Lodhi, Craig Saunders, John Shawe-Taylor, Nello Cristianini, and Chris Watkins. 2002. Text classification using string kernels. Journal of Machine Learning Research, 2(Feb):419–444. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412–1421, Lisbon, Portugal. Association for Computational Linguistics. Alessandro Moschitti. 2006. Efficient convolution kernels for dependency and constituent syntactic trees. In European Conference on Machine Learning, pages 318–329. Springer. Abiola Obamuyide and Andreas Vlachos. 2019. Model-agnostic meta-learning for relation classification with limited supervision. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5873–5879, Florence, Italy. Association for Computational Linguistics. Emanuel Parzen. 1962. On estimation of a probability density function and mode. The annals of mathematical statistics, 33(3):1065–1076. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. arXiv preprint arXiv:1912.01703. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683. Sachin Ravi and Hugo Larochelle. 2016. Optimization as a model for few-shot learning. Jake Russin, Jason Jo, Randall C O’Reilly, and Yoshua Bengio. 2019. Compositional generalization in a deep seq2seq model by separating syntax and semantics. arXiv preprint arXiv:1904.09708. Peter Shaw, Ming-Wei Chang, Panupong Pasupat, and Kristina Toutanova. 2020. Compositional generalization and natural language variation: Can a semantic parsing approach handle both? arXiv preprint arXiv:2010.12725. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762. Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. 2016. Matching networks for one shot learning. In Advances in neural information processing systems, pages 3630–3638. Bailin Wang, Mirella Lapata, and Ivan Titov. 2020a. Meta-learning for domain generalization in semantic parsing. arXiv preprint arXiv:2010.11988. 3333 Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020b. RATSQL: Relation-aware schema encoding and linking for text-to-SQL parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567–7578, Online. Association for Computational Linguistics. Kaizhong Zhang and Dennis Shasha. 1989. Simple fast algorithms for the editing distance between trees and related problems. SIAM journal on computing, 18(6):1245–1262. A Experiments A.1 Details of Base Parsers We implemented all models with Pytorch (Paszke et al., 2019). For the LSTM parsers, we use a twolayer encoder and one-layer decoder with attention (Bahdanau et al., 2014) and input-feeding (Luong et al., 2015). We only test bidirectional LSTM encoders, as unidirectional LSTM models do not perform very well in our preliminary experiments. For Transformer parsers, we use 2 encoder and decoder layers, 4 attention heads, and a feed-forward dimension of 1024. The hidden size for both LSTM and Transformer models are 256. The hyparameters of base parsers are mostly borrowed from related work and not tuned, as the primary goal of this work is the MAML training algorithm. To experiment with a wide variety of possible Seq2Seq models, we also try a Transformer encoder + LSTM decoder and find that this variant actually performs slightly better than both vanilla Transformer and LSTM models. Further exploration of this combination in pursuit of a better neural architecture for compositional generalization might be interesting for future work. A.2 Model Selection Protocol In our preliminary experiments on COGS, we find almost all the Seq2Seq models achieve > 99% in accuracy on the original Dev set. However, their performance on the Gen set diverge dramatically, ranging from 10% to 70%. The lack of an informative Dev set makes model selection extremely difficult and difficult to reproduce. This issue might also be one of the factors that results in the large variance of performance reported in previous work. Meanwhile, we found that some random seeds 10 yield consistently better performance than others across different conditions. For example, among 10Random seeds control the initialization of parameters and the order of training batches. the ten random seeds used for Lev-MAML + Transformer on COGS, the best performing seed obtains 73% whereas the lowest performing seed obtains 54%. Thus, it is important to compare different models using the same set of random seeds, and not to tune the random seeds in any model. To alleviate these two concerns, we choose the protocol that is mentioned in the main paper. This protocol helps to make the results reported in our paper reproducible. A.3 Details of Training and Evaluation Following Kim and Linzen (2020), we train all models from scratch using randomly initialized embeddings. For SCAN, models are trained for 1,000 steps with batch size 128. We choose model checkpoints based on their performance on the Dev set. For COGS, models are trained for 6,000 steps with batch size of 128. We choose the meta-train learning rate α in Equation 2, temperature η in Equation 4 based on the performance on the Gen Dev set. Finally we use the chosen α, η to train models with new random seeds, and only the last checkpoints (at step 6,000) are used for evaluation on the Test and Gen set. A.4 Other Splits of SCAN The SCAN dataset contains many splits, such as Add-Jump, Around Right, and Length split, each assessing a particular case of compositional generalization. We think that MCD splits are more representative of compositional generalization due to the nature of the principle of maximum compound divergence. Moreover, it is more challenging than other splits (except the Length split) according to Furrer et al. (2020). That GECA, which obtains 82% in accuracy on JUMP and Around Right splits, only obtains < 52% in accuracy on MCD splits in our experiments confirms that MCD splits are more challenging. A.5 Kernel Analysis The primary difference between the tree-kernel and string-kernel methods is in the diversity of the examples they select for the meta-test task. The tree kernel selects a broader range of lengths, often including atomic examples, a single word in length, matching a word in the original example from metatrain (see table 5). By design the partial tree kernel will always assign a non-zero value to an example that is an atom contained in the original sentence. We believe the diversity of the sentences selected 3334 Partial Tree Kernel top 10 100 1000 Mean Example Length (chars) 26.71 26.59 29.87 Std dev ± 6.80 ± 7.61 ± 8.85 Mean No. of Atoms 0.46 0.81 1.13 Std dev ± 0.67 ± 1.05 ± 0.81 LevDistance top 10 100 1000 Mean Example Length (chars) 31.04 30.45 29.28 Std dev ± 2.80 ± 3.77 ± 4.78 Mean No. of Atoms 0.00 0.00 0.02 Std dev ± 0.00 ± 0.02 ± 0.17 Table 5: Analyses of kernel diversity. Reporting mean example length and number of atoms for the top k highest scoring examples for each kernel. Note that atoms are only counted that also occur in the original example. Source Example: Emma lended the donut to the dog . Neighbours using Tree Kernel Similarity Emma was lended the donut . 0.74 The donut was lended to Emma . 0.62 Emma lended the donut to a dog . 0.55 Emma lended Liam the donut . 0.55 Emma lended a girl the donut . 0.55 Neighbours using String Kernel Emma lended the donut to a dog . 0.61 Emma lended the box to a dog . 0.36 Emma gave the cake to the dog . 0.33 Emma lended the cake to the girl . 0.33 Emma lended the liver to the girl . 0.33 Neighbours using LevDistance Emma lended the donut to a dog . -1.00 Emma loaned the donut to the teacher . -2.00 Emma forwarded the donut to the monster . -2.00 Emma gave the cake to the dog . -2.00 Charlotte lended the donut to the fish . -2.00 Source Example: The crocodile valued that a girl snapped . Neighbours using Tree Kernel Similarity A girl snapped . 0.55 A rose was snapped by a girl . 0.39 The cookie was snapped by a girl . 0.39 girl 0.32 value 0.32 Neighbours using String Kernel The crocodile liked a girl . 0.28 The girl snapped . 0.27 The crocodile hoped that a boy observed a girl . 0.26 The boy hoped that a girl juggled . 0.15 The cat hoped that a girl sketched . 0.15 Neighbours using LevDistance The crocodile liked a girl . -3.00 The boy hoped that a girl juggled . -3.00 The cat hoped that a girl sketched . -3.00 The cat hoped that a girl smiled . -3.00 Emma liked that a girl saw . -4.00 Table 6: Top scoring examples according to the tree kernel, string kernel and Levenshtein distance for two sentences and accompanying scores. by the tree kernel accounts for the superior performance of Tree-MAML compared with the other MAML conditions. The selection of a variety of lengths for meta-test constrains model updates on the meta-train task such that they must also accommodate the diverse and often atomic examples selected for meta-test. This constraint would seem to better inhibit memorizing large spans of the input unlikely to be present in meta-test. A.6 Meta-Test Examples In Table 6, we show top scoring examples retrieved by the similarity metrics for two sentences. We found that in some cases (e.g., the right part of Table 6), the tree-kernel can retrieve examples that diverge in length but are still semantically relevant. In contrast, string-based similarity metrics, especially LevDistance, tends to choose examples with similar lengths. A.7 COGS Subtask Analysis We notice distinct performance for different conditions on the different subtasks from the COGS dataset. In Figure 2 we show the performance of the Uni-MAML and Str-MAML conditions compared with the mean of those conditions. Where the bars are equal to zero the models’ performance on that task is roughly equal. Full task names for figure 2: (1) prim→subj proper, (2) active→passive, (3) only seen as unacc subj →unerg subj, (4) subj→obj proper, (5) only seen as unacc subj →obj omitted transitive subj, (6) pp recursion, (7) cp recursion, (8) obj pp→subj pp, (9) obj→subj common, (10) do dative→pp dative, (11) passive→active, 3335 Figure 2: Performance for the Uni-MAML and LevMAML conditions compared to the mean of those two conditions. (12) only seen as transitive subj →unacc subj, (13) obj omitted transitive→transitive, (14) subj→obj common, (15) prim→obj proper, (16) obj→subj proper, (17) pp dative→do dative, (18) unacc→transitive, (19) prim→subj common, (20) prim→obj common, (21) prim→inf arg.
|
2021
|
258
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3336–3349 August 1–6, 2021. ©2021 Association for Computational Linguistics 3336 Taming Pre-trained Language Models with N-gram Representations for Low-Resource Domain Adaptation Shizhe Diao♦, Ruijia Xu♦, Hongjin Su♣, Yilei Jiang♣ Yan Song♠♥, Tong Zhang♦ ♦The Hong Kong University of Science and Technology {sdiaoaa, rxuaq, tongzhang}@ust.hk ♣The Chinese University of Hong Kong ♠The Chinese University of Hong Kong (Shenzhen) ♥Shenzhen Research Institute of Big Data [email protected] Abstract Large pre-trained models such as BERT are known to improve different downstream NLP tasks, even when such a model is trained on a generic domain. Moreover, recent studies have shown that when large domain-specific corpora are available, continued pre-training on domain-specific data can further improve the performance of in-domain tasks. However, this practice requires significant domainspecific data and computational resources which may not always be available. In this paper, we aim to adapt a generic pretrained model with a relatively small amount of domain-specific data. We demonstrate that by explicitly incorporating the multi-granularity information of unseen and domain-specific words via the adaptation of (word based) ngrams, the performance of a generic pretrained model can be greatly improved. Specifically, we introduce a Transformer-based Domainaware N-gram Adaptor, T-DNA, to effectively learn and incorporate the semantic representation of different combinations of words in the new domain. Experimental results illustrate the effectiveness of T-DNA on eight lowresource downstream tasks from four domains. We show that T-DNA is able to achieve significant improvements compared to existing methods on most tasks using limited data with lower computational costs. Moreover, further analyses demonstrate the importance and effectiveness of both unseen words and the information of different granularities.1 1 Introduction Pre-trained language models have achieved great success and shown promise in various application scenarios across natural language understanding (Devlin et al., 2019; Liu et al., 2019; Tian et al., 2020a) and generation (Lewis et al., 2020; Zhang 1Our code is available at https://github.com/ shizhediao/T-DNA. et al., 2020; Yang et al., 2020). Normally applying pre-trained language models to different applications follows a two-stage paradigm: pre-training on a large unlabeled corpus and then fine-tuning on a downstream task dataset. However, when there are domain gaps between pre-training and fine-tuning data, previous studies (Beltagy et al., 2019; Lee et al., 2020) have observed a performance drop caused by the incapability of generalization to new domains. Towards filling the gaps, the main research stream (Beltagy et al., 2019; Alsentzer et al., 2019; Huang et al., 2019; Lee et al., 2020) on adapting pre-trained language models starts from a generic model (e.g., BERT, RoBERTa) and then continues pre-training with similar objectives on a large-scale domain-specific corpus. However, without providing sufficient understanding of the reason for the performance drop during the domain shift, it is prone to failure of adaptation. Therefore, many aspects of continuous pre-training are expected to be enhanced. First, although generic pre-trained models offer better initialization for continuous pre-training models, it still costs considerable time (and money) that are beyond the reach of many institutions.2 Second, it is clumsy to pre-train domain-specific models repeatedly for each domain on large-scale corpora.3 Therefore, it is helpful to have an efficient and flexible method for being able to adapt pre-trained language models to different domains requiring limited resources. Starting from the observed vocabulary mismatch problem (Gururangan et al., 2020), we further show empirically that the domain gap is largely caused by domain-specific n-grams.4 Motivated by this find2For example, BioBERT (Lee et al., 2020), initialized by generic BERT, was trained on biomedical corpora for 23 days on eight NVIDIA V100 GPUs. 3For example, SciBERT (Beltagy et al., 2019) needs to be trained from scratch if one wants to use a domain-specific vocabulary (i.e., SciVocab in their paper). 4We explain it in detail in the following section. 3337 ing, we propose a light-weight Transformer-based Domain-aware N-gram Adaptor (T-DNA) by incorporating n-gram representations to bridge the domain gap between source and target vocabulary. Specifically, the proposed model is able to explicitly learn and incorporate better representations of domain-specific words and phrases (in the form of n-grams) by the adaptor networks with only requiring small pieces of data. With this adaptor, once entering a new domain, one can choose to train the adaptor alone or train it with a Transformer-based backbone (e.g., BERT) together, where the joint training paradigm could provide more improvement. In addition, although it is designed for a lowresource setting, the adaptor is still able to work with enough data, which ensures its generalization ability in different scenarios. Experimental results demonstrate that T-DNA significantly improves domain adaptation performance based on a generic pre-trained model and outperforms all baselines on eight classification tasks (on eight datasets). The results confirm that incorporating domain-specific n-grams with the proposed T-DNA is an effective and efficient solution to domain adaptation, showing that the information carried by larger text granularity is highly important for language processing across domains. Moreover, further analyses investigate the factors that may influence the performance of our model, such as the amount of available data, the training time cost and efficiency, and the granularity of domain-specific information, revealing the best way and setting for using the model. 2 The Motivation As observed in Gururangan et al. (2020), the transfer gain of domain-specific pre-training becomes increasingly significant when the source and target domain are vastly dissimilar in terms of the vocabulary overlap. Motivated by this association between transfer gain and vocabulary distribution, we further investigate the shift of words and phrases across domains and attempt to alleviate the degradation of language models without large domainspecific corpora. In particular, we start with a RoBERTa-base model from the generic domain and then fine-tune it on the IMDB (Maas et al., 2011) dataset. We investigate the outputs predicted by the [CLS] embedding on the IMDB development set and divide them into two categories: correct predictions (true 1-gram 2-gram 3-gram 4-gram 5-gram Granularity 40 50 60 70 80 90 100 Ratio Label correct false Figure 1: The proportion of domain-specific n-grams in correct predictions and false predictions over 10 different random seeds. positive/negative) and false predictions (false positive/false negative). To examine the vocabulary mismatch problem during the domain shift, we extract the top 1K most frequent n-grams5 from these two categories respectively. We identify the n-grams not in the top 10K most frequent n-grams of source data6 as domain-specific n-grams. As revealed in Figure 1, a larger proportion of domainspecific n-grams are captured when the model is misled to make wrong predictions, which suggests that the shifts in semantic meaning for both words and phrases might account for the domain shift. Furthermore, we conjecture that the representations of domain-specific n-grams are unreliable, which exacerbates the model degradation. While more details will be presented in §6.3, we briefly mention here that the tokens usually improperly attend to other tokens in the sentence but omit the most important words and phrases. In light of this empirical evidence, we are motivated to design a framework to not only capture the domain-specific n-grams but also reliably embed them to extrapolate in the novel domain. 3 The T-DNA Our approach follows the standard recipe of pretraining and fine-tuning a language model, which receives a sentence X = t1t2 · · · ti · · · tT with ti indicating the i-th token, and outputs the representation of each token. The overall architecture of our approach is shown in Figure 2. In the middle, a generic pre-trained encoder, such 5Here we set n to 5. 6We sample a subset from English Wikipedia. 3338 Token Embedding Layer Input Subjective effects, psychomotor task performance, and physiological measures were … sub subjective effects jective effects , ps ych omo tor subjective psychomotor phychomotor task … physiological measures physiological psychomotor task performance task … Tokenization Positional Encoding + Add & Norm Feed Forward Add & Norm Multi-Head Attention ● ● + ! Add & Norm Feed Forward Add & Norm Multi-Head Attention ● ● N-gram Extraction Module N-gram Embedding Layer subjective psychomotor physiological Psychomotor task performance subjective effects … … Domain Lexicon " … !′ Tokenization performance … Figure 2: The overall architecture of our model. as BERT or RoBERTa, provides a representation at the subword-level without any target domain knowledge. The right-hand side shows the proposed T-DNA to enhance the backbone pre-trained encoder, where word based n-grams in X are extracted from a pre-constructed lexicon L, and are represented through n-gram attention module. The left-hand side shows the n-gram matching matrix and the integrating process of domain-specific representation and generic encoding. In this section, we start with a detailed description of lexicon construction, then introduce our n-gram encoding module and how to integrate ngram encoding with the backbone model to get domain-aware representation, and end with an illustration of two training strategies. 3.1 Lexicon Construction and N-gram Extraction To better represent and incorporate unseen and domain-specific n-grams, we first need to find and extract them. Here we propose to use an unsupervised method, pointwise mutual information (PMI), to find domain-specific words and phrases by collocations and associations between words. Given a sentence X = x1x2 · · · xK with K words, for any two adjacent words (e.g., ¯x, ex) within the sentence, their PMI is calculated by PMI(¯x, ex) = log p(¯xex) p(¯x)p(ex), (1) where p(x) is the probability of an n-gram x. When a high PMI score is detected between the adjacent ¯x and ex, it suggests they are good collocation pairs, because they have a high probability of cooccurrence and are more likely to form an n-gram. On the contrary, a delimiter is inserted between the two adjacent words if their PMI(¯x, ex) is less than a threshold σ, i.e., X = x1x2 · · · ¯x/ex · · · xK. As a result, those consecutive words without a delimiter are identified as candidate domain-specific n-grams. After using PMI to segment each sentence in the training set of a target task, we could select among candidate n-grams to obtain the final n-gram lexicon L, where each n-gram appears with a frequency of at least f. In light of this lexicon, for each training input sentence X = t1t2 · · · ti · · · tT with T tokens, where ti denotes the i-th token of X, we extract those sub-strings of X that exist in the lexicon to form domain-specific n-gram sequence S = s1s2, · · · , sj, · · · , sN, with sj indicating the j-th n-gram of X. At the same time, an n-gram matching matrix, M ∈RT×N, can be built to record the 3339 positions of the extracted domain-specific n-gram set and its associated tokens, where mij = 1 for ti ∈sj and mij = 0 for ti /∈sj. The matching matrix is shown in the left hand size of Figure 2. 3.2 Domain-aware Representation The backbone pre-trained encoder is a Transformer architecture (Vaswani et al., 2017) with L layers, S self-attention heads and H hidden dimensions initialized from any pre-trained encoder (e.g., BERT or RoBERTa). The input sentence is passed through it, resulting in a generic hidden state hi for each input token xi. To get the domain-aware hidden representation, the n-gram adaptor network is implemented by a Transformer encoder with l layers, S self-attention heads and H hidden dimensions. First, the embeddings of domain-specific n-grams could be obtained by an n-gram embedding layer and then they are fed into the n-gram encoder to get a sequence of hidden states g via a multi-head attention mechanism. The n-gram encoder is able to model the interactions among all extracted ngrams and dynamically weighs n-grams to emphasize truly useful n-grams and ignores noisy information. The combination of the generic representation and domain-specific n-gram representation are computed by h′ i = hi + X k gi,k, (2) where h′ i is the desired domain-aware representation, and gi,k is the resulting hidden state for the i-th token and the k-th n-gram associated with this token according to the matching matrix M. The ngram encoding process and hidden state integration is repeated layer-by-layer along with the generic encoder for l layers from the bottom. 3.3 Training Strategies Several training strategies could be used and we adopt two in our experiments: fine-tuning (FT) and task-adaptive pre-training (TAPT). For finetuning, we operate on the hidden state of the special classification token [CLS]. Following the tradition citation, we simply add a fully-connected layer as a classifier on top of the model and obtain the probabilities via a softmax layer. The classifier and the whole model are fine-tuned on the labeled task data in the target domain with cross-entropy loss. To inject unsupervised target domain knowledge, we leverage the task-adaptive pre-training proposed in (Gururangan et al., 2020) which strips the labels in downstream task training data and trains the model on this unlabeled data. We use the masked language model (MLM) as our objective and do not include the next sentence prediction (NSP) task following Liu et al. (2019); Lan et al. (2020). Note that, our model also supports other training strategies such as domain-adaptive pre-training, which proves to be effective in Gururangan et al. (2020). One can pre-train our model on a far larger domain corpus (normally beyond 10GB) at the beginning, and then do the task-adaptive pre-training and fine-tuning. Because our main goal is to adapt our model in a low-resource setting in terms of data size and time cost, we leave it for future research.7 4 Experiment Settings In this section, we first introduce eight benchmarking datasets. Then the baseline models, evaluation metrics, and implementation details are presented in the following three subsections, respectively. 4.1 Datasets Following Gururangan et al. (2020), we conduct our experiments on eight classification tasks from four domains including biomedical sciences, computer science, news and reviews. The datasets are described as follows. • CHEMPROT (Kringelum et al., 2016), a manually annotated chemical–protein interaction dataset extracted from 5,031 abstracts for relation classification. • RCT (Dernoncourt and Lee, 2017), which contains approximately 200,000 abstracts from public medicine with the role of each sentence clearly identified. • CITATIONINTENT (Jurgens et al., 2018), which contains around 2,000 citations annotated for their function. • SCIERC (Luan et al., 2018), which consists of 500 scientific abstracts annotated for relation classification. • HYPERPARTISAN (Kiesel et al., 2019), which contains 645 articles from Hyperpartisan news with either extreme left-wing or right-wing standpoint used for partisanship classification. • AGNEWS (Zhang et al., 2015), consisting of 127,600 categorized articles from more than 2000 news source for topic classification. 7We show some analyses and discussion of data size in Section 6.2. 3340 DOMAIN BIOMED CS NEWS REVIEWS DATASET CP RCT CI SE HP AG AM IMDB TRAIN S# 4.1K 1.8K 1.6K 3.2K 516 1.1K 1.1K 2.0K T# 895K 267K 376K 619K 1.7M 213K 1.0M 2.6M O.S# 4.1K 180K 1.6K 3.2K 516 115K 115K 20K O.T# 895K 27.4M 376K 619K 1.7M 21.4M 98.9M 25.9M DEV S# 2.4K 30K 114 455 64 5K 5K 5K T# 547K 4.6M 24K 89K 194K 929K 4.4M 6.6M TEST S# 3.4K 30K 139 974 65 7.6K 25K 25K T# 773K 4.6M 31K 187K 238K 1.4M 21.5M 31.8M CLASSES 13 5 6 7 2 4 2 2 Table 1: The statistics of the eight task datasets in four target domains. To limit the computational resources and maintain all datasets on thousand-level, we only take 10% of IMDB training set, and 1% of RCT, AG and AM training sets. O.S# and O.T# refer to the number of sentences and the number of tokens in the original datasets, respectively. S# denotes the number of sentences and T# is the number of tokens. CP, CI, SE, HP, AG and AM denote CHEMPROT, CITATIONINTENT, SCIERC, HYPERPARTISAN,AGNEWS and AMAZON, respectively. • AMAZON (McAuley et al., 2015), consisting of 145,251 reviews on Women’s and Men’s Clothing & Accessories, each representing users’ implicit feedback on items with a binary label signifying whether the majority of customers found the review helpful. • IMDB (Maas et al., 2011), 50,000 balanced positive and negative reviews from the Internet Movie Database for sentiment classification. To create a low-resource setting, we constrain the size of all datasets into thousand-level. To do so, we randomly select a subset for RCT, AG, Amazon, IMDB with the ratio 1%, 1%, 1%, 10%, respectively. The details can be found in Table 1. 4.2 Baselines In our experiments, the following two models serve as the main baselines. • ROBERTA+FT: fine-tuned off-the-shelf RoBERTa-base model for downstream tasks. • ROBERTA+TAPT: task-adaptive pre-trained on unlabeled task data starting from RoBERTa and then fine-tuned on labeled data. 4.3 Evaluation Metrics Following Beltagy et al. (2019), we adopt macroF1 for CitationIntent, SciERC, HyperPartisan, AGNews, Amazon, IMDB, and micro-F1 for ChemProt and RCT as evaluation metrics. MacroF1 will compute the F1 metric independently for each class and then take the average, whereas micro-F1 will aggregate the contributions of all classes to compute the average metric. In a multi-class classification setup, micro-F1 is preferable if there is class imbalance, which is true for ChemProt and RCT. 4.4 Implementation We implement the RoBERTa-base architecture and initialize it with pre-trained weights by Huggingface’s Transformers library8. In order to obtain a fast and warm start for n-gram representations, we utilize fastText (Bojanowski et al., 2017) to initialize n-gram embeddings. Considering the small amount of data and based on our experience, the number of N-gram encoding layers l is set to 1. For unsupervised task-adaptive pre-training (TAPT), the batch size is set to 16 and training epochs range from 10 to 15. We adopt Adam (Kingma and Ba, 2015) as the optimizer , where the corresponding learning rates of different datasets can be found in our code. The dropout rate is set to 0.5. For the task-specific fine-tuning (FT), we use similar hyperparameter settings and the details are elaborated in the Appendix. All the experiments are implemented on Nvidia V100 GPUs. 5 Experimental Results We compare the performance of the RoBERTa model with and without T-DNA on the aforementioned datasets. In both fine-tuning and task adaptive pre-training experiments, T-DNA shows significant improvements over the pre-trained generic RoBERTa. 8https://github.com/huggingface/transformers 3341 DOMAIN BIOMED CS NEWS REVIEWS DATASET CP RCT CI SE HP AG AM IMDB RoBERTa+FT 81.100.70 80.720.40 56.745.47 74.065.25 88.151.51 88.600.01 63.040.69 92.290.23 +T-DNA 82.660.31 81.520.41 64.954.98 78.612.00 92.490.69 88.910.06 63.920.62 92.910.71 RoBERTa+TAPT 82.241.33 82.730.23 63.442.30 77.851.12 92.700.73 88.840.01 64.130.22 92.770.25 +T-DNA 83.890.76 83.940.27 69.732.87 79.400.48 93.911.48 89.050.03 64.360.34 93.130.15 Table 2: The overall performance of T-DNA and the comparison against existing models on eight target downstream datasts. We report average scores across five random seeds, with standard deviations as subscripts. 5.1 Fine-Tuning The results of fine-tuning on eight datasets are reported in Table 4. In general, the RoBERTa model with T-DNA outperforms that without T-DNA on all datasets, clearly indicating the effectiveness of T-DNA by emphasizing multi-granularity information. On average, T-DNA is able to bring an improvement of performance by around 2.66%. Across all eight datasets, it is observed that TDNA achieves the greatest improvement (8.21%) on the CitationIntent dataset and the least improvement on the AGNews dataset. One reasonable explanation for different improvements is that the domain gap between the RoBERTa pre-training domain and the CS domain is the greatest so that far more gains could be obtained by an effective adaptation strategy. To confirm this, we follow Gururangan et al. (2020) to characterize the domain similarity by analyzing vocabulary overlap and we draw the same conclustion that RoBERTa’s pretraining domain has a similar vocabulary to News and Reviews, but far more dissimilar vocabulary to BioMed and CS. In light of this observation, we recognize that the proposed method is more applicable when the domain gap is large. In this scenario, the potential of incorporating multi-grained information by domain-specific n-grams is greatly exploited to boost the performance of adaptation. When comparing the improvements over four domains, T-DNA is able to offer 1.18%, 6.38%, 2.33%, 0.75% gains on BioMed, CS, News, Reviews, respectively. The improvement on the CS domain is the best while on the Reviews domain it is the poorest, which is consistent with previous analyses across datasets for similar reasons. 5.2 Task-Adaptive Pre-Training In the previous section, we show that T-DNA is helpful in fine-tuning. Additionally, we would like to explore whether T-DNA is complementary to more training strategies, such as task-adaptive pretraining (TAPT). TAPT has been shown useful for 0-gram 1-gram 2-gram 3-gram granularity of n-grams 55 60 65 70 75 80 85 90 performance CP RCT CI SE HP AG AM IMDB Figure 3: Effects of Different Granularities (N=0,1,2,3). pre-trained models in previous studies (Howard and Ruder, 2018; Gururangan et al., 2020), by pretraining on the unlabeled task dataset drawn from the task distribution. The experimental results of two models with and without T-DNA are reported in the bottom two rows in Table 4. From the results, we can clearly see that the model with TDNA achieves better performance on all datasets compared to the generic RoBERTa model without T-DNA. The T-DNA helps to improve the performance by approximately 1.59% on average, which shows that the effectiveness of T-DNA does not vanish when combined with TAPT. Instead, it further leads to a large performance boost for pre-trained models, indicating that T-DNA is a complementary approach, where explicitly modeling domain-specific information helps the unsupervised learning of representations (i.e., the masked language model (MLM) pre-training objective). Overall, for both FT and TAPT experiments, the results show that T-DNA significantly improves domain adaptation performance based on a generic pre-trained model. We attribute this improvement to the essential domain-specific semantic information that is carried by n-grams and the valid representation of n-grams from the T-DNA network. 6 Analyses We analyze several aspects of T-DNA, including the effects of different granularities and the effects 3342 Task RCT AG AM IMDB Model w.o w. w.o w. w.o w. w.o w. 10% 80.78 82.23↑1.45 90.11 92.01↑1.90 63.13 64.10↑0.97 92.29 92.91↑0.62 20% 85.22 86.16↑0.94 91.71 92.14↑0.43 64.01 65.12↑1.11 92.11 92.89↑0.78 50% 87.10 87.69↑0.59 92.17 92.58↑0.41 65.52 66.10↑0.58 93.13 93.32↑0.19 100% 87.31 87.69↑0.38 93.75 94.00↑0.25 66.79 67.14↑0.35 94.34 94.81↑0.47 Table 3: Performance gains of T-DNA w.r.t. different sampling ratios of RCT, AG, AM and IMDB datasets. w. and w.o indicate whether the model is equipped with T-DNA or not. The uparrow marks where a positive gain is obtained. of data size. In addition, we examine the attention mechanism to verify the effects of n-gram representations during the domain shift. The details are illustrated in this section. 6.1 Effects of Different Granularities The lexical unit in RoBERTa is a subword obtained from byte pair encoding (BPE) (Sennrich et al., 2016) tokenization, resulting in a smaller token space and more training data for each token. Our approach provides coarse-grained information carried by the larger lexical units, n-gram. To verify the contribution of larger granularity information, we compare the improvement brought by T-DNA with information of different granularities, for n from 0 to 3. Note that here n means that we extract and incorporate all n-grams with a length smaller or equal to n (within a certain granularity). For example, n = 3 means that we include all unigrams, bigrams and trigrams. Two consistent observations could be made. First, adding only 1-gram is able to bring improvements over 0-gram (i.e., without T-DNA) on all eight datasets, as shown in Figure 3. As we know, the tokens in the generic encoder are at the subword-level and our unigrams are at the word-level, which can be seen as a combination of subwords. Therefore, the results suggest that adding unseen words through our adaptor network is effective, which could enhance the interaction between subwords of the same word, especially for the new words in the target domain. Moreover, based on 1-gram, involving larger granularity offer further gains. Comparing 2-gram and 3-gram v.s. 1-gram, the consistent improvements of T-DNA demonstrate that the potential boundary information presented by n-grams plays an essential role in learning representations by providing explicit and better guidance. 6.2 Effects of Data Size In the previous section, we explored the virtue of incorporating multi-grained information under resource-limited settings, where only a small subset of specific datasets can be accessed. In addition, we are curious whether T-DNA could work well on a larger scale. To this end, we sample different ratios (i.e., 10%, 20%, 50%, 100%) of four datasets (i.e., RCT, AGNews, Amazon and IMDB) and investigate how T-DNA performs at different data scales. As shown in Table 3, the model with T-DNA always outperforms that without T-DNA w.r.t. any subsets of four datasets. This demonstrates that models with T-DNA could easily adapt to any size of dataset with the help of domainspecific n-gram information. However, it is also noted that the performance gains of our method decayed with the increase of the amount of training data, dropping from 1.24% (proportion=10%) to 0.36% (proportion=100%). It is not surprising because with adequate data, a model is able to learn a good representation with supervised learning without the need of prior knowledge. However, since sufficient data normally could not be accessed in reality, especially labeled data, we argue that T-DNA is desirable and necessary for domain adaptation. 6.3 Visualization of N-gram Representations To verify the effects of n-gram representations during the domain shift, we examine the attention mechanism of RoBERTa and T-DNA by plotting the attention maps and salience maps using the LIT tool (Tenney et al., 2020). In the attention map of RoBERTa without T-DNA, we found that the tokens usually improperly attend to other tokens in the sentence. For example, in Figure 4, “Barbie” attributes more attentions to “animated” and “scary” but omits “creepy” and fails to capture “scary as hell” as an integrated phase. In contrast, when the model is equipped with T-DNA, this variant will shift its attention to include “creepy” and 3343 model attention maps and salience maps prediction label RoBERTa positive negative RoBERTa +T-DNA negative negative That creepy animated Barbie is scary as hell ! I want to stop talking about her now That creepy animated Barbie is scary as hell ! I want to stop talking about her now . . That creepy animated Barbie is scary as hell ! I want to stop talking about her now . That creepy animated Barbie is scary as hell ! I want to stop talking about her now That creepy animated Barbie is scary as hell ! I want to stop talking about her now . . That creepy animated Barbie is scary as hell ! I want to stop talking about her now . Figure 4: The visualization of attention maps and salience maps of RoBERTa and RoBERTa+T-DNA. The upper region of each row shows the attention map, where thicker lines denote higher attention weights. The bottom region illustrates the salience map, where the darker color box denotes the more dominant weights for the prediction. force the model to focus on the informative phrase “scary as hell”. Furthermore, the salience map of RoBERTa without T-DNA suggests that “animated” and “scary” dominate its prediction while “creepy” and “scary as hell” are captured by our TDNA, which is consistent with the decision process of human beings. Due to the space limitations, more visualized examples are not shown here. However, based on considerable empirical evidence, we conclude that the unreliable representations of domain-specific n-grams (words and phrases) might be one of the main causes for model degradation. 7 Related Work A large performance drop of pre-trained models caused by domain shift has been observed and many domain-specific BERT models (Beltagy et al., 2019; Alsentzer et al., 2019; Huang et al., 2019; Lee et al., 2020) have been introduced to bridge the domain gap. For example, SciBERT (Beltagy et al., 2019) is trained on 1.14M scientific papers from Semantic Scholar corpus (Ammar et al., 2018) for 7 days on TPU v3-8 machine and BioBERT (Lee et al., 2020) is trained on PubMed abstracts and PMC full text articles for 23 days on eight NVIDIA V100 GPUs. ClinicalBERT (Alsentzer et al., 2019) is trained on about 2 million notes in the MIMIC-III v1.4 database (Johnson et al., 2016) for 17-18 days on a single GeForce GTX TITAN X 12 GB GPU. However, they all incur a huge computational cost, which is not affordable for many university labs or institutions. This is precisely why we believe that our efficient adaptor is useful to the community. Although Gururangan et al. (2020) introduced task-adaptive pre-training (TAPT) to save time by training on unlabeled downstream task data, we demonstrate that our plug-in adaptor is faster and more effective because of the explicit learning strategy and efficient model architecture. Out of vocabulary (OOV) words refer to those words that are not in the vocabulary list and have received a lot of attention in recent years. One way to handle OOV words is to simply utilize and learn an “unknown” embedding during training. Another way is to add in-domain words into the original vocabulary list and learn their representation by pretraining from scratch (Beltagy et al., 2019; Gu et al., 2020), which requires substantial resources and training data. Moreover, SciBERT (Beltagy et al., 2019) found that in-domain vocabulary is helpful but not significant while we attribute it to the inefficiency of implicit learning of in-domain vocabulary. To represent OOV words in multilingual settings, the mixture mapping method (Wang et al., 2019) utilized a mixture of English subwords embedding, but it has been shown useless for domain-specific 3344 words by Tai et al. (2020). ExBERT (Tai et al., 2020) applied an extension module to adapt an augmenting embedding for the in-domain vocabulary but it still needs large continuous pre-training. Similar to our work, they highlight the importance of the domain-specific words but all of these work neither explore the understanding of performance drop during a domain shift nor examine the importance of multi-grained information. Large granularity contextual information carried by spans or n-grams has proven to be helpful to enhance text representation for Chinese (Song et al., 2009; Song and Xia, 2012; Ouyang et al., 2017; Kim et al., 2018; Peng et al., 2018; Higashiyama et al., 2019; Tian et al., 2020e,b; Li et al., 2020; Diao et al., 2020; Song et al., 2021) and English (Joshi et al., 2020; Xiao et al., 2020; Tian et al., 2020c,d). In addition to text encoders on pre-training, the kNN-LM (Khandelwal et al., 2019) proposes to augment the language model for effective domain adaptation, by varying the nearest neighbor datastore of similar contexts without further training. However, all of the previous studies focused on either general pre-training procedures or different tasks (e.g., language modeling), and did not explore the effectiveness of multigrained information for domain adaptation. We hence view them as orthogonal to our work. 8 Conclusion In this work, we first reveal a novel discovery behind the performance drop during a domain shift, demonstrating that an unreliable representation of domain-specific n-grams causes the failure of adaptation. To this end, we propose an innovative adaptor network for generic pre-trained encoders, supporting many training strategies such as taskadaptive pre-training and fine-tuning, both leading to significant improvements to eight classification datasets from four domains (biomedical, computer science, news and reviews). Our method is easy to implement, simple but effective, implying that explicitly representing and incorporating domainspecific n-grams offer large gains. In addition, further analyses consistently demonstrate the importance and effectiveness of both unseen words and the information carried by coarse-grained n-grams. Acknowledgments This work was supported by the General Research Fund (GRF) of Hong Kong (No. 16201320). The authors also want to thank the Sinovation Ventures for their great support. Y. Song was supported by NSFC under the project “The Essential Algorithms and Technologies for Standardized Analytics of Clinical Texts” (12026610) and Shenzhen Institute of Artificial Intelligence and Robotics for Society under the project “Automatic Knowledge Enhanced Natural Language Understanding and Its Applications” (AC01202101001). R. Xu was supported by the Hong Kong PhD Fellowship Scheme (HKPFS). References Emily Alsentzer, John Murphy, William Boag, WeiHung Weng, Di Jindi, Tristan Naumann, and Matthew McDermott. 2019. Publicly Available Clinical BERT Embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop, pages 72–78. Waleed Ammar, Dirk Groeneveld, Chandra Bhagavatula, Iz Beltagy, Miles Crawford, Doug Downey, Jason Dunkelberger, Ahmed Elgohary, Sergey Feldman, Vu Ha, et al. 2018. Construction of the Literature Graph in Semantic Scholar. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 3 (Industry Papers), pages 84–91. Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A Pretrained Language Model for Scientific Text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3606–3611. Piotr Bojanowski, Édouard Grave, Armand Joulin, and Tomáš Mikolov. 2017. Enriching Word Vectors with Subword Information. Transactions of the Association for Computational Linguistics, 5:135–146. Franck Dernoncourt and Ji Young Lee. 2017. PubMed 200k RCT: a Dataset for Sequential Sentence Classification in Medical Abstracts. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 308–313. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Shizhe Diao, Jiaxin Bai, Yan Song, Tong Zhang, and Yonggang Wang. 2020. ZEN: Pre-training Chinese Text Encoder Enhanced by N-gram Representations. In Proceedings of the 2020 Conference on Empirical 3345 Methods in Natural Language Processing: Findings, pages 4729–4740. Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2020. DomainSpecific Language Model Pretraining for Biomedical Natural Language Processing. arXiv e-prints, pages arXiv–2007. Suchin Gururangan, Ana Marasovi´c, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8342–8360. Shohei Higashiyama, Masao Utiyama, Eiichiro Sumita, Masao Ideuchi, Yoshiaki Oida, Yohei Sakamoto, and Isaac Okada. 2019. Incorporating Word Attention into Character-Based Word Segmentation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2699– 2709, Minneapolis, Minnesota. Jeremy Howard and Sebastian Ruder. 2018. Universal Language Model Fine-tuning for Text Classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 328–339, Melbourne, Australia. Kexin Huang, Jaan Altosaar, and Rajesh Ranganath. 2019. ClinicalBERT: Modeling Clinical Notes and Predicting Hospital Readmission. arXiv preprint arXiv:1904.05342. AE Johnson, TJ Pollard, L Shen, LW Lehman, M Feng, M Ghassemi, B Moody, P Szolovits, LA Celi, and RG Mark. 2016. MIMIC-III, a Freely Accessible Critical Care Database. Scientific data, 3:160035– 160035. Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. 2020. Spanbert: Improving Pre-training by Representing and Predicting Spans. Transactions of the Association for Computational Linguistics, 8:64–77. David Jurgens, Srijan Kumar, Raine Hoover, Dan McFarland, and Dan Jurafsky. 2018. Measuring the Evolution of a Scientific Field through Citation Frames. Transactions of the Association for Computational Linguistics, 6:391–406. Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2019. Generalization through Memorization: Nearest Neighbor Language Models. In International Conference on Learning Representations. Johannes Kiesel, Maria Mestre, Rishabh Shukla, Emmanuel Vincent, Payam Adineh, David Corney, Benno Stein, and Martin Potthast. 2019. Semeval2019 Task 4: Hyperpartisan News Detection. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 829–839. Geewook Kim, Kazuki Fukui, and Hidetoshi Shimodaira. 2018. Word-like Character N-gram Embedding. In Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on Noisy Usergenerated Text, pages 148–152. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. In International Conference on Learning Representations. Jens Kringelum, Sonny Kim Kjaerulff, Søren Brunak, Ole Lund, Tudor I Oprea, and Olivier Taboureau. 2016. ChemProt-3.0: a Global Chemical Biology Diseases Mapping. Database, 2016. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. In International Conference on Learning Representations. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. BioBERT: A Pre-Trained Biomedical Language Representation Model for Biomedical Text Mining. Bioinformatics, 36(4):1234–1240. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pretraining for Natural Language Generation, Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880. Xiaonan Li, Hang Yan, Xipeng Qiu, and Xuanjing Huang. 2020. FLAT: Chinese NER using FlatLattice Transformer. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6836–6842, Online. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A Robustly Optimized BERT Pretraining Approach. arXiv preprint arXiv:1907.11692. Yi Luan, Luheng He, Mari Ostendorf, and Hannaneh Hajishirzi. 2018. Multi-Task Identification of Entities, Relations, and Coreference for Scientific Knowledge Graph Construction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3219–3232. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 3346 2011. Learning Word Vectors for Sentiment Analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon, USA. Association for Computational Linguistics. Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. 2015. Image-based Recommendations on Styles and Substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43–52. En Ouyang, Yuxi Li, Ling Jin, Zuofeng Li, and Xiaoyan Zhang. 2017. Exploring N-gram Character Presentation in Bidirectional RNN-CRF for Chinese Clinical Named Entity Recognition. In CEUR Workshop Proc, volume 1976, pages 37–42. Haiyun Peng, Yukun Ma, Yang Li, and Erik Cambria. 2018. Learning Multi-grained Aspect Target Sequence for Chinese Sentiment Analysis. KnowledgeBased Systems, 148:167–176. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural Machine Translation of Rare Words with Subword Units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725. Yan Song, Chunyu Kit, and Xiao Chen. 2009. Transliteration of Name Entity via Improved Statistical Translation on Character Sequences. In Proceedings of the 2009 Named Entities Workshop: Shared Task on Transliteration (NEWS 2009), pages 57–60, Suntec, Singapore. Yan Song and Fei Xia. 2012. Using a Goodness Measurement for Domain Adaptation: A Case Study on Chinese Word Segmentation. In LREC, pages 3853– 3860. Yan Song, Tong Zhang, Yonggang Wang, and Kai-Fu Lee. 2021. ZEN 2.0: Continue Training and Adaption for N-gram Enhanced Text Encoders. arXiv preprint arXiv:2105.01279. Wen Tai, HT Kung, Xin Luna Dong, Marcus Comiter, and Chang-Fu Kuo. 2020. exBERT: Extending Pretrained Models with Domain-specific Vocabulary Under Constrained Training Resources. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1433–1439. Ian Tenney, James Wexler, Jasmijn Bastings, Tolga Bolukbasi, Andy Coenen, Sebastian Gehrmann, Ellen Jiang, Mahima Pushkarna, Carey Radebaugh, Emily Reif, et al. 2020. The Language Interpretability Tool: Extensible, Interactive Visualizations and Analysis for NLP Models. arXiv preprint arXiv:2008.05122. Yuanhe Tian, Yan Song, Xiang Ao, Fei Xia, Xiaojun Quan, Tong Zhang, and Yonggang Wang. 2020a. Joint Chinese Word Segmentation and Partof-Speech Tagging via Two-way Attentions of Autoanalyzed Knowledge. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8286–8296. Yuanhe Tian, Yan Song, Xiang Ao, Fei Xia, Xiaojun Quan, Tong Zhang, and Yonggang Wang. 2020b. Joint Chinese Word Segmentation and Partof-speech Tagging via Two-way Attentions of Autoanalyzed Knowledge. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8286–8296, Online. Yuanhe Tian, Yan Song, and Fei Xia. 2020c. Supertagging combinatory categorial grammar with attentive graph convolutional networks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6037–6044. Yuanhe Tian, Yan Song, Fei Xia, and Tong Zhang. 2020d. Improving Constituency Parsing with Span Attention. In Findings of the 2020 Conference on Empirical Methods in Natural Language Processing. Yuanhe Tian, Yan Song, Fei Xia, Tong Zhang, and Yonggang Wang. 2020e. Improving Chinese Word Segmentation with Wordhood Memory Networks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8274–8285, Online. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. In Advances in neural information processing systems, pages 5998–6008. Hai Wang, Dian Yu, Kai Sun, Jianshu Chen, and Dong Yu. 2019. Improving Pre-Trained Multilingual Model with Vocabulary Expansion. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 316–327. Dongling Xiao, Yu-Kun Li, Han Zhang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. 2020. ERNIE-Gram: Pre-Training with Explicitly N-Gram Masked Language Modeling for Natural Language Understanding. arXiv preprint arXiv:2010.12148. Ze Yang, Wei Wu, Can Xu, Xinnian Liang, Jiaqi Bai, Liran Wang, Wei Wang, and Zhoujun Li. 2020. StyleDGPT: Stylized Response Generation with Pretrained Language Models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1548–1559. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level Convolutional Networks for Text Classification. Advances in neural information processing systems, 28:649–657. 3347 Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B Dolan. 2020. DIALOGPT: Large-Scale Generative Pre-training for Conversational Response Generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270–278. 3348 A Description of Computing Infrastructure All the experiments are implemented on Nvidia V100 GPUs with 32GB memory. B Run Time DOMAIN BIOMED CS NEWS REVIEWS DATASET CP RCT CI SE HP AG AM IMDB RoBERTa+FT 95 40 37 74 50 102 130 114 +T-DNA 93 39 40 72 52 104 131 113 RoBERTa+TAPT 300 132 117 234 285 389 402 392 +T-DNA 320 128 114 240 290 390 400 394 Table 4: Running time per epoch of models, in the unit of second. C Validation Performance DOMAIN BIOMED CS NEWS REVIEWS DATASET CP RCT CI SE HP AG AM IMDB RoBERTa+FT 80.08 81.21 58.06 75.33 93.50 88.70 62.50 93.04 +T-DNA 81.17 82.00 62.98 79.62 91.81 88.64 63.40 92.83 RoBERTa+TAPT 81.27 80.98 60.11 77.08 93.50 88.90 64.30 92.38 +T-DNA 82.58 83.24 67.89 80.69 93.74 89.31 64.27 93.11 Table 5: The validation performance. D Evaluation Measures We use manual tuning and adopt macro-F1 for CitationIntent, SciERC, HyperPartisan, AGNews, Amazon, IMDB, and micro-F1 for ChemProt and RCT as evaluation metrics. Macro-F1 will compute the F1 metric independently for each class and then take the average, whereas micro-F1 will aggregate the contributions of all classes to compute the average metric. In a multi-class classification setup, micro-F1 is preferable if there is class imbalance, which is true for ChemProt and RCT. E Bounds of Hyperparameters Hyperparameter Assaignment number of epochs 3(FT) or 15(TAPT) patience 1 batch size [4,8,16,32,64] learning rate [1e-5,1e-4] dropout 0.5 classification layer [1,2] learning rate optimizer Adam Adam epsilon 1e-8 Adam beta 0.9, 0.999 learning rate optimizer Adam Table 6: Bounds of hyperparameters. 3349 F Configuration of Best Model Hyperparameter Assaignment number of epochs 3(FT) or 15(TAPT) patience 1 batch size 32 learning rate 4e-5 dropout 0.5 classification layer 1 learning rate optimizer Adam Adam epsilon 1e-8 Adam beta 0.9, 0.999 learning rate optimizer Adam Table 7: Configuration of the best model.
|
2021
|
259
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 306–316 August 1–6, 2021. ©2021 Association for Computational Linguistics 306 Exploring the Efficacy of Automatically Generated Counterfactuals for Sentiment Analysis Linyi Yang 1,2,3,4, Jiazheng Li 2, P´adraig Cunningham 2, Yue Zhang 3,4 Barry Smyth 1,2, Ruihai Dong 1,2 1 The Insight Centre for Data Analytics, University College Dublin 2 School of Computer Science, University College Dublin 3 School of Engineering, Westlake University 4 Institute of Advanced Technology, Westlake Institute for Advanced Study {linyi.yang, ruihai.dong, barry.smyth}@insight-centre.org {padraig.cunningham}@ucd.ie {jiazheng.li}@ucdconnect.ie {yue.zhang}@westlake.edu.cn Abstract While state-of-the-art NLP models have been achieving the excellent performance of a wide range of tasks in recent years, important questions are being raised about their robustness and their underlying sensitivity to systematic biases that may exist in their training and test data. Such issues come to be manifest in performance problems when faced with out-ofdistribution data in the field. One recent solution has been to use counterfactually augmented datasets in order to reduce any reliance on spurious patterns that may exist in the original data. Producing high-quality augmented data can be costly and time-consuming as it usually needs to involve human feedback and crowdsourcing efforts. In this work, we propose an alternative by describing and evaluating an approach to automatically generating counterfactual data for the purpose of data augmentation and explanation. A comprehensive evaluation on several different datasets and using a variety of state-of-the-art benchmarks demonstrate how our approach can achieve significant improvements in model performance when compared to models training on the original data and even when compared to models trained with the benefit of human-generated augmented data. 1 Introduction Deep neural models have recently made remarkable advances on sentiment analysis (Devlin et al., 2018; Liu et al., 2019; Yang et al., 2019; Xie et al., 2020). However, their implementation in practical applications still encounters significant challenges. Of particular concern, these models tend to learn intended behavior that is often associated with spurious patterns (artifacts) (Jo and Bengio, 2017; Slack et al., 2020a). As an example, in the sentence “Nolan’s films always shock people, thanks to his superb directing skills”, the most influential word for the prediction of a positive sentiment should be “superb” instead of “Nolan” or “film”. The issue of spurious patterns also partially affects the out-ofdomain (OOD) generalization of the models trained on independent, identical distribution (IID) data, leading to performance decay under distribution shift (Quionero-Candela et al., 2009; Sugiyama and Kawanabe, 2012; Ovadia et al., 2019). Researchers have recently found that such concerns about model performance decay and social bias in NLP come about out-of-domain because of a sensitivity to semantically spurious signals (Gardner et al., 2020), and recent studies have uncovered a problematic tendency for gender bias in sentiment analysis (Zmigrod et al., 2019; Maudslay et al., 2019; Lu et al., 2020). To this end, one of the possible solutions is data augmentation with counterfactual examples (Kaushik et al., 2020) to ensure that models learn real causal associations between the input text and labels. For example, a sentiment-flipped counterfactual of last example could be “Nolan’s movies always bore people, thanks to his poor directorial skills.”. When added to the original set of training data, such kinds of counterfactually augmented data (CAD) have shown their benefits on learning real causal associations and improving the model robustness in recent studies (Kaushik et al., 2020, 2021; Wang and Culotta, 2021). Unlike gradient-based adversarial examples (Wang and Wan, 2019; Zhang et al., 2019; Zang et al., 2020), which cannot provide a clear boundary between positive and negative instances to humans, counterfactuals could provide “human-like” logic to show a modification to the 307 input that makes a difference to the output classification (Byrne, 2019). Recent attempts for generating counterfactual examples (also known as minimal pairs) rely on human-in-the-loop systems. Kaushik et al. (2020) proposed a human-in-the-loop method to generate CAD by employing human annotators to generate sentiment-flipped reviews. The human labeler is asked to make minimal and faithful edits to produce counterfactual reviews. Similarly, Srivastava et al. (2020) presented a framework to leverage strong prior (human) knowledge to understand the possible distribution shifts for a specific machine learning task; they use human commonsense reasoning as a source of information to build a more robust model against spurious patterns. Although useful for reducing sensitivity to spurious correlations, collecting enough high-quality human annotations is costly and time-consuming. The theory behind the ability of CAD to improve model robustness in sentiment analysis is discussed by Kaushik et al. (2021), where researchers present a theoretical characterization of the impact of noise in causal and non-causal features on model generalization. However, methods for automatically generating CAD have received less attention. The only existing approach (Wang and Culotta, 2021) has been tested on the logistic regression model only, despite the fact that recent state-of-the-art methods for sentiment classification are driven by neural models. Also, their automatically generated CAD cannot produce competitive performance compared to human-generated CAD. We believe that their method does not sufficiently leverage the power of pre-trained language models and fails to generate fluent and effective CAD. In addition, the relationships between out-of-domain generalization and sensitivity to spurious patterns were not explicitly investigated by Wang and Culotta (2021). To address these issues, we use four benchmark datasets (IMDB movie reviews as hold-out test while Amazon, Yelp, and Twitter datasets for outof-domain generalization test) to further explore the efficacy of CAD for sentiment analysis. First, we conduct a systematic comparison of several different state-of-the-art models (Wang and Culotta, 2021). This reveals how large Transformerbased models (Vaswani et al., 2017) with larger parameter sizes may improve the resilience of machine learning models. Specifically, we have found that for increasing parameter spaces, CAD’s performance benefit tends to decrease, regardless of whether CAD is controlled manually or automatically. Second, we introduce a novel masked language model for helping improve the fluency and grammar correctness of the generated CAD. Third, we add a fine-tuned model as a discriminator for automatically evaluating the edit-distance, using data generated with minimal and fluent edits (same requirements for human annotators in Kaushik et al. (2020)) to ensure the quality of generated counterfactuals. Experimental results show that it leads to significant prediction benefits using both hold-out tests and generalization tests. To the best of our knowledge, we are the first to automatically generate counterfactuals for use as augmented data to improve the robustness of neural classifiers, which can outperform existing, state-ofthe-art, human-in-the-loop approaches. We will release our code and datasets on GitHub 1. 2 Related Work This work mainly touches on three important areas: approaches to evaluation that go beyond traditional accuracy measures (Bender and Koller, 2020; Warstadt et al., 2020), the importance of counterfactuals in eXplainable AI (XAI) (Byrne, 2019; Keane and Smyth, 2020), and out-of-domain generalization in sentiment analysis (Kim and Hovy, 2004; Zhang et al., 2018; Zhang and Zhang, 2019). There has been an increasing interest in the role of Robustness Causal Thinking in ML, often by leveraging human feedback. Recently, some of the standard benchmark datasets have been challenged (Gardner et al., 2020; Ribeiro et al., 2020), in which the model performance is significantly lower on contrast sets than on original test sets; a difference of up to 25% in some cases. Researchers propose counterfactual data augmentation approaches for building robust models (Maudslay et al., 2019; Zmigrod et al., 2019; Lu et al., 2020), and find that spurious correlations threaten the model’s validity and reliability. In an attempt to address this problem, Kaushik et al. (2020) explore opportunities for developing human-in-the-loop systems by using crowd-sourcing to generate counterfactual data from original data, for data augmentation. Teney et al. (2020) shows the continuous effectiveness of CAD in computer vision (CV) and NLP. The idea of generating Counterfactuals in XAI 1https://github.com/lijiazheng99/Counterfactuals-forSentiment-Analysis 308 Original dataset Identify causal terms with MLM Classifiers Sentiment Dictionary Hierarchical RM-CT Hierarchical REP-CT MoverScore MoverScore is used to control the minimal edits of the automatically generated counterfactuals Classifiers Counterfactually augmented dataset Original Dataset Sampled Data points Human Annotators Human-generated Counterfactuals The approved counterfactual is added to the original dataset Our Method Kaushik, Hovy, and Lipton (2020) Wang and Culotta (2021) Original Dataset Identify likely causal features Replace causal features with antonyms Based on PyDictionary Original dataset Classifiers Counterfactually augmented dataset Our Method Sentiment Dictionary Hierarchical RM-CT Hierarchical REP-CT Identify causal terms with SCD MoverScore Classifiers MoverScore is used to control the minimal edits of the automatically generated CAD. Figure 1: Overview of previous CAD methods are shown on the left side, while the pipeline of our method is shown on the right. Hierarchical RM-CT (removing the casual terms) and Hierarchical REP-CT (replacing the casual terms) are our methods for automatically generating CAD, respectively. SCD denotes sampling and sensitivity of contextual decomposition. Sentiment Dictionary refers to the opinion lexicon published by (Hu and Liu, 2004). also shares important conceptual features with our work. Since human counterfactual explanations are minimal in the sense that they select a few relevant causes (Byrne, 2019; Keane and Smyth, 2020) as is the requirement of minimal edits in our generation process. This has been explored more in the field of CV (Goyal et al., 2019; Kenny and Keane, 2021), but investigated less in NLP. Recent work (Jacovi and Goldberg, 2020) highlight explanations of a given causal format, and Yang et al. (2020a) generate counterfactuals for explaining the prediction of financial text classification. We propose a similar but different research question, that is, whether the automatically generated counterfactual can be used for data augmentation to build more robust models, which has not been considered by the previous methods in XAI (Pedreschi et al., 2019; Slack et al., 2020b; Yang et al., 2020b; Ding et al., 2020). In the case of Sentiment Analysis, most of the previous works report experiments using a holdout test on the IID dataset (Liu, 2012; Yang et al., 2016; Johnson and Zhang, 2017). The current stateof-the-art methods make use of large pre-trained language models (e.g., BERT (Devlin et al., 2018), RoBERTa (Liu et al., 2019) and SMART-RoBERTa (Jiang et al., 2020)) for calculating input represntations. It has been shown that these methods can suffer from spurious patterns (Kaushik et al., 2020; Wang and Culotta, 2021). Very recently, Wang and Culotta (2021) provide a starting point for exploring the efficacy of automatically generated CAD for sentiment analysis, but it is still based on IID hold-out tests only. However, spurious patterns in the training and test sets could be tightly coupled, which may limit the possibility of observing their attendant accuracy issues using a hold-out test methodology. For this reason, we designed an indirect method for evaluating the robustness of models, by comparing the performance of models trained on original and augmented data using out-of-domain data. The prediction benefit for out-of-domain data should provide some evidence about whether a model’s sensitivity to spurious patterns has been successfully mitigated. The resulting counterfactuals can be used for data augmentation and can also provide contrastive explanations for classifiers, and important and desirable consideration for the recent move towards more XAI (Ribeiro et al., 2016; Lundberg and Lee, 2017; Lipton, 2018; Pedreschi et al., 2019; Slack et al., 2020b). 3 Detailed Implementation We propose a new approach for automatically generating counterfactuals to enhance the robustness of sentiment analysis models by inverting the sentiment of causally important terms according to Algorithm 1 and based on the following stages: 1. The identification of genuine causal terms using self-supervised contextual decomposition (Section 3.1). 2. Generating counterfactual samples by (a) RMCT (removing causal terms) and (b) REP-CT (replacing the causal terms) (Section 3.2). 3. Selecting the human-like counterfactuals using MoverScore. (Zhao et al., 2019) (Section 3.3). The end result will be a set of counterfactuals that can be used to augment an existing dataset. 309 3.1 Identifying Causal Terms To identify causally important terms, we propose a hierarchical method, based on the sampling and sensitivity of contextual decomposition technique from Jin et al. (2019), by incrementally removing words from a sentence in order to evaluate the model’s sensitivity to these words. Significant changes in model outputs suggest the removal of important terms. For example, removing the word “best” from “The movie is the best that I have ever seen.”, is likely to alter a model’s sentiment prediction more than the removal of other words from the sentence; thus “best” is an important word with respect to this sentence’s sentiment. In a similar way, phrases beginning with negative pronouns will likely be important; for instance, “not satisfy you” is important in “This movie could not satisfy you”. Given a word (or phrase starting with negative limitations) w in the sentence s, the importance of w can be calculated as in Equation 1 where s β\p denotes the sentence that resulting after masking out a single word (or a negative phrase as above). We use l (s β\p;bs) to represent the model prediction after replacing the masked-out context, while bsβ is a input sequence sampled from the input s. \p indicates the operation of masking out the phrase p in a input document D from the training set. The specific candidate causal terms found by this masking operation vary for different prediction models. φ(w,bs) = Esβ l (s β; bsβ) −l (s β\p; bsβ) l (s β; bsβ) (1) 3.2 Generating Human-like Counterfactuals This approach and the scoring function in Equation 1 is used in Algorithm 1 in two ways, to generate two types of plausible counterfactuals. First, it is used to identify words to remove from a sentence to produce a plausible counterfactual. This is referred to as RM-CT and is performed by lines 3–5 in Algorithm 1; for a sentence S(i), it’s correctly labeled sentiment words are identified (line 3), and sorted based on Equation 1 (line 4) with classifier C, and the most important of these words is removed from S(i) to produce S(i) rm (line 5). Second, the REP-CT technique instead replaces each causally important sentiment word in S(i) with an alternative word that has an opposing sentiment polarity (lines 6-11 in Algorithm 1). To do this the words in S(i) are each considered for replacement in order of their importance (lines 6 & 7) Algorithm 1 Generating plausible counterfactual instances. Input: Test document D(n)= {P1, P2, ..., Pn}, with corresponding ground-truth labels Y, pre-trained Mask Language Model MLM, fine-tuned transformer classifier C, Positive Word Dictionaries POS, Negative Word Dictionaries NEG. (pos and neg are predicates for positive and negative labels) Output: Plausible counterfactual D(k) cf = {D(k) rep, D(k) rm} 1: for Pk in D(n) do 2: for S(i), Yi in Pk do 3: bS(i) ← w ∈S(i) | (w ∈POS ∧Yi = pos) ∨(w ∈NEG ∧Yi = neg) 4: S(i) sorted ←sort bS(i), key = φ(w, bS(i)) (eq.1) 5: S(i) rm ←S(i) sorted[1 :] 6: S(i) rep ←S(i) sorted 7: for w ∈S(i) rep do 8: Wp ←MLM S(i) mask(w), S(i) rep 9: Wc ←{w ∈Wp | (w ∈POS ∧Yi! = pos) ∨(w ∈NEG ∧Yi! = neg) 10: S(i) rep(w) ←sort Wc, key = φ(w, Wc) [0] 11: end for 12: P (k) rm ←P (k) rm + S(i) rm 13: P (k) rep ←P (k) rep + S(i) rep 14: end for 15: D(n) rm ←D(n) rm + P (k) rm 16: D(n) rep ←D(n) rep + P (k) rep 17: end for 18: return D(n) rm, D(n) rep to create a new sentence S(i) rep. For each word w we use a masked language model (MLM) to generate a set of plausible replacements, Wp (line 8), and a subset of these, Wc, as replacement candidates if their sentiment is different from the sentiment of S(i), which is given by Yi (line 9). Here we are using the BERT-base-uncased as the pre-trained MLM for SVM and BiLSTM models 1. The size of candidate substitutions found by MLM output is set to 100 for all models.Then, Wc is sorted in descending order of importance using Equation 1 and the most important candidate is selected and used to replace w in S(i) rep (line 10). Algorithm 1 continues in this fashion to generate counterfactual sentences using RM-CT and REP-CT for each sentence in each paragraph of the target document 2. It returns two counterfactual documents, which correspond to documents produced from the RM-CT and REP-CT sentences; see lines 15–18. The above approach is not guaranteed to always generate counterfactuals. Typically, reviews that 1For Transformers-based models, we use their own pretrained MLM (e.g., RoBERTa and XLNet) as the generator. 2Generating one counterfactual edit for an IMDB instance takes an average of ≈3.4 seconds based on the RoBERTaLarge model. 310 cannot be transformed into plausible counterfactuals contain spurious associations that interfere with the model’s predictions. For example, in our method, the negative review “The film is pretty bad, and her performance is overacted” will be first modified as “The film is pretty good, and her performance is lifelike”. The revised review’s prediction will remain negative. Meanwhile, the word “her” will be identified as a potential causal term. To alleviate this problem, we further conduct the substitution of synonyms for those instances that have been already modified with antonym substitution by using causal terms. As an example, we will continue replacing the word “her” with “their” until the prediction has been flipped; see also Zmigrod et al. (2019) for related ideas. In conclusion, then, the final augmented dataset that is produced of three parts: (1) counterfactuals generated by RM-CT; (2) counterfactuals generated by REP-CT; (3) adversarial examples generated by synonym substitutions. 3.3 Ensuring Minimal Changes When generating plausible counterfactuals, it is desirable to make minimal changes so that the resulting counterfactual is as similar as possible to the original instance (Miller, 2019; Keane and Smyth, 2020). To evaluate this for the approach described we use the MoverScore (Zhao et al., 2019) – an edit-distance scoring metric originally designed for machine translation – which confirms that the MoverScore for the automatic CAD instances is marginally higher when compared to human-generated counterfactuals, indicated greater similarity between counterfactuals and their original instances. The MoverScore between humangenerated counterfactuals and original reviews is 0.74 on average (minimum value of 0.55) and our augmented data results in a slightly higher average score than human-generated data for all models. The generated counterfactuals and synonym substitutions that achieve a MoverScore above 0.55 are combined with the original dataset for training robust classifiers. 4 Datasets Our evaluation uses three different kinds of datasets, in-domain data, challenge data, and outof-domain data. State-of-the-art Models SST-2 IMDB SMART-RoBERTa (Jiang et al., 2020) 97.5 96.3 RoBERTa-Large (Liu et al., 2019) 96.7 96.3 RTC-attention (Zhang and Zhang, 2019) 90.3 88.7 Bi-LSTM 86.7 86.0 Table 1: The performance of state-of-the-art models in sentiment analysis. 4.1 In-domain Data We first adopt two of the most popular benchmark datasets – SST-2 and IMDB (Maas et al., 2011) – to show the recent advances on sentiment analysis with the benefit of pre-trained models. However, we mainly focus on the robustness of various models for sentiment analysis in this work, rather than in-domain accuracy. Hence, following Wang and Culotta (2021) and Kaushik et al. (2020), we perform binary sentiment classification experiments on the IMDB dataset sampled from Maas et al. (2011) that contains 1707 training, 245 validation, and 488 testing examples with challenge dataset (paired counterfactuals). 4.2 Challenge Data Based on the in-domain IMDB data, Kaushik et al. (2020) employ crowd workers not to label documents, but to revise movie review to reverse its sentiment, without making any gratuitous changes. We directly use human-generated counterfactuals by Kaushik et al. (2020) as our challenge data, enforcing a 50:50 class balance. 4.3 Out-of-domain Data We also evaluate our method on different out-ofdomain datasets, including Amazon reviews (Ni et al., 2019) from six genres: beauty, fashion, appliances, gift cards, magazines, and software, a Yelp review dataset, and the Semeval-2017 Twitter dataset (Rosenthal et al., 2017). These have all been sampled to provide a 50:50 label split. The size of the training data has been kept the same for all methods, and the results reported are the average from five runs to facilitate a direct comparison with baselines (Kaushik et al., 2020, 2021). 5 Results and Discussions We first describe the performance of the current state-of-the-art methods on sentiment analysis based on the SST-2 and IMDB benchmark datasets. Next, we will discuss the performance benefits by using our automatically generated counterfactuals 311 Models Parameter Training / Testing data AC: (Our method) O/O CF/O CF/CF O/CF C/O AC/O C/CF AC/CF SVM(TF-IDF) 80.0 58.3 91.2 51.0 83.7 84.8 87.3 86.1 Bi-LSTM 0.2M 79.3 62.5 89.1 55.7 81.5 82.2 92.0 88.5 Transformer-based Models BERT [ICLR,2021] 110M 87.4 80.4 90.8 82.2 88.5 90.6 95.1 92.2 WWM-BERT-Large 335M 91.2 86.9 96.9 93.0 91.0 91.8 95.3 94.1 XLNet-Large 340M 95.3 90.8 98.0 93.9 93.9 94.9 96.9 95.5 RoBERTa-Large 355M 93.4 91.6 96.9 93.0 93.6 94.1 96.7 94.3 Table 2: The accuracy of various models for sentiment analysis using different datasets, including the humangenerated counterfactual data and counterfactual samples generated by our pipeline. O denotes the original IMDB review dataset, CF represents the human-revised counterfactual samples, C denotes the combined dataset consisting of original and human-revised dataset, and AC denotes the original dataset combined with automatically generated counterfactuals. C and AC contain the same size of training samples (3.4K). Original Samples Original Robust Nolan’s film...superb directing skills (POS) superb:0.213 film:0.446 Nolan:0.028 0.627 0.019 0.029 It’s a poor film, but I must give it to the lead actress in this one (NEG) poor:-0.551 film:-0.257 actress:-0.02 -0.999 -7e-7 -1e-6 Table 3: Less sensitivity to spurious patterns has been shown in the robust BERT-base-uncased model. on an in-domain test. We further compare our method, human-label method, and two state-of-theart style-transfer methods (Sudhakar et al., 2019; Madaan et al., 2020) in terms of the model robustness on generalization test. Notably, we provide an ablation study lastly to discuss the influence of edit-distance for performance benefits. 5.1 State-of-the-art Models As the human-generated counterfactuals (Kaushik et al., 2020) are sampled from Maas et al. (2011), the results in Table 1 cannot be directly compared with Table 2 3. As shown in Table 1, by comparing BiLSTM to Transformer-base methods, it can be seen that remarkable advances in sentiment analysis have been achieved in recent years. On SST-2, SMART-RoBERTa (Jiang et al., 2020) outperforms Bi-LSTM by 10.8% (97.5% vs. 86.7%) accuracy, where a similar improvement is observed on IMDB (96.3% vs. 86.0%). According to the results, we select the following models for our experiments, which covers a spectrum of statistical, neural and pre-trained neural methods: SVM (Suykens and Vandewalle, 1999), Bi-LSTM (Graves and Schmidhuber, 2005), BERTBase (Devlin et al., 2018), RoBERTa-Large (Liu et al., 2019), and XLNet-Large (Yang et al., 2019). 3We can only get the human-generated counterfactual examples (Kaushik et al., 2020) sampled from the IMDB dataset. The SVM model for sentiment analysis is from scikit-learn and uses TF-IDF (Term FrequencyInverse Document Frequency) scores, while the Transformer-based models are built based on the Pytorch-Transformer package 4. We keep the prediction models the same as Kaushik et al. (2020), except for Naive Bayes, which has been abandoned due to its high-variance performance shown in our experiments. In the following experiments, we only care about whether the robustness of models has been improved when training on the augmented dataset (original data & CAD). Different counterfactual examples have been generated for different models in terms of their own causal terms in practice, while the hyper-parameters for different prediction models are all identified using a grid search conducted over the validation set. 5.2 Comparison with Original Data On the Influence of Spurious Patterns. As shown in Table 2, we find that the linear model (SVM) trained on the original and challenge (human-generated counterfactuals) data can achieve 80% and 91.2% accuracy testing on the IID hold-out data, respectively. However, the accuracy of the SVM model trained on the original set when testing on the challenge data drops dramatically (91.2% vs. 51%), and vice versa (80% vs. 58.3%). Similar findings were reported by Kaushik et al. (2020), where a similar pattern was observed in the Bi-LSTM model and BERT-base model. This provides further evidence supporting the idea that the spurious association in machine learning models is harmful to the performance on the challenge set for sentiment analysis. 4https://github.com/huggingface/ pytorch-transformers 312 On the Benefits of Robust BERT. As shown in Table 3, we also test whether the sensitivity to spurious patterns has been eliminated in the robust BERT model. We notice that the correlations of the real causal association “superb” and “poor” are improved from 0.213 to 0.627 and -0.551 to -0.999, respectively. While the correlation of spurious association “film” is decreased from 0.446 to 0.019 and -0.257 to -7e-7 on positive and the negative samples, respectively. This shows that the model trained with our CAD data does provide robustness against spurious patterns. On the Influence of Model Size. Previous works (Kaushik et al., 2021; Wang and Culotta, 2021) have not investigated the performance benefits on larger pre-trained models. While we further conduct experiments on various Transformer-based models with different parameter sizes to explore whether the larger transformer-based models can still enjoy the performance benefits of CAD (Table 2). We observe that although the test result can increase with the parameter size increasing (best for 94.9% using XLNet), the performance benefits brought by human-generated CAD and the autogenerated CAD declines continuously with the parameter size increase. For example, the BERT-baseuncased model trained on the auto-generated combined dataset can receive 3.2% (90.6% vs. 87.4%) improvement on accuracy while performance increases only 0.6% (91.8% vs. 91.2%) on accuracy for WWM-BERT-Large. It suggests that larger pretrained Transformer models may be less sensitive to spurious patterns. 5.3 Comparison with Human CAD Robustness in the In-domain Test. We can see that all of the models trained on automatic CAD – shown as AC in the Table 2 – can outperform the human-generated CAD varying with the models (AC/O vs. C/O) as follows: SVM (+1.1%), Bi-LSTM (+0.7%), BERT-base-uncased (+2.1%), BERT-Large (+0.8%), XLNet-Large (+1.0%), and RoBERTa-Large (+0.5%) when testing on the original data. If we adopt the automatic CAD (AC), we note a distinct improvement in Table 2 across all models trained on the challenge data in terms of 11.3% in average (AC/O vs. CF/O), whereas the human-generated CAD can achieve 10.2% accuracy improvement (C/O vs. CF/O) in average. It is noteworthy that the human-generated CAD can slightly outperform our method when testing on the Out-of-domain Test using Different Training Data SVM BERT Accuracy on Amazon Reviews Orig & CAD (Our Method) (3.4k) 78.6 84.7 Orig & CAD (By Human) (3.4k) 79.3 83.3 Orig. & (Sudhakar et al., 2019) 64.0 77.2 Orig. & (Madaan et al., 2020) 74.3 71.3 Orig. (3.4k) 74.5 80.0 Accuracy on Semeval 2017 Task B (Twitter) Orig & CAD (Our Method) (3.4k) 69.7 83.8 Orig & CAD (By Human) (3.4k) 66.8 82.8 Orig. & (Sudhakar et al., 2019) 59.4 72.5 Orig. & (Madaan et al., 2020) 62.8 79.3 Orig. (3.4k) 63.1 72.6 Accuracy on Yelp Reviews Orig & CAD (Our Method) (3.4k) 85.5 87.9 Orig & CAD (By Human) (3.4k) 85.6 86.6 Orig. & (Sudhakar et al., 2019) 69.4 84.5 Orig. & (Madaan et al., 2020) 81.3 78.8 Orig. (3.4k) 81.9 84.3 Table 4: Out-of-domain test accuracy of SVM and BERT-base-uncased models trained on the original (Orig.) IMDB review only, Counterfactually Augmented Data (CAD) combining with original data, and sentiment-flipped style-transfer examples. human-generated (CF) data, it may be because the training and test sets of the human-generated (CF) data are generated by the same group of labelers. Robustness in the Generalization Test. We explore how our approach makes prediction models more robust out-of-domain in Table 4. For direct comparison between our method and the humangenerated method, we adopt the fine-tuned BERTbase model trained with the augmented dataset (original & automatically revised data). The finetuned model is directly tested for out-of-domain data without any adjustment. As shown in Table 4, only our method and the human-label method can outperform the BERT model trained on the original data with average 6.5% and 5.3% accuracy improvements, respectively. Our method also offers performance benefits over three datasets even when compared to the human-label method on BERT. Neural Method vs. Statistical Method. As shown in Table 4, the performance of the SVM model with automatic CAD is more robust than other automated methods (Sudhakar et al., 2019; Madaan et al., 2020) across all datasets. However, the human-labeled CAD can improve Amazon reviews’ accuracy compared to our method using the SVM model by 0.7%. It indicates that humangenerated data may lead to more performance benefits on a statistical model. 313 Types of Algorithms Examples Ori: Some films just simply should not be remade. This is one of them. In and of itself it is not a bad film. Hierarchical RM-CT: Remove negative limitations Rev: Some films just simply should be remade. This is one of them. In and of itself it is a bad film. Ori: It is badly directed, badly acted and boring. Hierarchical RE-CT: Replacing the causal terms Rev: It is well directed, well acted and entertaining. Ori: This movie is so bad, it can only be compared to the all-time worst “comedy”: Police Academy 7. No laughs throughout the movie. Combined method: Rev: This movie is so good, it can only be compared to the all-time best “comedy”: Police Academy 7. Laughs throughout the movie. Table 5: Most prominent categories of edits for flipping the sentiment performed by our algorithms, namely hierarchical RM-CT and hierarchical REP-CT. 5.4 Comparison with Automatic Methods Automatic CAD vs. Style-transfer Methods. As shown in Table 4, the style-transfer results are consistent with Kaushik et al. (2021). We find that the sentiment-flipped instances generated by style-transfer methods degrade the test accuracy for all models on all kinds of datasets, whereas our method has achieved the best performance for all settings. It suggests that our method have its absolute advantage for data augmentation in sentiment analysis when compared to the state-of-theart style-transfer models. Our Methods vs. Implausible CAD. The authors of the only existing approach for automatically generating CAD (Wang and Culotta, 2021) report that their methods are not able to match the performance of human-generated CAD. Our methods consistently outperform human-labeled methods on both In-domain and Out-of-domain tests. To further provide quantitative evidence of the influence of the edit-distance in automatic CAD, we demonstrate an ablation study in Table 6. The result shows that the quality of the generated CAD, which is ignored in the previous work Wang and Culotta (2021), is crucial when training the robust classifiers. In particular, the BERT model finetuned with implausible CAD (below the threshold) can receive comparable negative results with the style-transfer samples, alongside the performance decrease on all datasets, except for Twitter. 5.5 Case Study and Limitations The three most popular kinds of edits are shown in Table 5. These are, negation words removal, sentiment words replacement, and the combination of these. It can be observed from these examples that we ensure the edits on original samples should be minimal and fluent as was required previously with human-annotated counterfactuals (Kaushik Training Data IMDB Out-of-domain Test BERT-base-uncased Orig. Amazon Twitter Yelp Orig. & CAD ↑(3.4K) 90.6 84.7 83.8 87.9 Orig. & CAD ↓(3.4K) 87.1 79.5 73.8 79.0 Orig. (1.7K) 87.4 80.0 72.6 84.3 Table 6: Ablation study on the influence of the editdistance controlled by the threshold of MoverScore. ↑ indicates the CAD (1.7K) above the threshold, while ↓ denotes the CAD (1.7K) below the threshold. et al., 2020). As shown in Table 5, we flipped the model’s prediction by replacing the causal terms in the phrase “badly directed, badly acted and boring” to “well directed, well acted and entertaining”, or removing “No laughs throughout the movie.” to “Laughs throughout the movie” for a movie review. We also noticed that our method may face the challenge when handling more complex reviews. For example, the sentence “Watch this only if someone has a gun to your head ... maybe.” is an apparent negative review for a human. However, our algorithm is hard to flip the sentiment of such reviews with no explicit casual terms. The technique on sarcasm and irony detection may have benefits for dealing with this challenge. 6 Conclusion We proposed a new framework to automatically generate counterfactual augmented data (CAD) for enhancing the robustness of sentiment analysis models. By combining the automatically generated CAD with the original training data, we can produce more robust classifiers. We further show that our methods can achieve better performance even when compared to models trained with humangenerated counterfactuals. More importantly, our evaluation based on several datasets has demonstrated that models trained on the augmented data (original & automatic CAD) appear to be less af314 fected by spurious patterns and generalize better to out-of-domain data. This suggests there exists a significant opportunity to explore the use of the CAD in a range of tasks (e.g., natural language inference, natural language understanding, and social bias correction.). Impact Statement Although the experiments in this paper are conducted only in the sentiment classification task, this study could be a good starting point to investigate the efficacy of automatically generated CAD for building robust systems in many NLP tasks, including Natural Language Inference (NLI), Named Entity Recognition (NER), Question Answering (QA) system, etc. Acknowledgment We would like to thank Eoin Kenny and Prof. Mark Keane from Insight Centre for their helpful advice and discussion during this work. Also, we would like to thank the anonymous reviewers for their insightful comments and suggestions to help improve the paper. This publication has emanated from research conducted with the financial support of Science Foundation Ireland under Grant number 12/RC/2289 P2. References Emily M. Bender and Alexander Koller. 2020. Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5185–5198, Online. Association for Computational Linguistics. Ruth MJ Byrne. 2019. Counterfactuals in explainable artificial intelligence (xai): evidence from human reasoning. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 6276–6282. AAAI Press. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Xiao Ding, Dingkui Hao, Yuewei Zhang, Kuo Liao, Zhongyang Li, Bing Qin, and Ting Liu. 2020. HITSCIR at SemEval-2020 task 5: Training pre-trained language model with pseudo-labeling data for counterfactuals detection. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 354–360, Barcelona (online). International Committee for Computational Linguistics. Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, et al. 2020. Evaluating models’ local decision boundaries via contrast sets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1307–1323. Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Counterfactual visual explanations. In ICML. Alex Graves and J¨urgen Schmidhuber. 2005. Framewise phoneme classification with bidirectional lstm and other neural network architectures. Neural networks, 18(5-6):602–610. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168–177. Alon Jacovi and Yoav Goldberg. 2020. Aligning faithful interpretations with their social attribution. arXiv preprint arXiv:2006.01067. Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Tuo Zhao. 2020. SMART: Robust and efficient fine-tuning for pretrained natural language models through principled regularized optimization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2177–2190, Online. Association for Computational Linguistics. Xisen Jin, Zhongyu Wei, Junyi Du, Xiangyang Xue, and Xiang Ren. 2019. Towards hierarchical importance attribution: Explaining compositional semantics for neural sequence models. In International Conference on Learning Representations. Jason Jo and Yoshua Bengio. 2017. Measuring the tendency of cnns to learn surface statistical regularities. arXiv preprint arXiv:1711.11561. Rie Johnson and Tong Zhang. 2017. Deep pyramid convolutional neural networks for text categorization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 562–570. Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. 2020. Learning the difference that makes a difference with counterfactually-augmented data. In International Conference on Learning Representations. Divyansh Kaushik, Amrith Setlur, Eduard Hovy, and Zachary C Lipton. 2021. Explaining the efficacy of counterfactually augmented data. In International Conference on Learning Representations. Mark T Keane and Barry Smyth. 2020. Good counterfactuals and where to find them: A case-based technique for generating counterfactuals for explainable ai (xai). In International Conference on Case-Based Reasoning (ICCBR). 315 Eoin M Kenny and Mark T Keane. 2021. On generating plausible counterfactual and semi-factual explanations for deep learning. In AAAI. Soo-Min Kim and Eduard Hovy. 2004. Determining the sentiment of opinions. In COLING 2004: Proceedings of the 20th International Conference on Computational Linguistics, pages 1367–1373. Zachary C Lipton. 2018. The mythos of model interpretability. Queue, 16(3):31–57. Bing Liu. 2012. Sentiment analysis and opinion mining. Synthesis lectures on human language technologies, 5(1):1–167. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Amancharla, and Anupam Datta. 2020. Gender bias in neural natural language processing. In Logic, Language, and Security, pages 189–202. Springer. Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions. In Advances in neural information processing systems, pages 4765–4774. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon, USA. Association for Computational Linguistics. Aman Madaan, Amrith Setlur, Tanmay Parekh, Barnabas Poczos, Graham Neubig, Yiming Yang, Ruslan Salakhutdinov, Alan W Black, and Shrimai Prabhumoye. 2020. Politeness transfer: A tag and generate approach. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1869–1881, Online. Association for Computational Linguistics. Rowan Hall Maudslay, Hila Gonen, Ryan Cotterell, and Simone Teufel. 2019. It’s all in the name: Mitigating gender bias with name-based counterfactual data substitution. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 5270–5278. Tim Miller. 2019. Explanation in artificial intelligence: Insights from the social sciences. Artificial Intelligence, 267:1–38. Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 188–197. Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. 2019. Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift. In Advances in Neural Information Processing Systems, pages 13991–14002. Dino Pedreschi, Fosca Giannotti, Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, and Franco Turini. 2019. Meaningful explanations of black box ai decision systems. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9780–9784. Joaquin Quionero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. 2009. Dataset shift in machine learning. The MIT Press. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. ” why should i trust you?” explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902– 4912, Online. Association for Computational Linguistics. Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. Semeval-2017 task 4: Sentiment analysis in twitter. In Proceedings of the 11th international workshop on semantic evaluation (SemEval-2017), pages 502– 518. Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju. 2020a. Fooling lime and shap: Adversarial attacks on post hoc explanation methods. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, pages 180–186. Dylan Slack, Sophie Hilgard, Sameer Singh, and Himabindu Lakkaraju. 2020b. How much should i trust you? modeling uncertainty of black box explanations. arXiv preprint arXiv:2008.05030. Megha Srivastava, Tatsunori Hashimoto, and Percy Liang. 2020. Robustness to spurious correlations via human annotations. In International Conference on Machine Learning, pages 9109–9119. PMLR. Akhilesh Sudhakar, Bhargav Upadhyay, and Arjun Maheswaran. 2019. “transforming” delete, retrieve, generate approach for controlled text style transfer. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 316 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3260– 3270. Masashi Sugiyama and Motoaki Kawanabe. 2012. Machine learning in non-stationary environments: Introduction to covariate shift adaptation. MIT press. Johan AK Suykens and Joos Vandewalle. 1999. Least squares support vector machine classifiers. Neural processing letters, 9(3):293–300. Damien Teney, Ehsan Abbasnedjad, and Anton van den Hengel. 2020. Learning what makes a difference from counterfactual examples and gradient supervision. arXiv preprint arXiv:2004.09034. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS. Ke Wang and Xiaojun Wan. 2019. Automatic generation of sentimental texts via mixture adversarial networks. Artificial Intelligence, 275:540–558. Zhao Wang and Aron Culotta. 2021. Robustness to spurious correlations in text classification via automatically generated counterfactuals. In AAAI. Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R Bowman. 2020. Blimp: The benchmark of linguistic minimal pairs for english. Transactions of the Association for Computational Linguistics, 8:377–392. Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems, 33. Linyi Yang, Eoin Kenny, Tin Lok James Ng, Yi Yang, Barry Smyth, and Ruihai Dong. 2020a. Generating plausible counterfactual explanations for deep transformers in financial text classification. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6150–6160. Xiaoyu Yang, Stephen Obadinma, Huasha Zhao, Qiong Zhang, Stan Matwin, and Xiaodan Zhu. 2020b. SemEval-2020 task 5: Counterfactual recognition. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 322–335, Barcelona (online). International Committee for Computational Linguistics. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5753–5763. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical attention networks for document classification. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 1480–1489. Yuan Zang, Fanchao Qi, Chenghao Yang, Zhiyuan Liu, Meng Zhang, Qun Liu, and Maosong Sun. 2020. Word-level textual adversarial attacking as combinatorial optimization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6066–6080. Huangzhao Zhang, Hao Zhou, Ning Miao, and Lei Li. 2019. Generating fluent adversarial examples for natural languages. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5564–5569. Yuan Zhang and Yue Zhang. 2019. Tree communication models for sentiment analysis. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3518–3527. Yue Zhang, Qi Liu, and Linfeng Song. 2018. Sentencestate lstm for text representation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 317–327. Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. 2019. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563–578. Ran Zmigrod, Sebastian J Mielke, Hanna Wallach, and Ryan Cotterell. 2019. Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1651–1661.
|
2021
|
26
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3350–3363 August 1–6, 2021. ©2021 Association for Computational Linguistics 3350 ERICA: Improving Entity and Relation Understanding for Pre-trained Language Models via Contrastive Learning Yujia Qin♣♠♦, Yankai Lin♦, Ryuichi Takanobu♣♦, Zhiyuan Liu♣∗, Peng Li♦, Heng Ji♠∗, Minlie Huang♣, Maosong Sun♣, Jie Zhou♦ ♣Department of Computer Science and Technology, Tsinghua University, Beijing, China ♠University of Illinois at Urbana-Champaign ♦Pattern Recognition Center, WeChat AI, Tencent Inc. [email protected] Abstract Pre-trained Language Models (PLMs) have shown superior performance on various downstream Natural Language Processing (NLP) tasks. However, conventional pre-training objectives do not explicitly model relational facts in text, which are crucial for textual understanding. To address this issue, we propose a novel contrastive learning framework ERICA to obtain a deep understanding of the entities and their relations in text. Specifically, we define two novel pre-training tasks to better understand entities and relations: (1) the entity discrimination task to distinguish which tail entity can be inferred by the given head entity and relation; (2) the relation discrimination task to distinguish whether two relations are close or not semantically, which involves complex relational reasoning. Experimental results demonstrate that ERICA can improve typical PLMs (BERT and RoBERTa) on several language understanding tasks, including relation extraction, entity typing and question answering, especially under low-resource settings.1 1 Introduction Pre-trained Language Models (PLMs) (Devlin et al., 2018; Yang et al., 2019; Liu et al., 2019) have shown superior performance on various Natural Language Processing (NLP) tasks such as text classification (Wang et al., 2018), named entity recognition (Sang and De Meulder, 2003), and question answering (Talmor and Berant, 2019). Benefiting from designing various effective self-supervised learning objectives, such as masked language modeling (Devlin et al., 2018), PLMs can effectively capture the syntax and semantics in text to generate informative language representations for downstream NLP tasks. ∗Corresponding author. 1Our code and data are publicly available at https:// github.com/thunlp/ERICA. [1] Culiacán is a city in northwestern Mexico. [2] Culiacán is the capital of the state of Sinaloa. [3] Culiacán is also the seat of Culiacán Municipality. [4] It had an urban population of 785,800 in 2015 while 905,660 lived in the entire municipality. [5] While Culiacán Municipality has a total area of 4,758 k!!, Culiacán itself is considerably smaller, measuring only. [6] Culiacán is a rail junction and is located on the Panamerican Highway that runs south to Guadalajara and Mexico City. [7] Culiacán is connected to the north with Los Mochis, and to the south with Mazatlán, Tepic. Culiacán Q: where is Guadalajara? Culiacán Mexico Panamerican Highway city of south to locate on A: Mexico. Culiacán Municipality Sinaloa Guadalajara Mexico City Los Mochis Figure 1: An example for a document “Culiacán”, in which all entities are underlined. We show entities and their relations as a relational graph, and highlight the important entities and relations to find out “where is Guadalajara”. However, conventional pre-training objectives do not explicitly model relational facts, which frequently distribute in text and are crucial for understanding the whole text. To address this issue, some recent studies attempt to improve PLMs to better understand relations between entities (Soares et al., 2019; Peng et al., 2020). However, they mainly focus on within-sentence relations in isolation, ignoring the understanding of entities, and the interactions among multiple entities at document level, whose relation understanding involves complex reasoning patterns. According to the statistics on a human-annotated corpus sampled from Wikipedia documents by Yao et al. (2019), at least 40.7% relational facts require to be extracted from multiple sentences. Specifically, we show an example in Figure 1, to understand that “Guadalajara is located in Mexico”, we need to consider the following clues jointly: (i) “Mexico” is the country of “Culiacán” from sentence 1; (ii) “Culiacán” is a rail junction lo3351 cated on “Panamerican Highway” from sentence 6; (iii) “Panamerican Highway” connects to “Guadalajara” from sentence 6. From the example, we can see that there are two main challenges to capture the in-text relational facts: 1. To understand an entity, we should consider its relations to other entities comprehensively. In the example, the entity “Culiacán”, occurring in sentence 1, 2, 3, 5, 6 and 7, plays an important role in finding out the answer. To understand “Culiacán”, we should consider all its connected entities and diverse relations among them. 2. To understand a relation, we should consider the complex reasoning patterns in text. For example, to understand the complex inference chain in the example, we need to perform multi-hop reasoning, i.e., inferring that “Panamerican Highway” is located in “Mexico” through the first two clues. In this paper, we propose ERICA, a novel framework to improve PLMs’ capability of Entity and RelatIon understanding via ContrAstive learning, aiming to better capture in-text relational facts by considering the interactions among entities and relations comprehensively. Specifically, we define two novel pre-training tasks: (1) the entity discrimination task to distinguish which tail entity can be inferred by the given head entity and relation. It improves the understanding of each entity via considering its relations to other entities in text; (2) the relation discrimination task to distinguish whether two relations are close or not semantically. Through constructing entity pairs with documentlevel distant supervision, it takes complex relational reasoning chains into consideration in an implicit way and thus improves relation understanding. We conduct experiments on a suite of language understanding tasks, including relation extraction, entity typing and question answering. The experimental results show that ERICA improves the performance of typical PLMs (BERT and RoBERTa) and outperforms baselines, especially under lowresource settings, which demonstrates that ERICA effectively improves PLMs’ entity and relation understanding and captures the in-text relational facts. 2 Related Work Dai and Le (2015) and Howard and Ruder (2018) propose to pre-train universal language representations on unlabeled text, and perform task-specific fine-tuning. With the advance of computing power, PLMs such as OpenAI GPT (Radford et al., 2018), BERT (Devlin et al., 2018) and XLNet (Yang et al., 2019) based on deep Transformer (Vaswani et al., 2017) architecture demonstrate their superiority in various downstream NLP tasks. Since then, numerous PLM extensions have been proposed to further explore the impacts of various model architectures (Song et al., 2019; Raffel et al., 2020), larger model size (Raffel et al., 2020; Lan et al., 2020; Fedus et al., 2021), more pre-training corpora (Liu et al., 2019), etc., to obtain better general language understanding ability. Although achieving great success, these PLMs usually regard words as basic units in textual understanding, ignoring the informative entities and their relations, which are crucial for understanding the whole text. To improve the entity and relation understanding of PLMs, a typical line of work is knowledgeguided PLM, which incorporates external knowledge such as Knowledge Graphs (KGs) into PLMs to enhance the entity and relation understanding. Some enforce PLMs to memorize information about real-world entities and propose novel pretraining objectives (Xiong et al., 2019; Wang et al., 2019; Sun et al., 2020; Yamada et al., 2020). Others modify the internal structures of PLMs to fuse both textual and KG’s information (Zhang et al., 2019; Peters et al., 2019; Wang et al., 2020; He et al., 2020). Although knowledge-guided PLMs introduce extra factual knowledge in KGs, these methods ignore the intrinsic relational facts in text, making it hard to understand out-of-KG entities or knowledge in downstream tasks, let alone the errors and incompleteness of KGs. This verifies the necessity of teaching PLMs to understand relational facts from contexts. Another line of work is to directly model entities or relations in text in pre-training stage to break the limitations of individual token representations. Some focus on obtaining better span representations, including entity mentions, via span-based pre-training (Sun et al., 2019; Joshi et al., 2020; Kong et al., 2020; Ye et al., 2020). Others learn to extract relation-aware semantics from text by comparing the sentences that share the same entity pair or distantly supervised relation in KGs (Soares et al., 2019; Peng et al., 2020). However, these methods only consider either individual entities or within-sentence relations, which limits the performance in dealing with multiple entities and relations at document level. In contrast, our ERICA considers the interactions among multiple entities 3352 Figure 2: An example of Entity Discrimination task. For an entity pair with its distantly supervised relation in text, the ED task requires the ground-truth tail entity to be closer to the head entity than other entities. and relations comprehensively, achieving a better understanding of in-text relational facts. 3 Methodology In this section, we introduce the details of ERICA. We first describe the notations and how to represent entities and relations in documents. Then we detail the two novel pre-training tasks: Entity Discrimination (ED) task and Relation Discrimination (RD) task, followed by the overall training objective. 3.1 Notations ERICA is trained on a large-scale unlabeled corpus leveraging the distant supervision from an external KG K. Formally, let D = {di}|D| i=1 be a batch of documents and Ei = {eij}|Ei| j=1 be all named entities in di, where eij is the j-th entity in di. For each document di, we enumerate all entity pairs (eij, eik) and link them to their corresponding relation ri jk in K (if possible) and obtain a tuple set Ti = {ti jk = (di, eij, ri jk, eik)|j ̸= k}. We assign no_relation to those entity pairs without relation annotation in K. Then we obtain the overall tuple set T = T1 S T2 S ... S T|D| for this batch. The positive tuple set T + is constructed by removing all tuples with no_relation from T . Benefiting from document-level distant supervision, T + includes both intra-sentence (relatively simple cases) and inter-sentence entity pairs (hard cases), whose relation understanding involves cross-sentence, multi-hop, or coreferential reasoning, i.e., T + = T + single S T + cross. 3.2 Entity & Relation Representation For each document di, we first use a PLM to encode it and obtain a series of hidden states {h1, h2, ..., h|di|}, then we apply mean pooling operation over the consecutive tokens that mention eij to obtain local entity representations. Note eij may appear multiple times in di, the k-th occurrence of eij, which contains the tokens from index nk start to nk end, is represented as: mk eij = MeanPool(hnk start, ..., hnk end). (1) To aggregate all information about eij, we average2 all representations of each occurrence mk eij as the global entity representation eij. Following Soares et al. (2019), we concatenate the final representations of two entities eij1 and eij2 as their relation representation, i.e., ri j1j2 = [eij1; eij2]. 3.3 Entity Discrimination Task Entity Discrimination (ED) task aims at inferring the tail entity in a document given a head entity and a relation. By distinguishing the ground-truth tail entity from other entities in the text, it teaches PLMs to understand an entity via considering its relations with other entities. As shown in Figure 2, in practice, we first sample a tuple ti jk = (di, eij, ri jk, eik) from T +, PLMs are then asked to distinguish the groundtruth tail entity eik from other entities in the document di. To inform PLMs of which head entity and relation to be conditioned on, we concatenate the relation name of ri jk, the mention of head entity eij and a separation token [SEP] in front of di, i.e., d∗ i =“relation_name entity_mention[SEP] di”3. The goal of entity discrimination task is equivalent to maximizing the posterior P(eik|eij, ri jk) = softmax(f(eik)) (f(·) indicates an entity classifier). However, we empirically find directly optimizing the posterior cannot well consider the relations among entities. Hence, we borrow the idea of contrastive learning (Hadsell et al., 2006) and push the representations of positive pair (eij, eik) closer than negative pairs, the loss function of ED task can be formulated as: LED = − X ti jk∈T + log exp(cos(eij, eik)/τ) |Ei| P l=1, l̸=j exp(cos(eij, eil)/τ) , (2) 2Although weighted summation by attention mechanism is an alternative, the specific method of entity information aggregation is not our main concern. 3Here we encode the modified document d∗ i to obtain the entity representations. The newly added entity_mention is not considered for head entity representation. 3353 Document 1 Document 2 Document 3 Document 3 single-sentence cross-sentence single-sentence cross-sentence founded by … Since 1773, when the Royal Swedish Opera was founded by Gustav III of Sweden … … Gates is an American business magnate, software developer, and philanthropist … He left his board positions at Microsoft … … Samarinda is the capital of East Kalimantan, Indonesia, on the island of Borneo … Samarinda is known for its traditional food amplang, as well as the cloth Sarung Samarinda … … Samarinda is the capital of East Kalimantan, Indonesia, on the island of Borneo … Samarinda is known for its traditional food amplang, as well as the cloth Sarung Samarinda … founded by capital of country Pre-trained Language Model Figure 3: An example of Relation Discrimination task. For entity pairs belonging to the same relations, the RD task requires their relation representations to be closer. where cos(·, ·) denotes the cosine similarity between two entity representations and τ (temperature) is a hyper-parameter. 3.4 Relation Discrimination Task Relation Discrimination (RD) task aims at distinguishing whether two relations are close or not semantically. Compared with existing relationenhanced PLMs, we employ document-level rather than sentence-level distant supervision to further make PLMs comprehend the complex reasoning chains in real-world scenarios and thus improve PLMs’ relation understanding. As depicted in Figure 3, we train the text-based relation representations of the entity pairs that share the same relations to be closer in the semantic space. In practice, we linearly4 sample a tuple pair tA = (dA, eA1, rA, eA2) and tB = (dB, eB1, rB, eB2) from T + s (T + single) or T + c (T + cross), where rA = rB. Using the method mentioned in Sec. 3.2, we obtain the positive relation representations rtA and rtB for tA and tB. To discriminate positive examples from negative ones, similarly, we adopt contrastive learning and define the loss function of RD task as follows: L T1,T2 RD = − X tA∈T1,tB∈T2 log exp(cos(rtA, rtB)/τ) Z , Z = N X tC∈T /{tA} exp(cos(rtA, rtC)/τ), LRD = LT + s ,T + s RD + L T + s ,T + c RD + L T + c ,T + s RD + L T + c ,T + c RD , (3) 4The sampling rate of each relation is proportional to its total number in the current batch. where N is a hyper-parameter. We ensure tB is sampled in Z and construct N −1 negative examples by sampling tC (rA ̸= rC) from T , instead of T +5. By additionally considering the last three terms of LRD in Eq.3, which require the model to distinguish complex inter-sentence relations with other relations in the text, our model could have better coverage and generality of the reasoning chains. PLMs are trained to perform reasoning in an implicit way to understand those “hard” inter-sentence cases. 3.5 Overall Objective Now we present the overall training objective of ERICA. To avoid catastrophic forgetting (McCloskey and Cohen, 1989) of general language understanding ability, we train masked language modeling task (LMLM) together with ED and RD tasks. Hence, the overall learning objective is formulated as follows: L = LED + LRD + LMLM. (4) It is worth mentioning that we also try to mask entities as suggested by Soares et al. (2019) and Peng et al. (2020), aiming to avoid simply relearning an entity linking system. However, we do not observe performance gain by such a masking strategy. We conjecture that in our document-level setting, it is hard for PLMs to overfit on memorizing entity mentions due to the better coverage and generality of document-level distant supervision. Besides, masking entities creates a gap between pre-training and fine-tuning, which may be a shortcoming of previous relation-enhanced PLMs. 4 Experiments In this section, we first describe how we construct the distantly supervised dataset and pre-training details for ERICA. Then we introduce the experiments we conduct on several language understanding tasks, including relation extraction (RE), entity typing (ET) and question answering (QA). We test ERICA on two typical PLMs, including BERT and RoBERTa (denoted as ERICABERT and ERICARoBERTa)6. We leave the training details 5In experiments, we find introducing no_relation entity pairs as negative samples further improves the performance and the reason is that increasing the diversity of training entity pairs is beneficial to PLMs. 6Since our main focus is to demonstrate the superiority of ERICA in improving PLMs to capture relational facts and advance further research explorations, we choose base models 3354 for downstream tasks and experiments on GLUE benchmark (Wang et al., 2018) in the appendix. 4.1 Distantly Supervised Dataset Construction Following Yao et al. (2019), we construct our pretraining dataset leveraging distant supervision from the English Wikipedia and Wikidata. First, we use spaCy7 to perform Named Entity Recognition, and then link these entity mentions as well as Wikipedia’s mentions with hyper-links to Wikidata items, thus we obtain the Wikidata ID for each entity. The relations between different entities are annotated distantly by querying Wikidata. We keep the documents containing at least 128 words, 4 entities and 4 relational triples. In addition, we ignore those entity pairs appearing in the test sets of RE and QA tasks to avoid test set leakage. In the end, we collect 1, 000, 000 documents (about 1G storage) in total with more than 4, 000 relations annotated distantly. On average, each document contains 186.9 tokens, 12.9 entities and 7.2 relational triples, an entity appears 1.3 times per document. Based on the human evaluation on a random sample of the dataset, we find that it achieves an F1 score of 84.7% for named entity recognition, and an F1 score of 25.4% for relation extraction. 4.2 Pre-training Details We initialize ERICABERT and ERICARoBERTa with bert-base-uncased and roberta-base checkpoints released by Google8 and Huggingface9. We adopt AdamW (Loshchilov and Hutter, 2017) as the optimizer, warm up the learning rate for the first 20% steps and then linearly decay it. We set the learning rate to 3 × 10−5, weight decay to 1 × 10−5, batch size to 2, 048 and temperature τ to 5 × 10−2. For LRD, we randomly select up to 64 negative samples per document. We train both models with 8 NVIDIA Tesla P40 GPUs for 2, 500 steps. 4.3 Relation Extraction Relation extraction aims to extract the relation between two recognized entities from a pre-defined relation set. We conduct experiments on both document-level and sentence-level RE. We test for experiments. 7https://spacy.io/ 8https://github.com/google-research/bert 9https://github.com/huggingface/ transformers Size 1% 10% 100% Metrics F1 IgF1 F1 IgF1 F1 IgF1 CNN 42.3 40.3 BILSTM 51.1 50.3 BERT 30.4 28.9 47.1 44.9 56.8 54.5 HINBERT 55.6 53.7 CorefBERT 32.8 31.2 46.0 43.7 57.0 54.5 SpanBERT 32.2 30.4 46.4 44.5 57.3 55.0 ERNIE 26.7 25.5 46.7 44.2 56.6 54.2 MTB 29.0 27.6 46.1 44.1 56.9 54.3 CP 30.3 28.7 44.8 42.6 55.2 52.7 ERICABERT 37.8 36.0 50.8 48.3 58.2 55.9 RoBERTa 35.3 33.5 48.0 45.9 58.5 56.1 ERICARoBERTa 40.1 38.0 50.3 48.3 59.0 56.6 Table 1: Results on document-level RE (DocRED). We report micro F1 (F1) and micro ignore F1 (IgF1) on test set. IgF1 metric ignores the relational facts shared by the train and dev/test sets. Dataset TACRED SemEval Size 1% 10% 100% 1% 10% 100% BERT 36.0 58.5 68.1 43.6 79.3 88.1 MTB 35.7 58.8 68.2 44.2 79.2 88.2 CP 37.1 60.6 68.1 40.3 80.0 88.5 ERICABERT 36.5 59.7 68.5 47.9 80.1 88.0 RoBERTa 26.3 61.2 69.7 46.0 80.3 88.8 ERICARoBERTa 40.0 61.9 69.8 46.3 80.4 89.2 Table 2: Results (test F1) on sentence-level RE (TACRED and SemEval-2010 Task8) on three splits (1%, 10% and 100%). three partitions of the training set (1%, 10% and 100%) and report results on test sets. Document-level RE For document-level RE, we choose DocRED (Yao et al., 2019), which requires reading multiple sentences in a document and synthesizing all the information to identify the relation between two entities. We encode all entities in the same way as in pre-training phase. The relation representations are obtained by adding a bilinear layer on top of two entity representations. We choose the following baselines: (1) CNN (Zeng et al., 2014), BILSTM (Hochreiter and Schmidhuber, 1997), BERT (Devlin et al., 2018) and RoBERTa (Liu et al., 2019), which are widely used as text encoders for relation extraction tasks; (2) HINBERT (Tang et al., 2020) which employs a hierarchical inference network to leverage the abundant information from different sources; (3) CorefBERT (Ye et al., 2020) which proposes a pre-training method to help BERT capture the coreferential relations in context; (4) SpanBERT (Joshi et al., 2020) which masks 3355 Metrics Macro F1 Micro F1 BERT 75.50 72.68 MTB 76.37 72.94 CP 76.27 72.48 ERNIE 76.51 73.39 ERICABERT 77.85 74.71 RoBERTa 79.24 76.38 ERICARoBERTa 80.77 77.04 Table 3: Results on entity typing (FIGER). We report macro F1 and micro F1 on the test set. and predicts contiguous random spans instead of random tokens; (5) ERNIE (Zhang et al., 2019) which incorporates KG information into BERT to enhance entity representations; (6) MTB (Soares et al., 2019) and CP (Peng et al., 2020) which introduce sentence-level relation contrastive learning for BERT via distant supervision. For fair comparison, we pre-train these baselines on our constructed pre-training data10 based on the implementation released by Peng et al. (2020)11. From the results shown in Table 1, we can see that: (1) ERICA outperforms all baselines significantly on each supervised data size, which demonstrates that ERICA could better understand the relations among entities in the document via implicitly considering their complex reasoning patterns in the pre-training; (2) both MTB and CP achieve worse results than BERT, which means sentence-level pre-training, lacking consideration for complex reasoning patterns, hurts PLM’s performance on document-level RE tasks to some extent; (3) ERICA outperforms baselines by a larger margin on smaller training sets, which means ERICA has gained pretty good document-level relation reasoning ability in contrastive learning, and thus obtains improvements more extensively under low-resource settings. Sentence-level RE For sentence-level RE, we choose two widely used datasets: TACRED (Zhang et al., 2017) and SemEval-2010 Task 8 (Hendrickx et al., 2019). We insert extra marker tokens to indicate the head and tail entities in each sentence. For baselines, we compare ERICA with BERT, RoBERTa, MTB and CP. From the results shown in Table 2, we observe that ERICA achieves almost comparable results on sentence-level RE tasks with CP, which means document-level pre-training in 10In practice, documents are split into sentences and we only keep within-sentence entity pairs. 11https://github.com/thunlp/ RE-Context-or-Names Setting Standard Masked Size 1% 10% 100% 1% 10% 100% FastQA 27.2 38.0 BiDAF 49.7 59.8 BERT 35.8 53.7 69.5 37.9 53.1 73.1 CorefBERT 38.1 54.4 68.8 39.0 53.5 70.7 SpanBERT 33.1 56.4 70.7 34.0 55.4 73.2 MTB 36.6 51.7 68.4 36.2 50.9 71.7 CP 34.6 50.4 67.4 34.1 47.1 69.4 ERICABERT 46.5 57.8 69.7 40.2 58.1 73.9 RoBERTa 37.3 57.4 70.9 41.2 58.7 75.5 ERICARoBERTa 47.4 58.8 71.2 46.8 63.4 76.6 Table 4: Results (accuracy) on the dev set of WikiHop. We test both the standard and masked settings on three splits (1%, 10% and 100%). Setting SQuAD TriviaQA NaturalQA Size 10% 100% 10% 100% 10% 100% BERT 79.7 88.9 60.8 70.7 68.4 78.4 MTB 63.5 87.1 52.0 67.8 61.2 76.7 CP 69.0 87.1 52.9 68.1 63.3 77.3 ERICABERT 81.8 88.9 63.5 71.9 70.2 79.1 RoBERTa 82.9 90.5 63.6 72.0 71.8 80.0 ERICARoBERTa 85.0 90.4 63.6 72.1 73.7 80.5 Table 5: Results (F1) on extractive QA (SQuAD, TriviaQA and NaturalQA) on two splits (10% and 100%). Results on 1% split are left in the appendix. ERICA does not impair PLMs’ performance on sentence-level relation understanding. 4.4 Entity Typing Entity typing aims at classifying entity mentions into pre-defined entity types. We choose FIGER (Ling et al., 2015), which is a sentencelevel entity typing dataset labeled with distant supervision. BERT, RoBERTa, MTB, CP and ERNIE are chosen as baselines. From the results listed in Table 3, we observe that, ERICA outperforms all baselines, which demonstrates that ERICA could better represent entities and distinguish them in text via both entity-level and relation-level contrastive learning. 4.5 Question Answering Question answering aims to extract a specific answer span in text given a question. We conduct experiments on both multi-choice and extractive QA. We test multiple partitions of the training set. Multi-choice QA For Multi-choice QA, we choose WikiHop (Welbl et al., 2018), which requires models to answer specific properties of an 3356 entity after reading multiple documents and conducting multi-hop reasoning. It has both standard and masked settings, where the latter setting masks all entities with random IDs to avoid information leakage. We first concatenate the question and documents into a long sequence, then we find all the occurrences of an entity in the documents, encode them into hidden representations and obtain the global entity representation by applying mean pooling on these hidden representations. Finally, we use a classifier on top of the entity representation for prediction. We choose the following baselines: (1) FastQA (Weissenborn et al., 2017) and BiDAF (Seo et al., 2016), which are widely used question answering systems; (2) BERT, RoBERTa, CorefBERT, SpanBERT, MTB and CP, which are introduced in previous sections. From the results listed in Table 4, we observe that ERICA outperforms baselines in both settings, indicating that ERICA can better understand entities and their relations in the documents and extract the true answer according to queries. The significant improvements in the masked setting also indicate that ERICA can better perform multi-hop reasoning to synthesize and analyze information from contexts, instead of relying on entity mention “shortcuts” (Jiang and Bansal, 2019). Extractive QA For extractive QA, we adopt three widely-used datasets: SQuAD (Rajpurkar et al., 2016), TriviaQA (Joshi et al., 2017) and NaturalQA (Kwiatkowski et al., 2019) in MRQA (Fisch et al., 2019) to evaluate ERICA in various domains. Since MRQA does not provide the test set for each dataset, we randomly split the original dev set into two halves and obtain the new dev/test set. We follow the QA setting of BERT (Devlin et al., 2018): we concatenate the given question and passage into one long sequence, encode the sequence by PLMs and adopt two classifiers to predict the start and end index of the answer. We choose BERT, RoBERTa, MTB and CP as baselines. From the results listed in Table 5, we observe that ERICA outperforms all baselines, indicating that through the enhancement of entity and relation understanding, ERICA is more capable of capturing in-text relational facts and synthesizing information of entities. This ability further improves PLMs for question answering. 5 Analysis In this section, we first conduct a suite of ablation studies to explore how LED and LRD contribute to Dataset DocRED FIGER WikiHop BERT 44.9 72.7 53.1 -NSP 45.2 72.6 53.6 -NSP+LED 47.6 73.8 59.8 -NSP+L T + c ,T + c RD 46.4 72.6 52.2 -NSP+L T + s,T + s RD 47.3 73.5 51.2 -NSP+LRD 48.0 74.0 52.0 ERICABERT 48.3 74.7 58.1 Table 6: Ablation study. We report test IgF1 on DocRED (10%), test micro F1 on FIGER and dev accuracy on the masked setting of WikiHop (10%). ERICA. Then we give a thorough analysis on how pre-training data’s domain / size and methods for entity encoding impact the performance. Lastly, we visualize the entity and relation embeddings learned by ERICA. 5.1 Ablation Study To demonstrate that the superior performance of ERICA is not owing to its longer pretraining (2500 steps) on masked language modeling, we include a baseline by optimizing LMLM only (removing the Next Sentence Prediction (-NSP) loss (Devlin et al., 2018)). In addition, to explore how LED and LRD impact the performance, we keep only one of these two losses and compare the results. Lastly, to evaluate how intra-sentence and inter-sentence entity pairs contribute to RD task, we compare the performances of only sampling intra-sentence entity pairs (L T + s ,T + s RD ) or inter-sentence entity pairs (L T + c ,T + c RD ), and sampling both of them (LRD) during pre-training. We conduct experiments on DocRED, WikiHop (masked version) and FIGER. For DocRED and WikiHop, we show the results on 10% splits and the full results are left in the appendix. From the results shown in Table 6, we can see that: (1) extra pretraining (-NSP) only contributes a little to the overall improvement. (2) For DocRED and FIGER, either LED or LRD is beneficial, and combining them further improves the performance; For WikiHop, LED dominates the improvement while LRD hurts the performance slightly, this is possibly because question answering more resembles the tail entity discrimination process, while the relation discrimination process may have conflicts with it. (3) For LRD, both intra-sentence and inter-sentence entity pairs contribute, which demonstrates that incorporating both of them is necessary for PLMs to understand relations between entities in text comprehensively. We also found empiri3357 Size 1% 10% 100% BERT 28.9 44.9 54.5 ERICABERT 36.0 48.3 55.9 ERICADocRED BERT 36.3 48.6 55.9 Table 7: Effects of pre-training data’s entity distribution shifting. We report test IgF1 on DocRED. 0% 30% 50% 70% 100% 1% DocRED 30 32 34 36 0% 30% 50% 70% 100% 10% DocRED 45 46 47 48 0% 30% 50% 70% 100% 100% DocRED 54.5 55.0 55.5 Figure 4: Impacts of relation distribution shifting. X axis denotes different ratios of relations, Y axis denotes test IgF1 on different partitions of DocRED. cally that when these two auxiliary objectives are only added into the fine-tuning stage, the model does not have performance gain. The reason is that the size and diversity of entities and relations in downstream training data are limited. Instead, pretraining with distant supervision on a large corpus provides a solution for increasing the diversity and quantity of training examples. 5.2 Effects of Domain Shifting We investigate two domain shifting factors: entity distribution and relation distribution, to explore how they impact ERICA’s performance. Entity Distribution Shifting The entities in supervised datasets of DocRED are recognized by human annotators while our pre-training data is processed by spaCy. Hence there may exist an entity distribution gap between pre-training and finetuning. To study the impacts of entity distribution shifting, we fine-tune a BERT model on training set of DocRED for NER tagging and re-tag entities in our pre-training dataset. Then we pre-train ERICA on the newly-labeled training corpus (denoted as ERICADocRED BERT ). From the results shown in Table 7, we observe that it performs better than the original ERICA, indicating that pre-training on a dataset that shares similar entity distributions with downstream tasks is beneficial. Relation Distribution Shifting Our pre-training data contains over 4, 000 Wikidata relations. To investigate whether training on a more diverse relation domain benefits ERICA, we train it with the pre-training corpus that randomly keeps only 30%, 50% and 70% the original relations, and compare 0% 10% 30% 50% 70%100% 1% DocRED 30 32 34 36 0% 10% 30% 50% 70%100% 10% DocRED 45 46 47 48 0% 10% 30% 50% 70%100% 100% DocRED 54.5 55.0 55.5 Figure 5: Impacts of pre-training data’s size. X axis denotes different ratios of pre-training data, Y axis denotes test IgF1 on different partitions of DocRED. Size 1% 10% 100% Metrics F1 IgF1 F1 IgF1 F1 IgF1 Mean Pool BERT 30.4 28.9 47.1 44.9 56.8 54.5 ERICABERT 37.8 36.0 50.8 48.3 58.2 55.9 ERICADocRED BERT 38.5 36.3 51.0 48.6 58.2 55.9 Entity Marker BERT 23.0 21.8 46.5 44.3 58.0 55.6 ERICABERT 34.9 33.0 50.2 48.0 59.9 57.6 ERICADocRED BERT 36.9 34.8 52.5 50.3 60.8 58.4 Table 8: Results (IgF1) on how entity encoding strategy influences ERICA’s performance on DocRED. We also show the impacts of entity distribution shifting (ERICADocRED BERT and ERICADocRED BERT ) as is mentioned in the main paper. their performances. From the results in Figure 4, we observe that the performance of ERICA improves constantly as the diversity of relation domain increases, which reveals the importance of using diverse training data on relation-related tasks. Through detailed analysis, we further find that ERICA is less competent at handling unseen relations in the corpus. This may result from the construction of our pre-training dataset: all the relations are annotated distantly through an existing KG with a pre-defined relation set. It would be promising to introduce more diverse relation domains during data preparation in future. 5.3 Effects of Pre-training Data’s Size To explore the effects of pre-training data’s size, we train ERICA on 10%, 30%, 50% and 70% of the original pre-training dataset, respectively. We report the results in Figure 5, from which we observe that with the scale of pre-training data becoming larger, ERICA is performing better. 5.4 Effects of Methods for Entity Encoding For all the experiments mentioned above, we encode each occurrence of an entity by mean pooling over all its tokens in both pre-training and downstream tasks. Ideally, ERICA should have consis3358 tent improvements on other kinds of methods for entity encoding. To demonstrate this, we try another entity encoding method mentioned by Soares et al. (2019) on three splits of DocRED (1%, 10% and 100%). Specifically, we insert a special start token [S] in front of an entity and an end token [E] after it. The representation for this entity is calculated by averaging the representations of all its start tokens in the document. To help PLMs discriminate different entities, we randomly assign different marker pairs ([S1], [E1]; [S2], [E2], ...) for each entity in a document in both pre-training and downstream tasks12. All occurrences of one entity in a document share the same marker pair. We show in Table 8 that ERICA achieves consistent performance improvements for both methods (denoted as Mean Pool and Entity Marker), indicating that ERICA is applicable to different methods for entity encoding. Specifically, Entity Marker achieves better performance when the scale of training data is large while Mean Pool is more powerful under low-resource settings. We also notice that training on a dataset that shares similar entity distributions is more helpful for Mean Pool, where ERICADocRED BERT achieves 60.8 (F1) and 58.4 (IgF1) on 100% training data. 5.5 Embedding Visualization In Figure 6, we show the learned entity and relation embeddings of BERT and ERICABERT on DocRED’s dev set by t-distributed stochastic neighbor embedding (t-SNE) (Hinton and Roweis, 2002). We label points with different colors to represent its corresponding category of entities or relations13 in Wikidata and only visualize the most frequent 10 relations. From the figure, we can see that jointly training LMLM with LED and LRD leads to a more compact clustering of both entities and relations belonging to the same category. In contrast, only training LMLM exhibits random distribution. This verifies that ERICA could better understand and represent both entities and relations in the text. 12In practice, we randomly initialize 100 entity marker pairs. 13(Key, value) pairs for relations defined in Wikidata are: (P176, manufacturer); (P150, contains administrative territorial entity); (P17, country); (P131, located in the administrative territorial entity); (P175, performer); (P27, country of citizenship); (P569, date of birth); (P1001, applies to jurisdiction); (P57, director); (P179, part of the series). BERT: entity entity MISC ORG PER LOC TIME NUM ERICA-BERT: entity entity TIME MISC ORG LOC PER NUM BERT: relation relation P17 P131 P1001 P27 P150 P175 P179 P57 P176 P569 ERICA-BERT: relation relation P176 P150 P17 P131 P175 P27 P569 P1001 P57 P179 Figure 6: t-SNE plots of learned entity and relation embeddings on DocRED comparing BERT and ERICABERT. 6 Conclusions In this paper, we present ERICA, a general framework for PLMs to improve entity and relation understanding via contrastive learning. We demonstrate the effectiveness of our method on several language understanding tasks, including relation extraction, entity typing and question answering. The experimental results show that ERICA outperforms all baselines, especially under low-resource settings, which means ERICA helps PLMs better capture the in-text relational facts and synthesize information about entities and their relations. Acknowledgments This work is supported by the National Key Research and Development Program of China (No. 2020AAA0106501) and Beijing Academy of Artificial Intelligence (BAAI). This work is also supported by the Pattern Recognition Center, WeChat AI, Tencent Inc. References Andrew M Dai and Quoc V Le. 2015. Semi-supervised sequence learning. In Advances in neural information processing systems, pages 3079–3087. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association 3359 for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Markus Eberts and Adrian Ulges. 2019. Span-based joint entity and relation extraction with transformer pre-training. CoRR, abs/1909.07755. William Fedus, Barret Zoph, and Noam Shazeer. 2021. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961. Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 shared task: Evaluating generalization in reading comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1–13, Hong Kong, China. Association for Computational Linguistics. Harsha Gurulingappa, Abdul Mateen Rajput, Angus Roberts, Juliane Fluck, Martin Hofmann-Apitius, and Luca Toldo. 2012. Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports. Journal of biomedical informatics, 45(5):885– 892. Raia Hadsell, Sumit Chopra, and Yann LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06), volume 2, pages 1735–1742. IEEE. Bin He, Di Zhou, Jinghui Xiao, Xin Jiang, Qun Liu, Nicholas Jing Yuan, and Tong Xu. 2020. BERTMK: Integrating graph contextualized knowledge into pre-trained language models. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2281–2290, Online. Association for Computational Linguistics. Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid O Séaghdha, Sebastian Padó, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2019. SemEval-2010 Task 8: Multi-way classification of semantic relations between pairs of nominals. In Proceedings of the Workshop on Semantic Evaluations: Recent Achievements and Future Directions (SEW-2009), pages 94– 99. Geoffrey E Hinton and Sam Roweis. 2002. Stochastic neighbor embedding. In Advances in neural information processing systems 15: 16th Annual Conference on Neural Information Processing Systems 2002. Proceedings of a meeting held September 12, 2002, Vancouver, British Columbia, Canada, volume 15, pages 857–864. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 328–339, Melbourne, Australia. Association for Computational Linguistics. Yichen Jiang and Mohit Bansal. 2019. Avoiding reasoning shortcuts: Adversarial evaluation, training, and model development for multi-hop qa. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, ACL 2019, July 28, 2019, Florence, Italy, pages 2726–2736. Association for Computational Linguistics. Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64–77. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–1611, Vancouver, Canada. Association for Computational Linguistics. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7, 2015, Conference Track Proceedings. Lingpeng Kong, Cyprien de Masson d’Autume, Lei Yu, Wang Ling, Zihang Dai, and Dani Yogatama. 2020. A mutual information maximization perspective of language representation learning. In Proceedings of 8th International Conference on Learning Representations, ICLR 2020, Virtual Conference, April 26, 2020, Conference Track Proceedings. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453–466. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A lite BERT for self-supervised learning of language representations. In Proceedings of 8th International Conference on Learning Representations, ICLR 2020, Virtual Conference, April 26, 2020, Conference Track Proceedings. Xiao Ling, Sameer Singh, and Daniel S Weld. 2015. Design challenges for entity linking. Transactions of the Association for Computational Linguistics, 3:315–328. 3360 Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. In Proceedings of 7th International Conference on Learning Representations, ICLR 2019. Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in connectionist networks: the sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109–165. Elsevier. Hao Peng, Tianyu Gao, Xu Han, Yankai Lin, Peng Li, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2020. Learning from context or names? an empirical study on neural relation extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3661–3672, Online. Association for Computational Linguistics. Matthew E Peters, Mark Neumann, Robert L Logan IV, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Association for Computational Linguistics. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics. Dan Roth and Wen-tau Yih. 2004. A linear programming formulation for global inference in natural language tasks. In Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004, pages 1–8, Boston, Massachusetts, USA. Association for Computational Linguistics. Erik F Sang and Fien De Meulder. 2003. Introduction to the conll-2003 shared task: Languageindependent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003. Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2016. Bidirectional attention flow for machine comprehension. In Proceedings of 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24, 2017, Con- ference Track Proceedings. Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2895–2905. Association for Computational Linguistics. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. 2019. Mass: Masked sequence to sequence pre-training for language generation. In Proceedings of International Conference on Machine Learning, pages 5926–5936. PMLR. Tianxiang Sun, Yunfan Shao, Xipeng Qiu, Qipeng Guo, Yaru Hu, Xuanjing Huang, and Zheng Zhang. 2020. CoLAKE: Contextualized language and knowledge embedding. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3660–3670, Barcelona, Spain (Online). International Committee on Computational Linguistics. Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223. Alon Talmor and Jonathan Berant. 2019. MultiQA: An empirical investigation of generalization and transfer in reading comprehension. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4911–4921. Association for Computational Linguistics. Hengzhu Tang, Yanan Cao, Zhenyu Zhang, Jiangxia Cao, Fang Fang, Shi Wang, and Pengfei Yin. 2020. Hin: Hierarchical inference network for documentlevel relation extraction. In Advances in Knowledge Discovery and Data Mining-24th Pacific-Asia Conference, PAKDD 2020, Singapore, May 11, 2020, Proceedings, Part I, volume 12084 of Lecture Notes in Computer Science, pages 197–209. Springer. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4 December 2017, Long Beach, CA, USA, pages 5998–6008. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Net3361 works for NLP1. Association for Computational Linguistics. Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Cuihong Cao, Daxin Jiang, Ming Zhou, et al. 2020. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808. Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhiyuan Liu, Juanzi Li, and Jian Tang. 2019. KEPLER: A unified model for knowledge embedding and pretrained language representation. Transactions of the Association for Computational Linguistics. Dirk Weissenborn, Georg Wiese, and Laura Seiffe. 2017. Making neural QA as simple as possible but not simpler. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 271–280. Association for Computational Linguistics. Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing datasets for multi-hop reading comprehension across documents. Transactions of the Association for Computational Linguistics, 6:287–302. Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. 2019. Pretrained encyclopedia: Weakly supervised knowledge-pretrained language model. In Proceedings of 8th International Conference on Learning Representations, ICLR 2020, Virtual Conference, April 26, 2020, Conference Track Proceedings. Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto. 2020. LUKE: Deep contextualized entity representations with entityaware self-attention. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada. Yuan Yao, Deming Ye, Peng Li, Xu Han, Yankai Lin, Zhenghao Liu, Zhiyuan Liu, Lixin Huang, Jie Zhou, and Maosong Sun. 2019. DocRED: A large-scale document-level relation extraction dataset. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 764–777. Deming Ye, Yankai Lin, Jiaju Du, Zhenghao Liu, Maosong Sun, and Zhiyuan Liu. 2020. Coreferential reasoning learning for language representation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7170–7186. Association for Computational Linguistics. Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2335–2344. Dublin City University and Association for Computational Linguistics. Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017. Positionaware attention and supervised data improve slot filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35–45. Association for Computational Linguistics. Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. 2019. ERNIE: Enhanced language representation with informative entities. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1441–1451. Association for Computational Linguistics. 3362 Appendices A Training Details for Downstream Tasks In this section, we introduce the training details for downstream tasks (relation extraction, entity typing and question answering). We implement all models based on Huggingface transformers14. A.1 Relation Extraction Document-level Relation Extraction For document-level relation extraction, we did experiments on DocRED (Yao et al., 2019). We modify the official code15 for implementation. For experiments on three partitions of the original training set (1%, 10% and 100%), we adopt batch size of 10, 32, 32 and training epochs of 400, 400, 200, respectively. We choose Adam optimizer (Kingma and Ba, 2014) as the optimizer and the learning rate is set to 4 × 10−5. We evaluate on dev set every 20/20/5 epochs and then test the best checkpoint on test set on the official evaluation server16. Sentence-level Relation Extraction For sentence-level relation extraction, we did experiments on TACRED (Zhang et al., 2017) and SemEval-2010 Task 8 (Hendrickx et al., 2019) based on the implementation of Peng et al. (2020)17. We did experiments on three partitions (1%, 10% and 100%) of the original training set. The relation representation for each entity pair is obtained in the same way as in pre-training phase. Other settings are kept the same as Peng et al. (2020) for fair comparison. A.2 Entity Tying For entity typing, we choose FIGER (Ling et al., 2015), whose training set is labeled with distant supervision. We modify the implementation of ERNIE (Zhang et al., 2019)18. In fine-tuning phrase, we encode the entities in the same way as in pre-training phase. We set the learning rate to 3 × 10−5 and batch size to 256, and fine-tune the 14https://github.com/huggingface/ transformers 15https://github.com/thunlp/DocRED 16https://competitions.codalab.org/ competitions/20717 17https://github.com/thunlp/ RE-Context-or-Names 18https://github.com/thunlp/ERNIE models for three epochs, other hyper-parameters are kept the same as ERNIE. A.3 Question Answering Multi-choice QA For multi-choice question answering, we choose WikiHop (Welbl et al., 2018). Since the standard setting of WikiHop does not provide the index for each candidate, we then find them by exactly matching them in the documents. We did experiments on three partitions of the original training data (1%, 10% and 100%). We set the batch size to 8 and learning rate to 5 × 10−5, and train for two epochs. Extractive QA For extractive question answering, we adopt MRQA (Fisch et al., 2019) as the testbed and choose three datasets: SQuAD (Rajpurkar et al., 2016), TriviaQA (Joshi et al., 2017) and NaturalQA (Kwiatkowski et al., 2019). We adopt Adam as the optimizer, set the learning rate to 3 × 10−5 and train for two epochs. In the main paper, we report results on two splits (10% and 100%) and results on 1% are listed in Table 11. B Generalized Language Understanding (GLUE) The General Language Understanding Evaluation (GLUE) benchmark (Wang et al., 2018) provides several natural language understanding tasks, which is often used to evaluate PLMs. To test whether LED and LRD impair the PLMs’ performance on these tasks, we compare BERT, ERICABERT, RoBERTa and ERICARoBERTa. We follow the widely used setting and use the [CLS] token as representation for the whole sentence or sentence pair for classification or regression. Table 9 shows the results on dev sets of GLUE Benchmark. It can be observed that both ERICABERT and ERICARoBERTa achieve comparable performance than the original model, which suggests that jointly training LED and LRD with LMLM does not hurt PLMs’ general ability of language understanding. C Full results of ablation study Full results of ablation study (DocRED, WikiHop and FIGER) are listed in Table 10. D Joint Named Entity Recognition and Relation Extraction Joint Named Entity Recognition (NER) and Relation Extraction (RE) aims at identifying entities in text and the relations between them. We 3363 Dataset MNLI(m/mm) QQP QNLI SST-2 CoLA STS-B MRPC RTE BERT 84.0/84.4 88.9 90.6 92.4 57.2 89.7 89.4 70.1 ERICABERT 84.5/84.7 88.3 90.7 92.8 57.9 89.5 89.5 69.6 RoBERTa 87.5/87.3 91.9 92.8 94.8 63.6 91.2 90.2 78.7 ERICARoBERTa 87.5/87.5 91.6 92.6 95.0 63.5 90.7 91.5 78.5 Table 9: Results on dev sets of GLUE Benchmark. We report matched/mismatched (m/mm) accuracy for MNLI, F1 score for QQP and MRPC, spearman correlation for STS-B and accuracy for other tasks. Dataset DocRED WikiHop (m) FIGER Size 1% 10% 100% 1% 10% 100% 100% BERT 28.9 44.9 54.5 37.9 53.1 73.1 72.7 -NSP 30.1 45.2 54.6 38.2 53.6 73.3 72.6 -NSP+LED 34.4 47.6 55.8 41.1 59.8 74.8 73.8 -NSP+L T + c ,T + c RD 34.8 46.4 54.7 37.4 52.2 72.8 72.6 -NSP+L T + s,T + s RD 33.9 47.3 55.5 38.0 51.2 72.5 73.5 -NSP+LRD 35.9 48.0 55.6 37.2 52.0 72.7 74.0 ERICABERT 36.0 48.3 55.9 40.2 58.1 73.9 74.7 Table 10: Full results of ablation study. We report test IgF1 on DocRED, dev accuracy on the masked (m) setting of WikiHop and test micro F1 on FIGER. Setting SQuAD TriviaQA NaturalQA BERT 15.8 28.7 31.5 MTB 11.2 22.0 28.4 CP 12.5 25.6 29.4 ERICABERT 51.3 51.4 42.9 RoBERTa 22.1 40.6 34.0 ERICARoBERTa 57.6 51.3 57.6 Table 11: Results (F1) on extractive QA (SQuAD, TriviaQA and NaturalQA) on 1% split. Model CoNLL04 ADE NER RE NER RE BERT 88.5 70.3 89.2 79.2 ERICABERT 89.3 71.5 89.5 80.2 RoBERTa 89.8 72.0 89.7 81.6 ERICARoBERTa 90.0 72.8 90.2 82.4 Table 12: Results (F1) on joint NER&RE. adopt SpERT (Eberts and Ulges, 2019) as the base model and conduct experiments on two datasets: CoNLL04 (Roth and Yih, 2004) and ADE (Gurulingappa et al., 2012) by replacing the base encoders (BERT and RoBERTa) with ERICABERT and ERICARoBERTa, respectively. We modify the implementation of SpERT19 and keep all the settings the same. From the results listed in Table 12, we can see that ERICA outperforms all baselines, which again demonstrates the superiority of ERICA in 19https://github.com/markus-eberts/spert helping PLMs better understand and represent both entities and relations in text.
|
2021
|
260
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3364–3375 August 1–6, 2021. ©2021 Association for Computational Linguistics 3364 Position Bias Mitigation: A Knowledge-Aware Graph Model for Emotion Cause Extraction Hanqi Yan, Lin Gui, Gabriele Pergola, Yulan He Department of Computer Science, University of Warwick {hanqi.yan, lin.gui, gabriele.pergola, yulan.he}@warwick.ac.uk Abstract The Emotion Cause Extraction (ECE) task aims to identify clauses which contain emotion-evoking information for a particular emotion expressed in text. We observe that a widely-used ECE dataset exhibits a bias that the majority of annotated cause clauses are either directly before their associated emotion clauses or are the emotion clauses themselves. Existing models for ECE tend to explore such relative position information and suffer from the dataset bias. To investigate the degree of reliance of existing ECE models on clause relative positions, we propose a novel strategy to generate adversarial examples in which the relative position information is no longer the indicative feature of cause clauses. We test the performance of existing models on such adversarial examples and observe a significant performance drop. To address the dataset bias, we propose a novel graph-based method to explicitly model the emotion triggering paths by leveraging the commonsense knowledge to enhance the semantic dependencies between a candidate clause and an emotion clause. Experimental results show that our proposed approach performs on par with the existing stateof-the-art methods on the original ECE dataset, and is more robust against adversarial attacks compared to existing models.1 1 Introduction Instead of detecting sentiment polarity from text, recent years have seen a surge of research activities that identify the cause of emotions expressed in text (Gui et al., 2017; Cheng et al., 2017a; Rashkin et al., 2018; Xia and Ding, 2019; Kim and Klinger, 2018; Oberl¨ander and Klinger, 2020). In a typical dataset for Emotion Cause Extract (ECE) (Gui 1Our code can be accessed at https://github.com /hanqi-qi/Position-Bias-Mitigation-in-Em otion-Cause-Analysis et al., 2017), a document consists of multiple clauses, one of which is the emotion clause annotated with a pre-defined emotion class label. In addition, one or more clauses are annotated as the cause clause(s) which expresses triggering factors leading to the emotion expressed in the emotion clause. An emotion extraction model trained on the dataset is expected to classify a given clause as a cause clause or not, given the emotion clause. 1.71 7.71 54.45 23.58 7.47 2.22 0.51 0 10 20 30 40 50 60 Prev3 Prev2 Prev1 emotion Next1 Next2 Next3 Percentage(%) Cause Position Figure 1: The distribution of positions of cause clauses relative to their corresponding emotion clauses in the ECE dataset (Gui et al., 2016). Nearly 87% of cause clauses are located near the emotion clause (About 55% are immediately preceding the emotion clause, 24% are the emotion clauses themselves and over 7% are immediately after the emotion clause). However, due to the difficulty in data collection, the ECE datasets were typically constructed by using emotion words as queries to retrieve relevant contexts as candidates for emotion cause annotation, which might lead to a strong positional bias (Ding and Kejriwal, 2020). Figure 1 depicts the distribution of positions of cause clauses relative to the emotion clause in the ECE dataset (Gui et al., 2016). Most cause clauses are either immediately preceding their corresponding emotion clauses or are the emotion clauses themselves. Existing ECE models tend to exploit such relative position information and have achieved good results on emotion cause detection. For example, The Rel3365 ative Position Augmented with Dynamic Global Labels (PAE-DGL) (Ding et al., 2019), RNNTransformer Hierarchical Network (RTHN) (Xia et al., 2019) and Multi-Attention-based Neural Network (MANN) (Li et al., 2019) all concatenate the relative position embeddings with clause semantic embeddings as the clause representations. We argue that models utilising clause relative positions would inherently suffer from the dataset bias, and therefore may not generalise well to unseen data when the cause clause is not in proximity to the emotion clause. For example, in a recently released emotion cause dataset, only 25-27% cause clauses are located immediately before the emotion clause (Poria et al., 2020). To investigate the degree of reliance of existing ECE models on clause relative positions, we propose a novel strategy to generate adversarial examples in which the relative position information is no longer the indicative feature of cause clauses. We test the performance of existing models on such adversarial examples and observe a significant performance drop. To alleviate the position bias problem, we propose to leverage the commonsense knowledge to enhance the semantic dependencies between a candidate clause and the emotion clause. More concretely, we build a clause graph, whose node features are initialised by the clause representations, and has two types of edges i.e., Sequence-Edge (SEdge) and Knowledge-Edge (K-Edge). A S-Edge links two consecutive clauses to capture the clause neighbourhood information, while a K-Edge links a candidate clause with the emotion clause if there exists a knowledge path extracted from the ConceptNet (Speer et al., 2017) between them. We extend Relation-GCNs (Schlichtkrull et al., 2018) to update the graph nodes by gathering information encoded in the two types of edges. Finally, the cause clause is detected by performing node (i.e., clause) classification on the clause graph. In summary, our contributions are three-fold: • We investigate the bias in the Emotion Cause Extraction (ECE) dataset and propose a novel strategy to generate adversarial examples in which the position of a candidate clause relative to the emotion clause is no longer the indicative feature for cause extraction. • We develop a new emotion cause extraction approach built on clause graphs in which nodes are clauses and edges linking two nodes capture the neighbourhood information as well as the implicit reasoning paths extracted from a commonsense knowledge base between clauses. Node representations are updated using the extended Relation-GCN. • Experimental results show that our proposed approach performs on par with the existing state-of-the-art methods on the original ECE dataset, and is more robust when evaluating on the adversarial examples. 2 Related Work The presented work is closely related to two lines of research in emotion cause extraction: positioninsensitive and position-aware models. Position-insensitive Models. A more traditional line of research exploited structural representations of textual units relying on rule-based systems (Lee et al., 2010) or incorporated commonsense knowledge bases (Gao et al., 2015) for emotion cause extraction. Machine learning methods leveraged text features (Gui et al., 2017) and combined them with multi-kernel Support Vector Machine (SVM) (Xu et al., 2017). More recent works developed neural architectures to generate effective semantic features. Cheng et al. (2017b) employed LSTM models, Gui et al. (2017) made use of memory networks, while Li et al. (2018) devised a Convolutional Neural Network (CNN) with a co-attention mechanism. (Chen et al., 2018) used the emotion classification task to enhance cause extraction results. Position-aware Models. More recent methodologies have started to explicitly leverage the positions of cause clauses with respect to the emotion clause. A common strategy is to concatenate the clause relative position embedding with the candidate clause representation (Ding et al., 2019; Xia et al., 2019; Li et al., 2019). The Relative Position Augmented with Dynamic Global Labels (PAE-DGL) (Ding et al., 2019) reordered clauses based on their distances from the target emotion clause, and propagated the information of surrounding clauses to the others. Xu et al. (2019) used emotion dependent and independent features to rank clauses and identify the cause. The RNN-Transformer Hierarchical Network (RTHN) (Xia et al., 2019) argued there exist relations between clauses in a document and proposed to classify multiple clauses simultaneously. Li et al. (2019) proposed a Multi-Attention-based Neural Network (MANN) to model the interactions between a candidate clause and the emotion clause. 3366 Bi-LSTM Bi-LSTM . . . Transformer 𝑝1 . . . 𝑝2 Bi-LSTM 𝑠1 𝛼1 𝛼𝐾 S-Edge K-Edge 𝐶)1 𝐶)𝐸 𝐶)8 . . . 𝐶)2 𝐶)6 C1 C8 C5 p1 p2 Bi-LSTM 𝛼2 𝐷 p1 p2 e15 𝐶)𝐸 ConceptNet ℎ/ ℎ0 Softmax 𝑦2𝟏 Document (b) Path Representations. (a) Document Encoding. (c) Clause graph update. (d) Classification. . . . . . . . . . 𝐶4 Bi-LSTM Figure 2: The framework of our proposed KAG. Given an input document consisting of eight clauses (C1 · · · C8), we first extract knowledge paths from ConceptNet between each candidate clause and the emotion clause (§3.1), e.g., two knowledge paths, p1 and p2, are extracted between C1 and the emotion clause C5. (a) Document Encoding. Clauses are fed into a word-level Bi-LSTM and a clause-level Transformer to obtain the clause representations ˆCi. The document embedding D is generated by Dot-Attention between the emotion embedding ˆ CE and clause embeddings. (b) Path Representations. The extracted knowledge paths are fed into Bi-LSTM to derive path representations. Multiple paths between a clause pair are aggregated into si based on their attention to the document representation D. (c) Clause Graph Update. A clause graph is built with the clause representations ˆCi used to initialise the graph nodes. The K-Edge weight eiE between a candidate clause ˆCi and the emotion clause ˆCE are measured by their distance along their path si. (d) Classification. Node representation hi of a candidate clause Ci is concatenated with the emotion node representation hE, and then fed to a softmax layer to yield the clause classification result ˆyi. The generated representations are fed to a CNN layer for emotion cause extraction. The Hierarchical Neural Network (Fan et al., 2019) aimed at narrowing the gap between the prediction distribution p and the true distribution of the cause clause relative positions. 3 Knowledge-Aware Graph (KAG) Model for Emotion Cause Extraction We first define the Emotion Cause Extraction (ECE) task here. A document D contains N clauses D = {Ci}N i=1, one of which is annotated as an emotion clause CE with a pre-defined emotion class label, Ew. The ECE task is to identify one or more cause clauses, Ct, 1 ≤t ≤N, that trigger the emotion expressed in CE. Note that the emotion clause itself can be a cause clause. We propose a Knowledge-Aware Graph (KAG) model as shown in Figure 2, which incorporates knowledge paths extracted from ConceptNet for emotion cause extraction. More concretely, for each document, a graph is first constructed by representing each clause in the document as a node. The edge linking two nodes captures the sequential relation between neighbouring clauses (called the Sequence Edge or S-Edge). In addition, to better capture the semantic relation between a candidate clause and the emotion clause, we identify keywords in the candidate clause which can reach the annotated emotion class label by following the knowledge paths in the ConceptNet. The extracted knowledge paths from ConceptNet are used to enrich the relationship between the candidate clause and the emotion clause and are inserted into the clause graph as the Knowledge Edge or K-Edge. We argue that by adding the K-Edges, we can better model the semantic relations between a candidate clause and the emotion clause, regardless of their relative positional distance. In what follows, we will first describe how to extract knowledge paths from ConceptNet, then present the incorporation of the knowledge paths into context modelling, and finally discuss the use of Graphical Convolutional Network (GCN) for learning node (or clause) representations and the prediction of the cause clause based on the learned node representations. 3.1 Knowledge Path Extraction from ConceptNet ConceptNet is a commonsense knowledge graph, which represents entities as nodes and relationship between them as edges. To explore the causal re3367 Bai Jinyue, an ordinary worker in XingTai Steel factory in HeBei province and the department leader replied to my mail when I found that my advice had been adopted I realized that I had made contributions to the country's development talked to the journalist with exicitment different departments, like the public security, have accepted his advice with a Thank You letter in his hands C2) C4) C3) C6) C7) C8) Since 27 years ago acceptance culture diffusion spread C1) C5) make better world happiness ConceptNet Figure 3: A document consisting of 8 clauses in the ECE dataset with extracted knowledge paths from the ConceptNet. Words in red are identified keywords. ‘happiness’ is the emotion label of the emotion clause C5. For better visualization, we only display two extracted knowledge paths between ‘adopt’ and ‘happiness’ in the ConceptNet. lation between a candidate clause and the emotion clause, we propose to extract cause-related paths linking a word in the candidate clause with the annotated emotion word or the emotion class label, Ew, in the emotion clause. More concretely, for a candidate clause, we first perform word segmentation using the Chinese segmentation tool, Jieba2, and then extract the top three keywords ranked by Text-Rank3. Based on the findings in (Fan et al., 2019) that sentiment descriptions can be relevant to the emotion cause, we also include adjectives in the keywords set. We regard each keyword in a candidate clause as a head entity, eh, and the emotion word or the emotion class label in the emotion clause as the tail entity, et. Similar to (Lin et al., 2019), we apply networkx4 to perform a depth-first search on the ConceptNet to identify the paths which start from eh and end at et, and only keep the paths which contain less than two intermediate entities. This is because shorter paths are more likely to offer reliable reasoning evidence (Xiong et al., 2017). Since not all relations in ConceptNet are related to or indicative of causal relations, we further remove the paths which contain any of these four relations: ‘antonym’, ‘distinct from’, ‘not desires’, and ‘not capable of’. Finally, we order paths by their lengths in an ascending order and choose the top K paths as the result for each candidateemotion clause pair5. An example is shown in Figure 3. The 5-th 2https://github.com/fxsjy/jieba 3We have also experimented with other keyword extraction strategies, such as extracting words with higher TFIDF values or keeping all words after removing the stop words. But we did not observe improved emotion cause detection results. 4http://networkx.github.io/ 5We set K to 15, which is the median of the number of paths between all the candidate-emotion clause pairs in our dataset. clause is annotated as the emotion clause and the emotion class label is ‘happiness’. For the keyword, ‘adopted’, in the first clause, we show two example paths extracted from ConceptNet, each of which links the word ‘adopted’ with ‘happiness’. One such a path is “adopted −related to→acceptance −has subevent→make better world −causes→happiness”. 3.2 Knowledge-Aware Graph (KAG) Model As shown in Figure 2, there are four components in our model: a document encoding module, a context-aware path representation learning module, a GCN-based graph representation updating module, and finally a softmax layer for cause clause classification. Initial Clause/Document Representation Learning For each clause Ci, we derive its representation, Ci, by using a Bi-LSTM operating on its constituent word vectors, where each word vector wi ∈Rd is obtained via an embedding layer. To capture the sequential relationship (S-Edges) between neighbouring clauses in a document, we feed the clause sequence into a transformer architecture. Similar to the original transformer incorporating the position embedding with the word embedding, we utilise the clause position information to enrich the clause representation. Here, the position embedding oi of each clause is concatenated with its representation Ci generated by Bi-LSTM. ˆ Ci = Transformer(Ci || oi) (1) We consider different ways for encoding position embeddings using either relative or absolute clause positions and explore their differences in the experiments section. In addition, we will also show the results without using position embeddings at all. 3368 Since the aim of our task is to identify the cause clause given an emotion clause, we capture the dependencies between each candidate clause and the emotion clause. Therefore, in the document context modelling, we consider the emotion clause ˆ CE, generated in a similar way as ˆ Ci, as the query vector, and the candidate clause representation ˆCi as both the key and value vectors, in order to derive the document representation, D ∈Rd. Context-Aware Path Representation In Section 3.1, we have chosen a maximum of K paths {pt}K t=1 linking each candidate Ci with the emotion clause. However, not every path correlates equally to the document context. Taking the document shown in Figure 3 as an example, the purple knowledge path is more closely related to the document context compared to the green path. As such, we should assign a higher weight to the purple path than the green one. We propose to use the document-level representation D obtained above as the query vector, and a knowledge path as both key and value vectors, in order to calculate the similarity between the knowledge path and the document context. For each pair of a candidate clause Ci and the emotion clause, we then aggregate the K knowledge paths to derive the contextaware path representation si ∈Rd below: si = K X t=1 αtpt αt = softmax( DT pt PK j=1 DT pj ) (2) where D is the document representation, pt is the path representation obtained from Bi-LSTM on a path expressed as an entity-relation word sequence. Update of Clause Representations by GCN After constructing a clause graph such as the one shown in Figure 2(c), we update the clause/node representations via S-Edges and K-Edges. Only clauses with valid knowledge paths to the emotion clause are connected with the emotion clause node. After initialising the node (or clause) in the clause graph with ˆCi and the extracted knowledge path with si, we update clause representation using an extended version of GCN, i.e. RelationGCNs (aka. R-GCNs) (Schlichtkrull et al., 2018), which is designed for information aggregation over multiple different edges: hℓ+1 i = σ( X r∈RNi X j∈Ni 1 ci,r W ℓ r hℓ j + W ℓ 0 hℓ i) (3) where W ℓ r hℓ j is the linear transformed information from the neighbouring node j with relation r at the ℓ-th layer, W ℓ r ∈Rd×d is relation-specific, Ni is the set of neighbouring nodes of the i-th node, RNj is the set of distinct edges linking the current node and its neighbouring nodes. When aggregating the neighbouring nodes information along the K-Edge, we leverage the path representation si to measure the node importance. This idea is inspired by the translation-based models in graph embedding methods (Bordes et al., 2013). Here, if a clause pair contains a possible reasoning process described by the K-Edge, then ˆhE ≈ˆhi + si holds. Otherwise, ˆhi + si should be far away from the emotion clause representation ˆhE.6 Therefore, we measure the importance of graph nodes according to the similarity between (hi + si) and hE. Here, we use the scaled DotAttention to calculate the similarity eiE and obtain the updated node representation zi. zi = softmax(eE)hℓ E eiE = (hi + si)T hE √ d (i ̸= E) (4) where eE is {eiE}N−1 i=1 . d is the dimension of graph node representations, and N rk is a set of neighbours by the K-Edge. Then, we combine the information encoded in SEdge with zi as in Eq. 3, and perform a non-linear transformation to update the graph node representation hℓ+1 i : hℓ+1 i = σ zℓ i + X j∈Nrs i (Wjhj) (5) where Nrs i is a set of i-th neighbours connected by the S-Edges. Cause Clause Detection Finally, we concatenate the candidate clause node hi and the emotion node representation he generated by the graph, and apply a softmax function to yield the predictive class distribution ˆyi. ˆyi = softmax W (hL i || hL E) + b , (6) 4 Experiments We conduct a thorough experimental assessment of the proposed approach against several state-of-theart models7. 6Here, we do not consider the cases when the candidate clause is the emotion clause (i.e., ˆhi = ˆhE), as the similarity between ˆhE + si and ˆhE will be much larger than the other pairs. 7Training and hyper-parameter details can be found in Appendix A. 3369 Methods P (%) R (%) F1 (%) W/O Pos RB 67.47 42.87 52.43 EMOCause 26.72 71.30 38.87 Ngrams+SVM 42.00 43.75 42.85 Multi-Kernel 65.88 69.27 67.52 CNN 62.15 59.44 60.76 CANN 77.21 68.91 72.66 Memnet 70.76 68.38 69.55 W. Pos HCS 73.88 71.54 72.69 MANN 78.43 75.87 77.06 LambdaMART 77.20 74.99 76.08 PAE-DGL 76.19 69.08 72.42 RTHN 76.97 76.62 76.77 Our KAG 79.12 75.81 77.43 : w/o R-GCNs 73.68 72.76 73.14 : w/o K-Edge 75.67 72.63 74.12 : w/o S-Edge 76.34 75.46 75.88 Table 1: Results of different models on the ECE dataset. Our model achieves the best Precision and F1 score. Dataset and Evaluation Metrics The evaluation dataset (Gui et al., 2016) consists of 2,105 documents from SINA city news. As the dataset size is not large, we perform 10-fold cross-validation and report results on three standard metrics, i.e. Precision (P), Recall (R), and F1-Measure, all evaluated at the clause level. Baselines We compare our model with the position-insensitive and position-aware baselines: RB (Lee et al., 2010) and EMOCause (Russo et al., 2011) are rules-based methods. Multi-Kernel (Gui et al., 2016) and Ngrams+SVM (Xu et al., 2017) leverage Support Vector Machines via different textual feature to train emotion cause classifiers. CNN (Kim, 2014) and CANN (Li et al., 2018) are vanilla or attention-enhanced approaches. Memnet (Gui et al., 2017) uses a deep memory network to re-frame ECE as a question-answering task. Position-aware models use the relative position embedding to enhance the semantic features. HCS (Yu et al., 2019) uses separate hierarchical and attention module to obtain context and information. Besides that, PAE-DGL (Ding et al., 2019) and RTHN (Xia et al., 2019) use similar Global Prediction Embedding (GPE) to twist the clauses’ first-round predictions. MANN (Li et al., 2019) performs multi-head attention in CNN to jointly encode the emotion and candidate clauses. LambdaMART (Xu et al., 2019) uses the relative position, word-embedding similarity and topic similarity as emotion-related feature to extract cause. 4.1 Main Results Table 1 shows the cause clause classification results on the ECE dataset. Two rule-based methods have poor performances, possibly due to their pre-defined rules. Multi-Kernel performs better than the vanilla SVM, being able to leverage more contextual information. Across the other three groups, the precision scores are higher than recall scores, and it is probably due to the unbalanced number of cause clauses (18.36%) and non-cause clauses (81.64%), leading the models to predict a clause as non-cause more often. Models in the position-aware group perform better than those in the other groups, indicating the importance of position information. Our proposed model outperforms all the other models except RHNN in which its recall score is slightly lower. We have also performed ablation studies by removing either K-Edge or S-Edge, or both of them (w/o R-GCNs). The results show that removing the RGCNs leads to a drop of nearly 4.3% in F1. Also, both the K-Edge and S-Edge contributes to emotion cause extraction. As contextual modelling has considered the position information, the removal of S-Edge leads to a smaller drop compared to the removal of K-Edge. 4.2 Impact of Encoding Clause Position Information In order to examine the impact of using the clause position information in different models, we replace the relative position information of the candidate clause with absolute positions. In the extreme case, we remove the position information from the models. The results are shown in Figure 4. It can be observed that the best results are achieved using relative positions for all models. Replacing relative positions using either absolution positions or no position at all results in a significant performance drop. In particular, MANN and PAE-DGL have over 50-54% drop in F1. The performance degradation is less significant for RTHN, partly due to its use of the Transformer architecture for context modeling. Nevertheless, we have observed a decrease in F1 score in the range of 20-35%. Our proposed model is less sensitive to the relative positions of candidate clauses. Its robust performance partly attributes to the use of (1) hierarchical contextual modeling via the Transformer structure, and (2) the K-Egde which helps explore causal links via commonsense knowledge regardless of a clause’s 3370 65.49 72.42 76.77 77.08 15.31 18.39 56.94 69.43 15.09 17.9 41.45 68.29 15 25 35 45 55 65 75 85 MANN PAE-DGL RTHN OURS F1(%) relative position absolute position no position Figure 4: Emotion cause extraction when using relative, absolute or no clause positional information. Our model demonstrates most stable performance without the relative position information. relative position. 4.3 Performance under Adversarial Samples In recent years, there have been growing interests in understanding vulnerabilities of NLP systems (Goodfellow et al., 2015; Ebrahimi et al., 2017; Wallace et al., 2019; Jin et al., 2020). Adversarial examples explore regions where the model performs poorly, which could help understanding and improving the model. Our purpose here is to evaluate if KAG is vulnerable as existing ECE models when the cause clauses are not in proximity to the emotion clause.Therefore, we propose a principled way to generate adversarial samples such that the relative position is no longer an indicative feature for the ECE task. Generation of adversarial examples We generate adversarial examples to trick ECE models, which relies on swapping two clauses Cr1 and Cr2, where r1 denotes the position of the most likely cause clause, while r2 denotes the position of the least likely cause clause. We identify r1 by locating the most likely cause clause based on its relative position with respect to the emotion clause in a document. As illustrated in Figure 1, over half of the cause clauses are immediately before the emotion clause in the dataset. We assume that the position of a cause clause can be modelled by a Gaussian distribution and estimate the mean and variance directly from the data, which are, {µ, σ2} = {−1, 0.5445}. The position index r1 can then be sampled from the Gaussian distribution. As the sampled value is continuous, we round the value to its nearest integer: r1 ←⌊g⌉, g ∽Gaussian(µ, σ2). (7) To locate the least likely cause clause, we propose to choose the value for r2 according to the attention score between a candidate clause and the emotion clause. Our intuition is that if the emotion clause has a lower score attended to a candidate clause, then it is less likely to be the cause clause. We use an existing emotion cause extraction model to generate contextual representations and use the Dot-Attention (Luong et al., 2015) to measure the similarity between each candidate clause and the emotion clause. We then select the index i which gives the lowest attention score and assign it to r2: r2 = arg min i {λi}N i=1, λi = Dot-Att.( ˆ Ci, ˆ CE), (8) where ˆ Ci is the representation of the i-th candidate clause, ˆ CE is the representation of the emotion clause, and N denotes a total of N clauses in a document. Here, we use existing ECE models as different discriminators to generate different adversarial samples.8 The desirable adversarial samples will fool the discriminator to predict the inverse label. We use leave-one-model-out to evaluate the performance of ECE models. In particular, one model is used as a Discriminator for generating adversarial samples which are subsequently used to evaluate the performance of other models. Results The results are shown in Table 2. The attacked ECE models are merely trained on the original dataset. The generated adversarial examples are used as the test set only. We can observe a significant performance drop of 23-32% for the existing ECE models, some of which even perform worse than the earlier rule-based methods, showing their sensitivity to the positional bias in the dataset. We also observe the performance degradation of our proposed KAG. But its performance drop is less significant compared to other models. The results verify the effectiveness of capturing the semantic dependencies between a candidate clause and the emotion clause via contextual and commonsense knowledge encoding. 4.4 Case Study and Error Analysis To understand how KAG aggregate information based on different paths, we randomly choose two examples to visualise the attention distributions (Eq. 4) on different graph nodes (i.e., clauses) 8The adversarial sample generation is independent from their training process. 3371 Discriminator Attacked ECE models PAEDGL MANN RTHN KAG PAEDGL 49.62 48.92 59.73 64.98 ↓31.76% ↓28.6% ↓22.20% ↓16.08% MANN 51.82 47.24 60.13 66.32 ↓28.45% ↓31.27% ↓21.65% ↓14.35% RTHN 48.63 49.63 57.78 63.47 ↓32.85% ↓27.64% ↓24.74% ↓18.03% KAG 48.52 48.24 59.53 62.39 ↓33.00% ↓29.67% ↓22.46% ↓19.42% Ave. Drop(%) ↓31.51% ↓29.29% ↓22.62% ↓16.97% Table 2: F1 score and relative drop (marked with ↓) of different ECE models on adversarial samples. The listed four ECE models are attacked by the adversarial samples generated from the respective discriminator. Our model shows the minimal drop rate comparing to other listed ECE models across all sets of adversarial samples. in Figure 5.9 These attention weights show the ‘distance’ between a candidate clause and the emotion clause during the reasoning process. The cause clauses are underlined, and keywords are in bold. Ci in brackets indicate the relative clause position to the emotion clause (which is denoted as C0). Ex.1 The crime that ten people were killed shocked the whole country (C−4). This was due to personal grievances (C−3). Qiu had arguments with the management staff (C−2), and thought the Taoist temple host had molested his wife (C−1). He became angry (C0), and killed the host and destroyed the temple (C1). In Ex.1, the emotion word is ‘angry’, the knowledge path identified by our model from ConceptNet is, “arguments →fight →angry” for Clause C−2, and “molest →irritate →exasperate→angry” for Clause C−1. Our model assigns the same attention weight to the clauses C−2, C−1 and the emotion clause, as shown in Figure 5. This shows that both paths are equally weighted by our model. Due to the K-Edge attention weights, our model can correctly identify both C−2 and C−1 clauses as the cause clauses. Ex.2 The LongBao Primary school locates between the two villages (C−2). Some unemployed people always cut through the school to take a shortcut (C−1). Liu Yurong worried that it would affect children’s study (C0). When he did not have teaching duties (C1), he stood guard outside the school gate (C2). In Ex.2, the path identified by our model from ConceptNet for Clause (C−1) is “unemployment →situation →trouble/danger→worried”. It has 9More cases can be found in the Appendix. been assigned the largest attention weight as shown in Figure 5. Note that the path identified is spurious since the emotion of ‘worried’ is triggered by ‘unemployment’ in the ConceptNet, while in the original text, ‘worried’ is caused by the event, ‘Unemployed people cut through the school’. This shows that simply using keywords or entities searching for knowledge paths from commonsense knowledge bases may lead to spurious knowledge extracted. We will leave the extraction of event-driven commonsense knowledge as future work. 0.0867 0.613 0.157 0.0869 0.056 0.082 0.052 0.262 0.262 0.262 0.082 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 pre4 pre3 pre3 pre1 emotion next1 next2 Attention Weights Clause Location Ex.1 Ex.2 Figure 5: Attention weights among different graph nodes/clauses on Ex.1 and Ex.2. 5 Conclusion and Future Work In this paper, we examine the positional bias in the annotated ECE dataset and investigate the degree of reliance of the clause position information in existing ECE models. We design a novel approach for generating adversarial samples. Moreover, we propose a graph-based model to enhance the semantic dependencies between a candidate clause and a given emotion clause by extracting relevant knowledge paths from ConceptNet. The experimental results show that our proposed method achieves comparative performance to the state-of-the-art methods, and is more robust against adversarial attacks. Our current model extracts knowledge paths linking two keywords identified in two separate clauses. In the future, we will exploit how to incorporate the event-level commonsense knowledge to improve the performance of emotion cause extraction. Acknowledgements This work was funded by the EPSRC (grant no. EP/T017112/1, EP/V048597/1). HY receives the PhD scholarship funded jointly by the University of Warwick and the Chinese Scholarship Council. YH is supported by a Turing AI Fellowship funded by the UK Research and Innovation (grant no. EP/V020579/1). We thank Yizhen Jia and 3372 Daoye Zhu for their valuable work on earlier code framework of this paper. We also thank the anonymous reviewers for their valuable comments. References Antoine Bordes, Nicolas Usunier, Alberto GarciaDuran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In C. J. C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 26, NIPS13, pages 2787–2795. Ying Chen, Wenjun Hou, Xiyao Cheng, and Shoushan Li. 2018. Joint learning for emotion classification and emotion cause detection. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 646–651, Brussels, Belgium. Association for Computational Linguistics. Xiyao Cheng, Ying Chen, Bixiao Cheng, Shoushan Li, and Guodong Zhou. 2017a. An emotion cause corpus for chinese microblogs with multiple-user structures. ACM Transactions on Asian and LowResource Language Information Processing, 17:1– 19. Xiyao Cheng, Ying Chen, Bixiao Cheng, Shoushan Li, and Guodong Zhou. 2017b. An emotion cause corpus for chinese microblogs with multiple-user structures. ACM Transaction Asian Low-Res. for Lang. Inf. Process., 17(1). Jiayuan Ding and Mayank Kejriwal. 2020. An experimental study of the effects of position bias on emotion causeextraction. CoRR, abs/2007.15066. Zixiang Ding, Huihui He, Mengran Zhang, and Rui Xia. 2019. From independent prediction to reordered prediction: Integrating relative position and global label information to emotion cause identification. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, pages 6343–6350. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751. Chuang Fan, Hongyu Yan, Jiachen Du, Lin Gui, Lidong Bing, Min Yang, Ruifeng Xu, and Ruibin Mao. 2019. A knowledge regularized hierarchical approach for emotion cause analysis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5614–5624, Hong Kong, China. Association for Computational Linguistics. Kai Gao, Hua Xu, and Jiushuo Wang. 2015. A rulebased approach to emotion cause detection for chinese micro-blogs. Expert Systems with Applications, 42(9):4517 – 4528. Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and harnessing adversarial examples. Lin Gui, Jiannan Hu, Yulan He, Ruifeng Xu, Qin Lu, and Jiachen Du. 2017. A question answering approach for emotion cause extraction. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017, pages 1593–1602. Lin Gui, Dongyin Wu, Ruifeng Xu, Qin Lu, and Yu Zhou. 2016. Event-driven emotion cause extraction with corpus construction. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 1639–1649. Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018–8025. Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1345–1359, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751, Doha, Qatar. Sophia Yat Mei Lee, Ying Chen, and Chu-Ren Huang. 2010. A text-driven rule-based system for emotion cause detection. In Proceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pages 45–53, Los Angeles, CA. Association for Computational Linguistics. Xiangju Li, Shi Feng, Daling Wang, and Yifei Zhang. 2019. Context-aware emotion cause analysis with multi-attention-based neural network. KnowledgeBased Systems, 174:205 – 218. Xiangju Li, Kaisong Song, Shi Feng, Daling Wang, and Yifei Zhang. 2018. A co-attention neural network model for emotion cause analysis with emotional context awareness. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 4752–4757. Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. Kagnet: Knowledge-aware graph networks for commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods 3373 in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 2829–2839. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, EMNLP 2015, Lisbon, Portugal, September 17-21, 2015, pages 1412–1421. Laura Oberl¨ander and Roman Klinger. 2020. Sequence labeling vs. clause classification for english emotion stimulus detection. In Proceedings of the 9th Joint Conference on Lexical and Computational Semantics (*SEM 2020), Barcelona, Spain. Association for Computational Linguistics. Soujanya Poria, Navonil Majumder, Devamanyu Hazarika, Deepanway Ghosal, Rishabh Bhardwaj, Samson Yu Bai Jian, Romila Ghosh, Niyati Chhaya, Alexander Gelbukh, and Rada Mihalcea. 2020. Recognizing emotion cause in conversations. arXiv preprint arXiv:2012.11820. Hannah Rashkin, Antoine Bosselut, Maarten Sap, Kevin Knight, and Yejin Choi. 2018. Modeling naive psychology of characters in simple commonsense stories. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2289–2299, Melbourne, Australia. Irene Russo, Tommaso Caselli, Francesco Rubino, Ester Boldrini, and Patricio Mart´ınez-Barco. 2011. Emocause: An easy-adaptable approach to extract emotion cause contexts. In Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis, WASSA@ACL 2011, Portland, OR, USA, June 24, 2011, pages 153–160. Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne vanden Berg, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In European Semantic Web Conference. Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4444–4451. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing nlp. arXiv preprint arXiv:1908.07125. Rui Xia and Zixiang Ding. 2019. Emotion-cause pair extraction: A new task to emotion analysis in texts. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 1003–1012. Rui Xia, Mengran Zhang, and Zixiang Ding. 2019. RTHN: A rnn-transformer hierarchical network for emotion cause extraction. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 5285–5291. Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. Deeppath: A reinforcement learning method for knowledge graph reasoning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing (EMNLP 2017), Copenhagen, Denmark. ACL. B. Xu, H. Lin, Y. Lin, Y. Diao, L. Yang, and K. Xu. 2019. Extracting emotion causes using learning to rank methods from an information retrieval perspective. IEEE Access, 7:15573–15583. Ruifeng Xu, Jiannan Hu, Qin Lu, Dongyin Wu, and Lin Gui. 2017. An ensemble approach for emotion cause detection with event extraction and multikernel svms. Tsinghua Science and Technology, 22(6):646–659. Xinyi Yu, Wenge Rong, Zhuo Zhang, Yuanxin Ouyang, and Zhang Xiong. 2019. Multiple level hierarchical network-based clause selection for emotion cause extraction. IEEE Access, 7:9071–9079. 3374 A Model Architecture In this section, we describe the details of the four main components in our model: contextual modelling, knowledge path encoding, clause graph update and cause clause classification. The dataset has 2,105 documents. The maximum number of clauses in a document is 75 and the maximum number of words per clause is 45. So we first pad the input documents into a matrix I with the shape of [2105, 75, 45]. A.1 Contextual Modelling a. token →clause We first apply a 1-layer BiLSTM of 100 hidden units to obtain word embeddings, w ∈R200. We then use two linear transformation layers (hidden units are [200,200],[200,1]) to map the original w to a scalar attention score α, then perform a weighted aggregation to generate the clause representation ˆCi ∈R200. b. clause →document We feed the clause representations into a Transformer. It has 3 stacked blocks, with the multi-head number set to 5, and the dimension of key, value, query is all set to 200. The query vector is the emotion clause representation ˆCE ∈R200, the key and value representations are candidate clause representations, also with 200 dimensions. Finally, the updated clause representations are aggregated via Dot-Attention to generate the document representation D ∈R200. A.2 Knowledge Path Encoding For each candidate clause and the emotion clause, we extract knowledge paths from ConceptNet and only select K paths. The values of K is set to 15, since the median of the number of paths between a candidate clause and the emotion clause is 15 in our dataset. We use the same Bi-LSTM described in Section A.1 to encode each knowledge path and generate the K number of path representations {pit}K t=1 between the i-th clause and the emotion clause. Then, the document representation D is applied as the query to attend to each path in {pit} to generate the final context-aware path representation si ∈R200. A.3 Clause Graph Update The graph nodes are initialised by clause presentations, with the feature dimension 200. To calculate the attention weights eiE in R-GCNs, We use the non-linearly transformed hi + si as the query, the non-linearly transformed hE as the value and key. The non-linear functions are independent Selu layers. A.4 Cause Clause Classification The MLP with [400,1] hidden units takes the concatenation of each candidate node {hL i }N i=1 and the emotion node representation hL E to predict the logit, after which, a softmax layer is applied to predict the probability of the cause clause. B Training Details for KAG We randomly split the datasets into 9:1 (train/test). For each split, we run 50 iterations to get the best model on the validation set, which takes an average time of around 23 minutes per split, when conducted on a NVIDIA GTX 1080Ti. For each split, we test the model on the test set at the end of each iteration and keep the best resulting F1 of the split. The number of model parameters is 1,133,002. Hyper-parameter Search We use the grid search to find the best parameters for our model on the validation data, and report in the following the hyper-parameter values providing the best performance. • The word embeddings used to initialise the Bi-LSTM is provided by NLPCC10. It was pre-trained on a 1.1 million Chinese Weibo corpora following the Word2Vec algorithm. The word embedding dimension is set to 200. • The position embedding dimension is set to 50, randomly initialised with the uniform distribution (-0.1,0.1). • The number of Transformer blocks is 2 and the number of graph layers is 3. • To regularise against over-fitting, we employ dropout (0.5 in the encoder, 0.2 in the graph layer). • The network is trained using the the Adam optimiser with a mini-batch size 64 and a learning rate η = 0.005. The parameters of our model are initialised with Glorot initialisation. C Error Analysis We perform error analysis to identify the limitations of the proposed model. In the following examples (Ex.1 and Ex.2), the cause clauses are in bold, our predictions are underlined. 10https://github.com/NUSTM/RTHN/tree/master/data 3375 Ex.1 Some kind people said (C−6), if Wu Xiaoli could find available kidneys (C−5), they would like to donate for her surgery (C−4). 4000RMB donation had been sent to Xiaoli (C−3), Qiu Hua said (C−2). The child’s desire to survival shocked us (C−1). The family’s companion was touching (C0). Wish kind people will be ready to give a helping hand (C1). Help the family in difficulty (C2). In the first example Ex.1, our model identifies the keyword survival in C−1 and extracts several paths from ‘survival’ to ‘touching’. However, the main event in clause C−1 concerns desire rather than survival. Our current model detects the emotion reasoning process from ConceptNet based on keywords identified in text, and inevitably introduces spurious knowledge paths to model learning. Ex.2 I have only one daughter (C0), and a granddaughter of 8 year-old (C−10). I would like to convey these memory to her (C−9). Last Spring Festival (C−8), I gave the DVD away to my granddaughter (C−7). I hope she can inherit my memory (C−6). Thus (C−5), I feel like that my ages become eternity (C−4). Sun Qing said (C−3). His father is a sensitive and has great passion for his life (C−2). He did so (C−1). Making me feel touched (C0). His daughter said (C1). In the Ex 2, our model detected the passion as a keyword and extracted knowledge paths between the clause C−2 and the emotion clause. However, it ignores the semantic dependency between the clause C−1 and the emotion clause. It is therefore more desirable to consider semantic dependencies or discourse relations between clauses/sentences for emotion reasoning path extraction from external commonsense knowledge sources. D Human Evaluation on the Generated Adversarial Samples The way adversarial examples generated changes the order of the original document clauses. Therefore, we would like to find out if such clause reordering changes the original semantic meaning and if these adversarial samples can be used to evaluate on the same emotion cause labels. We randomly selected 100 adversarial examples and ask two independent annotators to manually annotate emotion cause clauses based on the same annotation scheme of the ECE dataset. Compared to the original annotations, Annotator 1 achieved 0.954 agreement with the cohen’s kappa value of 0.79, while Annotator 2 achieved 0.938 agreement with the cohen’s kappa value of 0.72. This aligns with our intuition that an emotion expressed in text is triggered by a certain event, rather than determined by relative clause positions. A good ECE model should be able to learn a correlation between an event and its associated emotion. This also motivates our proposal of a knowledge-aware model which leverages commonsense knowledge to explicitly capture event-emotion relationships.
|
2021
|
261
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3376–3386 August 1–6, 2021. ©2021 Association for Computational Linguistics 3376 Every Bite Is an Experience: Key Point Analysis of Business Reviews Roy Bar-Haim, Lilach Eden, Yoav Kantor∗, Roni Friedman, Noam Slonim IBM Research {roybar,lilache,yoavka,roni.friedman-melamed,noams}@il.ibm.com Abstract Previous work on review summarization focused on measuring the sentiment toward the main aspects of the reviewed product or business, or on creating a textual summary. These approaches provide only a partial view of the data: aspect-based sentiment summaries lack sufficient explanation or justification for the aspect rating, while textual summaries do not quantify the significance of each element, and are not well-suited for representing conflicting views. Recently, Key Point Analysis (KPA) has been proposed as a summarization framework that provides both textual and quantitative summary of the main points in the data. We adapt KPA to review data by introducing Collective Key Point Mining for better key point extraction; integrating sentiment analysis into KPA; identifying good key point candidates for review summaries; and leveraging the massive amount of available reviews and their metadata. We show empirically that these novel extensions of KPA substantially improve its performance. We demonstrate that promising results can be achieved without any domain-specific annotation, while human supervision can lead to further improvement. 1 Introduction With their ever growing prevalence, online opinions and reviews have become essential for our everyday decision making. We turn to the wisdom of the crowd before buying a new laptop, choosing a restaurant or planning our next vacation. However, this abundance is often overwhelming: reading hundreds or thousands of reviews on a certain business or product is impractical, and users typically have to rely on aggregated numeric ratings, complemented by reading a small sample of reviews, which may not be representative. The vast majority of available information is left unexploited. ∗First three authors equally contributed to this work. Opinion summarization is a long-standing challenge, which has attracted a lot of research interest over the past two decades. Early works (Hu and Liu, 2004; Gamon et al., 2005; Snyder and Barzilay, 2007; Blair-goldensohn et al., 2008; Titov and McDonald, 2008) aimed to extract, aggregate and quantify the sentiment toward the main aspects or features of the reviewed entity (e.g., food, price, service, and ambience for restaurants). Such aspectbased sentiment summaries provide a high-level, quantitative view of the summarized opinions, but lack explanations and justifications for the assigned scores (Ganesan et al., 2010). An alternative line of work casts this problem as multi-document summarization, aiming to create a textual summary from the input reviews (Carenini et al., 2006; Ganesan et al., 2010; Chu and Liu, 2019; Braˇzinskas et al., 2020b). While such summaries provide more detail, they lack a quantitative view of the data. The salience of each element in the summary is not indicated, making it difficult to evaluate their relative significance. This is particularly important for the common case of conflicting opinions. In order to fully capture the controversy, the summary should ideally indicate the proportion of favorable vs. unfavorable reviews for the controversial aspect. Recently, Key Point Analysis (KPA) has been proposed as a novel extractive summarization framework that addresses the limitations of the above approaches (Bar-Haim et al., 2020a,b). KPA extracts the main points discussed in a collection of texts, and matches the input sentences to these key points (KPs). The salience of each KP corresponds to the number of its matching sentences. The set of key points is selected out of a set of candidates short input sentences with high argumentative quality, so that together they achieve high coverage, while aiming to avoid redundancy. The resulting summary provides both textual and quantitative 3377 Positive Key Points % Reviews Negative Key Points % Reviews Amazingly helpful and friendly staff. 8.6% Cons: poor customer service 9.8% Modern furnishings and very clean. 6.3% Food is way over priced. 3.5% The views are incredible. 5.2% Buffet was extremely disappointing. 3.4% The historic building is beautiful. 4.9% Plus it’s disgusting and unsanitary. 3.3% Rooms are nice and comfortable. 3.8% Employees are rude. 3.2% The rooftop pool/patio is superb. 3.6% Rooms had a foul odor. 3.1% Luxurious and spacious rooms. 2.7% Check-in took an hour. 3.0% The decor is very elegant. 2.6% Staff unhelpful and uncaring. 2.6% The food here is excellent. 2.4% Building is very dated. 2.3% Great location - walkable to anything. 2.2% Our room had mechanical issues. 1.8% Table 1: A sample summary produced by our system: Key Point Analysis of an hotel with 2,662 reviews from the Yelp dataset. Top 10 positive and negative key points are shown. The balanced mixture of positive and negative key points in this summary correlates with the hotel’s middling rating of 3.25 stars. Key Point: The views are incredible. Key Point: Cons: poor customer service The scenery is amazing. Service horrible from start to finish. Great view too, of the Bellagio fountains. The front desk was so rude to us. I love this place for the scenery. The people that check you in suck. Great room overlooking the pool. The guy at check in was far from friendly. All were beautifully appointed and had great views of the strip. Probably one of the worst customer experiences. Table 2: Sample matches of sentences to key points. views of the data, as illustrated in Table 1. Table 2 shows a few examples of matching sentences to KPs. Originally developed for argument summarization, KPA has also been applied to user reviews and municipal surveys, using the same supervised models that were only trained on argumentation data, and was shown to perform reasonably well. However, previous work only used KPA “out-ofthe-box”, and did not attempt to adapt it to different target domains. In this work we propose several improvements to KPA, in order to make it more suitable to review data, and in particular to large-scale review datasets: 1. We show how the massive amount of reviews available in datasets like Amazon and Yelp, as well as their meta-data, such as numeric rating, can be leveraged for this task. 2. We integrate sentiment classification into KPA, which is crucial for analyzing reviews. 3. We improve key point extraction by introducing Collective Key Point Mining: extracting a large, high-quality set of key points from a large collection of businesses in a given domain. 4. We define the desired properties of key points in the context of user reviews, and develop a classifier that detects such key points. We show empirically that these novel extensions of KPA substantially improve its performance. We demonstrate that promising results can be achieved without any domain-specific annotation, while human supervision can lead to further improvement. Overall, this work makes a dual contribution: first, it proposes a new framework for review summarization. Second, it advances the research on KPA, by introducing novel methods that may be applied not only to user reviews, but to other use cases as well. 2 Background: Key Point Analysis KPA was initially developed for summarizing large argument collections (Bar-Haim et al., 2020a). KPA matches the given arguments to a set of key points (KPs), defined as high-level arguments. The set of KPs can be either given as input, or automatically extracted from the data. The resulting summary includes the KPs, along with their salience, represented by the number (or fraction) of matching arguments. The user can also drill down from each KP to its associated arguments. Bar-Haim et al. (2020b) proposed the following method for automatic extraction of KPs from a set of arguments, opinions or views, which they refer to as comments: 1. Select short, high quality sentences as KP candidates. 3378 2. Map each comment to its best matching KP, if the match score exceeds some threshold tmatch. 3. Rank the candidates according to the number of their matches. 4. Remove candidates that are too similar to a higher-ranked candidate1. 5. Re-map the removed candidates and their matched comments to the remaining candidates. 6. Re-sort the candidates by the number of matches and output the top-k candidates. Given a set of KPs and a set of comments, a summary is created by mapping each comment to its best-matching KP, if the match score exceeds tmatch. The above method relies on two models: a matching model that assigns a match score for a (comment, KP) pair, and a quality model, that assigns a quality score for a given comment. The matching model was trained on the ArgKP dataset, which contains 24K (argument, KP) pairs labeled as matched/unmatched. The quality model was trained on the IBM-ArgQ-Rank-30kArgs dataset, which contains quality scores for 30K arguments (Gretz et al., 2020)2. The arguments in both datasets support or contest a variety of common controversial topics (e.g., “We should abolish capital punishment”), and were collected via crowdsourcing. Bar-Haim et al. showed that models trained on argumentation data not only perform well on arguments, but also achieve reasonable results on other domains, including survey data and sentences taken from user reviews. However, they did not attempt to adapt KPA to these domains. In the following sections we look more closely at applying KPA to business reviews. 3 Data and Task In this work we apply KPA to business reviews from the Yelp Open Dataset3. The dataset contains about 8 million reviews for 200K businesses. Each business is classified into multiple categories. 1That is, their match score with that candidate exceeds the threshold tmatch. 2Both datasets are available from https: //www.research.ibm.com/haifa/dept/vst/ debating_data.shtml 3https://www.yelp.com/dataset Businesses (%) Reviews Train 25% 1,289,754 Dev 25% 1,338,123 Test 50% 2,622,054 Table 3: Yelp dataset split RESTAURANTS is by far the most common category, comprising the majority of the reviews. Besides restaurants, the dataset contains a wide variety of other business types, from NAIL SALONS to DENTISTS. We focus on two business categories in our experiments: RESTAURANTS (4.9M reviews) and HOTELS (258K reviews). We will henceforth refer to these business categories as domains. Each review includes, in addition to the review text, several other attributes, most relevant for our work is the “star rating” on a 1-5 scale. We filtered and split the dataset as follows. First, we removed reviews with more than 15 sentences (10% of the reviews). Second, we removed businesses with less than 50 reviews. The remaining businesses were split into Train, Development (Dev) and Test set, as detailed in Table 3. Our goal is to create a summary of the reviews for a given business. The summary would list the top k positive and top k negative KPs, and indicate for each KP its salience in the reviews, represented by the percentage of reviews that match the KP. A review is matched to a KP if at least one of its sentences is matched to that KP. An example of such summary is given in Table 1. Table 2 shows a few examples of matching sentences to KPs. 4 Classification Models Our system employs several classification models: in addition to the matching and argument quality models discussed in Section 2, in this work we add a sentiment classification model and a KP quality model, to be discussed in the next sections. All four classifiers were trained by fine-tuning a RoBERTa-large model (Liu et al., 2019). Prior to the fine-tuning of each classifier, we adapted the model to the business reviews domain, by pretraining on the Yelp dataset. We performed Masked LM pertraining (Devlin et al., 2019; Liu et al., 2019) on 1.5 million sentences sampled from the train set with a length filter of 20-150 characters per sentence. The following parameters were used: learning rate - 1e-5; 2 epochs. Training took two days on a single v100 GPU. The matching model was then obtained by fine3379 tuning the pre-trained model on the ArgKP dataset, with the parameters specified by Bar-Haim et al. (2020b). The quality model was fine-tuned following the procedure described by Gretz et al. (2020), except for using RoBERTa-large instead of BERTbase, with learning rate of 1e-5. 5 Incorporating Sentiment into KPA Previous work on KPA has ignored the issue of sentiment (or stance) altogether. When applied to argumentation data, it was assumed that the stance of the arguments is known, and KPA was performed separately for pro and con arguments. Accordingly, the ArgKP dataset only contains (argument, KP) pairs having the same stance. There are, however, several advantages for incorporating sentiment into KPA, in particular when analyzing reviews: 1. Separating positive KPs from negative ones makes the summaries more readable. 2. Filtering neutral sentences, which are mostly irrelevant, may improve KPA quality. 3. Attempting to match only sentences and KPs with the same polarity may reduce both matching errors and run time. We developed a sentence-level sentiment classifier for Yelp data by leveraging the abundance of available star ratings for short reviews. We extracted from the entire train set reviews having at most 3 sentences and 64 tokens. Reviews with 12, 3 and 4-5 star rating were labeled as negative (NEG, 20% of the reviews), neutral (NEUT, 11%) and positive (POS, 69%), respectively. The reviews were divided into a training set, comprising 235,481 reviews, and a held-out set, comprising 26,166 reviews. The sentiment classifier was trained by finetuning the pre-trained model on the above training data, for 3 epochs. The first two rows in Table 4 show the classifier’s performance on the held-out set. Since we ultimately wish to apply the classifier to individual sentences, we also annotated a small sentence-level benchmark of 158 reviews from the held-out set, which contain 952 sentences. We selected a minimal threshold ts for predicting POS or NEG sentiment. If both POS and NEG predictions are below this threshold, the sentence is predicted as NEUT. The threshold was selected so that the recall of both POS and NEG is at least 70%, while POS NEG NEUT Reviews P 0.96 0.86 0.58 R 0.97 0.91 0.47 Sentences P 0.82 0.81 0.48 R 0.88 0.70 0.47 Table 4: Sentiment classification results on held-out data. Precision (P) and recall (R) per class are shown, for both complete reviews and individual sentences. aiming to maximize precision4. Sentence-level performance on the benchmark using this threshold is shown in the last two rows of Table 4. Almost all the errors involved neutral labels - confusion between positive and negative labels was very rare. We integrate sentiment into KPA as follows. We extract positive KPs from a set of sentences classified as positive, and likewise for negative KPs. In order to further improve precision, positive (negative) sentences are only selected from positive (negative) reviews. When matching sentences to the extracted KPs we filter out neutral sentences and match sentences only to KPs with the same polarity. However, at this stage we do not filter by the review polarity, since we would like to allow matching positive sentences in negative reviews and vice versa, as well as positive and negative sentences in neutral reviews. 6 Collective Key Point Mining KPA is an extractive summarization method: KPs are selected from the review sentences being summarized. When generating a summary for a business with just a few dozens of reviews, the input reviews may not have enough good KP candidates short sentences that concisely capture salient points in the reviews. This is a common problem for extractive summarization methods, where it is often difficult to find sentences that fit into the summary in their entirety. We propose to address this problem by mining KPs collectively for the whole domain (e.g., restaurants or hotels). The extracted set of domain KPs is then matched to the review sentences of each analyzed business. This method can extract KPs from reviews of thousands of businesses, rather than from a single business, and therefore is much more robust. It overcomes a fundamental limitation of extractive summarization - limited selection of candidate sentences, while sidestepping the com4The chosen threshold was 0.79. 3380 Sentences POS NEG Restaurants 49,685 48,751 Hotels 49,655 59,552 Table 5: Number of positive and negative sentences extracted for KP mining in each domain. plexity of sentence generation that exists in abstractive summarization. Using the same set of KPs for each business makes it easy to compare different businesses. For example, we can rank businesses by the prevalence of a certain KP of interest. For each domain, we sampled 12,000 positive reviews and 12,000 negative reviews from the train set, from which positive and negative KPs were extracted, respectively5. We extracted positive and negative sentences from the reviews using the sentiment classifier, as described in the previous section. We filtered sentences with less than 3 tokens or more than 36 tokens (not including punctuation), as well as sentences with less than 10 characters. The number of positive and negative sentences obtained for each domain is detailed in Table 5. We ran the KP extraction algorithm described in Section 2 separately for the positive and negative sentences in each domain. We used a matching threshold tmatch = 0.99. The length of KP candidates was constrained to 3-5 tokens, and their minimal quality score was tquality=0.426. For each run, we selected the resulting top 70 candidates. The number of RoBERTa predictions required by the algorithm is O(#KP-candidates × #sentences). While the input size in previous work was up to a few thousands of sentences, here we deal with 50K-60K sentences per run. In order to maintain reasonable run time, we had to constrain both the number of sentences and the number of KP candidates. We selected the top 25% sentences with the highest quality score. The maximal number of KP candidates was 1.5 × √Ns, where Ns is the number of input sentences, and the highest-quality candidates were selected. Each run took 3.5-4.5 hours using 10 v100 GPUs. 7 Improving Key Point Quality Previous work did not attempt to explicitly define the desired properties KPs should have, or to de5To ensure diversity over the businesses, we employed a two-step sampling process: first sampled a business and then sampled a review for the business. 6The threshold was selected by inspecting a sample of the training data. velop a model that identifies good KP candidates. Instead, KP candidates were selected based on their length and argument quality, using the quality model of Gretz et al. (2020). This quality model, however, is not ideally suited for selecting KP candidates for review summarization: first, it is trained on crowd-contributed arguments, rather than on sentences extracted from user reviews. Second, quality is determined based on whether the argument should be selected for a speech supporting or contesting a controversial topic, which is quite different from our use case. We fill this gap by defining the following requirements from a KP in review summarization: 1. VALIDITY: the KP should be a valid, understandable sentence. This would filter out sentences such as “It’s rare these days to find that!”. 2. SENTIMENT: it should have a clear sentiment (either positive or negative). This would exclude sentences like “I came for a company event”. 3. INFORMATIVENESS: it should discuss some aspect of the reviewed business. Statements such as “Love this place” or “We were very disappointed”, which merely express an overall sentiment should be discarded, as this information is already conveyed in the star rating. The KP should also be general enough to be relevant for other businesses in the domain. A common example of sentences that are too specific is mentioning the business name or a person’s name (“Byron at the front desk is the best!”). 4. SINGLE ASPECT: it should not discuss multiple aspects (e.g., “Decent price, respectable portions, good flavor”). As we show in Section 8, the method presented in the previous sections extracts many KPs that do not meet the above criteria. In order to improve this situation, we developed a new KP quality classifier. We created a labeled dataset for this task, as follows. We sampled from the restaurant and hotel reviews in the train set 2,000 sentences comprising 3-8 tokens and minimal argument quality of tquality. each sentence was annotated for each of the above criteria7 by 10 crowd annotators, using the Appen platform8. We took several measures 7The guidelines are included in the appendix. 8https://appen.com/ 3381 to ensure annotation quality, following Gretz et al. (2020) and Bar-Haim et al. (2020b). First, the annotation was performed by trusted annotators, who performed well on previous tasks. Second, we employed the Annotator-κ score (Toledo et al., 2019), which measures inter annotator agreement, and removed annotators whose annotator-κ was too low. The details are provided in the appendix. For each sentence and each criterion, the fraction of positive annotations was taken to be its confidence. The final dataset was created by setting upper and lower thresholds on the confidence value of each of the four criteria. Sentences that matched all the upper thresholds were considered positive. Sentences that matched any of the lower thresholds were considered negative. The rest of the sentences were discarded. The threshold values we used are given in the appendix. Overall, the dataset contains 404 positive examples and 1,291 negative examples. We trained a KP quality classifier by fine-tuning the pretrained RoBERTa model (cf. Section 4) on the above dataset (4 epochs, learning rate: 1e-05). Figure 1 shows that this classifier (denoted KP quality FT) performs reasonably well on the dataset, in a 4-fold cross-validation experiment. Unsurprisingly, the argument quality classifier trained on argumentation data is shown to perform poorly on this task. The classifier was used to filter bad KP candidates, as part of the KP mining algorithm (Section 6). Candidates that passed this filtering were filtered and ranked by the argument quality model as before. We selected a threshold of 0.4 for the classifier, which corresponds to keeping 32% of the candidates, with precision of 0.62 and recall of 0.82. 8 Evaluation 8.1 Experimental Setup Our evaluation follows Bar-Haim et al. (2020b), while making the necessary changes for our setting. Let D be a domain, K a set of positive and negative KPs for D, and B a sample of businesses in D. Applying KPA to a business b ∈B using the set of KPs K and a matching threshold tmatch creates a mapping from sentences in b’s reviews, denoted Rb, to KPs in K. By modifying tmatch we can explore the tradeoff between precision (fraction of correct matches) and coverage. Bar-Haim et al. performed KPA over individual sentences, and correspond0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Recall 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Precision Argument quality KP quality FT Figure 1: KP Quality Precision vs. Recall. The finetuned KP quality model (“KP quality FT”) and the original argument quality model are evaluated over the KP quality labeled dataset. ingly defined coverage as the fraction of matched sentences. We are more interested in review-level coverage, since not all the sentences in the review are necessarily relevant for the summary. Given KPA results for B, K and tmatch, we can compute the following measures: 1. Review Coverage: the fraction of reviews per business that are matched to at least one KP, macro-averaged over the businesses in B. 2. Mean Matches per Review: the average number of matched KPs per review, macroaveraged over the businesses in B. Computing precision requires a labeled sample. We create a sample S by repeating the following procedure until N samples are collected: 1. Sample a business b ∈B; a review r ∈Rb and a sentence s ∈r. 2. Let the KP k ∈K be the best match of s in K with match score m. 3. Add the tuple [(s, k), m] to S if m > tmin. The (s, k) pairs in S are annotated as correct/incorrect matches. We can then compute the precision for any threshold tmatch > tmin by considering the corresponding subset of the sample. We sampled for each domain 40 businesses from the test set, where each business has between 100 and 5,000 reviews. For each domain, and each evaluated set of KPs, we labeled a sample of 400 pairs. We experimented with several configurations of KPA adapted to Yelp reviews, as described in the previous sections. These configurations are denoted by the prefix RKPA. Each configuration only differs in the method it employs for creating the set 3382 of domain KPs (K): RKPA-BASE: This configuration filters KP candidates according to their length and quality, using the quality model trained on argumentation data. In each domain, the top 30 mined KPs for each polarity were selected. RKPA-FT: This configuration applies the finetuned KP quality model as an additional filter for KP candidates. As with the previous configuration, we take the top 30 KPs for each polarity, in each domain. RKPA-MANUAL: We also experimented with an alternative form of human supervision, where the set of automatically-extracted KPs obtained by the RKPA-BASE configuration is manually reviewed and edited. KPs may be rephrased, redundancies are removed and bad KPs are filtered out. While this kind of task is less suitable for crowdsourcing, it can be completed fairly quickly - about an hour per domain. The task was performed by two of the authors, each working on one domain and reviewing the results for the other domain. The final set includes: 18 positive and 15 negative KPs for restaurants; 20 positive and 20 negative KPs for hotels.9 In addition to the above configurations, we also experimented with a “vanilla” KPA configuration (denoted KPA), which replicates the system of BarHaim et al. (2020b), without any of the adaptations and improvements introduced in this work. No Yelp data was used for pretraining or fine-tuning the models; key points were extracted independently for each business in the test set; and no sentiment analysis was performed. Instead of taking the top 30 KPs for each polarity, we took the top 60 KPs. Sample labeling. Similar to the KP quality dataset, the eight samples of 400 pairs (two domains × four configurations) were annotated in the Appen crowdsourcing platform. The annotation guidelines are included in the appendix. Each instance was labeled by 8 trusted annotators, and annotators with Annotator-κ < 0.05 were removed (cf. Section 7). We set a high bar for labeling correct matches: at least 85% of the annotators had to agree that the match is correct, otherwise it was labeled as incorrect. 9The set of KPs for each configuration is provided as supplementary material. We verified the annotations consistency by sampling 250 pairs, and annotating each pair by 16 annotators. Annotations for each pair were randomly split into two sets of 8 annotations, and a binary label was derived from each set, as described above. The two sets of labels for the sample agreed on 85.2% of the pairs, with Cohen’s Kappa of 0.610. 8.2 Results Figure 2 shows the precision/coverage curves for the four configurations, where coverage is measured either as Review Coverage (left) or as Mean Matches per Review (right). We first note that all three configurations developed in this work outperform vanilla KPA by a large margin. The RKPA-BASE configuration, which is only trained on previously-available data, already achieves reasonable performance. For example, the precision at Review Coverage of 0.8 is 0.77 for hotels and 0.83 for restaurants. Applying human supervision for improving the set of key points, either by training a KP quality model on crowd labeling (RKPA-FT), or by employing a humanin-the loop approach (RKPA-MANUAL) leads to substantial improvement in both domains. While both alternatives perform well, RKPA-FT achieves better precision at higher coverage rates. Table 6 shows, for each configuration in the restaurants domain, the top 10 KPs ranked by their number of matches in the sample. The matching threshold for each configuration corresponds to Review Coverage of 0.75. For the RKPA-BASE configuration, we can see examples of KPs that discuss multiple aspects (rows 3, 4), are too general (row 8) or too specific (row 9). These issues are much improved by applying the KP quality classifier, as illustrated by the top 10 KPs for the RKPA-FT configuration. Table 7 provides a more systematic comparison of the KP quality in both configurations, based on the top 30 KPs for each polarity in each domain (120 in total per configuration). For each domain and configuration, the table shows the fraction of KPs that conform to our guidelines (Section 7). In both domains, KP quality is much improved for the RKPA-FT configuration. Error Analysis: By analyzing the top matching errors of both domains, we found several systematic patterns of errors. The most common type of 10This result is comparable to (Bar-Haim et al., 2020b), who reported Cohen’s Kappa of 0.63 in a similar experiment. 3383 0.0 0.2 0.4 0.6 0.8 1.0 Review Coverage 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Precision 0.0 0.5 1.0 1.5 2.0 2.5 Mean Matches per Review 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 RKPA-Base RKPA-FT RKPA-Manual KPA (a) Hotels 0.0 0.2 0.4 0.6 0.8 1.0 Review Coverage 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 Precision 0.0 0.5 1.0 1.5 2.0 2.5 Mean Matches per Review 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 RKPA-Base RKPA-FT RKPA-Manual KPA (b) Restaurants Figure 2: KPA Precision vs. Coverage # RKPA-Base RKPA-FT RKPA-Manual 1 The food here is superb. The food here is superb. Fresh and tasty ingredients 2 Service and quality was excellent. Customer service is consistently exceptional. Everything was delicious 3 Large portions and reasonable prices. Service is slow and inattentive. Quick and polite service. 4 Fantastic food, location, and ambiance. Service was friendly and welcoming. Service is slow and inattentive. 5 Staff is interactive and friendly. The food is very flavorful. Staff is interactive and friendly. 6 Again, flavorless and poor quality. Reasonably priced menu items. Very affordable prices 7 Ingredients where fresh and tasty. The restaurant is beautifully decorated. Atmosphere is fun and casual. 8 We’ll certainly be back again. Everything was cooked to perfection. The dishes are extremely overpriced. 9 Kevin, was rude and condescending. The overall ambience was pleasing. A lot of variety 10 Atmosphere is fun and casual. Staff are super nice & attentive. The food was flavorless Table 6: Top 10 key points for each configuration in the restaurants domain, ranked by their number of matches in the sample. The matching threshold for each configuration corresponds to Review Coverage of 0.75. RKPA-Base RKPA-FT Hotels 0.70 0.85 Restaurants 0.62 0.95 Table 7: Key point quality assessment. For each domain and configuration, the table shows the fraction of KPs that conform to our guidelines. error consisted of a KP and a sentence making the same claim towards different targets, e.g. “We had to refill our own wine and ask for refills of soda.” was matched to “Coffee was never even refilled.”. This usually stemmed from a too specific KP and was more common in the restaurants domain. In some cases, a sentence was matched to an unrelated KP with a shared concept or term. For example, “Cheap, easy, and filling” was matched to “Ordering is quick and easy”. Polarity errors were rare but present, e.g. “However she wasn’t the friendliest when she came to help us” and “The waitress was friendly though.”. 9 Related Work Previous work on review summarization was dominated by two paradigms: aspect-based sentiment summarization and multi-document opinion summarization. Aspect-based sentiment summarization. This line of work aims to create structured summaries that assign an aggregated sentiment score or rating to the main aspects of the reviewed entity (Hu and Liu, 2004; Gamon et al., 2005; Snyder and Barzilay, 2007; Blair-goldensohn et al., 2008; Titov and McDonald, 2008). Aspects typically comprise 1-2 words (e.g., service, picture quality), and are either predefined or extracted automatically. A core 3384 sub-task in this approach is Aspect-Based Sentiment Analysis: identification of aspect mentions in the text, which may be further classified into highlevel aspect categories, and classification of the sentiment towards these mentions. Recent examples are (Ma et al., 2019; Miao et al., 2020; Karimi et al., 2020). The main shortcoming of such summaries is the lack of detail, which makes it difficult for a user to understand why an aspect received a particular rating (Ganesan et al., 2010). Although some of these summaries include for each aspect a few supporting text snippets as “evidence”, these examples may be considered anecdotal rather than representative. Multi-document opinion summarization. This approach aims to create a fluent textual summary from the input reviews. A major challenge here is the limited amount of human-written summaries available for training. Recently, several abstractive neural summarization methods have shown promising results. These models require no summaries for training (Chu and Liu, 2019; Braˇzinskas et al., 2020b; Suhara et al., 2020), or only a handful of them (Braˇzinskas et al., 2020a). As discussed in the previous section, textual summaries provide more detail than aspect-based sentiment summaries, but lack a quantitative dimension. In addition, the assessment of such summaries is known to be difficult. As demonstrated in this work, KPA can be evaluated using straightforward measures such as precision and coverage. 10 Conclusion We introduced a novel paradigm for summarizing reviews, based on KPA. KPA addresses the limitations of previous approaches by generating summaries that combine both textual and quantitative views of the data. We presented several extensions to KPA, which make it more suitable for large-scale review summarization: collective key point mining for better key point extraction; integrating sentiment analysis into KPA; identifying good key point candidates for review summaries; and leveraging the massive amount of available reviews and their metadata. We achieved promising results over the Yelp dataset without requiring any domain-specific annotations. We also showed that performance can be substantially improved with human supervision. While we focused on user reviews, the methods introduced in this work may improve KPA performance in other domains as well. In future work we would like to generate richer summaries by combining domain level key points with “local” key points, individually extracted per business. It would also be interesting to adapt current methods for unsupervised abstractive summarization to generate key points. Ethical Considerations • Our use of the Yelp dataset has been reviewed and approved by both the data acquisition authority in our organization and the Yelp team. • We do not store or use any user information from the Yelp dataset. • We ensured fair compensation for crowd annotators as follows: we set a fair hourly rate according to our organization’s standards, and derived the payment per task from the hourly rate by estimating the expected time per task based on our own experience. • Regarding the potential use of the proposed method - one of the advantages of KPA is that it is transparent, verifiable and explainable the user can drill down from each key point to it matched sentences, which provide justification and supporting evidence for its inclusion in the summary. References Roy Bar-Haim, Lilach Eden, Roni Friedman, Yoav Kantor, Dan Lahav, and Noam Slonim. 2020a. From arguments to key points: Towards automatic argument summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4029–4039, Online. Association for Computational Linguistics. Roy Bar-Haim, Yoav Kantor, Lilach Eden, Roni Friedman, Dan Lahav, and Noam Slonim. 2020b. Quantitative argument summarization and beyond: Crossdomain key point analysis. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 39–49, Online. Association for Computational Linguistics. Sasha Blair-goldensohn, Tyler Neylon, Kerry Hannan, George A. Reis, Ryan Mcdonald, and Jeff Reynar. 2008. Building a sentiment summarizer for local service reviews. In NLP in the Information Explosion Era (NLPIX). 3385 Arthur Braˇzinskas, Mirella Lapata, and Ivan Titov. 2020a. Few-shot learning for opinion summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4119–4135, Online. Association for Computational Linguistics. Arthur Braˇzinskas, Mirella Lapata, and Ivan Titov. 2020b. Unsupervised opinion summarization as copycat-review generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5151–5169, Online. Association for Computational Linguistics. Giuseppe Carenini, Raymond Ng, and Adam Pauls. 2006. Multi-document summarization of evaluative text. In 11th Conference of the European Chapter of the Association for Computational Linguistics, Trento, Italy. Association for Computational Linguistics. Eric Chu and Peter Liu. 2019. MeanSum: A neural model for unsupervised multi-document abstractive summarization. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 1223–1232. PMLR. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Michael Gamon, Anthony Aue, Simon Corston-Oliver, and Eric Ringger. 2005. Pulse: Mining customer opinions from free text. In Advances in Intelligent Data Analysis VI, pages 121–132, Berlin, Heidelberg. Springer Berlin Heidelberg. Kavita Ganesan, ChengXiang Zhai, and Jiawei Han. 2010. Opinosis: A graph based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 340–348, Beijing, China. Coling 2010 Organizing Committee. Shai Gretz, Roni Friedman, Edo Cohen-Karlik, Assaf Toledo, Dan Lahav, Ranit Aharonov, and Noam Slonim. 2020. A large-scale dataset for argument quality ranking: Construction and analysis. In AAAI. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’04, pages 168–177, New York, NY, USA. ACM. Akbar Karimi, Leonardo Rossi, and Andrea Prati. 2020. Adversarial training for aspect-based sentiment analysis with bert. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. Dehong Ma, Sujian Li, Fangzhao Wu, Xing Xie, and Houfeng Wang. 2019. Exploring sequence-tosequence learning in aspect term extraction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3538– 3547, Florence, Italy. Association for Computational Linguistics. Zhengjie Miao, Yuliang Li, Xiaolan Wang, and WangChiew Tan. 2020. Snippext: Semi-supervised opinion mining with augmented data. Proceedings of The Web Conference 2020. Benjamin Snyder and Regina Barzilay. 2007. Multiple aspect ranking using the good grief algorithm. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Proceedings of the Main Conference, pages 300–307, Rochester, New York. Association for Computational Linguistics. Yoshihiko Suhara, Xiaolan Wang, Stefanos Angelidis, and Wang-Chiew Tan. 2020. OpinionDigest: A simple framework for opinion summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5789– 5798, Online. Association for Computational Linguistics. Ivan Titov and Ryan McDonald. 2008. A joint model of text and aspect ratings for sentiment summarization. In Proceedings of ACL-08: HLT, pages 308– 316, Columbus, Ohio. Association for Computational Linguistics. Assaf Toledo, Shai Gretz, Edo Cohen-Karlik, Roni Friedman, Elad Venezian, Dan Lahav, Michal Jacovi, Ranit Aharonov, and Noam Slonim. 2019. Automatic argument quality assessment - new datasets and methods. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 5624–5634, Hong Kong, China. Association for Computational Linguistics. Appendices A Key Point Quality Dataset A.1 Annotation Guidelines Below are the annotation guidelines for the KP quality annotation task: 3386 Positive Negative Validity Confidence >0.85 Confidence <0.8 Sentiment Clear sentiment with confidence >0.6 No sentiment or sentiment confidence <0.5 Informativeness Informative with confidence >0.6 Too specific\not informative; or doesn’t refer to an aspect with confidence >0.6 Multiple Aspects Confidence <= 0.57 confidence >= 0.85 Table 8: Criteria for creating the key point quality dataset from crowd annotations. Sentences that match all the positive criteria are labeled as valid key points; Sentences that match any of the negative criteria are labeled as invalid key points, and the rest are excluded. In the following you will be presented with a business category and a sentence extracted from a customer review on a certain business in that category. You will be asked to answer the following questions: 1. Is this a valid, understandable sentence? (Yes / No) 2. What is the sentiment this sentence expresses toward the reviewed business or aspect of that business? (Positive / Negative / Mixed sentiment / Neutral or unclear) 3. Can this sentence be used to review ASPECT(S) of another business under the same category? (No, it is too business specific / No, it does not refer to certain aspects of the business/ No, it is not informative / Yes) Note: An aspect of a business is a single attribute of its overall service/product. In hotels, for instance it could be the cleanliness of the room. In most businesses it could be the friendliness of the staff, the price, the conveniency of location etc. 4. Does this sentence discuss more than one independent aspect of the business? (Yes/No) A.2 Quality Control Annotators were excluded if their Annotator-κ score (Toledo et al., 2019), calculated for each question, was below any of these thresholds: • Question #3 (Informativeness): 0.3 • Question #4 (Multiple Aspects): 0.1 A.3 Final Dataset Generation Table 8 shows the criteria for the inclusion of a sentence in the KP Quality dataset. Sentences that match all the Positive criteria are considered valid key points; Sentences that match any of the Negative criteria are considered invalid key points, and the rest are excluded. The confidence of a criterion denotes the fraction of positive annotations in the case of a binary choice, or the fraction of annotations for a certain label otherwise. B Key Point Matching Annotation Guidelines Below are the match annotation guidelines for (sentence, KP) pairs: In this task you are presented with a business domain, a sentence taken from a review of a business in that domain and a key point. You will be asked to answer the following question: does the key point match the sentence? A key point matches a sentence if it captures the gist of the sentence, or is directly supported by a point made in the sentence. The options are: • Yes • No • Faulty key point (not a valid sentence or unclear)
|
2021
|
262
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3387–3402 August 1–6, 2021. ©2021 Association for Computational Linguistics 3387 Structured Sentiment Analysis as Dependency Graph Parsing Jeremy Barnes*, Robin Kurtz†, Stephan Oepen*, Lilja Øvrelid*and Erik Velldal* *University of Oslo, Department of Informatics †National Library of Sweden, KBLab { jeremycb | oe | liljao | erikve } @ifi.uio.no [email protected] Abstract Structured sentiment analysis attempts to extract full opinion tuples from a text, but over time this task has been subdivided into smaller and smaller sub-tasks, e.g., target extraction or targeted polarity classification. We argue that this division has become counterproductive and propose a new unified framework to remedy the situation. We cast the structured sentiment problem as dependency graph parsing, where the nodes are spans of sentiment holders, targets and expressions, and the arcs are the relations between them. We perform experiments on five datasets in four languages (English, Norwegian, Basque, and Catalan) and show that this approach leads to strong improvements over state-of-the-art baselines. Our analysis shows that refining the sentiment graphs with syntactic dependency information further improves results. 1 Introduction Structured1 sentiment analysis, i.e., the task of predicting a structured sentiment graph like the ones in Figure 1, can be theoretically cast as an information extraction problem in which one attempts to find all of the opinion tuples O = Oi, . . . , On in a text. Each opinion Oi is a tuple (h, t, e, p) where h is a holder who expresses a polarity p towards a target t through a sentiment expression e, implicitly defining pairwise relationships between elements of the same tuple. Liu (2012) argues that all of these elements2 are essential to fully resolve the sentiment analysis problem. 1We use the term ‘structured sentiment’ distinctly from Almars et al. (2017), who use it to refer to the latent hierarchical structure of sentiment aspects. We instead use ‘structured’ to refer to predicting sentiment graphs as a structured prediction task, as opposed to the many text classification task that are found in sentiment analysis. 2Liu (2012)’s definition replaces sentiment expression with the time when the opinion was expressed. However, most research on sentiment analysis focuses either on a variety of sub-tasks, which avoids performing the full task, or on simplified and idealized tasks, e.g., sentence-level binary polarity classification. We argue that the division of structured sentiment into these sub-tasks has become counterproductive, as reported experiments are often not sensitive to whether a given addition to the pipeline improves the overall resolution of sentiment, or do not take into account the inter-dependencies of the various sub-tasks. As such, we propose a unified approach to structured sentiment which jointly predicts all elements of an opinion tuple and their relations. Moreover, we cast sentiment analysis as a dependency graph parsing problem, where the sentiment expression is the root node, and the other elements have arcs which model the relationships between them. This methodology also enables us to take advantage of recent improvements in semantic dependency parsing (Dozat and Manning, 2018; Oepen et al., 2020; Kurtz et al., 2020) to efficiently learn a sentiment graph parser. This perspective also allows us to unify a number of approaches, including targeted, and opinion tuple mining. We aim to answer RQ1: whether graph-based approaches to structured sentiment outperform state-of-the-art sequence labeling approaches, and RQ2: how to best encode structured sentiment as parsing graphs. We perform experiments on five standard datasets in four languages (English, Norwegian, Basque, Catalan) and show that graph-based approaches outperform state-ofthe-art baselines on all datasets on several standard metrics, as well as our proposed novel (unlabeled and labeled) sentiment graph metrics. We further propose methods to inject linguistic structure into the sentiment graphs using syntactic dependencies. Our main contributions are therefore 1) proposing a holistic approach to structured sentiment through 3388 Some others give the new UMUC 5 stars - don't believe them . positive negative holder target expression target expression Figure 1: A structured sentiment graph is composed of a holder, target, sentiment expression, their relationships and a polarity attribute. Holders and targets can be null. sentiment graph parsing, 2) introducing new evaluation metrics for measuring model performance, and 3) extensive experimental results that outperform state-of-the-art baselines. Finally, we release the code and datasets3 to enable future work on this problem. 2 Related Work Structured sentiment analysis can be broken down into five sub-tasks: i) sentiment expression extraction, ii) sentiment target extraction, iii) sentiment holder extraction, iv) defining the relationship between these elements, and v) assigning polarity. Previous work on information extraction has used pipeline methods which first extract the holders, targets, and expressions (tasks i - iii) and subsequently predict their relations (task iv), mostly on the MPQA dataset (Wiebe et al., 2005). CRFs and a number of external resources (sentiment lexicons, dependency parsers, named-entity taggers) (Choi et al., 2006; Yang and Cardie, 2012) are strong baselines. Given the small size of the training data and the complicated task, these techniques often still outperform neural models, such as BiLSTMs (Katiyar and Cardie, 2016). Transition-based end-toend approaches have shown some potential (Zhang et al., 2019). However, all of this work ignores the polarity classification subtask. Targeted sentiment analysis only concentrates on extracting sentiment targets (task ii) and classifying the polarity directed towards them (task iv) (Jiang et al., 2011; Mitchell et al., 2013). Recent shared tasks on Aspect-Based Sentiment Analysis (ABSA) (Pontiki et al., 2014, 2015, 2016) also include target extraction and polarity classification subtasks. Joint approaches perform on par with pipeline methods (Li et al., 2019a) and multitask models can perform even better (He et al., 2019). Finally, pretrained language models (Devlin et al., 3Code and datasets available at https://github. com/jerbarnes/sentiment_graphs. 2019) can also lead to improvements on the ABSA data (Li et al., 2019b). End2End sentiment analysis is a recently proposed subtask which combines targeted sentiment (tasks ii and v) and sentiment expression extraction (task i), without requiring the resolution of relationships between targets and expressions. Wang et al. (2016) augment the ABSA datasets with sentiment expressions, but provide no details on the annotation process or any inter-annotator agreement. He et al. (2019) make use of this data and propose a multi-layer CNN (IMN) to create hidden representations h which are then fed to a target and opinion extraction module (AE), which is also a multi-layer CNN. This module predicts ˆyae, a sequence of BIO tags4 that predict the presence or absence of targets and expressions. After jointly predicting the targets and expressions, a second multi-layer CNN with a final self-attention network is used to classify the polarity, again as sequence labeling task (AS). This second module combines the information from h and ˆyae by incorporating the predicted probability of a token to be a target in the formulation of self-attention. Finally, an iterative message-passing algorithm updates h using the predictions from all the modules at the previous timestep. Chen and Qian (2020) instead propose RelationAware Collaborative Learning (RACL). This model creates task specific representations by first embedding a sentence, passing through a shared feed-forward network and finally a task-specific CNN. This approach then models interactions between each pair of sub-tasks (target extraction, expression extraction, sentiment classification) by creating pairwise weighted attention representations. These are then concatenated and used to create the task-specific predictions. The authors finally stack several RACL layers, using the output from the previous layer as input for the next. 4The tags include {BIO}-{target,expression} 3389 Both models perform well on the augmented SemEval data, but it is unlikely that these annotations are adequate for full structured sentiment, as Wang et al. (2016) only provide expression annotations for sentences that have targets, generally only include sentiment-bearing words (not phrases), and do not specify the relationship between target and expression. Finally, the recently proposed aspect sentiment triplet extraction (Peng et al., 2019; ?) attempts to extract targets, expressions and their polarity. However, the datasets used are unlikely to be adequate, as they augment available targeted datasets, but do not report annotation guidelines, procedure, or inter-annotator agreement. Graph parsing: Syntactic dependency graphs are regularly used in applications, supplying them with necessary grammatical information (Mintz et al., 2009; Cui et al., 2005; Bj¨orne et al., 2009; Johansson and Moschitti, 2012; Lapponi et al., 2012). The dependency graph structures used in these systems are predominantly restricted to trees. While trees are sufficient to encode syntactic dependencies, they are not expressive enough to handle meaning representations, that require nodes to have multiple incoming arcs, or having no incoming arcs at all (Kuhlmann and Oepen, 2016). While much of the early research on parsing these new structures (Oepen et al., 2014, 2015) focused on specialized decoding algorithms, Dozat and Manning (2018) presented a neural dependency parser that essentially relies only on its neural network structure to predict any type of dependency graph without restrictions to certain structures. Using the parser’s ability to learn arbitrary dependency graphs, Kurtz et al. (2020) phrased the task of negation resolution (Morante and Blanco, 2012; Morante and Daelemans, 2012) as a graph parsing task. This transformed the otherwise flat representations to dependency structures that directly encode the often overlapping relations between the building blocks of multiple negation instances at the same time. In a simpler fashion, Yu et al. (2020) exploit the parser of Dozat and Manning (2018) to predict spans of named entities. 3 Datasets We here focus on datasets that annotate the full task of structured sentiment as described initially. We perform experiments on five structured sentiment datasets in four languages, the statistics of which are shown in Table 1. The largest available structured sentiment dataset is the NoReCFine dataset (Øvrelid et al., 2020), a multi-domain dataset of professional reviews in Norwegian, annotated for structured sentiment. MultiBEU and MultiBCA (Barnes et al., 2018) are hotel reviews in Basque and Catalan, respectively. MPQA (Wiebe et al., 2005) annotates news wire text in English. Finally, DSUnis (Toprak et al., 2010) annotate English reviews of online universities and e-commerce. In our experiments, we use only the university reviews, as the e-commerce reviews have a large number of ‘polar targets’, i.e., targets with a polarity, but no accompanying sentiment expression. While all the datasets annotate holders, targets, and expressions, the frequency and distribution of these vary. Regarding holders, MPQA has the most (2,054) and DSUnis has the fewest (94), whereas NoReCFine has the largest proportion of targets (8,923) and expressions (11,115). The average length of holders (2.6 tokens) and targets (6.1 tokens) in MPQA is also considerably higher than the others. It is also worth pointing out that MPQA and DSUnis additionally include neutral polarity. In the case of MPQA the neutral class refers to verbs which are subjective but do not convey polarity, e.g., ‘say’, ‘opt for’. In DSUnis, however, the neutral label tends to indicate expressions that could entail mixed polarity or are polar under the right conditions, e.g., ‘the classes were not easy’ is considered neutral, as it is possible for difficult classes to be desirable at a university. MultiBEU, and MultiBCA also have labels for strong positive and strong negative, which we map to positive and negative, respectively. Finally, NoReCFine includes intensity annotations (strong, normal, slight), which we disregard for the purposes of these experiments. 4 Modeling This section describes how we define and encode sentiment graphs, detail the neural dependency graph models, as well as two state-of-the-art baselines for end-to-end sentiment analysis (target and expression extraction, plus polarity classification). 4.1 Graph Representations Structured sentiment graphs as in Figure 1 are directed graphs, that are made up of a set of labeled nodes and a set of unlabeled edges connecting pairs of nodes. Nodes in the structured sentiment graphs 3390 sentences holders targets expressions polarity # avg. # avg. max # avg. max # avg. max + neu − NoReCFine train 8634 16.7 898 1.1 12 6778 1.9 35 8448 4.9 40 5684 0 2756 dev 1531 16.9 120 1.0 3 1152 2.0 15 1432 5.1 31 988 0 443 test 1272 17.2 110 1.0 3 993 2.0 20 1235 4.9 30 875 0 358 MultiBCA train 1174 15.6 169 1.1 4 1695 2.4 18 1981 2.6 19 1272 0 708 dev 168 13.3 15 1.5 7 211 2.3 10 258 2.6 9 151 0 107 test 336 14.7 52 1.1 5 430 2.6 12 518 2.7 14 313 0 204 MultiBEU train 1064 10.5 205 1.1 6 1285 1.4 9 1684 2.2 10 1406 0 278 dev 152 10.7 33 1.1 2 153 1.3 6 204 2.5 8 168 0 36 test 305 10.7 58 1.1 2 337 1.4 8 440 2.2 9 375 0 65 MPQA train 4500 25 1306 2.6 27 1382 6.1 56 1656 2.4 14 675 271 658 dev 1622 23 377 2.6 16 449 5.3 41 552 2.1 8 241 105 202 test 1681 24 371 2.8 32 405 6.4 42 479 2.0 8 166 89 199 DSUnis train 2253 20 65 1.2 2 1252 1.2 5 837 1.9 9 495 149 610 dev 232 9 17 1.1 3 151 1.2 3 106 1.7 6 40 19 92 test 318 20 12 1.3 4 198 1.2 6 139 2.0 5 77 18 103 Table 1: Statistics of the datasets, including number of sentences and average length (in tokens) per split, as well as average and max lengths (in tokens) for holder, target, and expression annotations. Additionally, we include the distribution of polarity – restricted to positive, neutral, and negative – in each dataset. can span over multiple tokens and may have multiple incoming edges. The resulting graphs can have multiple entry points (roots), are not necessarily connected, and not every token is a node in the graph. The sentence’s sentiment expressions correspond to the roots of the graphs, connecting explicitly to their respective holders and targets. In order to apply the algorithm of Dozat and Manning (2018), we simplify these structures into bi-lexical dependency graphs visualized in Figure 2. Here, nodes correspond one-to-one to the tokens of the sequence and follow the same linear order. The edges are drawn as arcs in the half-plane above the sentence, connecting heads to dependents. Similarly to the source structures, the graphs can have multiple roots and nodes can have multiple or no incoming arcs. For some rare instances of structured sentiment graphs, the reduction to dependency graphs is lossy, as they do not allow multiple arcs to share the same head and dependent. This results in a slight mismatch of the learned and aimed-for representations. The choice of how to encode the sentiment graphs as parsing graphs opens for several alternate representations depending on the choice of head/dependent status of individual tokens in the target/holder/expression spans of the sentiment graph. We here propose two simple parsing graph representations: head-first and head-final, which Metric Name Level Strictness +/− Holder F1 Token-level Partial No Target F1 Token-level Partial No Exp. F1 Token-level Partial No Targeted F1 Token-level Exact Yes UF1 Graph arcs Exact No LF1 Graph arcs Exact Yes NSF1 Sentimentgraph Exact graph, partial token No SF1 Sentimentgraph Exact graph, partial token Yes Table 2: Metrics used to evaluate performance. Column +/−indicates whether polarity is included or not. The main metrics are Targeted F1, which allows us to compare to methods that do not perform the full task, and SF1, which best represents the full task. are shown in Figure 2. For head-first, we set the first token of the sentiment expression as a root node, and similarly set the first token in each holder and token span as the head of the span with all other tokens within that span as dependents. The labels simply denote the type of relation (target/holder) and for sentiment expressions, additionally encode the polarity. Head-final is similar, but instead sets the final token of spans as the heads, and the final token of the sentiment expression as the root node. 3391 Some others give the new UMUC 5 stars don’t believe them. exp:pos exp:neg target target target exp:pos holder holder target exp:neg (a) Some others give the new UMUC 5 stars don’t believe them. exp:pos exp:neg target target target exp:pos holder holder target exp:neg (b) Figure 2: Two parsing graph proposals to encode the sentiment graph: (a) head-first, where the first token of any span is the head, and (b) head-final, where the final token is the head. 4.2 Proposed model The neural graph parsing model used in this work is a reimplementation of the neural parser by Dozat and Manning (2018) which was used by Kurtz et al. (2020) for negation resolution. The parser learns to score each possible arc to then finally predict the output structure simply as a collection of all positively scored arcs. The base of the network structure is a bidirectional LSTM (BiLSTM), that processes the input sentence both from left-toright and right-to-left, to create contextualized representations c1, . . . , cn = BiLSTM(w1, . . . , wn) where wi is the concatenation of a word embedding, POS tag embedding, lemma embedding, and character embedding created by a character-based LSTM for the ith token. In our experiments, we further augment the token representations with pretrained contextualized embeddings from multilingual BERT (Xu et al., 2019). We use multilingual BERT as several languages did not have available monolingual BERT models at the time of the experiments (Catalan, Norwegian). The contextualized embeddings are then processed by two feedforward neural networks (FNN), creating specialized representations for potential heads and dependents, hi = FNNhead(ci) and di = FNNdep(ci). The scores for each possible arclabel combination are computed by a final bilinear transformation using the tensor U. Its inner dimension corresponds to the number of sentiment graph labels plus a special NONE label, indicating the absence of an arc, which allows the model to predict arcs and labels jointly, score(hi, dj) = h⊤ i Udj. 4.3 Baselines We compare our proposed graph prediction approach with three state-of-the-art baselines5 for extracting targets and expressions and predicting the polarity: IMN6, RACL7, as well as RACLBERT, which also incorporates contextualized embeddings. Instead of using BERTLarge, we use the cased BERT-multilingual-base in order to fairly compare with our own models. Note, however, that our model does not update the mBERT representations, putting it at a disadvantage to RACL-BERT. We also compare with previously reported extraction results from Barnes et al. (2018) and Øvrelid et al. (2020). 5 Evaluation As we are interested not only in extraction or classification, but rather in the full structured sentiment task, we propose metrics that capture the relations between all predicted elements, while enabling comparison with previous state-of-the-art models on different subtasks. The main metrics we use to rank models are Targeted F1 and Sentiment Graph F1. 5Despite having state-of-the-art results on MPQA, we do not compare with Katiyar and Cardie (2016) as they use different dataset splits, 10-fold cross-validation, and their code is not available. 6IMN code available at https://github.com/ ruidan/IMN-E2E-ABSA. 7https://github.com/NLPWM-WHU/RACL. 3392 Dataset Model Spans Targeted Parsing Graph Sent. Graph Holder F1 Target F1 Exp. F1 F1 UF1 LF1 NSF1 SF1 NoReCFine Øvrelid et al. (2020) 42.4 31.3 31.3 IMN 35.9 48.7 18.0 RACL 45.6 55.4 20.1 RACL-BERT 47.2 56.3 30.3 Head-first 51.1 50.1 54.4 30.5 39.2 31.5 37.0 29.5 Head-final 60.4∗ 54.8 55.5 31.9 48.0∗ 37.7∗ 39.2∗ 31.2∗ MultiBEU Barnes et al. (2018)† 54.0 57.0 54.0 IMN 48.2 65.2 39.5 RACL 55.4 70.7 48.2 RACL-BERT 59.9 72.6 56.8 Head-first 60.4 64.0 73.9 57.8 64.6 60.0 58.0 54.7 Head-final 60.5 64.0 72.1 56.9 60.8 56.0 58.0 54.7 MultiBCA Barnes et al. (2018)† 56.0 64.0 52.0 IMN 56.3 60.9 32.5 RACL 65.4 67.6 49.1 RACL-BERT 67.5 70.3 52.4 Head-first 43.0 72.5 71.1∗ 55.0∗ 66.8∗ 62.1∗ 62.0 56.8 Head-final 37.1 71.2 67.1 53.9 62.7 58.1 59.7 53.7 MPQA IMN 24.3 29.6 1.2 RACL 32.6 37.8 11.8 RACL-BERT 20.0 31.2 17.8 Head-first 43.8 51.0 48.1 33.5∗ 40.0 36.9 24.5 17.4 Head-final 46.3 49.5 46.0 18.6 41.4 38.0 26.1 18.8 DSUnis IMN 33.0 27.4 17.9 RACL 39.3 40.2 22.8 RACL-BERT 44.6 38.2 27.3 Head-first 28.0 39.9 40.3 26.7 35.3 31.4 31.0 25.0 Head-final 37.4 42.1 45.5∗ 29.6 38.1 33.9 34.3∗ 26.5 Table 3: Experiments comparing our sentiment graph approaches (Head-first/Head-final) using mBERT with the sequence-labeling baselines (IMN, RACL, RACL-BERT). Underlined numbers indicate the best result for the metric and dataset. ∗indicates approach is significantly better than second best (p < 0.05), as determined by a bootstrap with replacement test. † indicates results that are not comparable, as they were calculated with 10-fold cross-validation. Token-level F1 for Holders, Targets, and Expressions To easily compare our models to pipeline models, we evaluate how well these models are able to identify the elements of a sentiment graph with token-level F1. Targeted F1 This is a common metric in targeted sentiment analysis (also referred to as F1-i (He et al., 2019) or ABSA F1 (Chen and Qian, 2020)). A true positive requires the combination of exact extraction of the sentiment target, and the correct polarity. Parsing graph metrics We additionally compute graph-level metrics to determine how well the models predict the unlabeled and labeled arcs of the parsing graphs: Unlabeled F1 (UF1), Labeled F1 (LF1). These measure the amount of (in)correctly predicted arcs and labels, as the harmonic mean of precision and recall (Oepen et al., 2014). These metrics inform us of the local properties of the graph, and do not overly penalize a model if a few edges of a graph are incorrect. Sentiment graph metrics The two metrics that measure how well a model is able to capture the full sentiment graph (see Figure 1) are Non-polar Sentiment Graph F1 (NSF1) and Sentiment Graph F1 (SF1). For NSF1, each sentiment graph is a tuple of (holder, target, expression), while for SF1 we include polarity (holder, target, expression, polarity). A true positive is defined as an exact match at graph-level, weighting the overlap in predicted and gold spans for each element, averaged across all three spans. For precision we weight the number of correctly predicted tokens divided by the total number of predicted tokens (for recall, we divide instead by the number of gold tokens). We allow for empty holders and targets. 3393 6 Experiments All sentiment graph models use token-level mBERT representations in addition to word2vec skip-gram embeddings openly available from the NLPL vector repository8 (Fares et al., 2017). We train all models for 100 epochs and keep the model that performs best regarding LF1 on the dev set (Targeted F1 for the baselines). We use default hyperparameters from Kurtz et al. (2020) (see Appendix) and run all of our models five times with different random seeds and report the mean (standard deviation shown as well in Table 8 in the Appendix). We calculate statistical difference between the best and second best models through a bootstrap with replacement test (Berg-Kirkpatrick et al., 2012). As there are 5 runs, we require that 3 of 5 be statistically significant at p < 0.05. Table 3 shows the results for all datasets. On NoReCFine, the baselines IMN, RACL, and RACL-BERT perform well at extracting targets (35.9, 45.6, and 47.2 F1, respectively) and expressions (48.7/55.4/56.3), but struggle with the full targeted sentiment task (18.0/20.1/30.3). The graphbased models extract targets better (50.1/54.8) and have comparable scores for expressions (54.4/55.5). The holder extraction scores have a similar range (51.1/60.4). These patterns hold throughout the other datasets, where the proposed graph models nearly always perform best on extracting spans, although RACL-BERT achieves the best score on extracting targets on DSUnis (44.6 vs. 42.1). The graph models also outperform the strongest baseline (RACL-BERT) on targeted sentiment on all 5 datasets, although this difference is often not statistically significant (NoReCFine Head-first, MultiBEU Head-final) and RACL-BERT is better than Head-first on DSUnis. Regarding the Graph metrics, the results depend highly on the dataset, with UF1 and LF1 ranging from 35.3/31.4 (DSUnis Head-first) to 66.8/62.1 (MultiBCA Head-first). Sentiment Graph metrics NSF1 and SF1 have a similar, though slightly lower range (24.5/17.7 – 62.0/56.8). The graph and sentiment graph metrics do not correlate perfectly, however, as UF1 and LF1 on MPQA are relatively good 8Nordic Language Processing Laboratory vector repo.: http://vectors.nlpl.eu/repository/. We used 300-dimensional embeddings trained on English Wikipedia and Gigaword for English (model id 18 in the repo.), and 100dimensional embeddings trained on the 2017 CoNLL corpora for all others; Basque (id 32), Catalan (id 34), and Norwegian Bokm˚al (id 58). # H.first H.final RACL NoReCFine 147 63.3 67.8 65.6 MultiBEU 45 68.9 65.9 29.2 MultiBCA 74 72.2 73.7 28.2 MPQA 40 55.4 58.5 28.8 DSUnis 10 56.9 43.1 31.4 Table 4: Number of sentences with multiple targets (#) and Macro F1 on the target extraction task for Headfinal and RACL. Head-final is consistently better than RACL on extracting multiple targets. (40.0/36.9 and 41.4/38.0 for Head-first and Headfinal, respectively), but the NSF1 and SF1 are poor (24.5/17.4 and 26.1/18.8). On average IMN is the weakest baseline, followed by RACL and then RACL-BERT. The main improvement that RACL-BERT gives over RACL on these datasets is seen in the Targeted metric, i.e., the contextualized representations improve the polarity classification more than the extraction task. The proposed graph-based models are consistently the best models across the metrics and datasets. Regarding graph representations, the differences between Head-first and Head-final are generally quite small. Head-first performs better on MultiBCA and slightly better on MultiBEU, while for the others (NoReCFine, MPQA, and DSUnis) Head-final is better. This suggests that the main benefit is the joint prediction of all spans and relationships, and that the specific graph representation matters less. 7 Analysis In this section we perform a deeper analysis of the models in order to answer the research questions. 7.1 Do syntactically informed sentiment graphs improve results? Our two baseline graph representations, Head-first and Head-final, are crude approximations of linguistic structure. In syntactic and semantic dependency graphs, heads are often neither the first or last word, but rather the most salient word according to various linguistic criteria. First, we enrich the dependency labels to distinguish edges that are internal to a holder/target/expression span from those that are external and perform experiments by adding an ‘in label’ to non-head nodes within the graph, which we call +inlabel. We further inform the head selection of the parsing graphs with syntactic information in the Dep. edges parsing 3394 Spans Targeted Graph Sent. Graph Holder F1 Target F1 Exp. F1 F1 UF1 LF1 NSF1 SF1 NoReCFine 1.2 5.0 3.4 4.2 2.8 2.7 4.6 4.0 MultiBEU 2.9 0.6 0.8 1.1 1.0 1.4 1.2 1.4 MultiBCA 0.4 1.6 1.6 2.1 2.0 1.8 3.3 2.8 MPQA 8.2 8.8 5.2 7.2 6.6 7.3 5.4 5.1 DSUnis 7.9 1.2 4.3 6.4 3.9 5.7 3.6 6.0 Table 5: Average gains in percentage points by including mBERT representations. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 3: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis) in percentage points. The results are averaged across datasets. graphs, where we compute the dependency graph for each sentence9 and set the head of each span to be the node that has an outgoing edge in the corresponding syntactic graph. As there can be more than one such edge, we default to the first. A manual inspection showed that this approach sometimes set unlikely dependency label types as heads, e.g., punct, obl. Therefore, we suggest a final approach, Dep. labels, which filters out these unlikely heads. The full results are shown in Table 8 in the Appendix. The implementation of the graph structure has a large effect on all metrics, although the specific results depend on the dataset. We plot the average effect of each implementation across all datasets in Figure 3, as well as each individual dataset (Figures 4–8 in the Appendix). +inlabel tends to improve results on the non-English datasets, consistently increasing target and expression extraction and targeted sentiment. It also generally improves the graph scores UF1 and LF1 on the non-English datasets. 9We use SpaCy (Honnibal et al., 2020) for English, Stanza (Qi et al., 2020) for Basque and Catalan and UDPipe (Straka and Strakov´a, 2017) for Norwegian. Dep. edges has the strongest positive effect on the NSF1 and SF1 (an avg. 2.52 and 2.22 percentage point (pp) over Head-final, respectively). However, this average is pulled down by poorer performance on the English datasets. Removing these two, the average benefit is 5.2 and 4.2 for NSF1 and SF1, respectively. On span extraction and targeted sentiment, however, Dep. edges leads to poorer scores overall. Dep. labels does not lead to any consistent improvements. These results indicate that incorporating syntactic dependency information is particularly helpful for the full structured sentiment task, but that these benefits do not always show at a more local level, i.e., span extraction. 7.2 Do graph models perform better on sentences with multiple targets? We hypothesize that predicting the full sentiment graph may have a larger effect on sentences with multiple targets. Therefore, we create a subset of the test data containing sentences with multiple targets and reevaluate Head-first, Head-final, and RACL-BERT on the target extraction task. Table 4 shows the number of sentences with multiple targets and the Target span extraction score for each model. On this subset, Head-first and Head-final outperform RACL-BERT on 9 of 10 experiments, confirming the hypothesis that the graph models improve on examples with multiple targets. 7.3 How much does mBERT contribute? We also perform experiments without mBERT (shown in Table 7 in the Appendix) and show the average gains (over all 6 graph setups) of including it in Table 5. Adding the mBERT features leads to average improvements in all experiments: for extracting spans an average gain of 4.1 pp for holders, 3.4 for targets, and 3.1 for expressions. For targeted sentiment there is a larger gain of 4.2 pp, while for the parsing graph metrics UF1 and lF1 the gains are more limited (3.3 pp/ 3.8 pp) and similarly for NSF1 and SF1 (3.6 pp/ 3.9 pp). The gains are 3395 NoReCFine 57.0 (1.5) MultiBEU 75.7 (0.8) MultiBCA 71.7 (2.4) MPQA 38.5 (1.4) DSUnis 44.5 (2.4) Table 6: Polarity F1 scores (unweighted and weighted) of models augmented with mBERT on the head-final setup. We report average and standard deviation over 5 runs. largest for the English datasets (MPQA, DSUnis) followed by NoReCFine, and finally MultiBCA and MultiBEU. This corroborates the bias towards English and similar languages that has been found in multilingual language models (Artetxe et al., 2020; Conneau et al., 2020) and motivates the need for language-specific contextualized embeddings. 7.4 Analysis of polarity predictions In this section we zoom in on polarity, in order to quantify how well models perform at predicting only polarity. As the polarity annotations are bound to the expressions, we consider true positives to be any expression that overlaps the gold expression and has the same polarity. Table 6 shows that the polarity predictions are best on and MultiBCA, followed by NoReCFine and DSUnis, and finally MPQA. This is likely due to the number of domains and characteristics of the data. NoReCFine contains many domains and has longer expressions, while MPQA contains many highly ambiguous polar expressions, e.g., ‘said’, ‘asked’, which have different polarity depending on the context. 8 Conclusion In this paper, we have proposed a dependency graph parsing approach to structured sentiment analysis and shown that these models outperform state-of-the-art sequence labeling models on five benchmark datasets. Using parse trees as input has shown promise for sentiment analysis in the past, either to guide a tree-based algorithm (Socher et al., 2013; Tai et al., 2015) or to create features for sentiment models (Nakagawa et al., 2010; Almeida et al., 2015). However, to the authors’ knowledge, this is the first attempt to directly predict dependencybased sentiment graphs. In the future, we would like to better exploit the similarities between dependency parsing and sentiment graph parsing, either by augmenting the token-level representations with contextualized vectors from their heads in a dependency tree (Kurtz et al., 2020) or by multi-task learning to dependency parse. We would also like to explore different graph parsing approaches, e.g., PERIN (Samuel and Straka, 2020). Acknowledgements This work has been carried out as part of the SANT project (Sentiment Analysis for Norwegian Text), funded by the Research Council of Norway (grant number 270908). The computations were performed on resources provided by UNINETT Sigma2 - the National Infrastructure for High Performance Computing and Data Storage in Norway. References Abdulqader Almars, Xue Li, Xin Zhao, Ibrahim A. Ibrahim, Weiwei Yuan, and Bohan Li. 2017. Structured sentiment analysis. In Advanced Data Mining and Applications, pages 695–707, Cham. Springer International Publishing. Mariana S. C. Almeida, Cl´audia Pinto, Helena Figueira, Pedro Mendes, and Andr´e F. T. Martins. 2015. Aligning opinions: Cross-lingual opinion mining with dependencies. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 408–418, Beijing, China. Association for Computational Linguistics. Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics. Jeremy Barnes, Toni Badia, and Patrik Lambert. 2018. MultiBooked: A corpus of basque and Catalan hotel reviews annotated for aspect-level sentiment classification. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018), Miyazaki, Japan. European Languages Resources Association (ELRA). Taylor Berg-Kirkpatrick, David Burkett, and Dan Klein. 2012. An Empirical Investigation of Statistical Significance in NLP. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 995–1005, Jeju Island, Korea. Association for Computational Linguistics. 3396 Jari Bj¨orne, Juho Heimonen, Filip Ginter, Antti Airola, Tapio Pahikkala, and Tapio Salakoski. 2009. Extracting Complex Biological Events with Rich Graph-based Feature Sets. In Proceedings of the Workshop on Current Trends in Biomedical Natural Language Processing: Shared Task, BioNLP ’09, pages 10–18, Stroudsburg, PA, USA. Association for Computational Linguistics. Zhuang Chen and Tieyun Qian. 2020. Relation-aware collaborative learning for unified aspect-based sentiment analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3685–3694, Online. Association for Computational Linguistics. Yejin Choi, Eric Breck, and Claire Cardie. 2006. Joint extraction of entities and relations for opinion recognition. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, pages 431–439, Sydney, Australia. Association for Computational Linguistics. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440– 8451, Online. Association for Computational Linguistics. Hang Cui, Renxu Sun, Keya Li, Min-Yen Kan, and TatSeng Chua. 2005. Question answering passage retrieval using dependency relations. In Proceedings of the 28th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’05, pages 400–407, Salvador, Brazil. Association for Computing Machinery. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Timothy Dozat and Christopher D. Manning. 2018. Simpler but more accurate semantic dependency parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 484–490, Melbourne, Australia. Association for Computational Linguistics. Murhaf Fares, Andrey Kutuzov, Stephan Oepen, and Erik Velldal. 2017. Word vectors, reuse, and replicability: Towards a community repository of large-text resources. In Proceedings of the 21st Nordic Conference on Computational Linguistics, pages 271– 276, Gothenburg, Sweden. Association for Computational Linguistics. Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Dahlmeier. 2019. An interactive multi-task learning network for end-to-end aspect-based sentiment analysis. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 504–515, Florence, Italy. Association for Computational Linguistics. Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. spaCy: Industrial-strength Natural Language Processing in Python. Long Jiang, Mo Yu, Ming Zhou, Xiaohua Liu, and Tiejun Zhao. 2011. Target-dependent twitter sentiment classification. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 151–160, Portland, Oregon, USA. Association for Computational Linguistics. Richard Johansson and Alessandro Moschitti. 2012. Relational Features in Fine-Grained Opinion Analysis. Computational Linguistics, 39(3):473–509. Arzoo Katiyar and Claire Cardie. 2016. Investigating LSTMs for joint extraction of opinion entities and relations. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 919–929, Berlin, Germany. Association for Computational Linguistics. Marco Kuhlmann and Stephan Oepen. 2016. Towards a Catalogue of Linguistic Graph Banks. Computational Linguistics, 42(4):819–827. Robin Kurtz, Stephan Oepen, and Marco Kuhlmann. 2020. End-to-end negation resolution as graph parsing. In Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020 Shared Task on Parsing into Enhanced Universal Dependencies, pages 14–24, Online. Association for Computational Linguistics. Emanuele Lapponi, Erik Velldal, Lilja Øvrelid, and Jonathon Read. 2012. UiO2: Sequence-labeling Negation Using Dependency Features. In Proceedings of the First Joint Conference on Lexical and Computational Semantics - Volume 1: Proceedings of the Main Conference and the Shared Task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation, SemEval ’12, pages 319–327, Stroudsburg, PA, USA. Association for Computational Linguistics. Xin Li, Lidong Bing, Piji Li, and Wai Lam. 2019a. A unified model for opinion target extraction and target sentiment prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6714– 6721. Xin Li, Lidong Bing, Wenxuan Zhang, and Wai Lam. 2019b. Exploiting BERT for end-to-end aspectbased sentiment analysis. In Proceedings of the 5th 3397 Workshop on Noisy User-generated Text (W-NUT 2019), pages 34–41, Hong Kong, China. Association for Computational Linguistics. Bing Liu. 2012. Sentiment Analysis and Opinion Mining. Morgan & Claypool Publishers. Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1003–1011, Suntec, Singapore. Association for Computational Linguistics. Margaret Mitchell, Jacqui Aguilar, Theresa Wilson, and Benjamin Van Durme. 2013. Open domain targeted sentiment. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1643–1654, Seattle, Washington, USA. Association for Computational Linguistics. Roser Morante and Eduardo Blanco. 2012. *SEM 2012 Shared Task: Resolving the Scope and Focus of Negation. In *SEM 2012: The First Joint Conference on Lexical and Computational Semantics – Volume 1: Proceedings of the Main Conference and the Shared Task, and Volume 2: Proceedings of the Sixth International Workshop on Semantic Evaluation (SemEval 2012), pages 265–274, Montr´eal, Canada. Association for Computational Linguistics. Roser Morante and Walter Daelemans. 2012. ConanDoyle-neg: Annotation of negation cues and their scope in Conan Doyle stories. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 1563–1568, Istanbul, Turkey. European Language Resources Association (ELRA). Tetsuji Nakagawa, Kentaro Inui, and Sadao Kurohashi. 2010. Dependency tree-based sentiment classification using CRFs with hidden variables. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 786–794, Los Angeles, California. Association for Computational Linguistics. Stephan Oepen, Omri Abend, Lasha Abzianidze, Johan Bos, Jan Hajic, Daniel Hershcovich, Bin Li, Tim O’Gorman, Nianwen Xue, and Daniel Zeman. 2020. MRP 2020: The Second Shared Task on Cross-Framework and Cross-Lingual Meaning Representation Parsing. In Proceedings of the CoNLL 2020 Shared Task: Cross-Framework Meaning Representation Parsing, pages 1–22, Online. Association for Computational Linguistics. Stephan Oepen, Marco Kuhlmann, Yusuke Miyao, Daniel Zeman, Silvie Cinkova, Dan Flickinger, Jan Hajic, and Zdenka Uresova. 2015. SemEval 2015 Task 18: Broad-Coverage Semantic Dependency Parsing. Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 915–926. Stephan Oepen, Marco Kuhlmann, Yusuke Miyao, Daniel Zeman, Dan Flickinger, Jan Hajic, Angelina Ivanova, and Yi Zhang. 2014. SemEval 2014 Task 8: Broad-Coverage Semantic Dependency Parsing. Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 63–72. Lilja Øvrelid, Petter Mæhlum, Jeremy Barnes, and Erik Velldal. 2020. A fine-grained sentiment dataset for Norwegian. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5025– 5033, Marseille, France. European Language Resources Association. Haiyun Peng, Lu Xu, Lidong Bing, Fei Huang, Wei Lu, and Luo Si. 2019. Knowing what, how and why: A near complete solution for aspect-based sentiment analysis. Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad AL-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orph´ee De Clercq, V´eronique Hoste, Marianna Apidianaki, Xavier Tannier, Natalia Loukachevitch, Evgeniy Kotelnikov, Nuria Bel, Salud Mar´ıa Jim´enez-Zafra, and G¨uls¸en Eryi˘git. 2016. SemEval-2016 task 5: Aspect based sentiment analysis. In Proceedings of the 10th International Workshop on Semantic Evaluation (SemEval2016), pages 19–30, San Diego, California. Association for Computational Linguistics. Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. SemEval-2015 task 12: Aspect based sentiment analysis. In Proceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015), pages 486–495, Denver, Colorado. Association for Computational Linguistics. Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. SemEval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 27–35, Dublin, Ireland. Association for Computational Linguistics. Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101– 108, Online. Association for Computational Linguistics. David Samuel and Milan Straka. 2020. ´UFAL at MRP 2020: Permutation-invariant semantic parsing in PERIN. In Proceedings of the CoNLL 2020 3398 Shared Task: Cross-Framework Meaning Representation Parsing, pages 53–64, Online. Association for Computational Linguistics. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631–1642, Seattle, Washington, USA. Association for Computational Linguistics. Milan Straka and Jana Strakov´a. 2017. Tokenizing, pos tagging, lemmatizing and parsing ud 2.0 with udpipe. In Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 88–99, Vancouver, Canada. Association for Computational Linguistics. Kai Sheng Tai, Richard Socher, and Christopher D. Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1556–1566, Beijing, China. Association for Computational Linguistics. Cigdem Toprak, Niklas Jakob, and Iryna Gurevych. 2010. Sentence and expression level annotation of opinions in user-generated discourse. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 575–584, Uppsala, Sweden. Association for Computational Linguistics. Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, and Xiaokui Xiao. 2016. Recursive neural conditional random fields for aspect-based sentiment analysis. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 616–626, Austin, Texas. Association for Computational Linguistics. Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39(2-3):165–210. Hu Xu, Bing Liu, Lei Shu, and Philip Yu. 2019. BERT post-training for review reading comprehension and aspect-based sentiment analysis. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2324–2335, Minneapolis, Minnesota. Association for Computational Linguistics. Bishan Yang and Claire Cardie. 2012. Extracting opinion expressions with semi-Markov conditional random fields. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1335–1345, Jeju Island, Korea. Association for Computational Linguistics. Juntao Yu, Bernd Bohnet, and Massimo Poesio. 2020. Named Entity Recognition as Dependency Parsing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6470–6476, Online. Association for Computational Linguistics. Meishan Zhang, Qiansheng Wang, and Guohong Fu. 2019. End-to-end neural opinion extraction with a transition-based model. Information Systems, 80:56 – 63. 3399 A Appendix Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 4: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis) for NoReCFine. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 5: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis) for MultiBEU. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 6: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis) in percentage points for MultiBCA. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 7: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis) in percentage points for MPQA. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 8: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (x-axis in percentage points) for DSUnis. Holder F1 Target F1 Exp F1 Targeted F1 UF1 LF1 NSF1 SF1 Dep. labels Dep. edges Head-final+inlabel Head-final Head-first+inlabel Head-first 4 2 0 2 4 Figure 9: Average benefit of each graph annotation scheme (y-axis) on the evaluation metrics (xaxis) in percentage points. The results on NoReCFine, MultiBEU, MultiBCA. 3400 Spans Targeted Graph Sent. Graph Holder F1 Target F1 Exp. F1 F1 UF1 LF1 NSF1 SF1 NoReCFine IMN 35.9 48.7 18.0 RACL 45.6 55.4 20.1 Head-first 48.4 (2.2) 47.1 (1.6) 52.0 (1.6) 33.0 (1.4) 37.6 (0.5) 29.8 (0.4) 32.9 (1.6) 26.1 (1.5) +inlabel 50.4 (4.0) 47.6 (2.5) 51.0 (1.3) 27.3 (1.1) 36.9 (0.5) 29.4 (0.8) 32.9 (1.1) 25.8 (0.5) Head-final 57.0 (3.3) 49.4 (0.9) 52.1 (1.8) 26.0 (0.6) 45.1 (1.2) 35.2 (1.1) 34.4 (0.7) 27.2 (0.9) +inlabel 57.9 (1.7) 50.1 (1.3) 52.6 (0.4) 29.6 (0.6) 45.0 (1.0) 35.2 (0.5) 35.1 (1.6) 27.0 (1.3) Dep. edges 54.4 (3.9) 49.0 (2.5) 51.4 (1.7) 26.7 (3.1) 39.3 (1.1) 31.5 (1.3) 47.2 (0.9) 36.0 (1.1) Dep. labels 51.6 (2.6) 46.5 (3.0) 50.7 (2.7) 26.7 (1.9) 36.7 (1.1) 28.3 (0.8) 33.4 (1.8) 25.4 (1.8) MultiBEU IMN 48.2 65.2 39.5 RACL 55.4 70.7 48.2 Head-first 60.8 (3.8) 64.1 (1.4) 72.2 (0.7) 53.9 (1.8) 62.9 (0.6) 58.2 (0.3) 58.5 (2.3) 54.7 (2.6) +inlabel 59.8 (1.6) 64.3 (0.9) 71.9 (0.8) 57.9 (1.9) 62.6 (0.6) 57.5 (1.1) 57.3 (1.5) 53.6 (1.3) Head-final 57.0 (2.0) 66.0 (1.6) 72.2 (0.6) 55.5 (1.7) 60.2 (0.8) 55.5 (0.9) 59.6 (0.8) 56.3 (1.0) +inlabel 53.7 (1.2) 64.0 (2.4) 72.9 (0.6) 54.9 (2.0) 60.1 (1.5) 54.9 (1.7) 57.1 (3.2) 53.5 (3.3) Dep. edges 53.1 (1.9) 63.8 (1.7) 71.0 (1.1) 53.7 (1.7) 59.0 (1.3) 54.5 (1.6) 59.0 (1.6) 55.6 (1.8) Dep. labels 52.0 (3.8) 63.0 (1.1) 71.3 (1.9) 54.0 (1.0) 59.5 (1.1) 54.9 (1.1) 58.6 (2.8) 54.6 (2.4) MultiBCA IMN 56.3 60.9 32.5 RACL 65.4 67.6 53.1 Head-first 41.9 (2.8) 69.8 (1.7) 68.9 (1.4) 57.3 (2.0) 64.2 (0.7) 59.9 (0.8) 58.2 (2.3) 53.3 (2.2) +inlabel 42.4 (2.6) 70.9 (0.8) 69.9 (0.9) 50.9 (1.3) 64.4 (0.9) 59.6 (0.7) 55.7 (2.0) 50.7 (2.1) Head-final 40.4 (2.5) 69.9 (1.5) 66.8 (0.8) 50.8 (2.6) 60.9 (0.6) 57.1 (0.8) 57.7 (1.0) 53.3 (1.3) +inlabel 36.4 (2.2) 69.1 (1.1) 65.4 (0.6) 52.9 (0.9) 60.6 (0.9) 57.0 (0.6) 58.0 (1.9) 53.5 (2.0) Dep. edges 42.6 (6.1) 69.1 (0.5) 67.3 (0.6) 50.6 (1.3) 59.3 (0.7) 55.7 (1.1) 57.5 (2.1) 52.8 (1.8) Dep. labels 43.8 (3.4) 70.3 (0.8) 67.2 (1.1) 50.6 (1.8) 61.0 (0.5) 57.1 (0.8) 57.8 (1.4) 52.7 (1.9) MPQA IMN 24.3 29.6 1.2 RACL 32.6 37.8 11.8 Head-first 35.2 (1.1) 40.5 (1.8) 41.7 (1.7) 22.6 (3.1) 32.2 (1.3) 28.2 (1.4) 19.4 (1.5) 12.4 (1.7) +inlabel 35.6 (1.4) 41.6 (1.1) 42.4 (2.3) 14.0 (0.9) 32.9 (0.8) 28.9 (0.9) 20.4 (1.0) 13.2 (1.2) Head-final 37.1 (1.3) 42.1 (1.0) 41.9 (0.8) 13.3 (1.9) 35.7 (0.8) 31.7 (0.5) 18.7 (0.7) 12.5 (1.6) +inlabel 37.0 (0.5) 42.3 (1.2) 41.6 (1.6) 15.2 (2.5) 35.5 (0.7) 31.7 (0.6) 19.6 (0.6) 12.6 (0.8) Dep. edges 35.4 (1.5) 39.1 (2.0) 41.6 (1.1) 12.3 (0.8) 28.9 (1.7) 24.8 (1.5) 19.0 (0.9) 11.9 (1.1) Dep. labels 36.7 (0.6) 39.2 (2.3) 40.4 (2.1) 12.6 (1.2) 28.9 (1.6) 24.5 (1.1) 18.9 (0.8) 11.6 (1.0) DSUnis IMN 33.0 27.4 17.9 RACL 39.3 40.2 22.8 Head-first 25.6 (5.1) 36.8 (3.5) 39.0 (1.5) 23.4 (1.8) 32.9 (1.7) 25.9 (1.7) 27.8 (1.4) 18.8 (2.0) +inlabel 22.9 (5.7) 38.6 (3.9) 38.6 (2.8) 18.3 (2.7) 33.7 (2.8) 26.4 (1.6) 25.9 (3.4) 15.2 (2.2) Head-final 29.2 (8.4) 38.1 (2.0) 39.5 (2.4) 21.8 (1.1) 31.1 (2.1) 26.0 (1.0) 29.1 (3.3) 20.4 (2.0) +inlabel 30.2 (8.9) 38.2 (2.9) 38.8 (2.0) 24.4 (2.7) 32.4 (2.1) 28.4 (2.1) 28.1 (2.9) 22.4 (3.4) Dep. edges 33.9 (5.4) 39.2 (2.9) 39.3 (3.4) 22.2 (2.7) 32.4 (2.1) 26.8 (2.0) 29.3 (1.5) 19.8 (1.2) Dep. labels 21.3 (18.1) 40.0 (1.5) 38.4 (2.4) 21.4 (4.3) 32.1 (1.5) 27.2 (1.5) 28.2 (1.1) 20.5 (3.1) Table 7: Experiments without contextualized embeddings. 3401 Spans Targeted Graph Sent. Graph Holder F1 Target F1 Exp. F1 F1. UF1 LF1 NSF1 SF1 NoReCFine RACL-BERT 47.2 56.3 30.3 Head-first 51.1 (3.2) 50.1 (3.4) 54.4 (1.6) 30.5 (2.3) 39.2 (0.5) 31.5 (0.5) 37.0 (2.6) 29.5 (2.4) +inlabel 51.6 (2.8) 52.7 (0.7) 54.6 (1.4) 32.2 (1.4) 39.6 (0.8) 32.0 (0.7) 37.6 (1.2) 29.5 (1.2) Head-final 60.4 (1.2) 54.8 (1.6) 55.5 (1.5) 31.9 (1.3) 48.0 (1.3) 37.7 (1.4) 39.2 (1.7) 31.2 (1.6) +inlabel 57.1 (3.0) 55.2 (1.0) 56.3 (1.3) 34.8 (1.0) 48.7 (1.2) 38.3 (1.0) 40.5 (1.1) 31.7 (1.1) Dep. edges 54.0 (3.4) 53.6 (1.5) 55.0 (0.9) 32.7 (1.6) 41.5 (0.7) 33.8 (0.4) 50.9 (0.3) 39.4 (0.4) Dep. labels 52.7 (5.6) 53.6 (0.3) 54.4 (1.5) 32.7 (1.6) 40.7 (0.8) 32.2 (0.5) 38.2 (1.4) 30.0 (1.2) MultiBEU RACL-BERT 59.9 72.6 56.8 Head-first 60.4 (2.2) 64.0 (2.4) 73.9 (1.0) 57.8 (2.4) 64.6 (1.0) 60.0 (1.6) 58.0 (1.1) 54.7 (1.6) +inlabel 59.6 (1.9) 65.9 (0.9) 74.2 (0.7) 59.2 (0.9) 64.7 (0.7) 60.3 (1.1) 59.8 (1.1) 56.1 (1.6) Head-final 60.5 (2.2) 64.0 (2.3) 72.1 (1.2) 56.9 (1.7) 60.8 (0.8) 56.0 (1.1) 58.0 (2.1) 54.7 (1.8) +inlabel 58.1 (2.4) 64.7 (1.1) 72.0 (0.7) 58.5 (1.4) 60.6 (1.1) 56.6 (0.7) 59.8 (1.6) 56.9 (1.8) Dep. edges 58.8 (4.2) 64.8 (1.4) 71.2 (0.8) 54.0 (1.8) 59.9 (0.4) 55.5 (0.7) 60.9 (1.6) 57.4 (1.6) Dep. labels 56.3 (2.1) 65.4 (0.9) 72.9 (1.1) 54.9 (0.8) 60.0 (0.9) 55.6 (0.8) 60.5 (1.1) 57.1 (1.1) MultiBCA RACL-BERT 67.5 70.3 52.4 Head-first 43.0 (1.3) 72.5 (1.0) 71.1 (0.8) 55.0 (0.9) 66.8 (0.5) 62.1 (0.5) 62.0 (1.1) 56.8 (0.7) +inlabel 43.1 (2.2) 73.4 (1.0) 70.3 (1.0) 55.8 (1.8) 66.2 (0.3) 61.5 (0.6) 61.1 (1.0) 56.0 (1.0) Head-final 37.1 (4.2) 71.2 (0.6) 67.1 (1.7) 53.9 (2.2) 62.7 (0.4) 58.1 (0.8) 59.7 (1.1) 53.7 (2.4) +inlabel 34.9 (4.1) 70.7 (1.4) 68.2 (1.0) 53.5 (0.7) 63.4 (0.5) 58.7 (0.6) 60.9 (1.1) 55.1 (1.2) Dep. edges 46.3 (3.1) 70.3 (0.6) 69.2 (1.4) 53.4 (1.5) 60.8 (0.4) 57.5 (0.6) 60.7 (1.0) 55.6 (0.9) Dep. labels 45.6 (2.9) 70.3 (1.1) 69.1 (1.7) 53.9 (1.5) 62.5 (0.6) 59.1 (0.6) 60.4 (1.0) 55.8 (1.2) MPQA RACL-BERT 20.0 31.2 17.8 Head-first 43.8 (1.8) 51.0 (1.9) 48.1 (0.8) 33.5 (3.1) 40.0 (1.0) 36.9 (1.2) 24.5 (2.3) 17.4 (2.7) +inlabel 43.1 (1.5) 51.5 (1.0) 47.5 (1.1) 21.3 (0.4) 40.6 (0.5) 37.5 (0.5) 24.5 (1.3) 17.3 (1.0) Head-final 46.3 (1.8) 49.5 (0.9) 46.0 (1.1) 21.9 (1.4) 41.4 (0.7) 38.0 (0.5) 26.1 (0.7) 18.8 (0.7) +inlabel 45.6 (2.5) 49.4 (2.1) 45.6 (1.1) 20.7 (1.0) 40.4 (1.5) 37.2 (1.9) 25.2 (1.7) 17.8 (1.3) Dep. edges 44.0 (1.5) 48.5 (1.2) 46.3 (1.9) 18.9 (2.3) 35.4 (1.3) 31.9 (1.2) 24.2 (1.6) 16.3 (1.9) Dep. labels 43.7 (0.9) 47.7 (2.3) 47.5 (0.8) 21.9 (0.7) 35.6 (1.2) 32.0 (1.3) 24.0 (0.8) 17.2 (0.8) DSUnis RACL-BERT 44.6 38.2 27.3 Head-first 28.0 (7.7) 39.9 (2.2) 40.3 (0.6) 26.7 (2.1) 35.3 (0.9) 31.4 (1.3) 31.0 (1.4) 25.0 (1.3) +inlabel 30.9 (9.9) 38.4 (3.3) 40.6 (2.9) 26.7 (2.4) 34.2 (2.1) 30.7 (2.5) 30.5 (2.1) 25.4 (2.3) Head-final 37.4 (11.6) 42.1 (2.7) 45.5 (2.4) 29.6 (1.7) 38.1 (1.9) 33.9 (2.3) 34.3 (4.2) 26.5 (3.5) +inlabel 30.6 (16.4) 38.9 (3.1) 45.2 (2.7) 28.1 (3.7) 37.3 (2.7) 33.3 (2.1) 29.4 (2.8) 23.7 (2.4) Dep. edges 32.7 (12.1) 39.9 (2.8) 44.8 (4.0) 28.9 (3.6) 37.3 (2.5) 33.8 (2.7) 33.2 (4.5) 27.3 (4.1) Dep. labels 30.8 (5.8) 38.9 (0.9) 43.1 (1.2) 27.8 (1.9) 35.7 (1.5) 32.1 (1.6) 31.3 (1.1) 25.3 (2.0) Table 8: Experiments with mBERT. 3402 GPU Infrastructure NVIDIA P100, 16 GiB RAM CPU Infrastructure Intel Xeon-Gold 6138 2.0 GHz Training duration 00:31:43 (MultiBEU) – 07:40:54 (NoReCFine) Model implementation https://github.com/jerbarnes/sentiment_graphs/src Hyperparameter Best assignment embedding Word2Vec SkipGram 100D contexualized embedding mBERT embeddings trainable False number of epochs 100 batch size 50 beta1 0 beta2 0.95 l2 3e-09 hidden lstm 200 hidden char lstm 100 layers lstm 3 dim mlp 200 dim embedding 100 dim char embedding 80 early stopping 0 pos style xpos attention bilinear model interpolation 0.5 loss interpolation 0.025 lstm implementation drop connect char implementation convolved emb dropout type replace bridge dpa+ dropout embedding 0.2 dropout edge 0.2 dropout label 0.3 dropout main recurrent 0.2 dropout recurrent char 0.3 dropout main ff 0.4 dropout char ff 0.3 dropout char linear 0.3
|
2021
|
263
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3403–3417 August 1–6, 2021. ©2021 Association for Computational Linguistics 3403 Consistency Regularization for Cross-Lingual Fine-Tuning Bo Zheng†∗, Li Dong‡, Shaohan Huang‡, Wenhui Wang‡, Zewen Chi‡∗ Saksham Singhal‡, Wanxiang Che†, Ting Liu†, Xia Song‡, Furu Wei‡ †Harbin Institute of Technology ‡Microsoft Corporation {bzheng,car,tliu}@ir.hit.edu.cn {lidong1,shaohanh,wenwan,saksingh,xiaso,fuwei}@microsoft.com Abstract Fine-tuning pre-trained cross-lingual language models can transfer task-specific supervision from one language to the others. In this work, we propose to improve cross-lingual finetuning with consistency regularization. Specifically, we use example consistency regularization to penalize the prediction sensitivity to four types of data augmentations, i.e., subword sampling, Gaussian noise, code-switch substitution, and machine translation. In addition, we employ model consistency to regularize the models trained with two augmented versions of the same training set. Experimental results on the XTREME benchmark show that our method1 significantly improves crosslingual fine-tuning across various tasks, including text classification, question answering, and sequence labeling. 1 Introduction Pre-trained cross-lingual language models (Conneau and Lample, 2019; Conneau et al., 2020a; Chi et al., 2020) have shown great transferability across languages. By fine-tuning on labeled data in a source language, the models can generalize to other target languages, even without any additional training. Such generalization ability reduces the required annotation efforts, which is prohibitively expensive for low-resource languages. Recent work has demonstrated that data augmentation is helpful for cross-lingual transfer, e.g., translating source language training data into target languages (Singh et al., 2019), and generating codeswitch data by randomly replacing input words in the source language with translated words in target languages (Qin et al., 2020). By populating the dataset, their fine-tuning still treats training ∗Contribution during internship at Microsoft Research. 1The code is available at https://github.com/ bozheng-hit/xTune. instances independently, without considering the inherent correlations between the original input and its augmented example. In contrast, we propose to utilize consistency regularization to better leverage data augmentation for cross-lingual fine-tuning. Intuitively, for a semantic-preserving augmentation strategy, the predicted result of the original input should be similar to its augmented one. For example, the classification predictions of an English sentence and its translation tend to remain consistent. In this work, we introduce a cross-lingual finetuning method XTUNE that is enhanced by consistency regularization and data augmentation. First, example consistency regularization enforces the model predictions to be more consistent for semantic-preserving augmentations. The regularizer penalizes the model sensitivity to different surface forms of the same example (e.g., texts written in different languages), which implicitly encourages cross-lingual transferability. Second, we introduce model consistency to regularize the models trained with various augmentation strategies. Specifically, given two augmented versions of the same training set, we encourage the models trained on these two datasets to make consistent predictions for the same example. The method enforces the corpus-level consistency between the distributions learned by two models. Under the proposed fine-tuning framework, we study four strategies of data augmentation, i.e., subword sampling (Kudo, 2018), code-switch substitution (Qin et al., 2020), Gaussian noise (Aghajanyan et al., 2020), and machine translation. We evaluate XTUNE on the XTREME benchmark (Hu et al., 2020), including three different tasks on seven datasets. Experimental results show that our method outperforms conventional fine-tuning with data augmentation. We also demonstrate that XTUNE is flexible to be plugged in various 3404 tasks, such as classification, span extraction, and sequence labeling. We summarize our contributions as follows: • We propose XTUNE, a cross-lingual finetuning method to better utilize data augmentations based on consistency regularization. • We study four types of data augmentations that can be easily plugged into cross-lingual fine-tuning. • We give instructions on how to apply XTUNE to various downstream tasks, such as classification, span extraction, and sequence labeling. • We conduct extensive experiments to show that XTUNE consistently improves the performance of cross-lingual fine-tuning. 2 Related Work Cross-Lingual Transfer Besides learning crosslingual word embeddings (Mikolov et al., 2013; Faruqui and Dyer, 2014; Guo et al., 2015; Xu et al., 2018; Wang et al., 2019), most recent work of cross-lingual transfer is based on pre-trained cross-lingual language models (Conneau and Lample, 2019; Conneau et al., 2020a; Chi et al., 2020). These models generate multilingual contextualized word representations for different languages with a shared encoder and show promising cross-lingual transferability. Cross-Lingual Data Augmentation Machine translation has been successfully applied to the cross-lingual scenario as data augmentation. A common way to use machine translation is to finetune models on both source language training data and translated data in all target languages. Furthermore, Singh et al. (2019) proposed to replace a segment of source language input text with its translation in another language. However, it is usually impossible to map the labels in source language data into target language translations for token-level tasks. Zhang et al. (2019) used code-mixing to perform the syntactic transfer in cross-lingual dependency parsing. Fei et al. (2020) constructed pseudo translated target corpora from the gold-standard annotations of the source languages for cross-lingual semantic role labeling. Fang et al. (2020) proposed an additional Kullback-Leibler divergence self-teaching loss for model training, based on autogenerated soft pseudo-labels for translated text in the target language. Besides, Qin et al. (2020) finetuned models on multilingual code-switch data, which achieves considerable improvements. Consistency Regularization One strand of work in consistency regularization focused on regularizing model predictions to be invariant to small perturbations on image data. The small perturbations can be random noise (Zheng et al., 2016), adversarial noise (Miyato et al., 2019; Carmon et al., 2019) and various data augmentation approaches (Hu et al., 2017; Ye et al., 2019; Xie et al., 2020). Similar ideas are used in the natural language processing area. Both adversarial noise (Zhu et al., 2020; Jiang et al., 2020; Liu et al., 2020) and sampled Gaussian noise (Aghajanyan et al., 2020) are adopted to augment input word embeddings. Another strand of work focused on consistency under different model parameters (Tarvainen and Valpola, 2017; Athiwaratkun et al., 2019), which is complementary to the first strand. We focus on the cross-lingual setting, where consistency regularization has not been fully explored. 3 Methods Conventional cross-lingual fine-tuning trains a pretrained language model on the source language and directly evaluates it on other languages, which is also known as the setting of zero-shot cross-lingual fine-tuning. Specifically, given a training corpus D in the source language (typically in English), and a model f(·; θ) that predicts task-specific probability distributions, we define the loss of cross-lingual fine-tuning as: Ltask(D, θ) = X x∈D ℓ(f(x; θ), G(x)), where G(x) denotes the ground-truth label of example x, ℓ(·, ·) is the loss function depending on the downstream task. Apart from vanilla cross-lingual fine-tuning on the source language, recent work shows that data augmentation is helpful to improve performance on the target languages. For example, Conneau and Lample (2019) add translated examples to the training set for better cross-lingual transfer. Let A(·) be a cross-lingual data augmentation strategy (such as code-switch substitution), and DA = D ∪ {A(x) | x ∈D} be the augmented training corpus, the fine-tuning loss is Ltask(DA, θ). Notice that it is non-trivial to apply some augmentations for tokenlevel tasks directly. For instance, in part-of-speech 3405 Figure 1: Overview of our two-stage fine-tuning algorithm. The model parameters f(·; θ∗) in the second stage are copied from the first stage. tagging, the labels of source language examples can not be mapped to the translated examples because of the lack of explicit alignments. 3.1 XTUNE: Cross-Lingual Fine-Tuning with Consistency Regularization We propose to improve cross-lingual fine-tuning with two consistency regularization methods, so that we can effectively leverage cross-lingual data augmentations. 3.1.1 Example Consistency Regularization In order to encourage consistent predictions for an example and its semantically equivalent augmentation, we introduce example consistency regularization, which is defined as follows: R1(D, θ, A) = X x∈D KLS(f(x; θ)∥f(A(x); θ)), KLS(P, Q) = KL(stopgrad(P)∥Q)+ KL(stopgrad(Q)∥P) where KLS(·) is the symmertrical Kullback-Leibler divergence. The regularizer encourages the predicted distributions f(x; θ) and f(A(x); θ) to agree with each other. The stopgrad(·) operation2 is used to stop back-propagating gradients, which is also employed in (Jiang et al., 2020; Liu et al., 2020). The ablation studies in Section 4.2 empirically show that the operation improves fine-tuning performance. 2Implemented by .detach() in PyTorch. 3.1.2 Model Consistency Regularization While the example consistency regularization is conducted at the example level, we propose the model consistency to further regularize the model training at the corpus level. The regularization is conducted at two stages. First, we obtain a finetuned model θ∗on the training corpus D: θ∗= arg min θ1 Ltask(D, θ1). In the second stage, we keep the parameters θ∗ fixed. The regularization term is defined as: R2(DA, θ, θ∗) = X x∈DA KL(f(x; θ∗)∥f(x; θ)) where DA is the augmented training corpus, and KL(·) is Kullback-Leibler divergence. For each example x of the augmented training corpus DA, the model consistency regularization encourages the prediction f(x; θ) to be consistent with f(x; θ∗). The regularizer enforces the corpus-level consistency between the distributions learned by two models. An unobvious advantage of model consistency regularization is the flexibility with respect to data augmentation strategies. For the example of partof-speech tagging, even though the labels can not be directly projected from an English sentence to its translation, we are still able to employ the regularizer. Because the term R2 is put on the same example x ∈DA, we can always align the tokenlevel predictions of the models θ and θ∗. 3406 I love to eat apples. Ich mag es, Äpfel zu essen J'adore manger des pommes. 我喜欢吃苹果。 _I/_love/_to/_eat/_apple/s/. _I/_love/_to/_e/a/t/_app/l/es/. _/I/_lo/ve/_to/_e/at/_app/l/es/. I 喜欢to essen apples. I liebe to eat 苹果. Ich 喜欢to eat pommes. Gaussian Noise Machine Translation CodeSwitch Subword Sampling Embedding Layer I love to eat apples. Figure 2: Cross-lingual data augmentation strategies. 3.1.3 Full XTUNE Fine-Tuning As shown in Figure 1, we combine example consistency regularization R1 and model consistency regularization R2 as a two-stage fine-tuning process. Formally, we fine-tune a model with R1 in the first stage: θ∗= arg min θ1 Ltask(D, θ1) + R1(D, θ1, A∗) where the parameters θ∗are kept fixed for R2 in the second stage. Then the final loss is computed via: LXTUNE = Ltask(DA, θ) + λ1R1(DA, θ, A′) + λ2R2(DA, θ, θ∗) where λ1 and λ2 are the corresponding weights of two regularization methods. Notice that the data augmentation strategies A, A′, and A∗can be either different or the same, which are tuned as hyper-parameters. 3.2 Data Augmentation We consider four types of data augmentation strategies in this work, which are shown in Figure 2. We aim to study the impact of different data augmentation strategies on cross-lingual transferability. 3.2.1 Subword Sampling Representing a sentence in different subword sequences can be viewed as a data augmentation strategy (Kudo, 2018; Provilkov et al., 2020). We utilize XLM-R (Conneau et al., 2020a) as our pre-trained cross-lingual language model, while it applies subword tokenization directly on raw text data using SentencePiece (Kudo and Richardson, 2018) with a unigram language model (Kudo, 2018). As one of our data augmentation strategies, we apply the on-the-fly subword sampling algorithm in the unigram language model to generate multiple subword sequences. 3.2.2 Gaussian Noise Most data augmentation strategies in NLP change input text discretely, while we directly add random perturbation noise sampled from Gaussian distribution on the input embedding layer to conduct data augmentation. When combining this data augmentation with example consistency R1, the method is similar to the stability training (Zheng et al., 2016), random perturbation training (Miyato et al., 2019) and the R3F method (Aghajanyan et al., 2020). We also explore Gaussian noise’s capability to generate new examples on continuous input space for conventional fine-tuning. 3.2.3 Code-Switch Substitution Anchor points have been shown useful to improve cross-lingual transferability. Conneau et al. (2020b) analyzed the impact of anchor points in pre-training cross-lingual language models. Following Qin et al. (2020), we generate code-switch data in multiple languages as data augmentation. We randomly select words in the original text in the source language and replace them with target language words in the bilingual dictionaries to obtain code-switch data. Intuitively, this type of data augmentation explicitly helps pre-trained cross-lingual models align the multilingual vector space by the replaced anchor points. 3.2.4 Machine Translation Machine translation has been proved to be an effective data augmentation strategy (Singh et al., 2019) under the cross-lingual scenario. However, the ground-truth labels of translated data can be unavailable for token-level tasks (see Section 3), which disables conventional fine-tuning on the augmented data. Meanwhile, our proposed model consistency R2 can not only serve as consistency regularization but also can be viewed as a self-training objective to enable semi-supervised training on the unlabeled target language translations. 3407 3.3 Task Adaptation We give instructions on how to apply XTUNE to various downstream tasks, i.e., classification, span extraction, and sequence labeling. By default, we use model consistency R2 in full XTUNE. We describe the usage of example consistency R1 as follows. 3.3.1 Classification For classification task, the model is expected to predict one distribution per example on nlabel types, i.e., model f(·; θ) should predict a probability distribution pcls ∈Rnlabel. Thus we can directly use example consistency R1 to regularize the consistency of the two distributions for all four types of our data augmentation strategies. 3.3.2 Span Extraction For span extraction task, the model is expected to predict two distributions per example pstart, pend ∈ Rnsubword, indicating the probability distribution of where the answer span starts and ends, nsubword denotes the length of the tokenized input text. For Gaussian noise, the subword sequence remains unchanged so that example consistency R1 can be directly applied to the two distributions. Since subword sampling and code-switch substitution will change nsubword, we control the ratio of words to be modified and utilize example consistency R1 on unchanged positions only. We do not use the example consistency R1 for machine translation because it is impossible to explicitly align the two distributions. 3.3.3 Sequence Labeling Recent pre-trained language models generate representations at the subword-level. For sequence labeling tasks, these models predict label distributions on each word’s first subword. Therefore, the model is expected to predict nword probability distributions per example on nlabel types. Unlike span extraction, subword sampling, code-switch substitution, and Gaussian noise do not change nword. Thus the three data augmentation strategies will not affect the usage of example consistency R1. Although word alignment is a possible solution to map the predicted label distributions between translation pairs, the word alignment process will introduce more noise. Therefore, we do not employ machine translation as data augmentation for the example consistency R1. 4 Experiments 4.1 Experiment Setup Datasets For our experiments, we select three types of cross-lingual understanding tasks from XTREME benchmark (Hu et al., 2020), including two classification datasets: XNLI (Conneau et al., 2018), PAWS-X (Yang et al., 2019), three span extraction datasets: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), TyDiQA-GoldP (Clark et al., 2020), and two sequence labeling datasets: NER (Pan et al., 2017), POS (Nivre et al., 2018). The statistics of the datasets are shown in the supplementary document. Fine-Tuning Settings We consider two typical fine-tuning settings from Conneau et al. (2020a) and Hu et al. (2020) in our experiments, which are (1) cross-lingual transfer: the models are finetuned on English training data without translation available, and directly evaluated on different target languages; (2) translate-train-all: translationbased augmentation is available, and the models are fine-tuned on the concatenation of English training data and its translated data on all target languages. Since the official XTREME repository3 does not provide translated target language data for POS and NER, we use Google Translate to obtain translations for these two datasets. Implementation Details We utilize XLMR (Conneau et al., 2020a) as our pre-trained cross-lingual language model. The bilingual dictionaries we used for code-switch substitution are from MUSE (Lample et al., 2018).4 For languages that cannot be found in MUSE, we ignore these languages since other bilingual dictionaries might be of poorer quality. For the POS dataset, we use the average-pooling strategy on subwords to obtain word representation since part-of-speech is related to different parts of words, depending on the language. We tune the hyper-parameter and select the model with the best average results over all the languages’ development set. There are two datasets without development set in multi-languages. For XQuAD, we tune the hyper-parameters with the development set of MLQA since they share the same training set and have a higher degree of overlap in languages. For TyDiQA-GoldP, we use the English test set 3github.com/google-research/xtreme 4github.com/facebookresearch/MUSE 3408 Model Pair Sentence Structure Prediction Question Answering XNLI PAWS-X POS NER XQuAD MLQA TyDiQA Metrics Acc. Acc. F1 F1 F1/EM F1/EM F1/EM Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) mBERT 65.4 81.9 70.3 62.2 64.5/49.4 61.4/44.2 59.7/43.9 63.1 XLM 69.1 80.9 70.1 61.2 59.8/44.3 48.5/32.6 43.6/29.1 58.6 X-STILTs (Phang et al., 2020) 80.4 87.7 74.4 63.4 77.2/61.3 72.3/53.5 76.0/59.5 72.3 VECO (Luo et al., 2020) 79.9 88.7 75.1 65.7 77.3/61.8 71.7/53.2 67.6/49.1 71.4 XLM-Rlarge 79.2 86.4 72.6 65.4 76.6/60.8 71.6/53.2 65.1/45.0 70.0 XTUNE 82.6 89.8 78.5 69.3 79.4/64.4 74.4/56.2 74.8/59.4 74.9 Translate-train-all (translation-based augmentation is available for English training data) VECO (Luo et al., 2020) 83.0 91.1 75.1 65.7 79.9/66.3 73.1/54.9 75.0/58.9 74.1 FILTER (Fang et al., 2020) 83.9 91.4 76.2 67.7 82.4/68.0 76.2/57.7 68.3/50.9 74.4 XLM-Rlarge 82.6 90.4 80.2/65.9 72.8/54.3 66.5/47.7 XTUNE 84.8 91.6 79.3 69.9 82.5/69.0 75.0/57.1 75.4/60.8 76.5 Table 1: Evaluation results on the XTREME benchmark. Results of mBERT (Devlin et al., 2019), XLM (Conneau and Lample, 2019) and XLM-Rlarge (Conneau et al., 2020a) are taken from (Hu et al., 2020). Results of XLM-Rlarge under the translate-train-all setting are from FILTER (Fang et al., 2020). The results of XTUNE are from the best models selected with the performance on the corresponding development set. Model Pair Sentence Structure Prediction Question Answering XNLI PAWS-X POS NER XQuAD MLQA TyDiQA Metrics Acc. Acc. F1 F1 F1/EM F1/EM F1/EM Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rbase 74.9 84.9 75.6 61.8 71.9/56.4 65.0/47.1 55.4/38.3 XTUNE 77.7 87.5 76.5 63.0 73.9/59.0 68.1/50.2 61.2/45.2 with only example consistency R1 77.6 87.2 76.3 62.4 73.6/58.6 67.6/49.7 60.7/44.4 with only model consistency R2 76.6 86.3 76.3 63.0 73.2/58.1 66.7/49.0 59.2/42.3 Translate-train-all (translation-based augmentation is available for English training data) XLM-Rbase 78.8 88.4 75.2/61.4 67.8/50.1 63.7/47.7 XTUNE 80.6 89.4 77.8 63.7 78.1/64.4 69.7/52.1 65.9/51.1 with only example consistency R1 80.5 89.3 76.1/62.5 69.1/51.6 65.1/50.3 with only model consistency R2 78.9 88.5 76.6 63.5 77.4/63.4 68.7/51.1 64.5/48.7 remove stopgrad in R1 80.2 89.1 76.8 63.4 77.3/63.4 69.9/52.1 65.1/50.5 Table 2: Ablation studies on the XTREME benchmark. All numbers are averaged over five random seeds. as the development set. In order to make a fair comparison, the ratio of data augmentation in DA is all set to 1.0. The detailed hyper-parameters are shown in the supplementary document. 4.2 Results Table 1 shows our results on XTREME. For the cross-lingual transfer setting, we outperform previous works on all seven cross-lingual language understanding datasets.5 Compared to XLM-Rlarge baseline, we achieve an absolute 4.9-point improvement (70.0 vs. 74.9) on average over seven datasets. For the translate-train-all setting, we achieved stateof-the-art results on six of the seven datasets. Com5X-STILTs (Phang et al., 2020) uses additional SQuAD v1.1 English training data for the TyDiQA-GoldP dataset, while we prefer a cleaner setting here. pared to FILTER,6 we achieve an absolute 2.1point improvement (74.4 vs. 76.5), and we do not need English translations during inference. Table 2 shows how the two regularization methods affect the model performance separately. For the cross-lingual transfer setting, XTUNE achieves an absolute 2.8-point improvement compared to our implemented XLM-Rbase baseline. Meanwhile, fine-tuning with only example consistency R1 and model consistency R2 degrades the averaged results by 0.4 and 1.0 points, respectively. For the translate-train-all setting, our proposed model consistency R2 enables training on POS and NER even if labels of target language translations 6FILTER directly selects the best model on the test set of XQuAD and TyDiQA-GoldP. Under this setting, we can obtain 83.1/69.7 for XQuAD, 75.5/61.1 for TyDiQA-GoldP. 3409 Model en ar bg de el es fr hi ru sw th tr ur vi zh Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) R3F (Aghajanyan et al., 2020) 89.4 80.6 84.6 83.7 83.6 85.1 84.2 77.3 82.3 72.6 79.4 80.7 74.2 81.1 80.1 81.2 R4F (Aghajanyan et al., 2020) 89.6 80.5 84.6 84.2 83.6 85.2 84.7 78.2 82.5 72.7 79.2 80.3 73.9 80.9 80.6 81.4 XLM-Rlarge 88.7 77.2 83.0 82.5 80.8 83.7 82.2 75.6 79.1 71.2 77.4 78.0 71.7 79.3 78.2 79.2 XTUNE 89.6 81.6 85.9 84.8 84.3 86.5 85.4 80.5 82.8 73.3 80.3 82.1 77.1 83.0 82.3 82.6 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 89.5 83.6 86.4 85.6 85.4 86.6 85.7 81.1 83.7 78.7 81.7 83.2 79.1 83.9 83.8 83.9 XLM-Rlarge 88.6 82.2 85.2 84.5 84.5 85.7 84.2 80.8 81.8 77.0 80.2 82.1 77.7 82.6 82.7 82.6 XTUNE 89.9 84.0 87.0 86.5 86.2 87.4 86.6 83.2 85.2 80.0 82.7 84.1 79.6 84.8 84.3 84.8 Table 3: XNLI accuracy scores for each language. XLM-Rlarge under the cross-lingual transfer setting are from (Hu et al., 2020). Results of XLM-Rlarge under the translate-train-all setting are from (Fang et al., 2020). Method Model XNLI POS MLQA Baseline XLM-Rbase 74.9 75.6 65.0/47.1 Subword Sampling Data Aug. 75.3 75.8 64.7/46.7 XTUNER1 76.5 76.3 67.4/49.5 XTUNER2 75.8 76.3 66.7/49.0 Gaussian Noise Data Aug. 74.7 75.6 64.2/46.1 XTUNER1 76.3 75.7 66.7/48.9 XTUNER2 75.5 76.2 66.3/48.5 CodeSwitch Data Aug. 76.5 75.1 63.8/45.9 XTUNER1 77.6 75.8 67.6/49.7 XTUNER2 76.8 76.1 66.3/48.6 Machine Translation Data Aug. 78.8 67.8/50.1 XTUNER1 79.7 XTUNER2 78.9 76.6 68.7/51.1 Table 4: Comparison between different data augmentation strategies. “Data Aug.” uses data augmentation for conventional fine-tuning. “XTUNER1” denotes finetuning with only example consistency R1. “XTUNER2” denotes fine-tuning with only model consistency R2. are unavailable in these two datasets. To make a fair comparison in the translate-train-all setting, we augment the English training corpus with target language translations when fine-tuning with only example consistency R1. Otherwise, we only use the English training corpus in the first stage, as shown in Figure 1(a). Compared to XTUNE, the performance drop on two classification datasets under this setting is relatively small since R1 can be directly applied between translation-pairs in any languages. However, the performance is significantly degraded in three question answering datasets, where we can not align the predicted distributions between translation-pairs in R1. We use subword sampling as the data augmentation strategy in R1 for this situation. Fine-tuning with only model consistency R2 degrades the overall performance by 1.1 points. These results demonstrate that the two consistency regularization methods complement each other. BeModel Tatoeba BUCC XLM-Rbase (cross-lingual transfer) 74.2 78.2 XLM-Rbase (translate-train-all) 79.7 79.7 XTUNE (translate-train-all) 82.3 82.2 with only example consistency R1 82.0 82.1 with only model consistency R2 79.5 79.0 Table 5: Results of cross-lingual retrieval with the models fine-tuned on XNLI. sides, we observe that removing stopgrad degrades the overall performance by 0.5 points. Table 3 provides results of each language on the XNLI dataset. For the cross-lingual transfer setting, we utilize code-switch substitution as data augmentation for both example consistency R1 and model consistency R2. We utilize all the bilingual dictionaries, except for English to Swahili and English to Urdu, which MUSE does not provide. Results show that our method outperforms all baselines on each language, even on Swahili (+2.2 points) and Urdu (+5.4 points), indicating our method can be generalized to low-resource languages even without corresponding machine translation systems or bilingual dictionaries. For translate-train-all setting, we utilize machine translation as data augmentation for both example consistency R1 and model consistency R2. We improve the XLM-Rlarge baseline by +2.2 points on average, while we still have +0.9 points on average compared to FILTER. It is worth mentioning that we do not need corresponding English translations during inference. Complete results on other datasets are provided in the supplementary document. 4.3 Analysis It is better to employ data augmentation for consistency regularization than for conventional fine-tuning. As shown in Table 4, com3410 (a) cross-lingual transfer (b) translate-train-all (c) xTune Label neutral contradiction entailment Language de en fr zh Figure 3: t-SNE visualization of 100 examples in four languages from the XNLI development set (best viewed in color). We fine-tune the XLM-Rbase model on XNLI and use the hidden states of [CLS] symbol in the last layer. Examples with different labels are represented with different colors. Examples in different languages are represented with different markers. The red lines connect English examples and their translations in target languages. pared to employing data augmentation for conventional fine-tuning (Data Aug.), our regularization methods (XTUNER1, XTUNER2) consistently improve the model performance under all four data augmentation strategies. Since there is no labeled data on translations in POS and the issue of distribution alignment in example consistency R1, when machine translation is utilized as data augmentation, the results for Data Aug. and XTUNER1 in POS, as well as XTUNER1 in MLQA, are unavailable. We observe that Data Aug. can enhance the overall performance for coarse-grained tasks like XNLI, while our methods can further improve the results. However, Data Aug. even causes the performance to degrade for fine-grained tasks like MLQA and POS. In contrast, our proposed two consistency regularization methods improve the performance by a large margin (e.g., for MLQA under code-switch data augmentation, Data Aug. decreases baseline by 1.2 points, while XTUNER1 increases baseline by 2.6 points). We give detailed instructions on how to choose data augmentation strategies for XTUNE in the supplementary document. XTUNE improves cross-lingual retrieval. We fine-tune the models on XNLI with different settings and compare their performance on two crosslingual retrieval datasets. Following Chi et al. (2020) and Hu et al. (2020), we utilize representations averaged with hidden-states on the layer 8 of XLM-Rbase. As shown in Table 5, we observe significant improvement from the translatetrain-all baseline to fine-tuning with only example consistency R1, this suggests regularizing the taskspecific output of translation-pairs to be consistent also encourages the model to generate languageinvariant representations. XTUNE only slightly improves upon this setting, indicating R1 between translation-pairs is the most important factor to improve cross-lingual retrieval task. XTUNE improves decision boundaries as well as the ability to generate language-invariant representations. As shown in Figure 3, we present t-SNE visualization of examples from the XNLI development set under three different settings. We observe the model fine-tuned with XTUNE significantly improves the decision boundaries of different labels. Besides, for an English example and its translations in other languages, the model fine-tuned with XTUNE generates more similar representations compared to the two baseline models. This observation is also consistent with the cross-lingual retrieval results in Table 5. 5 Conclusion In this work, we present a cross-lingual fine-tuning framework XTUNE to make better use of data augmentation. We propose two consistency regularization methods that encourage the model to make consistent predictions for an example and its semantically equivalent data augmentation. We explore four types of cross-lingual data augmentation strategies. We show that both example and model consistency regularization considerably boost the performance compared to directly fine-tuning on data augmentations. Meanwhile, model consistency regularization enables semi-supervised training on the unlabeled target language translations. XTUNE combines the two regularization methods, and the experiments show that it can improve the performance by a large margin on the XTREME benchmark. 3411 Acknowledgments Wanxiang Che is the corresponding author. This work was supported by the National Key R&D Program of China via grant 2020AAA0106501 and the National Natural Science Foundation of China (NSFC) via grant 61976072 and 61772153. References Armen Aghajanyan, Akshat Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. 2020. Better fine-tuning by reducing representational collapse. CoRR, abs/2008.03156. Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4623–4637. Association for Computational Linguistics. Ben Athiwaratkun, Marc Finzi, Pavel Izmailov, and Andrew Gordon Wilson. 2019. There are many consistent explanations of unlabeled data: Why you should average. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, John C. Duchi, and Percy Liang. 2019. Unlabeled data improves adversarial robustness. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pages 11190–11201. Zewen Chi, Li Dong, Furu Wei, Nan Yang, Saksham Singhal, Wenhui Wang, Xia Song, Xian-Ling Mao, Heyan Huang, and Ming Zhou. 2020. InfoXLM: An information-theoretic framework for cross-lingual language model pre-training. CoRR, abs/2007.07834. Jonathan H. Clark, Jennimaria Palomaki, Vitaly Nikolaev, Eunsol Choi, Dan Garrette, Michael Collins, and Tom Kwiatkowski. 2020. Tydi QA: A benchmark for information-seeking question answering in typologically diverse languages. Trans. Assoc. Comput. Linguistics, 8:454–470. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020a. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8440–8451. Association for Computational Linguistics. Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pages 7057–7067. Alexis Conneau, Ruty Rinott, Guillaume Lample, Adina Williams, Samuel R. Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2475–2485. Association for Computational Linguistics. Alexis Conneau, Shijie Wu, Haoran Li, Luke Zettlemoyer, and Veselin Stoyanov. 2020b. Emerging cross-lingual structure in pretrained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 6022–6034. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics. Yuwei Fang, Shuohang Wang, Zhe Gan, Siqi Sun, and Jingjing Liu. 2020. FILTER: An enhanced fusion method for cross-lingual language understanding. CoRR, abs/2009.05166. Manaal Faruqui and Chris Dyer. 2014. Improving vector space word representations using multilingual correlation. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2014, April 26-30, 2014, Gothenburg, Sweden, pages 462–471. The Association for Computer Linguistics. Hao Fei, Meishan Zhang, and Donghong Ji. 2020. Cross-lingual semantic role labeling with highquality translated training corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7014–7026. Association for Computational Linguistics. Jiang Guo, Wanxiang Che, David Yarowsky, Haifeng Wang, and Ting Liu. 2015. Cross-lingual dependency parsing based on distributed representations. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 1234– 1244. The Association for Computer Linguistics. 3412 Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalisation. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 1318 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 4411– 4421. PMLR. Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. 2017. Learning discrete representations via information maximizing self-augmented training. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 1558–1567. PMLR. Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Tuo Zhao. 2020. SMART: Robust and efficient fine-tuning for pretrained natural language models through principled regularized optimization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 2177–2190. Association for Computational Linguistics. Taku Kudo. 2018. Subword regularization: Improving neural network translation models with multiple subword candidates. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 1520, 2018, Volume 1: Long Papers, pages 66–75. Association for Computational Linguistics. Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018: System Demonstrations, Brussels, Belgium, October 31 - November 4, 2018, pages 66–71. Association for Computational Linguistics. Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc’Aurelio Ranzato. 2018. Unsupervised machine translation using monolingual corpora only. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. Patrick S. H. Lewis, Barlas Oguz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7315–7330. Association for Computational Linguistics. Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. CoRR, abs/2004.08994. Fuli Luo, W. Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and L. Si. 2020. VECO: Variable encoder-decoder pre-training for cross-lingual understanding and generation. ArXiv, abs/2010.16046. Tomas Mikolov, Quoc V. Le, and Ilya Sutskever. 2013. Exploiting similarities among languages for machine translation. CoRR, abs/1309.4168. Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. 2019. Virtual adversarial training: A regularization method for supervised and semisupervised learning. IEEE Trans. Pattern Anal. Mach. Intell., 41(8):1979–1993. Joakim Nivre, Rogier Blokland, Niko Partanen, and Michael Rießler. 2018. Universal dependencies 2.2. Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Crosslingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 1946–1958. Association for Computational Linguistics. Jason Phang, Phu Mon Htut, Yada Pruksachatkun, Haokun Liu, Clara Vania, Katharina Kann, Iacer Calixto, and Samuel R. Bowman. 2020. English intermediate-task training improves zero-shot crosslingual transfer too. CoRR, abs/2005.13013. Ivan Provilkov, Dmitrii Emelianenko, and Elena Voita. 2020. BPE-Dropout: Simple and effective subword regularization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 1882–1892. Association for Computational Linguistics. Libo Qin, Minheng Ni, Yue Zhang, and Wanxiang Che. 2020. CoSDA-ML: Multi-lingual code-switching data augmentation for zero-shot cross-lingual NLP. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 3853–3860. ijcai.org. Jasdeep Singh, Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. 2019. XLDA: cross-lingual data augmentation for natural language inference and question answering. CoRR, abs/1905.11471. Antti Tarvainen and Harri Valpola. 2017. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net. Yuxuan Wang, Wanxiang Che, Jiang Guo, Yijia Liu, and Ting Liu. 2019. Cross-lingual BERT transformation for zero-shot dependency parsing. In 3413 Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 5720– 5726. Association for Computational Linguistics. Qizhe Xie, Zihang Dai, Eduard H. Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Ruochen Xu, Yiming Yang, Naoki Otani, and Yuexin Wu. 2018. Unsupervised cross-lingual transfer of word embedding spaces. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 November 4, 2018, pages 2465–2474. Association for Computational Linguistics. Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3685– 3690. Association for Computational Linguistics. Mang Ye, Xu Zhang, Pong C. Yuen, and Shih-Fu Chang. 2019. Unsupervised embedding learning via invariant and spreading instance feature. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 1620, 2019, pages 6210–6219. Computer Vision Foundation / IEEE. Meishan Zhang, Yue Zhang, and Guohong Fu. 2019. Cross-lingual dependency parsing using code-mixed treebank. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 997– 1006. Association for Computational Linguistics. Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. 2016. Improving the robustness of deep neural networks via stability training. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 4480–4488. IEEE Computer Society. Chen Zhu, Yu Cheng, Zhe Gan, Siqi Sun, Tom Goldstein, and Jingjing Liu. 2020. FreeLB: Enhanced adversarial training for natural language understanding. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Appendix A Statistics of XTREME Datasets Task Dataset |Train| |Lang| Classification XNLI 392K 15 PAWS-X 49.4K 7 Structured POS 21K 33 Prediction NER 20K 40 Question Answering XQuAD 87K 11 MLQA 87K 7 TyDiQA 3.7K 9 Table 6: Statistics for the datasets in the XTREME benchmark. we report the number of training examples (|Train|), and the number of languages (|Lang|). B Hyper-Parameters For XNLI, PAWS-X, POS and NER, we fine-tune 10 epochs. For XQuAD and MLQA, we fine-tune 4 epochs. For TyDiQA-GoldP, we fine-tune 20 epochs and 10 epochs for base and large model, respectively. We select λ1 in [1.0, 2.0, 5.0], λ2 in [0.3, 0.5, 1.0, 2.0, 5.0]. For learning rate, we select in [5e-6, 7e-6, 1e-5, 1.5e-5] for large models, [7e-6, 1e-5, 2e-5, 3e-5] for base models. We use batch size 32 for all datasets and 10% of total training steps for warmup with a linear learning rate schedule. Our experiments are conducted with a single 32GB Nvidia V100 GPU, and we use gradient accumulation for large-size models. The other hyper-parameters for the two-stage XTUNE training are shown in Table 7 and Table 8. C Results for Each Dataset and Language We provide detailed results for each dataset and language below. We compare our method against XLM-Rlarge for cross-lingual transfer setting, FILTER (Fang et al., 2020) for translate-train-all setting. D How to Select Data Augmentation Strategies in XTUNE We give instructions on selecting a proper data augmentation strategy depending on the corresponding task. 3414 Variable XNLI PAWS-X POS NER XQuAD MLQA TyDiQA Stage 1 A∗ CS CS SS SS CS CS SS Stage 2 A CS CS SS SS SS SS SS A′ CS CS SS SS SS SS SS Hyper-parameters λ1 5.0 5.0 5.0 5.0 5.0 5.0 5.0 λ2 5.0 2.0 0.3 5.0 5.0 5.0 5.0 Table 7: The best hyper-parameters used for XTUNE under the cross-lingual transfer setting. “SS”, “CS”, “MT” denote the data augmentation methods: subword sampling, code-switch substitution, and machine translation, respectively. Variable XNLI PAWS-X POS NER XQuAD MLQA TyDiQA Stage 1 A∗ MT MT SS SS CS CS SS Stage 2 A MT MT MT MT MT MT MT A′ MT MT SS SS SS SS SS Hyper-parameters λ1 5.0 5.0 5.0 5.0 5.0 5.0 5.0 λ2 1.0 1.0 0.3 1.0 0.1 0.5 0.3 Table 8: The best hyper-parameters used for XTUNE under the translate-train-all setting. “SS”, “CS”, “MT” denote the data augmentation methods subword sampling, code-switch substitution, and machine translation, respectively. Method Model XNLI POS MLQA Avg. XLM-Rbase 10.6 20.8 20.3 17.2 Subword Sampling Data Aug. 10.5 20.5 20.2 17.1 XTUNER1 10.2 20.2 19.6 16.7 XTUNER2 10.6 20.1 19.8 16.8 Gaussian Noise Data Aug. 10.8 20.6 19.8 17.1 XTUNER1 10.5 20.7 19.8 17.0 XTUNER2 10.8 20.2 19.7 16.9 CodeSwitch Data Aug. 9.2 21.1 20.5 16.9 XTUNER1 9.1 20.7 19.4 16.4 XTUNER2 8.8 20.2 20.0 16.3 Machine Translation Data Aug. 7.2 17.9 XTUNER1 6.9 XTUNER2 7.2 19.6 17.1 14.6 Table 9: Cross-lingual transfer gap, i.e., averaged performance drop between English and other languages in zero-shot transfer. A smaller gap indicates better transferability. For MLQA, we report the average of F1-scores and exact match scores. D.1 Classification The two distribution in example consistency R1 can always be aligned. Therefore, we recommend using machine translation as data augmentation if the machine translation systems are available. Otherwise, the priority of our data augmentation strategies is code-switch substitution, subword sampling and Gaussian noise. D.2 Span Extraction The two distribution in example consistency R1 can not be aligned in translation-pairs. Therefore, it is impossible to use machine translation as data augmentation in example consistency R1. We prefer to use code-switch when applying example consistency R1 individually. However, when the training corpus is augmented with translations, since the bilingual dictionaries between arbitrary language pairs may not be available, we recommend using subword sampling in example consistency R1. D.3 Sequence Labeling Similar to span extraction, the two distribution in example consistency R1 can not be aligned in translation-pairs. Therefore, we do not use machine translation in example consistency R1. Unlike classification and span extraction, sequence labeling requires finer-grained information and is more sensitive to noise. We found code-switch is worse than subword sampling as data augmentation in both example consistency R1 and model consistency R2, it will even degrade performance for certain hyperparameters. Thus we recommend using subword sampling in example consistency R1, and use machine translation to augment the English training corpus if machine translation systems are available, otherwise subword sampling. 3415 E Cross-Lingual Transfer Gap As shown in Table 9, the cross-lingual transfer gap can be reduced under all four data augmentation strategies. Meanwhile, we observe machine translation and code-switch substitution achieve a smaller cross-lingual transfer gap than the other two data augmentation methods. This suggests the data augmentation methods with cross-lingual knowledge have a greater improvement in crosslingual transferability. Although code-switch significantly reduces the transfer gap on XNLI, the improvement is relatively small on POS and MLQA under the cross-lingual transfer setting, indicating the noisy code-switch substitution will harm the cross-lingual transferability on finer-grained tasks. 3416 Model en de es fr ja ko zh Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 94.7 89.7 90.1 90.4 78.7 79.0 82.3 86.4 XTUNE 96.0 92.5 92.2 92.7 84.9 84.2 86.6 89.8 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 95.9 92.8 93.0 93.7 87.4 87.6 89.6 91.5 XTUNE 96.1 92.6 93.1 93.9 87.8 89.0 88.8 91.6 Table 10: PAWSX results (accuracy scores) for each language. Model en ar de el es hi ru th tr vi zh Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 86.5/75.7 68.6/49.0 80.4/63.4 79.8/61.7 82.0/63.9 76.7/59.7 80.1/64.3 74.2/62.8 75.9/59.3 79.1/59.0 59.3/50.0 76.6/60.8 XTUNE 88.9/78.6 77.1/60.0 83.1/67.2 82.6/66.0 83.0/65.1 77.8/61.8 80.8/64.8 73.5/62.1 77.6/62.0 81.8/62.5 67.7/58.4 79.4/64.4 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 86.4/74.6 79.5/60.7 83.2/67.0 83.0/64.6 85.0/67.9 83.1/66.6 82.8/67.4 79.6/73.2 80.4/64.4 83.8/64.7 79.9/77.0 82.4/68.0 XTUNE 88.8/78.1 79.7/63.9 83.7/68.2 83.0/65.7 84.7/68.3 80.7/64.9 82.2/66.6 81.9/76.1 79.3/65.0 82.7/64.5 81.3/78.0 82.5/69.0 Table 11: XQuAD results (F1/EM scores) for each language. Model en ar de es hi vi zh Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 83.5/70.6 66.6/47.1 70.1/54.9 74.1/56.6 70.6/53.1 74.0/52.9 62.1/37.0 71.6/53.2 XTUNE 85.2/72.6 67.9/47.7 72.2/56.8 75.5/57.9 73.2/55.1 75.9/54.7 71.1/48.6 74.4/56.2 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 84.0/70.8 72.1/51.1 74.8/60.0 78.1/60.1 76.0/57.6 78.1/57.5 70.5/47.0 76.2/57.7 XTUNE 85.3/72.9 69.7/50.1 72.3/57.3 76.3/58.8 74.0/56.0 76.5/55.9 70.8/48.3 75.0/57.1 Table 12: MLQA results (F1/EM scores) for each language. Model en ar bn fi id ko ru sw te Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 71.5/56.8 67.6/40.4 64.0/47.8 70.5/53.2 77.4/61.9 31.9/10.9 67.0/42.1 66.1/48.1 70.1/43.6 65.1/45.0 XTUNE 75.3/63.6 77.4/60.3 72.4/58.4 75.5/60.2 81.5/68.5 68.6/58.3 71.1/48.8 73.3/56.7 78.4/60.1 74.8/59.4 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 72.4/59.1 72.8/50.8 70.5/56.6 73.3/57.2 76.8/59.8 33.1/12.3 68.9/46.6 77.4/65.7 69.9/50.4 68.3/50.9 XTUNE 73.8/61.6 77.8/60.2 73.5/61.1 77.0/62.2 80.8/68.1 66.9/56.5 72.1/51.9 77.9/65.3 77.6/60.7 75.3/60.8 Table 13: TyDiQA-GolP results (F1/EM scores) for each language. 3417 Model af ar bg de el en es et eu fa fi fr he hi hu id it Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 89.8 67.5 88.1 88.5 86.3 96.1 88.3 86.5 72.5 70.6 85.8 87.2 68.3 76.4 82.6 72.4 89.4 XTUNE 90.4 72.8 89.0 89.4 87.0 96.1 88.8 88.1 73.1 74.7 87.2 89.5 83.5 77.7 83.6 73.2 90.5 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 88.7 66.1 88.5 89.2 88.3 96.0 89.1 86.3 78.0 70.8 86.1 88.9 64.9 76.7 82.6 72.6 89.8 XTUNE 90.7 74.2 89.9 90.2 87.4 96.1 90.5 88.4 75.9 74.2 87.9 90.2 85.9 79.3 83.2 73.3 91.0 Model ja kk ko mr nl pt ru ta te th tl tr ur vi yo zh Avg. Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 15.9 78.1 53.9 80.8 89.5 87.6 89.5 65.2 86.6 47.2 92.2 76.3 70.3 56.8 24.6 25.7 73.8 XTUNE 62.7 78.3 55.7 82.4 90.2 88.5 90.5 63.6 88.3 61.8 94.5 76.9 72.0 57.8 24.4 69.4 78.5 Fine-tune multilingual model on all target language target language training sets (translate-train-all) FILTER (Fang et al., 2020) 40.4 80.4 53.3 86.4 89.4 88.3 90.5 65.3 87.3 57.2 94.1 77.0 70.9 58.0 43.1 53.1 76.9 XTUNE 65.3 79.8 56.0 85.5 89.7 89.3 90.8 65.7 85.5 61.4 93.8 78.3 74.0 57.5 27.9 68.8 79.3 Table 14: POS results (accuracy) for each language. Model en af ar bg bn de el es et eu fa fi fr he hi hu id it ja jv Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 84.7 78.9 53.0 81.4 78.8 78.8 79.5 79.6 79.1 60.9 61.9 79.2 80.5 56.8 73.0 79.8 53.0 81.3 23.2 62.5 XTUNE 85.0 80.4 59.1 84.8 79.1 80.5 82.0 78.1 81.5 64.5 65.9 82.2 81.9 62.0 75.0 82.8 55.8 83.1 30.5 65.9 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 83.5 80.4 60.7 83.5 78.4 80.4 80.7 74.0 81.0 66.9 71.3 80.2 79.9 57.4 74.3 82.2 54.0 81.9 24.3 63.5 XTUNE 84.4 81.7 59.7 85.3 80.8 80.9 82.0 74.1 83.4 69.9 63.6 82.5 80.6 64.0 76.3 83.8 57.9 83.3 26.5 69.8 Model ka kk ko ml mr ms my nl pt ru sw ta te th tl tr ur vi yo zh Cross-lingual-transfer (models are fine-tuned on English training data without translation available) XLM-Rlarge 71.6 56.2 60.0 67.8 68.1 57.1 54.3 84.0 81.9 69.1 70.5 59.5 55.8 1.3 73.2 76.1 56.4 79.4 33.6 33.1 XTUNE 76.7 57.5 65.9 68.1 73.3 67.2 63.7 85.3 84.0 73.6 70.1 66.1 60.1 1.8 76.9 83.6 76.0 80.3 44.4 38.7 Translate-train-all (translation-based augmentation is available for English training data) FILTER (Fang et al., 2020) 71.0 51.1 63.8 70.2 69.8 69.3 59.0 84.6 82.1 71.1 70.6 64.3 58.7 2.4 74.4 83.0 73.4 75.8 42.9 35.4 XTUNE 76.3 56.9 67.1 72.6 71.5 72.5 66.7 85.8 82.1 75.2 72.4 66.0 61.8 1.1 77.5 83.7 75.6 80.8 44.9 36.5 Table 15: NER results (F1 scores) for each language.
|
2021
|
264
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3418–3430 August 1–6, 2021. ©2021 Association for Computational Linguistics 3418 Improving Pretrained Cross-Lingual Language Models via Self-Labeled Word Alignment Zewen Chi†∗, Li Dong‡, Bo Zheng‡∗, Shaohan Huang‡ Xian-Ling Mao†, Heyan Huang†, Furu Wei‡ †Beijing Institute of Technology ‡Microsoft Research {czw,maoxl,hhy63}@bit.edu.cn {lidong1,v-zhebo,shaohanh,fuwei}@microsoft.com Abstract The cross-lingual language models are typically pretrained with masked language modeling on multilingual text or parallel sentences. In this paper, we introduce denoising word alignment as a new cross-lingual pre-training task. Specifically, the model first self-labels word alignments for parallel sentences. Then we randomly mask tokens in a bitext pair. Given a masked token, the model uses a pointer network to predict the aligned token in the other language. We alternately perform the above two steps in an expectationmaximization manner. Experimental results show that our method improves cross-lingual transferability on various datasets, especially on the token-level tasks, such as question answering, and structured prediction. Moreover, the model can serve as a pretrained word aligner, which achieves reasonably low error rates on the alignment benchmarks. The code and pretrained parameters are available at github.com/CZWin32768/XLM-Align. 1 Introduction Despite the current advances in NLP, most applications and resources are still English-centric, making non-English users hard to access. Therefore, it is essential to build cross-lingual transferable models that can learn from the training data in highresource languages and generalize on low-resource languages. Recently, pretrained cross-lingual language models have shown their effectiveness for cross-lingual transfer. By pre-training on monolingual text and parallel sentences, the models provide significant improvements on a wide range of crosslingual end tasks (Conneau and Lample, 2019; Conneau et al., 2020; Liu et al., 2020; Chi et al., 2021b). Cross-lingual language model pre-training is typically achieved by learning various pretext tasks on ∗Contribution during internship at Microsoft Research. monolingual and parallel corpora. By simply learning masked language modeling (MLM; Devlin et al. 2019) on monolingual text of multiple languages, the models surprisingly achieve competitive results on cross-lingual tasks (Wu and Dredze, 2019; K et al., 2020). Besides, several pretext tasks are proposed to utilize parallel corpora to learn better sentence-level cross-lingual representations (Conneau and Lample, 2019; Chi et al., 2021b; Hu et al., 2020a). For example, the translation language modeling (TLM; Conneau and Lample 2019) task performs MLM on the concatenated parallel sentences, which implicitly enhances cross-lingual transferability. However, most pretext tasks either learn alignment at the sentence level or implicitly encourage cross-lingual alignment, leaving explicit fine-grained alignment task not fully explored. In this paper, we introduce a new cross-lingual pre-training task, named as denoising word alignment. Rather than relying on external word aligners trained on parallel corpora (Cao et al., 2020; Zhao et al., 2020; Wu and Dredze, 2020), we utilize self-labeled alignments in our task. During pretraining, we alternately self-label word alignments and conduct the denoising word alignment task in an expectation-maximization manner. Specifically, the model first self-labels word alignments for a translation pair. Then we randomly mask tokens in the bitext sentence, which is used as the perturbed input for denosing word alignment. For each masked token, the model learns a pointer network to predict the self-labeled alignments in the other language. We repeat the above two steps to iteratively boost the bitext alignment knowledge for cross-lingual pre-training. We conduct extensive experiments on a wide range of cross-lingual understanding tasks. Experimental results show that our model outperforms the baseline models on various datasets, particularly on the token-level tasks such as question answer3419 ing and structured prediction. Moreover, our model can also serve as a multilingual word aligner, which achieves reasonable low error rates on the bitext alignment benchmarks. Our contributions are summarized as follows: • We present a cross-lingual pre-training paradigm that alternately self-labels and predicts word alignments. • We introduce a pre-training task, denoising word alignment, which predicts word alignments from perturbed translation pairs. • We propose a word alignment algorithm that formulates the word alignment problem as optimal transport. • We demonstrate that our explicit alignment objective is effective for cross-lingual transfer. 2 Related Work Cross-lingual LM pre-training Pretrained with masked language modeling (MLM; Devlin et al. 2019) on monolingual text, multilingual BERT (mBERT; Devlin et al. 2019) and XLM-R (Conneau et al., 2020) produce promising results on cross-lingual transfer benchmarks (Hu et al., 2020b). mT5 (Xue et al., 2020) learns a multilingual version of T5 (Raffel et al., 2020) with text-totext tasks. In addition to monolingual text, several methods utilize parallel corpora to improve crosslingual transferability. XLM (Conneau and Lample, 2019) presents the translation language modeling (TLM) task that performs MLM on concatenated translation pairs. ALM (Yang et al., 2020) introduces code-switched sequences into cross-lingual LM pre-training. Unicoder (Huang et al., 2019) employs three cross-lingual tasks to learn mappings among languages. From an information-theoretic perspective, InfoXLM (Chi et al., 2021b) proposes the cross-lingual contrastive learning task to align sentence-level representations. Additionally, AMBER (Hu et al., 2020a) introduces an alignment objective that minimizes the distance between the forward and backward attention matrices. More recently, Ernie-M (Ouyang et al., 2020) presents the back-translation masked language modeling task that generates pseudo parallel sentence pairs for learning TLM, which provides better utilization of monolingual corpus. VECO (Luo et al., 2020) pretrains a unified cross-lingual language model for both NLU and NLG. mT6 (Chi et al., 2021a) improves the multilingual text-to-text transformer with translation pairs. Notably, Word-aligned BERT models (Cao et al., 2020; Zhao et al., 2020) finetune mBERT by an explicit alignment objective that minimizes the distance between aligned tokens. Wu and Dredze (2020) exploit contrastive learning to improve the explicit alignment objectives. However, Wu and Dredze (2020) show that these explicit alignment objectives do not improve cross-lingual representations under a more extensive evaluation. Moreover, these models are restricted to stay close to their original pretrained values, which is not applicable for large-scale pre-training. On the contrary, we demonstrate that employing our explicit alignment objective in large-scale pre-training can provide consistent improvements over baseline models. Word alignment The IBM models (Brown et al., 1993) are statistical models for modeling the translation process that can extract word alignments between sentence pairs. A large number of word alignment models are based on the IBM models (Och and Ney, 2003; Mermer and Sarac¸lar, 2011; Dyer et al., 2013; ¨Ostling and Tiedemann, 2016). Recent studies have shown that word alignments can be extracted from neural machine translation models (Ghader and Monz, 2017; Koehn and Knowles, 2017; Li et al., 2019) or from pretrained cross-lingual LMs (Jalili Sabet et al., 2020; Nagata et al., 2020). 3 Method Figure 1 illustrates an overview of our method for pre-training our cross-lingual LM, which is called XLM-ALIGN. XLM-ALIGN is pretrained in an expectation-maximization manner with two alternating steps, which are word alignment selflabeling and denoising word alignment. We first formulate word alignment as an optimal transport problem, and self-label word alignments of the input translation pair on-the-fly. Then, we update the model parameters with the denoising word alignment task, where the model uses a pointer network (Vinyals et al., 2015) to predict the aligned tokens from the perturbed translation pair. 3.1 Word Alignment Self-Labeling The goal of word alignment self-labeling is to estimate the word alignments of the input translation pair on-the-fly, given the current XLM-ALIGN model. Given a source sentence 3420 (a) Word alignment self-labeling XLM-Align Encoder 你好 世界 。 Hello world . 你好<-> Hello 世界<-> world 。<-> . Self-Labeled Word Alignments XLM-Align Encoder Alignment Probability 你好 [M] 。 [M] world . 你好 世界 。 Hello world . Alignment as Optimal Transport Translation Pair Noisy Translation Pair (Random Masks) Pointer Network query keys DWA Loss (b) Denoising word alignment Figure 1: An overview of our method. XLM-ALIGN is pretrained in an expectation-maximization manner with two alternating steps. (a) Word alignment self-labeling: we formulate word alignment as an optimal transport problem, and self-labels word alignments of the input translation pair on-the-fly; (b) Denoising word alignment: we update the model parameters with the denoising word alignment task, where the model uses a pointer network to predict the aligned tokens from the perturbed translation pair. S = s1 . . . si . . . sn and a target sentence T = t1 . . . tj . . . tm, we model the word alignment between S and T as a doubly stochastic matrix A ∈Rn×m + such that the rows and the columns all sum to 1, where Aij stands for the probability of the alignment between si and tj. The rows and the columns of A represent probability distributions of the forward alignment and the backward alignment, respectively. To measure the similarity between two tokens from S and T , we define a metric function fsim by using cross-lingual representations produced by XLM-ALIGN: fsim(si, tj) = −log max(ϵ, h⊤ i hj) (1) where ϵ is a constant to avoid negative values in the log function, and hi is the hidden vector of the i-th token by encoding the concatenated sequence of S and T with XLM-ALIGN. Empirically, the metric function produces a high similarity score if the two input tokens are semantically similar. The word alignment problem is formulated as finding A that maximizes the sentence similarity between S and T : max A n X i=1 m X j=1 Aijfsim(si, tj) (2) We can find that Eq. (2) is identical to the regularized optimal transport problem (Peyr´e et al., 2019), if we add an entropic regularization to A: max A n X i=1 m X j=1 Aijfsim(si, tj) −µAij log Aij (3) Eq. (3) has a unique solution A∗such that A∗= diag(u)Kdiag(v) (4) Kij = efsim(si,tj)/µ (5) where u ∈Rn +, v ∈Rm +, K ∈Rn×m + . According to Sinkhorn’s algorithm (Peyr´e et al., 2019), the variables u and v can be calculated by the following iterations: ut+1 = 1n Kvt , vt+1 = 1m K⊤ut+1 (6) where vt can be initialized by vt=0 = 1m. With the solved stochastic matrix A∗, we can produce the forward word alignments −→ A by applying argmax over rows: −→ A = {(i, j) | j = arg max k A∗ ik} (7) Similarly, the backward word alignments ←− A can be computed by applying argmax over columns. To obtain high-precision alignment labels, we adopt an iterative alignment filtering operation. We initialize 3421 the alignment labels A as ∅. In each iteration, we follow the procedure of Itermax (Jalili Sabet et al., 2020) that first computes −→ A and ←− A by Eq. (7). Then, the alignment labels are updated by: A ←A ∪(−→ A ∩←− A) (8) Finally, A∗is updated by: A∗ ij ← 0, (i, j) ∈A αA∗ ij, ∃k (i, k) ∈A ∨(k, j) ∈A A∗ ij, others (9) where α is a discount factor. After several iterations, we obtain the final self-labeled word alignments A. 3.2 Denoising Word Alignment After self-labeling word alignments, we update the model parameters with the denoising word alignment (DWA) task. The goal of DWA is to predict the word alignments from the perturbed version of the input translation pair. Consider the perturbed version of the input translation pair (S∗, T ∗) constructed by randomly replacing the tokens with masks. We first encode the translation pair into hidden vectors h∗with the XLM-ALIGN encoder: h∗ 1 . . . h∗ n+m = encoder([S∗, T ∗]) (10) where [S∗, T ∗] is the concatenated sequence of S∗ and T ∗with the length of n + m. Then, we build a pointer network upon the XLM-ALIGN encoder that predicts the word alignments. Specifically, for the i-th source token, we use h∗ i as the query vector and h∗ n+1, . . . , h∗ n+m as the key vectors. Given the query and key vectors, the forward alignment probability ai is computed by the scaled dot-product attention (Vaswani et al., 2017): ai = softmax(q⊤ i K √dh ) (11) qi = linear(h∗ i ) (12) K = linear([h∗ n+1 . . . h∗ n+m]) (13) where dh is the dimension of the hidden vectors. Similarly, the backward alignment probability can be computed by above equations if we use target tokens as the query vectors and h∗ 1 . . . h∗ n as key vectors. Notice that we only consider the self-labeled and masked positions as queries. Formally, we use the following query positions in the pointer network: P = {i|(i, ·) ∈A ∨(·, i) ∈A} ∩M (14) where M is the set of masked positions. The training objective is to minimize the cross-entropy between the alignment probabilities and the selflabeled word alignments: LDWA = X i∈P CE(ai, A(i)) (15) where CE(·, ·) stands for the cross-entropy loss, and A(i) is the self-labeled aligned position of the i-th token. Algorithm 1 Pre-training XLM-ALIGN Input: Multilingual corpus Dm, parallel corpus Dp, learning rate τ Output: XLM-ALIGN parameters θ 1: Initialize θ with cold-start pre-training 2: while not converged do 3: X ∼Dm, (S, T ) ∼Dp 4: A ←fself-labeling(S, T ; θ) 5: g ←∇θLMLM(X) + ∇θLTLM(S, T ) + ∇θLDWA(S, T , A) 6: θ ←θ −τg 3.3 Pre-training XLM-ALIGN We illustrate the pre-training procedure of XLMALIGN in Algorithm 1. In addition to DWA, we also include MLM and TLM for pre-training XLMALIGN, which implicitly encourage the crosslingual alignment. The overall loss function is defined as: LMLM(X) + LTLM(S, T ) + LDWA(S, T , A) In each iteration, we first sample monolingual text X, and parallel text (S, T ). Then, we self-label word alignments and update the model parameters by learning pretext tasks. Notice that the model parameters are initialized by a cold-start pre-training to avoid producing low-quality alignment labels. The cold-start pre-training can be accomplished by using a pretrained LM as the model initialization. 4 Experiments 4.1 Pre-training Following previous cross-lingual pretrained models (Conneau and Lample, 2019; Conneau et al., 3422 Model Structured Prediction Question Answering Sentence Classification Avg POS NER XQuAD MLQA TyDiQA XNLI PAWS-X Metrics F1 F1 F1 / EM F1 / EM F1 / EM Acc. Acc. MBERT* 70.3 62.2 64.5 / 49.4 61.4 / 44.2 59.7 / 43.9 65.4 81.9 63.1 XLM* 70.1 61.2 59.8 / 44.3 48.5 / 32.6 43.6 / 29.1 69.1 80.9 58.6 MT5base 56.6 67.0 / 49.0 64.6 / 45.0 58.1 / 42.8 75.4 87.4 XLM-Rbase 75.6 61.8 71.9 / 56.4 65.1 / 47.2 55.4 / 38.3 75.0 84.9 66.4 XLM-ALIGN 76.0 63.7 74.7 / 59.0 68.1 / 49.8 62.1 / 44.8 76.2 86.8 68.9 Table 1: Evaluation results on XTREME structured prediction, question answering, and sentence classification tasks. We adopt the cross-lingual transfer setting, where models are only fine-tuned on the English training data but evaluated on all target languages. Results with “*” are taken from (Hu et al., 2020b). Results of XLM-ALIGN and XLM-Rbase are averaged over five runs. 2020; Chi et al., 2021b), we use raw sentences from the Wikipedia dump and CCNet (Wenzek et al., 2019) for MLM, including 94 languages. For TLM and DWA, we use parallel corpora from MultiUN (Ziemski et al., 2016), IIT Bombay (Kunchukuttan et al., 2018), OPUS (Tiedemann, 2012), and WikiMatrix (Schwenk et al., 2019), including 14 English-centric language pairs. We pretrain a Transformer with 12 layers and the hidden size of 768, where the parameters are initialized with XLM-R (Conneau et al., 2020). The model is optimized with the Adam optimizer (Kingma and Ba, 2015) for 150K steps with batch size of 2, 048. Notice that TLM and DWA share the same forward procedure for encoding the perturbed sentence pair. The pre-training of XLMALIGN takes about six days with two Nvidia DGX2 stations. More details of the training data and the hyperparameters are in supplementary document. 4.2 XTREME Benchmark XTREME is a multilingual benchmark for evaluating cross-lingual generalization. We evaluate our model on 7 cross-lingual downstream tasks included by XTREME, which can be grouped into 3 categories: (1) Structured prediction: part-ofspeech tagging on the Universal Dependencies v2.5 (Zeman et al., 2019), and named entity recognition on the WikiAnn (Pan et al., 2017; Rahimi et al., 2019) dataset; (2) Question answering: crosslingual question answering on MLQA (Lewis et al., 2020) and XQuAD (Artetxe et al., 2020), and gold passage of typologically diverse question answering (TyDiQA-GoldP; Clark et al. 2020); (3) Sentence classification: cross-lingual natural language inference (XNLI; Conneau et al. 2018), and crosslingual paraphrase adversaries from word scrambling (PAWS-X; Yang et al. 2019). Baselines We use the following pretrained crosslingual LMs as baselines. (1) Multilingual BERT (MBERT; Devlin et al. 2019) is pretrained with masked language modeling (MLM) and next sentence prediction on Wikipedia of 104 languages; (2) XLM (Conneau and Lample, 2019) is jointly pretrained with MLM on 100 languages and translation language modeling (TLM) on 14 language pairs; (3) MT5 (Xue et al., 2020) is the multilingual version of T5 pretrained with text-to-text tasks; (4) XLM-R (Conneau et al., 2020) is pretrained with MLM on large-scale CC-100 dataset with long training steps. Fine-tuning Following Hu et al. (2020b), we adopt the zero-shot transfer setting for evaluation, where the models are only fine-tuned on English training data but evaluated on all target languages. Besides, we only use one model for evaluation on all target languages, rather than selecting different models for each language. The detailed fine-tuning hyperparameters can be found in supplementary document. Results In Table 1, we present the evaluation results on XTREME structured prediction, question answering, and sentence classification tasks. It can be observed that our XLM-ALIGN obtains the best average score over all the baseline models, improving the previous score from 66.4 to 68.9. It demonstrates that our model learns more transferable representations for the cross-lingual tasks, which is beneficial for building more accessible multilingual NLP applications. It is worth mentioning that our method brings noticeable improvements on the question answering and the structured prediction tasks. Compared with XLM-Rbase, XLM-ALIGN provides 6.7% and 1.9% F1 improvements on TyDiQA and NER. The improvements show that the 3423 Alignment Method Pretrained Alignment Error Rate ↓ Avg Model en-de en-fr en-hi en-ro fast align (Dyer et al., 2013) 32.14 19.46 59.90 SimAlign - Argmax (Jalili Sabet et al., 2020) XLM-R 19. 7. 39. 29. 24. SimAlign - Itermax (Jalili Sabet et al., 2020) XLM-R 20. 9. 39. 28. 24. SimAlign - Itermax (reimplementation) XLM-R 20.15 10.05 38.72 27.41 24.08 Ours - Optimal Transport (Section 3.1) XLM-R 17.74 7.54 37.79 27.49 22.64 SimAlign (reimplementation) XLM-ALIGN 18.93 10.33 33.84 27.09 22.55 Ours - Optimal Transport (Section 3.1) XLM-ALIGN 16.63 6.61 33.98 26.97 21.05 Table 2: Evaluation results for word alignment on four English-centric language pairs. We report the alignment error rate scores (lower is better). For both SimAlign (Jalili Sabet et al., 2020) and our optimal-transport alignment method, we use the hidden vectors from the 8-th layer produced by XLM-Rbase or XLM-ALIGN. “(reimplementation)” is our reimplementation of SimAlign-Itermax. 3 4 5 6 7 8 9 10 11 12 Layer 15 20 25 30 35 Alignment Error Rate XLM-R XLM-Align Figure 2: Evaluation results on word alignment across different layers. We illustrate the averaged AER scores on the test sets of four language pairs. The results of the first two layers are not included due to the high AER. pretrained XLM-ALIGN benefits from the explicit word alignment objective, particularly on the structured prediction and question answering tasks that require token-level cross-lingual transfer. In terms of sentence classification tasks, XLM-ALIGN also consistently outperforms XLM-Rbase. 4.3 Word Alignment Word alignment is the task of finding corresponding word pairs in a parallel sentence. We conduct evaluations with golden alignments of four language pairs from EuroParl1, WPT20032, and WPT20053, containing 1,244 annotated sentence pairs in total. We use alignment error rate (AER; Och and Ney 1www-i6.informatik.rwth-aachen.de/ goldAlignment/ 2web.eecs.umich.edu/˜mihalcea/wpt/ 3web.eecs.umich.edu/˜mihalcea/wpt05/ 2003) as the evaluation metrics. Results We first explore whether our word alignment self-labeling method is effective for generating high-quality alignment labels. Thus, we compare our method with (1) fast align (Dyer et al., 2013), a widely-used implementation of IBM Model 2 (Och and Ney, 2003); (2) SimAlign (Jalili Sabet et al., 2020), state-of-theart unsupervised word alignment method. For a fair comparison, we use the same pretrained LM and hidden layer as in SimAlign to produce sentence representations. In specific, we take the hidden vectors from the 8-th layer of XLM-Rbase or XLMALIGN, and obtain the alignments following the procedure as described in Section 3.1. Since the produced alignments are subword-level, we convert the alignments into word-level by the following rule that “if two subwords are aligned, the words they belong to are also aligned”. As shown in Table 2, we report the AER scores on the four language pairs. It can be observed that our optimal-transport method outperforms fast align and SimAlign, demonstrating that our method can produce high-quality alignment labels, which is helpful for the DWA task. Moreover, our method consistently outperforms SimAlign when using hidden vectors from both XLM-Rbase and XLM-ALIGN. Then, we compare our XLM-ALIGN with XLMRbase on the word alignment task. Empirically, a lower AER indicates that the model learns better cross-lingual representations. From Table 2, XLM-ALIGN obtains the best AER results over all the four language pairs, reducing the averaged AER from 22.64 to 21.05. Besides, un3424 Models XNLI POS NER MLQA Avg XLM-R* 74.6 75.7 61.6 65.7 69.4 XLM-ALIGN 75.2 75.6 62.6 66.7 70.0 −DWA 75.1 75.2 62.0 65.8 69.5 −TLM 74.4 76.0 60.4 66.0 69.2 Table 3: Ablation studies on the components of XLMALIGN. XLM-R* stands for continue-training XLMRbase with MLM for fair comparisons. Results are averaged over five runs. der both SimAlign and our optimal-transport method, XLM-ALIGN provides consistent reduction of AER, demonstrating the effectiveness of our method for learning fine-grained cross-lingual representations. We also compare XLM-ALIGN with XLM-Rbase using the hidden vectors from the 3-th layer to the 12-th layer. We illustrate the averaged AER scores in Figure 2. Notice that the results on the first two layers are not presented in the figure because of the high AER. It can be observed that XLM-ALIGN consistently improves the results over XLM-Rbase across these layers. Moreover, it shows a parabolic trend across the layers of XLM-Rbase, which is consistent with the results in (Jalili Sabet et al., 2020). In contrast to XLM-Rbase, XLM-ALIGN alleviates this trend and greatly reduces AER in the last few layers. We believe this property of XLMALIGN brings better cross-lingual transferability on the end tasks. 5 Analysis In this section, we conduct comprehensive ablation studies for a better understanding of our XLMALIGN. To reduce the computational cost, we reduce the batch size to 256, and pretrain models with 50K steps in the following experiments. 5.1 Ablation Studies We perform ablation studies to understand the components of XLM-ALIGN, by removing the denoising word alignment loss (−DWA), the TLM loss (−TLM), or removing both (XLM-R*), which is identical to continue-training XLM-Rbase with MLM. We evaluate the models on XNLI, POS, NER, and MLQA, and present the results in Table 3. Comparing −TLM with −DWA, we find that DWA is more effective for POS and MLQA, while TLM performs better on XNLI and NER. Comparing −TLM with XLM-R*, it shows that directly learning DWA slightly harms the perforLayer XNLI POS NER MLQA Avg Layer-8 75.1 75.3 61.9 66.7 69.8 Layer-10 75.2 75.6 62.6 66.7 70.0 Layer-12 75.2 75.8 62.3 67.0 70.1 Table 4: Results of XLM-ALIGN with different layers used for word alignment self-labeling during pretraining. Results are averaged over five runs. Layer XNLI POS NER MLQA Avg Layer-8 75.4 75.3 61.7 66.2 69.7 Layer-10 75.1 75.6 62.5 66.3 69.9 Layer-12 75.2 75.8 62.3 67.0 70.1 Table 5: Results of XLM-ALIGN with different layers used for denoising word alignment during pre-training. Results are averaged over five runs. mance. However, jointly learning DWA with TLM provides remarkable improvements over −DWA, especially on the question answering and the structure prediction tasks that requires token-level crosslingual transfer. This indicates that TLM potentially improves the quality of self-labeled word alignments, making DWA more effective for crosslingual transfer. 5.2 Word Alignment Self-Labeling Layer It has been shown that the word alignment performance has a parabolic trend across the layers of mBERT and XLM-R (Jalili Sabet et al., 2020). It indicates that the middle layers produce higherquality word alignments than the bottom and the top layers. To explore which layer produces better alignment labels for pre-training, we pretrain three variants of XLM-ALIGN, where we use the hidden vectors from three different layers for word alignment self-labeling. We use the 8-th, 10-th, and 12-th layers for word alignment self-labeling during the pre-training. We present the evaluation results in Table 4. Surprisingly, although Layer8 produces higher-quality alignment labels at the beginning of the pre-training, using the alignment labels from the 12-th layer learns a more transferable XLM-ALIGN model for cross-lingual end tasks. 5.3 Denoising Word Alignment Layer Beyond the self-labeling layer, we also investigate which layer is better for learning the denoising word alignment task. Recent studies have shown 3425 Filtering XNLI POS NER MLQA Avg Enable 75.2 75.6 62.6 66.7 70.0 Disable 74.2 75.3 61.6 65.3 69.1 Table 6: Effects of alignment filtering in word alignment self-labeling. Results are averaged over five runs. that it is beneficial to learn sentence-level crosslingual alignment at a middle layer (Chi et al., 2021b). Therefore, we pretrain XLM-ALIGN models by using three different layers for DWA, that is, using the hidden vectors of middle layers as the input of the pointer network. We compare the evaluation results of the three models in Table 5. It can be found that learning DWA at Layer-8 improves XNLI while learning DWA at higher layers produces better performance on the other three tasks. It suggests that, compared with sentence-level pretext tasks that prefers middle layers, the DWA task should be applied at top layers. 5.4 Effects of Alignment Filtering Although our self-labeling method produces highquality alignment labels, the alignment filtering operation can potentially make some of the tokens unaligned, which reduces the example efficiency. Thus, we explore whether the alignment filtering is beneficial for pre-training XLM-ALIGN. To this end, we pretrain an XLM-ALIGN model without alignment filtering. In specific, we use the union set of the forward and backward alignments as the selflabeled alignments so that all tokens are aligned at least once. The forward and backward alignments are obtained by applying the argmax function over rows and columns of A∗, respectively. Empirically, the alignment filtering operation generates high-precision yet fewer labels, while removing the filtering promises more labels but introduces low-confident labels. In Table 6, we compare the results of the models with or without alignment filtering. It can be observed that the alignment filtering operation improves the performance on the end tasks. This demonstrates that it is necessary to use high-precision labels for learning the denoising word alignment task. On the contrary, using perturbed alignment labels in pre-training harms the performance on the end tasks. 5.5 Effects of DWA Query Positions In the denoising word alignment task, we always use the hidden vectors of the masked positions Position XNLI POS NER MLQA Avg masked 75.2 75.6 62.6 66.7 70.0 unmasked 75.5 75.5 62.0 66.5 69.8 all-aligned 75.3 75.9 61.6 66.7 69.9 no-query 75.1 75.2 62.0 65.8 69.5 Table 7: Effects of the query positions in the pointer network for denoising word alignment. Results are averaged over five runs. as the query vectors in the pointer network. To explore the impact of the DWA query positions, we compare three different query positions in Table 7: (1) masked: only using the masked tokens as queries; (2) unmasked: randomly using 15% of the unmasked tokens as queries; (3) all-aligned: for each self-labeled aligned pair, randomly using one of the two tokens as a query. Also, we include the no-query baseline that does not use any queries, which is identical to removing DWA. It can be observed that using all the three query positions improves the performance over the no-query baseline. Moreover, using the masked positions as queries achieves better results than the other two positions, demonstrating the effectiveness of the masked query positions. 6 Discussion In this paper, we introduce denoising word alignment as a new cross-lingual pre-training task. By alternately self-labeling and predicting word alignments, our XLM-ALIGN model learns transferable cross-lingual representations. Experimental results show that our method improves the cross-lingual transferability on a wide range of tasks, particularly on the token-level tasks such as question answering and structured prediction. Despite the effectiveness for learning crosslingual transferable representations, our method also has the limitation that requires a cold-start pre-training to prevent the model from producing low-quality alignment labels. In our experiments, we also try to pretrain XLM-ALIGN from scratch, i.e., without cold-start pre-training. However, the DWA task does not work very well due to the lowquality of self-labeled alignments. Thus, we recommend continue-training XLM-ALIGN on the basis of other pretrained cross-lingual language models. For future work, we would like to research on removing this restriction so that the model can learn word alignments from scratch. 3426 7 Ethical Considerations Despite the current advances in NLP, most NLP research works and applications are English-centric, making none-English users hard to access to NLPrelated services. Our method aims to pretrain cross-lingual language models that transfer supervision signals from high-resource languages to lowresource languages, which makes the NLP services and applications more accessible for low-resourcelanguage speakers. Furthermore, our method can build multilingual models that serve on different languages at the same time, reducing the computational resources for building multilingual models separately for each language. Acknowledgements Heyan Huang is the corresponding author. The work is supported by National Key R&D Plan (No. 2018YFB1005100), National Natural Science Foundation of China (No. 61751201, 61602197 and 61772076), Natural Science Fund of Beijing (No. Z181100008918002), and the funds of Beijing Advanced Innovation Center for Language Resources (No. TYZ19005). References Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics. Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263– 311. Steven Cao, Nikita Kitaev, and Dan Klein. 2020. Multilingual alignment of contextual word representations. In International Conference on Learning Representations. Zewen Chi, Li Dong, Shuming Ma, Shaohan Huang Xian-Ling Mao, Heyan Huang, and Furu Wei. 2021a. mt6: Multilingual pretrained text-to-text transformer with translation pairs. arXiv preprint arXiv:2104.08692. Zewen Chi, Li Dong, Furu Wei, Nan Yang, Saksham Singhal, Wenhui Wang, Xia Song, Xian-Ling Mao, Heyan Huang, and Ming Zhou. 2021b. InfoXLM: An information-theoretic framework for cross-lingual language model pre-training. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3576–3588, Online. Association for Computational Linguistics. Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics, 8:454– 470. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440– 8451, Online. Association for Computational Linguistics. Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. In Advances in Neural Information Processing Systems, pages 7057–7067. Curran Associates, Inc. Alexis Conneau, Ruty Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475–2485, Brussels, Belgium. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Chris Dyer, Victor Chahuneau, and Noah A Smith. 2013. A simple, fast, and effective reparameterization of ibm model 2. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 644–648. Hamidreza Ghader and Christof Monz. 2017. What does attention in neural machine translation pay attention to? In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 30–39, Taipei, Taiwan. Asian Federation of Natural Language Processing. Junjie Hu, Melvin Johnson, Orhan Firat, Aditya Siddhant, and Graham Neubig. 2020a. Explicit alignment objectives for multilingual bidirectional encoders. arXiv preprint arXiv:2010.07972. 3427 Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020b. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalization. arXiv preprint arXiv:2003.11080. Haoyang Huang, Yaobo Liang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, and Ming Zhou. 2019. Unicoder: A universal language encoder by pretraining with multiple cross-lingual tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2485–2494, Hong Kong, China. Association for Computational Linguistics. Masoud Jalili Sabet, Philipp Dufter, Franc¸ois Yvon, and Hinrich Sch¨utze. 2020. SimAlign: High quality word alignments without parallel training data using static and contextualized embeddings. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1627–1643, Online. Association for Computational Linguistics. Karthikeyan K, Zihan Wang, Stephen Mayhew, and Dan Roth. 2020. Cross-lingual ability of multilingual bert: An empirical study. In International Conference on Learning Representations. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, San Diego, CA. Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39, Vancouver. Association for Computational Linguistics. Anoop Kunchukuttan, Pratik Mehta, and Pushpak Bhattacharyya. 2018. The IIT Bombay English-Hindi parallel corpus. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, Miyazaki, Japan. European Language Resources Association. Patrick Lewis, Barlas Oguz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315– 7330, Online. Association for Computational Linguistics. Xintong Li, Guanlin Li, Lemao Liu, Max Meng, and Shuming Shi. 2019. On the word alignment from neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1293–1303. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. arXiv preprint arXiv:2001.08210. Fuli Luo, Wei Wang, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, and Luo Si. 2020. Veco: Variable encoder-decoder pre-training for cross-lingual understanding and generation. arXiv preprint arXiv:2010.16046. Cos¸kun Mermer and Murat Sarac¸lar. 2011. Bayesian word alignment for statistical machine translation. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 182–187. Masaaki Nagata, Katsuki Chousa, and Masaaki Nishino. 2020. A supervised word alignment method based on cross-language span prediction using multilingual BERT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 555–565, Online. Association for Computational Linguistics. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational linguistics, 29(1):19–51. Robert ¨Ostling and J¨org Tiedemann. 2016. Efficient word alignment with markov chain monte carlo. The Prague Bulletin of Mathematical Linguistics, 106(1):125–146. Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Erniem: Enhanced multilingual representation by aligning cross-lingual semantics with monolingual corpora. arXiv preprint arXiv:2012.15674. Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Crosslingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1946–1958, Vancouver, Canada. Association for Computational Linguistics. Gabriel Peyr´e, Marco Cuturi, et al. 2019. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355–607. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-totext transformer. Journal of Machine Learning Research, 21(140):1–67. Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 151–164, Florence, Italy. Association for Computational Linguistics. Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzm´an. 2019. WikiMatrix: Mining 135M parallel sentences in 1620 language pairs from wikipedia. arXiv preprint arXiv:1907.05791. 3428 J¨org Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation, pages 2214–2218, Istanbul, Turkey. European Language Resources Association. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008. Curran Associates, Inc. Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In Advances in Neural Information Processing Systems, volume 28, pages 2692–2700. Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzman, Armand Joulin, and Edouard Grave. 2019. CCNet: Extracting high quality monolingual datasets from web crawl data. arXiv preprint arXiv:1911.00359. Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 833–844, Hong Kong, China. Association for Computational Linguistics. Shijie Wu and Mark Dredze. 2020. Do explicit alignments robustly improve multilingual encoders? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4471–4482, Online. Association for Computational Linguistics. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2020. mt5: A massively multilingual pre-trained text-to-text transformer. arXiv preprint arXiv:2010.11934. Jian Yang, Shuming Ma, Dongdong Zhang, Shuangzhi Wu, Zhoujun Li, and Ming Zhou. 2020. Alternating language modeling for cross-lingual pre-training. In Thirty-Fourth AAAI Conference on Artificial Intelligence. Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3687– 3692, Hong Kong, China. Association for Computational Linguistics. Daniel Zeman, Joakim Nivre, Mitchell Abrams, and et al. 2019. Universal dependencies 2.5. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics ( ´UFAL), Faculty of Mathematics and Physics, Charles University. Wei Zhao, Steffen Eger, Johannes Bjerva, and Isabelle Augenstein. 2020. Inducing languageagnostic multilingual representations. arXiv preprint arXiv:2008.09112. Michał Ziemski, Marcin Junczys-Dowmunt, and Bruno Pouliquen. 2016. The united nations parallel corpus v1. 0. In LREC, pages 3530–3534. A Pre-Training Data We use raw sentences from the Wikipedia dump and CCNet4 as monolingual corpora. The CCNet corpus we use is reconstructed following (Conneau et al., 2020) to reproduce the CC-100 corpus. The resulting corpus contains 94 languages. Table 8 and Table 9 report the language codes and data size of CCNet and Wikipedia dump. Notice that several languages share the same ISO language codes, e.g., zh represents both Simplified Chinese and Traditional Chinese. Besides, Table 10 shows the statistics of our parallel corpora. Code Size (GB) Code Size (GB) Code Size (GB) af 0.2 hr 1.4 pa 0.8 am 0.4 hu 9.5 pl 28.6 ar 16.1 hy 0.7 ps 0.4 as 0.1 id 17.2 pt 39.4 az 0.8 is 0.5 ro 11.0 ba 0.2 it 47.2 ru 253.3 be 0.5 ja 86.8 sa 0.2 bg 7.0 ka 1.0 sd 0.2 bn 5.5 kk 0.6 si 1.3 ca 3.0 km 0.2 sk 13.6 ckb 0.6 kn 0.3 sl 6.2 cs 14.9 ko 40.0 sq 3.0 cy 0.4 ky 0.5 sr 7.2 da 6.9 la 0.3 sv 60.4 de 99.0 lo 0.2 sw 0.3 el 13.1 lt 2.3 ta 7.9 en 731.6 lv 1.3 te 2.3 eo 0.5 mk 0.6 tg 0.7 es 85.6 ml 1.3 th 33.0 et 1.4 mn 0.4 tl 1.2 eu 1.0 mr 0.5 tr 56.4 fa 19.0 ms 0.7 tt 0.6 fi 5.9 mt 0.2 ug 0.2 fr 89.9 my 0.4 uk 13.4 ga 0.2 ne 0.6 ur 3.0 gl 1.5 nl 25.9 uz 0.1 gu 0.3 nn 0.4 vi 74.5 he 4.4 no 5.5 yi 0.3 hi 5.0 or 0.3 zh 96.8 Table 8: The statistics of CCNet used for pre-training. B Hyperparameters for Pre-Training As shown in Table 11, we present the hyperparameters for pre-training XLM-ALIGN. We use the same vocabulary with XLM-R (Conneau et al., 2020). 4https://github.com/facebookresearch/ cc_net 3429 Code Size (GB) Code Size (GB) Code Size (GB) af 0.12 hr 0.28 pa 0.10 am 0.01 hu 0.80 pl 1.55 ar 1.29 hy 0.60 ps 0.04 as 0.04 id 0.52 pt 1.50 az 0.24 is 0.05 ro 0.42 ba 0.13 it 2.70 ru 5.63 be 0.31 ja 2.65 sa 0.04 bg 0.62 ka 0.37 sd 0.02 bn 0.41 kk 0.29 si 0.09 ca 1.10 km 0.12 sk 0.21 ckb 0.00 kn 0.25 sl 0.21 cs 0.81 ko 0.56 sq 0.11 cy 0.06 ky 0.10 sr 0.74 da 0.33 la 0.05 sv 1.70 de 5.43 lo 0.01 sw 0.03 el 0.73 lt 0.19 ta 0.46 en 12.58 lv 0.12 te 0.45 eo 0.25 mk 0.34 tg 0.04 es 3.38 ml 0.28 th 0.52 et 0.23 mn 0.05 tl 0.04 eu 0.24 mr 0.10 tr 0.43 fa 0.66 ms 0.20 tt 0.09 fi 0.68 mt 0.01 ug 0.03 fr 4.00 my 0.15 uk 2.43 ga 0.03 ne 0.06 ur 0.13 gl 0.27 nl 1.38 uz 0.06 gu 0.09 nn 0.13 vi 0.76 he 1.11 no 0.54 yi 0.02 hi 0.38 or 0.04 zh 1.08 Table 9: The statistics of Wikipedia dump used for pretraining. ISO Code Size (GB) ISO Code Size (GB) en-ar 5.88 en-ru 7.72 en-bg 0.49 en-sw 0.06 en-de 4.21 en-th 0.47 en-el 2.28 en-tr 0.34 en-es 7.09 en-ur 0.39 en-fr 7.63 en-vi 0.86 en-hi 0.62 en-zh 4.02 Table 10: Parallel data used for pre-training. C Hyperparameters for Fine-Tuning In Table 12, we present the hyperparameters for fine-tuning XLM-Rbase and XLM-ALIGN on the XTREME end tasks. For each task, the hyperparameters are searched on the joint validation set of all languages. D Detailed Results on XTREME We present the detailed results of XLM-ALIGN on XTREME in Table 13-19. Hyperparameters Value Layers 12 Hidden size 768 FFN inner hidden size 3,072 Attention heads 12 Training steps 150K Batch size 2,048 Adam ϵ 1e-6 Adam β (0.9, 0.98) Learning rate 2e-4 Learning rate schedule Linear Warmup steps 10,000 Gradient clipping 1.0 Weight decay 0.01 Self-labeling layer 10 Entropic regularization µ 1.0 Sinkhorn iterations 2 Alignment filtering iterations 2 Alignment filtering α 0.9 Table 11: Hyperparameters used for pre-training XLMALIGN. 3430 POS NER XQuAD MLQA TyDiQA XNLI PAWS-X Batch size {8,16,32} 8 32 32 32 32 32 Learning rate {1,2,3}e-5 {5,...,9}e-6 {2,3,4}e-5 {2,3,4}e-5 {2,3}e-5 {5,...,8}e-6 {1,2}e-5 LR schedule Linear Linear Linear Linear Linear Linear Linear Warmup 10% 10% 10% 10% 10% 12,500 steps 10% Weight decay 0 0 0 0 0 0 0 Epochs 10 10 4 {2,3,4} {5,10,15,20} 10 10 Table 12: Hyperparameters used for fine-tuning XLM-Rbase and XLM-ALIGN on the XTREME end tasks. Model af ar bg de el en es et eu fa fi fr he hi hu id it XLM-ALIGN 88.5 69.1 88.8 88.8 85.8 95.9 88.5 84.9 68.3 70.9 84.8 88.1 79.6 71.6 83.3 72.3 89.4 Model ja kk ko mr nl pt ru ta te th tl tr ur vi yo zh Avg XLM-ALIGN 51.1 75.3 53.8 80.3 89.3 87.6 88.9 62.3 85.9 60.2 90.1 74.8 63.3 55.9 24.2 67.9 76.0 Table 13: Results on part-of-speech tagging. Model ar he vi id jv ms tl eu ml ta te af nl en de el bn hi mr ur XLM-ALIGN 57.7 54.3 72.5 49.7 56.9 68.3 72.0 53.1 68.6 58.0 54.6 76.3 82.1 84.2 77.9 76.4 73.1 69.2 64.9 65.8 Model fa fr it pt es bg ru ja ka ko th sw yo my zh kk tr et fi hu Avg XLM-ALIGN 53.2 79.0 79.4 78.8 73.8 78.9 66.2 23.0 70.6 56.6 2.2 69.3 43.8 56.5 28.3 49.2 77.5 73.3 77.0 77.0 63.7 Table 14: Results on WikiAnn named entity recognition. Model en es de el ru tr ar vi th zh hi Avg XLM-ALIGN 85.7 / 74.6 70.3 / 52.5 76.6 / 60.3 75.5 / 56.8 79.4 / 60.8 71.8 / 54.7 75.4 / 59.4 72.1 / 61.0 70.9 / 55.5 76.7 / 56.9 67.3 / 56.8 74.7 / 59.0 Table 15: Results on XQuAD question answering. Model en es de ar hi vi zh Avg XLM-ALIGN 81.5 / 68.3 70.3 / 52.2 64.5 / 49.8 60.7 / 41.2 65.2 / 47.5 69.8 / 48.9 64.4 / 40.4 68.1 / 49.8 Table 16: Results on MLQA question answering. Model en ar bn fi id ko ru sw te Avg XLM-ALIGN 69.4 / 56.2 68.7 / 49.4 56.0 / 38.9 64.2 / 47.2 73.9 / 57.9 53.0 / 40.4 62.3 / 38.0 60.1 / 42.8 51.0 / 31.9 62.1 / 44.8 Table 17: Results on TyDiQA question answering. Model en fr es de el bg ru tr ar vi th zh hi sw ur Avg XLM-ALIGN 86.7 80.6 81.0 78.8 77.4 78.8 77.4 75.2 73.9 76.9 73.8 77.0 71.9 67.1 66.6 76.2 Table 18: Results on XNLI natural language inference. Model en fr de es ja ko zh Avg XLM-ALIGN 95.1 89.3 90.5 90.7 79.1 79.5 83.2 86.8 Table 19: Results on PAWS-X cross-lingual paraphrase adversaries.
|
2021
|
265
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3431–3441 August 1–6, 2021. ©2021 Association for Computational Linguistics 3431 Rejuvenating Low-Frequency Words: Making the Most of Parallel Data in Non-Autoregressive Translation Liang Ding∗ The University of Sydney [email protected] Longyue Wang∗ Tencent AI Lab [email protected] Xuebo Liu University of Macau [email protected] Derek F. Wong University of Macau [email protected] Dacheng Tao JD Explore Academy, JD.com [email protected] Zhaopeng Tu Tencent AI Lab [email protected] Abstract Knowledge distillation (KD) is commonly used to construct synthetic data for training non-autoregressive translation (NAT) models. However, there exists a discrepancy on lowfrequency words between the distilled and the original data, leading to more errors on predicting low-frequency words. To alleviate the problem, we directly expose the raw data into NAT by leveraging pretraining. By analyzing directed alignments, we found that KD makes low-frequency source words aligned with targets more deterministically but fails to align sufficient low-frequency words from target to source. Accordingly, we propose reverse KD to rejuvenate more alignments for lowfrequency target words. To make the most of authentic and synthetic data, we combine these complementary approaches as a new training strategy for further boosting NAT performance. We conduct experiments on five translation benchmarks over two advanced architectures. Results demonstrate that the proposed approach can significantly and universally improve translation quality by reducing translation errors on low-frequency words. Encouragingly, our approach achieves 28.2 and 33.9 BLEU points on the WMT14 English-German and WMT16 Romanian-English datasets, respectively. Our code, data, and trained models are available at https://github.com/ longyuewangdcu/RLFW-NAT. 1 Introduction Recent years have seen a surge of interest in nonautoregressive translation (NAT, Gu et al., 2018), which can improve the decoding efficiency by predicting all tokens independently and simultaneously. The non-autoregressive factorization breaks conditional dependencies among output tokens, ∗Liang Ding and Longyue Wang contributed equally to this work. Work was done when Liang Ding and Xuebo Liu were interning at Tencent AI Lab. which prevents a model from properly capturing the highly multimodal distribution of target translations. As a result, the translation quality of NAT models often lags behind that of autoregressive translation (AT, Vaswani et al., 2017) models. To balance the trade-off between decoding speed and translation quality, knowledge distillation (KD) is widely used to construct a new training data for NAT models (Gu et al., 2018). Specifically, target sentences in the distilled training data are generated by an AT teacher, which makes NAT easily acquire more deterministic knowledge and achieve significant improvement (Zhou et al., 2020). Previous studies have shown that distillation may lose some important information in the original training data, leading to more errors on predicting low-frequency words. To alleviate this problem, Ding et al. (2021b) proposed to augment NAT models the ability to learn lost knowledge from the original data. However, their approach relies on external resources (e.g. word alignment) and human-crafted priors, which limits the applicability of the method to a broader range of tasks and languages. Accordingly, we turn to directly expose the raw data into NAT by leveraging pretraining without intensive modification to model architectures (§2.2). Furthermore, we analyze bilingual links in the distilled data from two alignment directions (i.e. source-to-target and target-to-source). We found that KD makes low-frequency source words aligned with targets more deterministically but fails to align low-frequency words from target to source due to information loss. Inspired by this finding, we propose reverse KD to recall more alignments for low-frequency target words (§2.3). We then concatenate two kinds of distilled data to maintain advantages of deterministic knowledge and low-frequency information. To make the most of authentic and synthetic data, we combine three complementary approaches (i.e. raw pretraining, 3432 bidirectional distillation training and KD finetuning) as a new training strategy for further boosting NAT performance (§2.4). We validated our approach on five translation benchmarks (WMT14 En-De, WMT16 Ro-En, WMT17 Zh-En, WAT17 Ja-En and WMT19 EnDe) over two advanced architectures (Mask Predict, Ghazvininejad et al., 2019; Levenshtein Transformer, Gu et al., 2019). Experimental results show that the proposed method consistently improve translation performance over the standard NAT models across languages and advanced NAT architectures. Extensive analyses confirm that the performance improvement indeed comes from the better lexical translation accuracy especially on low-frequency tokens. Contributions Our main contributions are: • We show the effectiveness of rejuvenating lowfrequency information by pretraining NAT models from raw data. • We provide a quantitative analysis of bilingual links to demonstrate the necessity to improve low-frequency alignment by leveraging both KD and reverse KD. • We introduce a simple and effective training recipe to accomplish this goal, which is robustly applicable to several model structures and language pairs. 2 Rejuvenating Low-Frequency Words 2.1 Preliminaries Non-Autoregressive Translation Given a source sentence x, an AT model generates each target word yt conditioned on previously generated ones y<t, leading to high latency on the decoding stage. In contrast, NAT models break this autoregressive factorization by producing target words in parallel. Accordingly, the probability of generating y is computed as: p(y|x) = T Y t=1 p(yt|x; θ) (1) where T is the length of the target sequence, and it is usually predicted by a separate conditional distribution. The parameters θ are trained to maximize the likelihood of a set of training examples according to L(θ) = arg maxθ log p(y|x; θ). Typically, most NAT models are implemented upon the framework of Transformer (Vaswani et al., 2017). Knowledge Distillation Gu et al. (2018) pointed out that NAT models suffer from the multimodality problem, where the conditional independence assumption prevents a model from properly capturing the highly multimodal distribution of target translations. Thus, the sequence-level knowledge distillation is introduced to reduce the modes of training data by replacing their original target-side samples with sentences generated by an AT teacher (Gu et al., 2018; Zhou et al., 2020; Ren et al., 2020). Formally, the original parallel data Raw and the distilled data −→ KD can be defined as follows: Raw = {(xi, yi)}N i=1 (2) −→ KD = {(xi, fs7→t(xi))|xi ∈Raws}N i=1 (3) where fs7→t represents an AT-based translation model trained on Raw data for translating text from the source to the target language. N is the total number of sentence pairs in training data. As shown in Figure 1 (a), well-performed NAT models are generally trained on −→ KD data instead of Raw. 2.2 Pretraining with Raw Data Motivation Gao et al. (2018) showed that more than 90% of words are lower than 10e-4 frequency in WMT14 En-De dataset. This token imbalance problem biases translation models towards overfitting to frequent observations while neglecting those low-frequency observations (Gong et al., 2018; Nguyen and Chiang, 2018; Gu et al., 2020). Thus, the AT teacher fs7→t tends to generate more high-frequency tokens and less low-frequency tokens during constructing distilled data −→ KD. On the one hand, KD can reduce the modes in training data (i.e. multiple lexical choices for a source word), which lowers the intrinsic uncertainty (Ott et al., 2018) and learning difficulty for NAT (Zhou et al., 2020; Ren et al., 2020), making it easily acquire more deterministic knowledge. On the other hand, KD aggravates the imbalance of high-frequency and low-frequency words in training data and lost some important information originated in raw data. Ding et al. (2021b) revealed the side effect of distilled training data, which cause lexical choice errors for low-frequency words in NAT models. Accordingly, they introduced an extra bilingual data-dependent prior objective to augments NAT models the ability to learn the lost knowledge from raw data. We use their findings as our departure point, but rejuvenate low-frequency 3433 (a) Traditional Training (b) Raw Pretraining (c) Bidirectional Distillation Training Figure 1: An illustration of different strategies for training NAT models. “distill” and “reverse distill” indicate sequence-level knowledge distillation with forward and backward AT teachers, respectively. The data block in transparent color means source- or target-side data are synthetically generated. Best view in color. Data s 7→t LFW Links t 7→s LFW Links R P F1 R P F1 Raw 66.4 81.9 73.3 72.3 80.6 76.2 −→ KD 73.4 89.2 80.5 69.9 79.1 74.2 ←− KD 61.2 79.4 69.1 82.9 83.1 83.0 Table 1: Evaluation on aligned links between sourceand target-side low-frequency words (LFW). A directed line indicates aligning bilingual words from the source to the target side (s 7→t) or in an opposite way (t 7→s). R, P and F1 are recall, precision and F1-score. words in a more simple and direct way: directly exposing raw data into NAT via pretraining. Our Approach Many studies have shown that pretraining could transfer the knowledge and data distribution, especially for rare categories, hence improving the model robustness (Hendrycks et al., 2019; Mathis et al., 2021). Here we want to transfer the distribution of lost information, e.g. lowfrequency words. As illustrated in Figure 1(b), we propose to first pretrain NAT models on Raw data and then continuously train them on −→ KD data. The raw data maintain the original distribution especially on low-frequency words. Although it is difficult for NAT to learn high-mode data, the pretraining can acquire general knowledge from authentic data, which may help better and faster learning further tasks. Thus, we early stop pretraining when the model can achieve 90% of the best performance of raw data in terms of BLEU score (Platanios et al., 2019)1. In order to keep the merits of low-modes, 1In preliminary experiments, we tried another simple strategy: early-stop at fixed step according to the size of training data (e.g. training 70K En-De and early stop at 20K / 30K / 40K, respectively). We found that both strategies achieve Data Sentence RawS 海克曼和奥德海姆提出... 模型 RawT Hackman and Oldham propose ... model −→ KDT Heckman and Oddheim propose ... model ←− KDS 哈克曼和奥尔德姆提出... 模式 Table 2: An example in different kinds of data. “Raw” means the original data while “−→ KD” and “←− KD” indicate syntactic data distilled by KD and reverse KD, respectively. The subscript “S” or “T” is short for source- or target-side. The low-frequency words are highlighted with colors and italics are incorrect translations. we further train the pretrained model on distilled data −→ KD. As it is easy for NAT to learn deterministic knowledge, we finetune the model for the rest steps. For fair comparison, the total training steps of the proposed method are same as the traditional one. In general, we expect that this training recipe can provide a good trade-off between raw and distilled data (i.e. high-modes and complete vs. low-modes and incomplete). 2.3 Bidirectional Distillation Training Analyzing Bilingual Links in Data KD simplifies the training data by replacing low-frequency target words with high-frequency ones (Zhou et al., 2020). This is able to facilitate easier aligning source words to target ones, resulting in high bilingual coverage (Jiao et al., 2020). Due to the information loss, we argue that KD makes lowfrequency target words have fewer opportunities to align with source ones. To verify this, we propose a method to quantitatively analyze bilingual links from two directions, where low-frequency words similar performance. 3434 are aligned from source to target (s 7→t) or in an opposite direction (t 7→s). The method can be applied to different types of data. Here we take s 7→t links in Raw data as an example to illustrate the algorithm. Given the WMT14 En-De parallel corpus, we employ an unsupervised word alignment method2 (Och and Ney, 2003) to produce a word alignment, and then we extract aligned links whose source words are low-frequency (called s 7→t LFW Links). Second, we randomly select a number of samples from the parallel corpus. For better comparison, the subset should contains the same i in Equation (2) as that of other type of datasets (e.g. i in Equation (3) for −→ KD). Finally, we calculate recall, precision, F1 scores based on low-frequency bilingual links for the subset. Recall (R) represents how many low-frequency source words can be aligned to targets. Precision (P) means how many aligned low-frequency links are correct according to human evaluation. F1 is the harmonic mean between precision and recall. Similarly, we can analyze t 7→s LFW Links by considering low-frequency targets. Table 1 shows the results on low-frequency links. Compared with Raw, −→ KD can recall more s 7→t LFW links (73.4 vs. 66.4) with more accurate alignment (89.2 vs. 73.3). This demonstrates the effectiveness of KD for NAT models from the bilingual alignment perspective. However, in the t 7→s direction, there are fewer LFW links (69.9 vs. 72.3) with worse alignment quality (79.1 vs. 80.6) in −→ KD than those in Raw. This confirms our claim that KD harms NAT models due to the loss of lowfrequency target words. Inspired by these findings, it is natural to assume that reverse KD exhibits complementary properties. Accordingly, we conduct the same analysis method on ←− KD data, and found better t 7→s links but worse s 7→t links compared with Raw. Take the Zh-En sentence pair in Table 2 for example, −→ KD retains the source side lowfrequency Chinese words “海克曼” (RawS) but generates the high-frequency English words “Heckman” instead of the golden “Hackman” (−→ KDT). On the other hand, ←− KD preserves the low-frequency English words “Hackman” (RawT) but produces the high-frequency Chinese words “哈克曼” (←− KDS). Our Approach Based on analysis results, we propose to train NAT models on bidirectional distil2The FastAlign (Dyer et al., 2013) was employed to build word alignments for the training datasets. lation by concatenating two kinds of distilled data. The reverse distillation is to replace the source sentences in the original training data with synthetic ones generated by a backward AT teacher.3 According to Equation 3, ←− KD can be formulated as: ←− KD = {(yi, ft7→s(yi))|yi ∈Rawt}N i=1 (4) where ft7→s represents an AT-based translation model trained on Raw data for translating text from the target to the source language. Figure 1(c) illustrates the training strategy. First, we employ both fs7→t and ft7→s AT models to generate −→ KD and ←− KD data, respectively. Considering complementarity of two distilled data, we combine −→ KD and ←− KD as a new training data for training NAT models. We expect that 1) distilled data can maintain advantages of low-modes; 2) bidirectinoal distillation can recall more LFW links on two directions with better alignment quality, leading to the overall improvements. Besides, Nguyen et al. (2020) claimed that combining different distilled data (generated by various models trained with different seeds) improves data diversification for NMT, and we leave this for future work. 2.4 Combining Both of Them: Low-Frequency Rejuvenation (LFR) We have proposed two parallel approaches to rejuvenate low-frequency knowledge from authentic (§2.2) and synthetic (§2.3) data, respectively. Intuitively, we combine both of them to further improve the model performance. From data view, two presented training strategies are: Raw →−→ KD (Raw Pretraining) and −→ KD + ←− KD (Bidirectional Distillation Training). Considering the effectiveness of pretraining (Mathis et al., 2021) and clean finetuning (Wu et al., 2019), we introduce a combined pipeline: Raw →−→ KD + ←− KD →−→ KD as out best training strategy. There are many possible ways to implement the general idea of combining two approaches. The aim of this paper is not to explore the whole space but simply to show that one fairly straightforward implementation works well and the idea is reasonable. Nonetheless, we compare possible strategies of combination two approaches as well as demonstrate their complementarity in §3.3. While in main experiments (in §3.2), we valid the combination strategy, namely Low-Frequency Rejuvenation (LFR). 3This is different from back-translation (Edunov et al., 2018), which is an alternative to leverage monolingual data. 3435 Model Iteration Speed En-De Ro-En BLEU ALF BLEU ALF AT Models Transformer-BASE (Ro-En Teacher) n/a 1.0× 27.3 70.5 34.1 73.6 Transformer-BIG (En-De Teacher) n/a 0.8× 29.2 73.0 n/a n/a Existing NAT Models NAT (Gu et al., 2018) 1.0 2.4× 19.2 n/a 31.4 n/a Iterative NAT (Lee et al., 2018) 10.0 2.0× 21.6 30.2 DisCo (Kasai et al., 2020) 4.8 3.2× 26.8 33.3 Mask-Predict (Ghazvininejad et al., 2019) 10.0 1.5× 27.0 33.3 Levenshtein (Gu et al., 2019) 2.5 3.5× 27.3 33.3 Our NAT Models Mask-Predict (Ghazvininejad et al., 2019) 10.0 1.5× 27.0 68.4 33.3 70.9 +Low-Frequency Rejuvenation 27.8† 72.3 33.9† 72.4 Levenshtein (Gu et al., 2019) 2.5 3.5× 27.4 69.2 33.2 71.1 +Low-Frequency Rejuvenation 28.2† 72.8 33.8† 72.7 Table 3: Comparison with previous work on WMT14 En-De and WMT16 Ro-En. “Iteration” indicates the number of iterative refinement while “Speed” shows the speed-up ratio of decoding. “ALF” is the translation accuracy on low-frequency words. “†” indicates statistically significant difference (p < 0.05) from corresponding baselines. 3 Experiment 3.1 Setup Data Main experiments are conducted on four widely-used translation datasets: WMT14 EnglishGerman (En-De, Vaswani et al. 2017), WMT16 Romanian-English (Ro-En, Gu et al. 2018), WMT17 Chinese-English (Zh-En, Hassan et al. 2018), and WAT17 Japanese-English (Ja-En, Morishita et al. 2017), which consist of 4.5M, 0.6M, 20M, and 2M sentence pairs, respectively. We use the same validation and test datasets with previous works for fair comparison. To prove the universality of our approach, we further experiment on different data volumes, which are sampled from WMT19 En-De.4 The Small and Medium corpora respectively consist of 1.0M and 4.5M sentence pairs, and Large one is the whole dataset which contains 36M sentence pairs. We preprocess all data via BPE (Sennrich et al., 2016) with 32K merge operations. We use tokenized BLEU (Papineni et al., 2002) as the evaluation metric, and sign-test (Collins et al., 2005) for statistical significance test. The translation accuracy of lowfrequency words is measured by AoLC (Ding et al., 2021b), where word alignments are established 4http://www.statmt.org/wmt19/ translation-task.html based on the widely-used automatic alignment tool GIZA++ (Och and Ney, 2003). Models We validated our research hypotheses on two state-of-the-art NAT models: • Mask-Predict (MaskT, Ghazvininejad et al. 2019) that uses the conditional mask LM (Devlin et al., 2019) to iteratively generate the target sequence from the masked input. We followed its optimal settings to keep the iteration number as 10 and length beam as 5. • Levenshtein Transformer (LevT, Gu et al. 2019) that introduces three steps: deletion, placeholder and token prediction. The decoding iterations adaptively depends on certain conditions. We closely followed previous works to apply sequence-level knowledge distillation to NAT (Kim and Rush, 2016). Specifically, we train both BASE and BIG Transformer as the AT teachers. For BIG model, we adopt large batch strategy (i.e. 458K tokens/batch) to optimize the performance. Most NAT tasks employ Transformer-BIG as their strong teacher except for Ro-En and Small En-De, which are distilled by Transformer-BASE. Training Traditionally, NAT models are usually trained for 300K steps on regular batch size (i.e. 3436 Model Zh-En Ja-En BLEU ALF BLEU ALF AT 25.3 66.2 29.8 70.8 MaskT 24.2 61.5 28.9 66.9 +LFR 25.1† 64.8 29.6† 68.9 LevT 24.4 62.7 29.1 66.8 +LFR 25.1† 65.3 29.7 69.2 Table 4: Performance on other language pairs, including WMT17 Zh-En and WAT17 Ja-En. “†” indicates statistically significant difference (p < 0.05) from corresponding baselines. 128K tokens/batch). In this work, we empirically adopt large batch strategy (i.e. 480K tokens/batch) to reduce the training steps for NAT (i.e. 70K). Accordingly, the learning rate warms up to 1 × 10−7 for 10K steps, and then decays for 60k steps with the cosine schedule (Ro-En models only need 4K and 21K, respectively). For regularization, we tune the dropout rate from [0.1, 0.2, 0.3] based on validation performance in each direction, and apply weight decay with 0.01 and label smoothing with ϵ = 0.1. We use Adam optimizer (Kingma and Ba, 2015) to train our models. We followed the common practices (Ghazvininejad et al., 2019; Kasai et al., 2020) to evaluate the performance on an ensemble of top 5 checkpoints to avoid stochasticity. Note that the total training steps of the proposed approach (in §2.2∼2.4) are identical with those of the standard training (in §2.1). Taking the best training strategy (Raw →−→ KD + ←− KD →−→ KD) for example, we empirically set the training step for each stage is 20K, 20K and 30K, respectively. And Ro-En models respectively need 8K, 8K and 9K steps in corresponding training stage. 3.2 Results Comparison with Previous Work Table 3 lists the results of previous competitive NAT models (Gu et al., 2018; Lee et al., 2018; Kasai et al., 2020; Gu et al., 2019; Ghazvininejad et al., 2019) on the WMT16 Ro-En and WMT14 En-De benchmark. We implemented our approach on top of two advanced NAT models (i.e. Mask-Predict and Levenshtein Transformer). Compared with standard NAT models, our training strategy significantly and consistently improves translation performance (BLEU↑) across different language pairs and NAT models. Besides, the improvements on translation Model Law Med. IT Kor. Sub. AT 41.5 30.8 27.5 8.6 15.4 MaskT 37.3 28.2 24.6 7.3 11.2 +LFR 38.1† 28.8 25.4† 8.9† 14.3† LevT 37.5 28.4 24.7 7.5 12.4 +LFR 38.5† 29.4† 25.9† 8.4† 14.5† Table 5: Performance on domain shift setting. Models are trained on WMT14 En-De news domain but evaluated on out-of-domain test sets, including law, medicine, IT, koran and subtitle. “†” indicates statistically significant difference (p < 0.05) from corresponding baselines. performance are mainly due to a increase of translation accuracy on low-frequency words (ALF↑), which reconfirms our claims. For instance, our method significantly improves the standard MaskPredict model by +0.8 BLEU score with a substantial +3.6 increase in ALF score. Encouragingly, our approach push the existing NAT models to achieve new SOTA performances (i.e. 28.2 and 33.9 BLEU on En-De and Ro-En, respectively). It is worth noting that our data-level approaches neither modify model architecture nor add extra training loss, thus do not increase any latency (“Speed”), maintaining the intrinsic advantages of non-autoregressive generation. We must admit that our strategy indeed increase the amount of computing resources due to that we should train ft7→s AT teachers for building ←− KD data. Results on Other Language Pairs Table 4 lists the results of NAT models on Zh-En and Ja-En language pairs, which belong to different language families (i.e. Indo-European, Sino-Tibetan and Japonic). Compared with baselines, our method significantly and incrementally improves the translation quality in all cases. For Zh-En, LFR achieves on average +0.8 BLEU improvement over the traditional training, along with increasing on average +3.0% accuracy on low-frequency word translation. For long-distance language pair Ja-En, our method still improves the NAT model by on average +0.7 BLEU point with on average +2.2 ALF. Furthermore, NAT models with the proposed training strategy perform closely to their AT teachers (i.e. 0.2 ∆BLEU). This shows the effectiveness and universality of our method across language pairs. 3437 Model BLEU 1.0M 4.5M 36.0M AT 25.5 37.6 40.2 MaskT 23.7 35.4 36.8 +LFR 24.3† 36.2† 37.7† Table 6: Performance on different scale of training data. The small and medium datasets are sampled from the large WMT19 En-De dataset, and evaluations are conducted on the same testset. “†” indicates statistically significant difference (p < 0.05) from corresponding baselines. Results on Domain Shift Scenario The lexical choice must be informed by linguistic knowledge of how the translation model’s input data maps onto words in the target domain. Since low-frequency words get lost in traditional NAT models, the problem of lexical choice is more severe under domain shift scenario (i.e. models are trained on one domain but tested on other domains). Thus, we conduct evaluation on WMT14 En-De models over five out-of-domain test sets (M¨uller et al., 2020), including law, medicine, IT, Koran and movie subtitle domains. As shown in Table 5, standard NAT models suffer large performance drops in terms of BLEU score (i.e. on average -2.9 BLEU over AT model). By observing these outputs, we found a large amount of translation errors on low-frequency words, most of which are domain-specific terminologies. In contrast, our approach improves translation quality (i.e. on average -1.4 BLEU over AT model) by rejuvenating low-frequency words to a certain extent, showing that LFR increases the domain robustness of NAT models. Results on Different Data Scales To confirm the effectiveness of our method across different data sizes, we further experiment on three En-De datasets at different scale. The small- and mediumscale training data are randomly sampled from WM19 En-De corpus, containing about 1.0M and 4.5M sentence pairs, respectively. The large-scale one is collected from WMT19, which consists of 36M sentence pairs. We report the BLEU scores on same testset newstest2019 for fair comparison. We employs base model to train the small-scale AT teacher, and big model with large batch strategy (i.e. 458K tokens/batch) to build the AT teachers for medium- and large-scale. As seen in Table 6, our simple training recipe boost performances for Model BLEU ALF Mask-Predict 27.0 68.4 +Raw Data Prior 27.8 72.4 +Low-Frequency 27.8 72.3 +Combination 28.1 72.9 Table 7: Complementary to other work. “Combination” indicates combining “+Raw Data Prior” proposed by Ding et al. (2021b) with our “+Low-Frequency”. Experiments are conducted on WMT14 En-De. NAT models across different size of datasets, especially on large scale (+0.9), showing the robustness and effectiveness of our approach. Complementary to Related Work Ding et al. (2021b) is relevant to our work, which introduced an extra bilingual data-dependent prior objective to augment NAT models the ability to learn lowfrequency words in raw data. Our method is complementary to theirs due to that we only change data and training strategies (model-agnostic). As shown in Table 7, two approaches yield comparable performance in terms of BLEU and ALF. Besides, combination can further improve BLEU as well as ALF scores (i.e. +0.3 and +0.6). This illustrates the complementarity of model-level and data-level approaches on rejuvenating low-frequency knowldege for NAT models. 3.3 Analysis We conducted extensive analyses to better understand our approach. All results are reported on the Mask-Predict models. Accuracy of Lexical Choice To understand where the performance gains come from, we conduct fine-grained analysis on lexical choice. We divide “All” tokens into three categories based on their frequency, including “High”, “Medium” and “Low”. Following Ding et al. (2021b), we measure the accuracy of lexical choice on different frequency of words. Table 8 shows the results. Takeaway: The majority of improvements on translation accuracy is from the low-frequency words, confirming our hypothesis. Low-Frequency Words in Output We expect to recall more low-frequency words in translation output. As shown in Table 9, we calculate the ratio of low-frequency words in generated sentences. As seen, KD biases the NAT model towards gen3438 Model En-De Zh-En Ja-En All High Med. Low All High Med. Low All High Med. Low MaskT (Raw) 74.3 75.9 74.6 72.5 68.5 71.5 68.3 65.1 73.1 75.5 74.7 69.1 MaskT (KD) 76.3 82.4 78.3 68.4 72.7 81.4 75.2 61.5 75.3 82.8 76.3 66.9 +Raw-Pretrain 77.7 83.1 78.4 71.9 73.4 81.6 75.3 64.1 76.1 83.4 76.7 68.3 +Bi-Distillation 77.9 83.1 78.5 72.3 73.7 81.7 75.3 64.8 76.5 83.5 76.7 68.9 Table 8: Analysis on different frequency words in terms of accuracy of lexical choice. We split “All” words into “High”, “Medium” and “Low” categories. Shades of cell color represent differences between ours and KD. Model En-De Zh-En Ja-En MaskT (Raw) 10.3% 6.7% 9.4% MaskT (KD) 7.6% 4.2% 6.9% +Raw-Pretrain 9.3% 5.6% 8.4% +Bi-Distillation 9.7% 6.8% 8.7% Table 9: Ratio of low-frequency target words in output. # Strategy BLEU ALF 1 Raw 24.1 69.3 2 −→ KD 25.4 66.4 3 Raw+−→ KD 25.6 67.7 4 Raw→−→ KD 25.9 68.2 5 Raw+←− KD+−→ KD 25.7 67.9 6 Raw→←− KD+−→ KD 25.7 68.3 7 Raw→←− KD+−→ KD→−→ KD 26.3 69.5 Table 10: Performances of different strategies. The models are trained and tested on WMT14 En-De. “A+B” means concatenate A and B while “A→B” indicates pretraining on A and then finetuning on B. erating high-frequency tokens (Low freq.↓) while our method can not only correct this bias (on average +18% and +26% relative changes for +rawpretrain and +Bi-distillation), but also enhance translation (BLEU↑in Table 4). Takeaway: Our method generates translations that contain more low-frequency words. Effects of Variant Training Strategies As discussed in §2.4, we carefully investigate alternative training approaches in Table 10. We make the total training step identical to that of vanilla NAT models, and report both BLEU and ALF scores. As seen, all variant strategies perform better than the standard KD method in terms both BLEU and Model All High Med. Low Training on Raw Data AT-Teacher 79.3 84.7 80.2 73.0 AT-Student 76.8 80.2 77.4 72.8 Training on Distilled Data AT-Student 77.3 82.5 78.6 70.9 +LFT 78.1 83.2 78.7 72.5 Table 11: Analysis on AT models in term of the accuracy of lexical choice on WMT14 En-De. We split “All” words into “High”, “Medium” and “Low” categories. ALF scores, confirming the necessity of our work. Takeaway: 1) Pretraining is more effective than combination on utilizing data manipulation strategies; 2) raw data and bidirectional distilled data are complementary to each other; 3) it is indispensable to finetune models on −→ KD in the last stage. Our Approach Works for AT Models Although our work is designed for NAT models, we also investigated whether our LFT method works for general cases, e.g. autoregressive models. We used Transformer-BIG as the teacher model. For fair comparison, we leverage the TransformerBASE as the student model, which shares the same model capacity with NAT student (i.e. MaskT). The result lists in Table 11. As seen, AT models also suffer from the problem of low-frequency words when using knowledge distillation, and our approach also works for them. Takeaway: Our method works well for general cases through rejuvenating more low-frequency words. 4 Related Work Low-Frequency Words Benefiting from continuous representation learned from the training data, NMT models have shown the promising performance. However, Koehn and Knowles (2017) point 3439 that low-frequency words translation is still one of the key challenges for NMT according to the Zipf’s law (Zipf, 1949). For AT models, Arthur et al. (2016) address this problem by integrating a count-based lexicon, and Nguyen and Chiang (2018) propose an additional lexical model, which is jointly trained with the AT model. Recently, Gu et al. (2020) adaptively re-weight the rare words during training. The lexical choice problem is more serious for NAT models, since 1) the lexical choice errors (low-resource words in particular) of AT distillation will propagate to NAT models; and 2) NAT lacks target-side dependencies thus misses necessary target-side context. In this work, we alleviate this problem by solving the first challenge. Data Manipulation Our work is related to previous studies on manipulating training data for NMT. Bogoychev and Sennrich (2019) show that forwardand backward-translations (FT/ BT) could both boost the model performances, where FT plays the role of domain adaptation and BT makes the translation fluent. Fadaee and Monz (2018) sample the monolingual data with more difficult words (e.g. rare words) to perform BT, achieving significant improvements compared with randomly sampled BT. Nguyen et al. (2020) diversify the data by applying FT and BT multiply times. However, different from AT, the prerequisite of training a well-performed NAT model is to perform KD. We compared with related works in Table 10 and found that our approach consistently outperforms them. Note that all the ablation studies focus on exploiting the parallel data without augmenting additional data. Non-Autoregressive Translation A variety of approaches have been exploited to bridge the performance gap between NAT and AT models. Some researchers proposed new model architectures (Lee et al., 2018; Ghazvininejad et al., 2019; Gu et al., 2019; Kasai et al., 2020), aided with additional signals (Wang et al., 2019; Ran et al., 2019; Ding et al., 2020), introduced sequential information (Wei et al., 2019; Shao et al., 2019; Guo et al., 2020; Hao et al., 2021), and explored advanced training objectives (Ghazvininejad et al., 2020; Du et al., 2021). Our work is close to the research line on training methods. Ding et al. (2021b) revealed the low-frequency word problem in distilled training data, and introduced an extra Kullback-Leibler divergence term derived by comparing the lexical choice of NAT model and that embedded in the raw data. Ding et al. (2021a) propose a simple and effective training strategy, which progressively feeds different granularity of data into NAT models by leveraging curriculum learning. 5 Conclusion In this study, we propose simple and effective training strategies to rejuvenate the low-frequency information in the raw data. Experiments show that our approach consistently and significantly improves translation performance across language pairs and model architectures. Notably, domain shift is an extreme scenario to diagnose low-frequency translation, and our method significant improves them. Extensive analyses reveal that our method improves the accuracy of lexical choices for low-frequency source words, recalling more low-frequency words in translations as well, which confirms our claim. Acknowledgments We are grateful to the anonymous reviewers and the area chair for their insightful comments and suggestions. Xuebo Liu and Derek F. Wong were supported in part by the Science and Technology Development Fund, Macau SAR (Grant No. 0101/2019/A2), and the Multi-year Research Grant from the University of Macau (Grant No. MYRG2020-00054-FST). References Philip Arthur, Graham Neubig, and Satoshi Nakamura. 2016. Incorporating discrete translation lexicons into neural machine translation. In EMNLP. Nikolay Bogoychev and Rico Sennrich. 2019. Domain, translationese and noise in synthetic data for neural machine translation. ArXiv. Michael Collins, Philipp Koehn, and Ivona Kuˇcerov´a. 2005. Clause restructuring for statistical machine translation. In ACL. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL. Liang Ding, Longyue Wang, Xuebo Liu, Derek F Wong, Dacheng Tao, and Zhaopeng Tu. 2021a. Progressive multi-granularity training for nonautoregressive translation. In ACL. Liang Ding, Longyue Wang, Xuebo Liu, Derek F. Wong, Dacheng Tao, and Zhaopeng Tu. 2021b. Understanding and improving lexical choice in nonautoregressive translation. In ICLR. 3440 Liang Ding, Longyue Wang, Di Wu, Dacheng Tao, and Zhaopeng Tu. 2020. Context-aware cross-attention for non-autoregressive translation. In COLING. Cunxiao Du, Zhaopeng Tu, and Jing Jiang. 2021. Order-agnostic cross entropy for non-autoregressive machine translation. In ICML. Chris Dyer, Victor Chahuneau, and Noah A Smith. 2013. A simple, fast, and effective reparameterization of ibm model 2. In NAACL. Sergey Edunov, Myle Ott, Michael Auli, and David Grangier. 2018. Understanding back-translation at scale. In EMNLP. Marzieh Fadaee and Christof Monz. 2018. Backtranslation sampling by targeting difficult words in neural machine translation. In EMNLP. Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tieyan Liu. 2018. Representation degeneration problem in training natural language generation models. In ICLR. Marjan Ghazvininejad, Vladimir Karpukhin, Luke Zettlemoyer, and Omer Levy. 2020. Aligned cross entropy for non-autoregressive machine translation. In ICML. Marjan Ghazvininejad, Omer Levy, Yinhan Liu, and Luke Zettlemoyer. 2019. Mask-predict: Parallel decoding of conditional masked language models. In EMNLP. Chengyue Gong, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2018. Frage: Frequency-agnostic word representation. NeurIPS. Jiatao Gu, James Bradbury, Caiming Xiong, Victor OK Li, and Richard Socher. 2018. Non-autoregressive neural machine translation. In ICLR. Jiatao Gu, Changhan Wang, and Junbo Zhao. 2019. Levenshtein transformer. In NeurIPS. Shuhao Gu, Jinchao Zhang, Fandong Meng, Yang Feng, Wanying Xie, Jie Zhou, and Dong Yu. 2020. Token-level adaptive training for neural machine translation. In EMNLP. Junliang Guo, Xu Tan, Linli Xu, Tao Qin, Enhong Chen, and Tie-Yan Liu. 2020. Fine-tuning by curriculum learning for non-autoregressive neural machine translation. In AAAI. Yongchang Hao, Shilin He, Wenxiang Jiao, Zhaopeng Tu, Michael Lyu, and Xing Wang. 2021. Multi-task learning with shared encoder for non-autoregressive machine translation. In NAACL. Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, William Lewis, Mu Li, et al. 2018. Achieving human parity on automatic chinese to english news translation. arXiv. Dan Hendrycks, Kimin Lee, and Mantas Mazeika. 2019. Using pre-training can improve model robustness and uncertainty. In ICML. Wenxiang Jiao, Xing Wang, Shilin He, Irwin King, Michael R. Lyu, and Zhaopeng Tu. 2020. Data rejuvenation: Exploiting inactive training examples for neural machine translation. In EMNLP. Jungo Kasai, James Cross, Marjan Ghazvininejad, and Jiatao Gu. 2020. Parallel machine translation with disentangled context transformer. In arXiv. Yoon Kim and Alexander M Rush. 2016. Sequencelevel knowledge distillation. In EMNLP. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR. Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In WMT. Jason Lee, Elman Mansimov, and Kyunghyun Cho. 2018. Deterministic non-autoregressive neural sequence modeling by iterative refinement. In EMNLP. Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. 2021. Pretraining boosts out-of-domain robustness for pose estimation. In WACV. Makoto Morishita, Jun Suzuki, and Masaaki Nagata. 2017. Ntt neural machine translation systems at wat 2017. In IJCNLP. Mathias M¨uller, Annette Rios, and Rico Sennrich. 2020. Domain Robustness in Neural Machine Translation. In AMTA. Toan Nguyen and David Chiang. 2018. Improving lexical choice in neural machine translation. In NAACL. Xuan-Phi Nguyen, Joty Shafiq, Kui Wu, and Ai Ti Aw. 2020. Data diversification: A simple strategy for neural machine translation. In NeurIPS. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational linguistics. Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. 2018. Analyzing uncertainty in neural machine translation. In ICML. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL. Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabas Poczos, and Tom Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In NAACL. Qiu Ran, Yankai Lin, Peng Li, and Jie Zhou. 2019. Guiding non-autoregressive neural machine translation decoding with reordering information. arXiv. 3441 Yi Ren, Jinglin Liu, Xu Tan, Zhou Zhao, Sheng Zhao, and Tie-Yan Liu. 2020. A study of nonautoregressive model for sequence generation. In ACL. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In ACL. Chenze Shao, Jinchao Zhang, Yang Feng, Fandong Meng, and Jie Zhou. 2019. Minimizing the bag-ofngrams difference for non-autoregressive neural machine translation. In AAAI. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS. Yiren Wang, Fei Tian, Di He, Tao Qin, ChengXiang Zhai, and Tie-Yan Liu. 2019. Non-autoregressive machine translation with auxiliary regularization. In AAAI. Bingzhen Wei, Mingxuan Wang, Hao Zhou, Junyang Lin, and Xu Sun. 2019. Imitation learning for nonautoregressive neural machine translation. In ACL. Lijun Wu, Yiren Wang, Yingce Xia, QIN Tao, Jianhuang Lai, and Tie-Yan Liu. 2019. Exploiting monolingual data at scale for neural machine translation. In EMNLP. Chunting Zhou, Graham Neubig, and Jiatao Gu. 2020. Understanding knowledge distillation in nonautoregressive machine translation. In ICLR. George K. Zipf. 1949. Human behavior and the principle of least effort.
|
2021
|
266
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3442–3455 August 1–6, 2021. ©2021 Association for Computational Linguistics 3442 G-Transformer for Document-level Machine Translation Guangsheng Bao1,2, Yue Zhang∗,1,2, Zhiyang Teng1,2, Boxing Chen3 and Weihua Luo3 1 School of Engineering, Westlake University 2 Institute of Advanced Technology, Westlake Institute for Advanced Study 3 DAMO Academy, Alibaba Group Inc. {baoguangsheng, zhangyue, tengzhiyang}@westlake.edu.cn {boxing.cbx, weihua.luowh}@alibaba-inc.com Abstract Document-level MT models are still far from satisfactory. Existing work extend translation unit from single sentence to multiple sentences. However, study shows that when we further enlarge the translation unit to a whole document, supervised training of Transformer can fail. In this paper, we find such failure is not caused by overfitting, but by sticking around local minima during training. Our analysis shows that the increased complexity of target-to-source attention is a reason for the failure. As a solution, we propose G-Transformer, introducing locality assumption as an inductive bias into Transformer, reducing the hypothesis space of the attention from target to source. Experiments show that G-Transformer converges faster and more stably than Transformer, achieving new state-of-the-art BLEU scores for both nonpretraining and pre-training settings on three benchmark datasets. 1 Introduction Document-level machine translation (MT) has received increasing research attention (Gong et al., 2011; Hardmeier et al., 2013; Garcia et al., 2015; Miculicich et al., 2018a; Maruf et al., 2019; Liu et al., 2020). It is a more practically useful task compared to sentence-level MT because typical inputs in MT applications are text documents rather than individual sentences. A salient difference between document-level MT and sentence-level MT is that for the former, much larger inter-sentential context should be considered when translating each sentence, which include discourse structures such as anaphora, lexical cohesion, etc. Studies show that human translators consider such contexts when conducting document translation (Hardmeier, 2014; L¨aubli et al., 2018). Despite that neural models achieve competitive performances on sentence∗* Corresponding author. Context Encoder Source Encoder Target Decoder Source Translation Context (a) Sentence-by-sentence Translation Source Encoder Target Decoder Source1 Translation1 Source2 … Translation2 … (b) Multi-sentence Translation Source1 Translation1 Source2 … Translation2 … … … ① ② ③ ① ② ③ (c) G-Transformer (Doc-by-doc Translation) Figure 1: Overview of model structures for documentlevel machine translation. level MT, the performance of document-level MT is still far from satisfactory. Existing methods can be mainly classified into two categories. The first category translates a document sentence by sentence using a sequence-tosequence neural model (Zhang et al., 2018; Miculicich et al., 2018b; Maruf et al., 2019; Zheng et al., 2020). Document-level context is integrated into sentence-translation by introducing additional context encoder. The structure of such a model is shown in Figure 1(a). These methods suffer from two limitations. First, the context needs to be encoded separately for translating each sentence, which adds to the runtime complexity. Second, more importantly, information exchange cannot be made between the current sentence and its document context in the same encoding module. The second category extends the translation unit from a single sentence to multiple sentences (Tiedemann and Scherrer, 2017; Agrawal et al., 3443 2018; Zhang et al., 2020) and the whole document (Junczys-Dowmunt, 2019; Liu et al., 2020). Recently, it has been shown that when the translation unit increases from one sentence to four sentences, the performance improves (Zhang et al., 2020; Scherrer et al., 2019). However, when the whole document is encoded as a single unit for sequence to sequence translation, direct supervised training has been shown to fail (Liu et al., 2020). As a solution, either large-scale pre-training (Liu et al., 2020) or data augmentation (Junczys-Dowmunt, 2019) has been used as a solution, leading to improved performance. These methods are shown in Figure 1(b). One limitation of such methods is that they require much more training time due to the necessity of data augmentation. Intuitively, encoding the whole input document as a single unit allows the best integration of context information when translating the current sentence. However, little work has been done investigating the underlying reason why it is difficult to train such a document-level NMT model. One remote clue is that as the input sequence grows larger, the input becomes more sparse (Pouget-Abadie et al., 2014; Koehn and Knowles, 2017). To gain more understanding, we make dedicated experiments on the influence of input length, data scale and model size for Transformer (Section 3), finding that a Transformer model can fail to converge when training with long sequences, small datasets, or big model size. We further find that for the failed cases, the model gets stuck at local minima during training. In such situation, the attention weights from the decoder to the encoder are flat, with large entropy values. This can be because that larger input sequences increase the challenge for focusing on a local span to translate when generating each target word. In other words, the hypothesis space for target-to-source attention is increased. Given the above observations, we investigate a novel extension of Transformer, by restricting selfattention and target-to-source attention to a local context using a guidance mechanism. As shown in Figure 1(c), while we still encode the input document as a single unit, group tags 1⃝ 2⃝ 3⃝are assigned to sentences to differentiate their positions. Target-to-source attention is guided by matching the tag of target sentence to the tags of source sentences when translating each sentence, so that the hypothesis space of attention is reduced. Intuitively, the group tags serve as a constraint on attention, which is useful for differentiating the current sentence and its context sentences. Our model, named G-Transformer, can be thus viewed as a combination of the method in Figure 1(a) and Figure 1(b), which fully separate and fully integrates a sentence being translated with its document level context, respectively. We evaluate our model on three commonly used document-level MT datasets for EnglishGerman translation, covering domains of TED talks, News, and Europarl from small to large. Experiments show that G-Transformer converges faster and more stably than Transformer on different settings, obtaining the state-of-the-art results under both non-pretraining and pre-training settings. To our knowledge, we are the first to realize a truly document-by-document translation model. We release our code and model at https://github.com/baoguangsheng/g-transformer. 2 Experimental Settings We evaluate Transformer and G-Transformer on the widely adopted benchmark datasets (Maruf et al., 2019), including three domains for EnglishGerman (En-De) translation. TED. The corpus is transcriptions of TED talks from IWSLT 2017. Each talk is used as a document, aligned at the sentence level. tst2016-2017 is used for testing, and the rest for development. News. This corpus uses News Commentary v11 for training, which is document-delimited and sentence-aligned. newstest2015 is used for development, and newstest2016 for testing. Europarl. The corpus is extracted from Europarl v7, where sentences are segmented and aligned using additional information. The train, dev and test sets are randomly split from the corpus. The detailed statistics of these corpora are shown in Table 1. We pre-process the documents by splitting them into instances with up-to 512 tokens, taking a sentence as one instance if its length exceeds 512 tokens. We tokenize and truecase the sentences with MOSES (Koehn et al., 2007) tools, applying BPE (Sennrich et al., 2016) with 30000 merging operations. We consider three standard model configurations. Base Model. Following the standard Transformer base model (Vaswani et al., 2017), we use 6 layers, 8 heads, 512 dimension outputs, and 2048 3444 Language Dataset #Sentences #Documents #Instances Avg #Sents/Inst Avg #Tokens/Inst train/dev/test train/dev/test train/dev/test train/dev/test train/dev/test En-De TED 0.21M/9K/2.3K 1.7K/92/22 11K/483/123 18.3/18.5/18.3 436/428/429 News 0.24M/2K/3K 6K/80/154 18.5K/172/263 12.8/12.6/11.3 380/355/321 Europarl 1.67M/3.6K/5.1K 118K/239/359 162K/346/498 10.3/10.4/10.3 320/326/323 Table 1: En-De datasets for evaluation. -5 0 5 10 15 20 25 30 35 d-BLEU Tokens 64 128 256 512 1024 (a) Input Length (Base model with filtered data.) -5 0 5 10 15 20 25 30 35 d-BLEU Instances 1.25K 2.5K 5K 10K 20K 40K 80K 160K (b) Data Scale (Base model with 512 tokens input.) Figure 2: Transformer on various input length and data scale. dimension hidden vectors. Big Model. We follow the standard Transformer big model (Vaswani et al., 2017), using 6 layers, 16 heads, 1024 dimension outputs, and 4096 dimension hidden vectors. Large Model. We use the same settings of BART large model (Lewis et al., 2020), which involves 12 layers, 16 heads, 1024 dimension outputs, and 4096 dimension hidden vectors. We use s-BLEU and d-BLEU (Liu et al., 2020) as the metrics. The detailed descriptions are in Appendix A. 3 Transformer and Long Inputs We empirically study Transformer (see Appendix B) on the datasets. We run each experiment five times using different random seeds, reporting the average score for comparison. 3.1 Failure Reproduction Input Length. We use the Base model and fixed dataset for this comparison. We split both the training and testing documents from Europarl dataset into instances with input length of 64, 128, 256, 512, and 1024 tokens, respectively. For fair comparison, we remove the training documents with a length of less than 768 tokens, which may favour small input length. The results are shown in Figure 2a. When the input length increases from 256 tokens to 512 tokens, the BLEU score drops dramatically from 30.5 to 2.3, indicating failed training with 512 and 1024 tokens. It demonstrates the difficulty when dealing with long inputs of Trans2 4 6 8 10 12 0K 2K 4K 6K 8K 10K 12K Loss Steps Train Valid (a) Failed Model 2 4 6 8 10 12 0K 10K 20K 30K 40K 50K 60K Loss Steps Train Valid (b) Successful Model Figure 3: Loss curve of the models and the local minima. former. Data Scale. We use the Base model and a fixed input length of 512 tokens. For each setting, we randomly sample a training dataset of the expected size from the full dataset of Europarl. The results are shown in Figure 2b. The performance increases sharply when the data scale increases from 20K to 40K. When data scale is equal or less than 20K, the BLEU scores are under 3, which is unreasonably low, indicating that with a fixed model size and input length, the smaller dataset can also cause the failure of the training process. For data scale more than 40K, the BLEU scores show a wide dynamic range, suggesting that the training process is unstable. Model Size. We test Transformer with different model sizes, using the full dataset of Europarl and a fixed input length of 512 tokens. Transformer-Base can be trained successfully, giving a reasonable BLEU score. However, the training of the Big and Large models failed, resulting in very low BLEU scores under 3. It demonstrates that the increased model size can also cause the failure with a fixed input length and data scale. The results confirm the intuition that the performance will drop with longer inputs, smaller datasets, or bigger models. However, the BLEU scores show a strong discontinuity with the change of input length, data scale, or model size, falling into two discrete clusters. One is successfully trained cases with d-BLEU scores above 10, and the other is failed cases with d-BLEU scores under 3. 3445 7.6 7.8 8 8.2 8.4 0K 2K 4K 6K 8K 10K 12K Entropy (bit) Steps Train Valid (a) Failed Model 5 6 7 8 9 0K 10K 20K 30K 40K 50K 60K Entropy (bit) Steps Train Valid (b) Successful Model Figure 4: Cross-attention distribution of Transformer shows that the failed model sticks at the local minima. 4 5 6 7 8 9 0K 10K 20K 30K 40K 50K 60K Entropy (bit) Steps Train Valid (a) Encoder 4.6 5 5.4 5.8 6.2 6.6 7 0K 10K 20K 30K 40K 50K 60K Entropy (bit) Steps Train Valid (b) Decoder Figure 5: For the successful model, the attention distribution shrinks to narrow range (low entropy) and then expands to wider range (high entropy). 3.2 Failure Analysis Training Convergence. Looking into the failed models, we find that they have a similar pattern on loss curves. As an example of the model trained on 20K instances shown in Figure 3a, although the training loss continually decreases during training process, the validation loss sticks at the level of 7, reaching a minimum value at around 9K training steps. In comparison, the successfully trained models share another pattern. Taking the model trained on 40K instances as an example, the loss curves demonstrate two stages, which is shown in Figure 3b. In the first stage, the validation loss similar to the failed cases has a converging trend to the level of 7. In the second stage, after 13K training steps, the validation loss falls suddenly, indicating that the model may escape successfully from local minima. From the two stages of the learning curve, we conclude that the real problem, contradicting our first intuition, is not about overfitting, but about local minima. Attention Distribution. We further look into the attention distribution of the failed models, observing that the attentions from target to source are widely spread over all tokens. As Figure 4a shows, the distribution entropy is high for about 8.14 bits on validation. In contrast, as shown in Figure 4b, the successfully trained model has a much lower attention entropy of about 6.0 bits on validation. Furthermore, we can see that before 13K training Source: <s> the Commission shares ... of the European Union institutional framework . </s> 1 <s> Commission participation is expressly provided for ... of all its preparatory bodies . </s> 2 <s> only in exceptional circumstances ... be excluded from these meetings . </s> 3 ... Target: <s> die Kommission teilt die Ansicht ... des institutionellen Rahmens der Europischen Union ist . </s> 1 <s> die Geschftsordnung des Rates ... der Kommission damit ausdrcklich vor . </s> 2 <s> die Kommission kann nur ... wobei fallweise zu entscheiden ist . </s> 3 ... Figure 6: Example of English-German translation with group alignments. steps, the entropy sticks at a plateau, confirming with the observation of the local minima in Figure 3b. It indicates that the early stage of the training process for Transformer is difficult. Figure 5 shows the self-attention distributions of the successfully trained models. The attention entropy of both the encoder and the decoder drops fast at the beginning, leading to a shrinkage of the attention range. But then the attention entropy gradually increases, indicating an expansion of the attention range. Such back-and-forth oscillation of the attention range may also result in unstable training and slow down the training process. 3.3 Conclusion The above experiments show that training failure on Transformer can be caused by local minima. Additionally, the oscillation of attention range may make it worse. During training process, the attention module needs to identify relevant tokens from whole sequence to attend to. Assuming that the sequence length is N, the complexity of the attention distribution increases when N grows from sentence-level to document-level. We propose to use locality properties (Rizzi, 2013; Hardmeier, 2014; Jawahar et al., 2019) of both the language itself and the translation task as a constraint in Transformer, regulating the hypothesis space of the self-attention and target-to-source attention, using a simple group tag method. 4 G-Transformer An example of G-Transformer is shown in Figure 6, where the input document contains more than 3 sentences. As can be seen from the figure, G-Transformer extends Transformer by augmenting the input and output with group tags (Bao and Zhang, 2021). In particular, each token is assigned a group tag, indicating its sentential index. While 3446 source group tags can be assigned deterministically, target tags are assigned dynamically according to whether a generated sentence is complete. Starting from 1, target words copy group tags from its predecessor unless the previous token is </s>, in which case the tag increases by 1. The tags serve as a locality constraint, encouraging target-to-source attention to concentrate on the current source sentence being translated. Formally, for a source document X and a target document Y , the probability model of Transformer can be written as ˆY = arg max Y P(Y |X), (1) and G-Transformer extends it by having ˆY = argY max Y,GY P(Y, GY |X, GX), (2) where GX and GY denotes the two sequences of group tags GX = {gi = k if wi ∈sentX k else 0}||X| i=1, GY = {gj = k if wj ∈sentY k else 0}||Y | j=1, (3) where sentk represents the k-th sentence of X or Y . For the example shown in Figure 6, GX = {1, ..., 1, 2, ..., 2, 3, ..., 3, 4, ...} and GY = {1, ..., 1, 2, ..., 2, 3, ..., 3, 4, ...}. Group tags influence the auto-regressive translation process by interfering with the attention mechanism, which we show in the next section. In GTransformer, we use the group-tag sequence GX and GY for representing the alignment between X and Y , and for generating the localized contextual representation of X and Y . 4.1 Group Attention An attention module can be seen as a function mapping a query and a set of key-value pairs to an output (Vaswani et al., 2017). The query, key, value, and output are all vectors. The output is computed by summing the values with corresponding attention weights, which are calculated by matching the query and the keys. Formally, given a set of queries, keys, and values, we pack them into matrix Q, K, and V , respectively. We compute the matrix outputs Attention(Q, K, V ) = softmax QKT √dk V, (4) where dk is the dimensions of the key vector. Attention allows a model to focus on different positions. Further, multi-head attention (MHA) allows a model to gather information from different representation subspaces MHA(Q, K, V ) = Concat(head1, ..., headh)W O, headi = Attention(QW Q i , KW K i , V W V i ), (5) where the projections of W O, W Q i , W K i , and W V i are parameter matrices. We update Eq 4 using group-tags, naming it group attention (GroupAttn). In addition to inputs Q, K, and V , two sequences of group-tag inputs are involved, where GQ corresponds to Q and GK corresponds to K. We have args = (Q, K, V, GQ, GK), GroupAttn(args) = softmax QKT √dk + M(GQ, GK) V, (6) where function M(·) works as an attention mask, excluding all tokens outside the sentence. Specifically, M(·) gives a big negative number γ to make softmax close to 0 for the tokens with a different group tag compared to current token M(GQ, GK) = min(1, abs(GQIT K −IQGT K)) ∗γ, (7) where IK and IQ are constant vectors with value 1 on all dimensions, that IK has dimensions equal to the length of GK and IQ has dimensions equal to the length of GQ. The constant value γ can typically be −1e8. Similar to Eq 5, we use group multi-head attention args = (Q, K, V, GQ, GK), GroupMHA(args) = Concat(head1, ..., headh)W O, (8) where headi = GroupAttn(QW Q i , KW K i , V W V i , GQ, GK), (9) and the projections of W O, W Q i , W K i , and W V i are parameter matrices. Encoder. For each layer a group multi-head attention module is used for self-attention, assigning the same group-tag sequence for the key and the value that GQ = GK = GX. Decoder. We use one group multi-head attention module for self-attention and another group multihead attention module for cross-attention. Similar to the encoder, we assign the same group-tag sequence to the key and value of the self-attention, that GQ = GK = GY , but use different group-tag sequences for cross-attention that GQ = GY and GK = GX. 3447 Method TED News Europarl s-BLEU d-BLEU s-BLEU d-BLEU s-BLEU d-BLEU SENTNMT (Vaswani et al., 2017) 23.10 22.40 29.40 HAN (Miculicich et al., 2018b) 24.58 25.03 28.60 SAN (Maruf et al., 2019) 24.42 24.84 29.75 Hybrid Context (Zheng et al., 2020) 25.10 24.91 30.40 Flat-Transformer (Ma et al., 2020) 24.87 23.55 30.09 Transformer on sent (baseline) 24.82 25.19 31.37 Transformer on doc (baseline) 0.76 0.60 33.10 G-Transformer random initialized (ours) 23.53 25.84* 23.55 25.23* 32.18* 33.87* G-Transformer fine-tuned on sent Transformer (ours) 25.12 27.17* 25.52 27.11* 32.39* 34.08* Fine-tuning on Pre-trained Model Flat-Transformer+BERT (Ma et al., 2020) 26.61 24.52 31.99 G-Transformer+BERT (ours) 26.81 26.14 32.46 Transformer on sent fine-tuned on BART (baseline) 27.78 29.90 31.87 Transformer on doc fine-tuned on BART (baseline) 28.29 30.49 34.00 G-Transformer fine-tuned on BART (ours) 28.06 30.03* 30.34* 31.71* 32.74* 34.31* Table 2: Case-sensitive BLEU scores on En-De translation. “*” indicates statistically significant at p < 0.01 compared to the Transformer baselines. Complexity. Consider a document with M sentences and N tokens, where each sentence contains N/M tokens on average. The complexities of both the self-attention and cross-attention in Transformer are O(N2). In contrast, the complexity of group attention in G-Transformer is O(N2/M) given the fact that the attention is restricted to a local sentence. Theoretically, since the average length N/M of sentences tends to be constant, the time and memory complexities of group attention are approximately O(N), making training and inference on very long inputs feasible. 4.2 Combined Attention We use only group attention on lower layers for local sentence representation, and combined attention on top layers for integrating local and global context information. We use the standard multihead attention in Eq 5 for global context, naming it global multi-head attention (GlobalMHA). Group multi-head attention in Eq 8 and global multi-head attention are combined using a gate-sum module (Zhang et al., 2016; Tu et al., 2017) HL = GroupMHA(Q, K, V, GQ, GK), HG = GlobalMHA(Q, K, V ), g = sigmoid([HL, HG]W + b), H = HL ⊙g + HG ⊙(1 −g), (10) where W and b are linear projection parameters, and ⊙denotes element-wise multiplication. Previous study (Jawahar et al., 2019) shows that the lower layers of Transformer catch more local syntactic relations, while the higher layers represent longer distance relations. Based on these findings, we use combined attention only on the top layers for integrating local and global context. By this design, on lower layers, the sentences are isolated from each other, while on top layers, the crosssentence interactions are enabled. Our experiments show that the top 2 layers with global attention are sufficient for document-level NMT, and more layers neither help nor harm the performance. 4.3 Inference During decoding, we generate group-tag sequence GY according to the predicted token, starting with 1 at the first <s> and increasing 1 after each </s>. We use beam search and apply the maximum length constraint on each sentence. We generate the whole document from start to end in one beam search process, using a default beam size of 5. 5 G-Transformer Results We compare G-Transformer with Transformer baselines and previous document-level NMT models on both non-pretraining and pre-training settings. The detailed descriptions about these training settings are in Appendix C.1. We make statistical significance test according to Collins et al. (2005). 5.1 Results on Non-pretraining Settings As shown in Table 2, the sentence-level Transformer outperforms previous document-level models on News and Europarl. Compared to this strong baseline, our randomly initialized model of G-Transformer improves the s-BLEU by 0.81 point on the large dataset Europarl. The results on the small datasets TED and News are worse, indicating overfitting with long inputs. When GTransformer is trained by fine-tuning the sentence3448 level Transformer, the performance improves on the three datasets by 0.3, 0.33, and 1.02 s-BLEU points, respectively. Different from the baseline of document-level Transformer, G-Transformer can be successfully trained on small TED and News. On Europarl, G-Transformer outperforms Transformer by 0.77 d-BLEU point, and G-Transformer fine-tuned on sentence-level Transformer enlarges the gap to 0.98 d-BLEU point. G-Transformer outperforms previous documentlevel MT models on News and Europarl with a significant margin. Compared to the best recent model Hyrbid-Context, G-Transformer improves the s-BLEU on Europarl by 1.99. These results suggest that in contrast to previous short-context models, sequence-to-sequence model taking the whole document as input is a promising direction. 5.2 Results on Pre-training Settings There is relatively little existing work about document-level MT using pre-training. Although Flat-Transformer+BERT gives a state-of-the-art scores on TED and Europarl, the score on News is worse than previous non-pretraining model HAN (Miculicich et al., 2018b). G-Transformer+BERT improves the scores by margin of 0.20, 1.62, and 0.47 s-BLEU points on TED, News, and Europarl, respectively. It shows that with a better contextual representation, we can further improve documentlevel MT on pretraining settings. We further build much stronger Transformer baselines by fine-tuning on mBART25 (Liu et al., 2020). Taking advantage of sequence-to-sequence pre-training, the sentence-level Transformer gives much better s-BLEUs of 27.78, 29.90, and 31.87, respectively. G-Transformer fine-tuned on mBART25 improves the performance by 0.28, 0.44, and 0.87 s-BLEU, respectively. Compared to the document-level Transformer baseline, GTransformer gives 1.74, 1.22, and 0.31 higher d-BLEU points, respectively. It demonstrates that even with well-trained sequence-to-sequence model, the locality bias can still enhance the performance. 5.3 Convergence We evaluate G-Transformer ad Transformer on various input length, data scale, and model size to better understand that to what extent it has solved the convergence problem of Transformer. -5 5 15 25 35 45 d-BLEU Tokens 64 128 256 512 1024 Transformer G-Transformer (a) Input Length -5 5 15 25 35 45 d-BLEU Instances 1.25K 2.5K 5K 10K 20K 40K 80K 160K Transformer G-Transformer (b) Data Scale Figure 7: G-Transformer compared with Transformer. 3 4 5 6 7 8 9 0K 10K 20K 30K 40K 50K 60K Entropy (bit) Steps Transformer Group Attention Global Attention (a) Cross-Attention 3 4 5 6 7 8 9 0K 10K 20K 30K 40K 50K 60K Entropy (bit) Steps Transformer Group Attention Global Attention (b) Encoder Self-Attention Figure 8: Comparison on the development of crossattention and encoder self-attention. Input Length. The results are shown in Figure 7a. Unlike Transformer, which fails to train on long input, G-Transformer shows stable scores for inputs containing 512 and 1024 tokens, suggesting that with the help of locality bias, a long input does not impact the performance obviously. Data Scale. As shown in Figure 7b, overall GTransformer has a smooth curve of performance on the data scale from 1.25K to 160K. The variances of the scores are much lower than Transformer, indicating stable training of G-Transformer. Additionally, G-Transformer outperforms Transformer by a large margin on all the settings. Model Size. Unlike Transformer, which fails to train on Big and Large model settings, GTransformer shows stable scores on different model sizes. As shown in Appendix C.2, although performance on small datasets TED and News drops largely for Big and Large model, the performance on large dataset Europarl only decreases by 0.10 d-BLEU points for the Big model and 0.66 for the Large model. Loss. Looking into the training process of the above experiments, we see that both the training and validation losses of G-Transformer converge much faster than Transformer, using almost half time to reach the same level of loss. Furthermore, the validation loss of G-Transformer converges to much lower values. These observations demonstrate that G-Transformer converges faster and better. Attention Distribution. Benefiting from the separate group attention and global attention, GTransformer avoids the oscillation of attention 3449 Method TED News Europarl Drop G-Transformer (fnt.) 25.12 25.52 32.39 - target-side context 25.05 25.41 32.16 -0.14 - source-side context 24.56 24.58 31.39 -0.70 Table 3: Impact of source-side and target-side context reporting in s-BLEU. Here, fnt. denotes the model finetuned on sentence-level Transformer. Method deixis el.infl. el.VP CADec (Voita et al., 2019b) 81.6 72.2 80.0 LSTM-Tran (Zhang et al., 2020) 91.0 82.2 78.2 sent (Voita et al., 2019b) 50.0 53.0 28.4 concat (Voita et al., 2019b) 83.5 76.2 76.6 G-Transformer 89.9 84.8 82.4 Table 4: Impact on discourse by the source-side context, in accuracy of correctly identifying the discourse phenomena. Here, el. means ellipsis. LSTM-Tran denotes LSTM-Transformer. range, which happens to Transformer. As shown in Figure 8a, Transformer sticks at the plateau area for about 13K training steps, but G-Transformer shows a quick and monotonic convergence, reaching the stable level using about 1/4 of the time that Transformer takes. Through Figure 8b, we can find that G-Transformer also has a smooth and stable curve for the convergence of self-attention distribution. These observations imply that the potential conflict of local sentence and document context can be mitigated by G-Transformer. 5.4 Discussion of G-Transformer Document Context. We study the contribution of the source-side and target-side context by removing the cross-sentential attention in Eq 10 from the encoder and the decoder gradually. The results are shown in Table 3. We take the G-Transformer fine-tuned on the sentence-level Transformer as our starting point. When we disable the targetside context, the performance decreases by 0.14 s-BLEU point on average, which indicates that the target-side context does impact translation performance significantly. When we further remove the source-side context, the performance decrease by 0.49, 0.83, and 0.77 s-BLEU point on TED, News, and Europarl, respectively, which indicates that the source-side context is relatively more important for document-level MT. To further understand the impact of the sourceside context, we conduct an experiment on automatic evaluation on discourse phenomena which rely on source context. We use the human labeled evaluation set (Voita et al., 2019b) on EnglishMethod TED News Europarl Drop G-Transformer (rnd.) 25.84 25.23 33.87 - word-dropout 25.49 24.65 33.70 -0.37 - language locality 22.47 22.41 33.63 -1.78 - translation locality 0.76 0.60 33.10 -14.68 Table 5: Contribution of locality bias and word-dropout reporting in d-BLEU. Here, rnd. denotes the model trained using randomly initialized parameters. Method TED News Europarl Drop G-Transformer (rnd.) Combined attention 25.84 25.23 33.87 Only group attention 25.62 25.14 33.12 -0.35 Only global attention 25.00 24.54 32.87 -0.84 Table 6: Separate effect of group and global attention reporting in d-BLEU. Here, rnd. denotes the model trained using randomly initialized parameters. Russion (En-Ru) for deixis and ellipsis. We follow the Transformer concat baseline (Voita et al., 2019b) and use both 6M sentence pairs and 1.5M document pairs from OpenSubtitles2018 (Lison et al., 2018) to train our model. The results are shown in Table 4. G-Transformer outperforms Transformer baseline concat (Voita et al., 2019b) with a large margin on three discourse features, indicating a better leverage of the source-side context. When compared to previous model LSTM-T, G-Transformer achieves a better ellipsis on both infl. and VP. However, the score on deixis is still lower, which indicates a potential direction that we can investigate in further study. Word-dropout. As shown in Table 5, worddropout (Appendix C.1) contributes about 0.37 dBLEU on average. Its contribution to TED and News is obvious in 0.35 and 0.58 d-BLEU, respectively. However, for large dataset Europarl, the contribution drops to 0.17, suggesting that with sufficient data, word-dropout may not be necessary. Locality Bias. In G-Transformer, we introduce locality bias to the language modeling of source and target, and locality bias to the translation between source and target. We try to understand these biases by removing them from G-Transformer. When all the biases removed, the model downgrades to a document-level Transformer. The results are shown in Table 5. Relatively speaking, the contribution of language locality bias is about 1.78 d-BLEU on average. While the translation locality bias contributes for about 14.68 d-BLEU on average, showing critical impact on the model convergence on small datasets. These results suggest that the locality bias may be the key to train 3450 whole-document MT models, especially when the data is insufficient. Combined Attention. In G-Transformer, we enable only the top K layers with combined attention. On Europarl7, G-Transformer gives 33.75, 33.87, and 33.84 d-BLEU with top 1, 2, and 3 layers with combined attention, respectively, showing that K = 2 is sufficient. Furthermore, we study the effect of group and global attention separately. As shown in Table 6, when we replace the combined attention on top 2 layers with group attention, the performance drops by 0.22, 0.09, and 0.75 d-BLEU on TED, News, and Europarl, respectively. When we replace the combined attention with global attention, the performance decrease is enlarged to 0.84, 0.69, and 1.00 d-BLEU, respectively. These results demonstrate the necessity of combined attention for integrating local and global context information. 6 Related Work The unit of translation has evolved from word (Brown et al., 1993; Vogel et al., 1996) to phrase (Koehn et al., 2003; Chiang, 2005, 2007) and further to sentence (Kalchbrenner and Blunsom, 2013; Sutskever et al., 2014; Bahdanau et al., 2014) in the MT literature. The trend shows that larger units of translation, when represented properly, can lead to improved translation quality. A line of document-level MT extends translation unit to multiple sentences (Tiedemann and Scherrer, 2017; Agrawal et al., 2018; Zhang et al., 2020; Ma et al., 2020). However, these approaches are limited within a short context of maximum four sentences. Recent studies extend the translation unit to whole document (Junczys-Dowmunt, 2019; Liu et al., 2020), using large augmented dataset or pretrained models. Liu et al. (2020) shows that Transformer trained directly on documentlevel dataset can fail, resulting in unreasonably low BLEU scores. Following these studies, we also model translation on the whole document. We solve the training challenge using a novel locality bias with group tags. Another line of work make document-level machine translation sentence by sentence, using additional components to represent the context (Maruf and Haffari, 2018; Zheng et al., 2020; Zhang et al., 2018; Miculicich et al., 2018b; Maruf et al., 2019; Yang et al., 2019). Different from these approaches, G-Transformer uses a generic design for both source and context, translating whole document in one beam search instead of sentence-by-sentence. Some methods use a two-pass strategy, generating sentence translation first, integrating context information through a post-editing model (Voita et al., 2019a; Yu et al., 2020). In contrast, G-Transformer uses a single model, which reduces the complexity for both training and inference. The locality bias we introduce to G-Transformer is different from the ones in Longformer (Beltagy et al., 2020) and Reformer (Kitaev et al., 2020) in the sense that we discuss locality in the context of representing the alignment between source sentences and target sentences in document-level MT. Specifically, Longformer introduces locality only to self-attention, while G-Transformer also introduces locality to cross-attention, which is shown to be the key for the success of G-Transformer. Reformer, basically same as Transformer, searches for attention targets in the whole sequence, while G-Transformer mainly restricts the attention inside a local sentence. In addition, the motivations are different. While Longformer and Reformer focus on the time and memory complexities, we focus on attention patterns in cases where a translation model fails to converge during training. 7 Conclusion We investigated the main reasons for Transformer training failure in document-level MT, finding that target-to-source attention is a key factor. According to the observation, we designed a simple extension of the standard Transformer architecture, using group tags for attention guiding. Experiments show that the resulting G-Transformer converges fast and stably on small and large data, giving the state-of-the-art results compared to existing models under both pre-training and random initialization settings. Acknowledgments We would like to thank the anonymous reviewers for their valuable feedback. We thank Westlake University High-Performance Computing Center for supporting on GPU resources. This work is supported by grants from Alibaba Group Inc. and Sichuan Lan-bridge Information Technology Co.,Ltd. 3451 References Ruchit Rajeshkumar Agrawal, Marco Turchi, and Matteo Negri. 2018. Contextual handling in neural machine translation: Look behind, ahead and on both sides. In 21st Annual Conference of the European Association for Machine Translation, pages 11–20. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. Guangsheng Bao and Yue Zhang. 2021. Contextualized rewriting for text summarization. In The Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv:2004.05150. Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew Dai, Rafal Jozefowicz, and Samy Bengio. 2016. Generating sentences from a continuous space. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 10–21, Berlin, Germany. Association for Computational Linguistics. Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263– 311. David Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 263–270, Ann Arbor, Michigan. Association for Computational Linguistics. David Chiang. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Michael Collins, Philipp Koehn, and Ivona Kuˇcerov´a. 2005. Clause restructuring for statistical machine translation. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL05), pages 531–540. Eva Mart´ınez Garcia, Cristina Espa˜na-Bonet, and Llu´ıs M`arquez. 2015. Document-level machine translation with word vector models. In Proceedings of the 18th Annual Conference of the European Association for Machine Translation, pages 59–66, Antalya, Turkey. Zhengxian Gong, Min Zhang, and Guodong Zhou. 2011. Cache-based document-level statistical machine translation. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 909–919, Edinburgh, Scotland, UK. Association for Computational Linguistics. Christian Hardmeier. 2014. Discourse in statistical machine translation. Ph.D. thesis, Acta Universitatis Upsaliensis. Christian Hardmeier, Sara Stymne, J¨org Tiedemann, and Joakim Nivre. 2013. Docent: A document-level decoder for phrase-based statistical machine translation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 193–198, Sofia, Bulgaria. Association for Computational Linguistics. Ganesh Jawahar, Benoˆıt Sagot, and Djam´e Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651–3657, Florence, Italy. Association for Computational Linguistics. Marcin Junczys-Dowmunt. 2019. Microsoft translator at WMT 2019: Towards large-scale document-level neural machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 225–233, Florence, Italy. Association for Computational Linguistics. Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1700–1709, Seattle, Washington, USA. Association for Computational Linguistics. Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. In International Conference on Learning Representations. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague, Czech Republic. Association for Computational Linguistics. Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39, Vancouver. Association for Computational Linguistics. Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics, pages 127–133. Samuel L¨aubli, Rico Sennrich, and Martin Volk. 2018. Has machine translation achieved human parity? a case for document-level evaluation. In Proceedings of the 2018 Conference on Empirical Methods 3452 in Natural Language Processing, pages 4791–4796, Brussels, Belgium. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Pierre Lison, J¨org Tiedemann, and Milen Kouylekov. 2018. OpenSubtitles2018: Statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726–742. Shuming Ma, Dongdong Zhang, and Ming Zhou. 2020. A simple and effective unified encoder for documentlevel machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3505–3511, Online. Association for Computational Linguistics. Sameen Maruf and Gholamreza Haffari. 2018. Document context neural machine translation with memory networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1275– 1284, Melbourne, Australia. Association for Computational Linguistics. Sameen Maruf, Andr´e F. T. Martins, and Gholamreza Haffari. 2019. Selective attention for context-aware neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3092–3102, Minneapolis, Minnesota. Association for Computational Linguistics. Lesly Miculicich, Dhananjay Ram, Nikolaos Pappas, and James Henderson. 2018a. Document-level neural machine translation with hierarchical attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2947–2954, Brussels, Belgium. Association for Computational Linguistics. Lesly Miculicich, Dhananjay Ram, Nikolaos Pappas, and James Henderson. 2018b. Document-level neural machine translation with hierarchical attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2947–2954, Brussels, Belgium. Association for Computational Linguistics. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53, Minneapolis, Minnesota. Association for Computational Linguistics. Jean Pouget-Abadie, Dzmitry Bahdanau, Bart van Merri¨enboer, Kyunghyun Cho, and Yoshua Bengio. 2014. Overcoming the curse of sentence length for neural machine translation using automatic segmentation. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 78–85, Doha, Qatar. Association for Computational Linguistics. Luigi Rizzi. 2013. Locality. Lingua, 130:169–186. Yves Scherrer, J¨org Tiedemann, and Sharid Lo´aiciga. 2019. Analysing concatenation approaches to document-level NMT in two different domains. In Proceedings of the Fourth Workshop on Discourse in Machine Translation (DiscoMT 2019), pages 51–61, Hong Kong, China. Association for Computational Linguistics. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725, Berlin, Germany. Association for Computational Linguistics. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215. J¨org Tiedemann and Yves Scherrer. 2017. Neural machine translation with extended context. In Proceedings of the Third Workshop on Discourse in Machine Translation, pages 82–92, Copenhagen, Denmark. Association for Computational Linguistics. Zhaopeng Tu, Yang Liu, Zhengdong Lu, Xiaohua Liu, and Hang Li. 2017. Context gates for neural machine translation. Transactions of the Association for Computational Linguistics, 5:87–99. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6000–6010. Stephan Vogel, Hermann Ney, and Christoph Tillmann. 1996. HMM-based word alignment in statistical translation. In COLING 1996 Volume 2: The 16th International Conference on Computational Linguistics. 3453 Elena Voita, Rico Sennrich, and Ivan Titov. 2019a. Context-aware monolingual repair for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 877–886, Hong Kong, China. Association for Computational Linguistics. Elena Voita, Rico Sennrich, and Ivan Titov. 2019b. When a good translation is wrong in context: Context-aware machine translation improves on deixis, ellipsis, and lexical cohesion. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1198–1212, Florence, Italy. Association for Computational Linguistics. Zhengxin Yang, Jinchao Zhang, Fandong Meng, Shuhao Gu, Yang Feng, and Jie Zhou. 2019. Enhancing context modeling with a query-guided capsule network for document-level translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1527– 1537, Hong Kong, China. Association for Computational Linguistics. Lei Yu, Laurent Sartran, Wojciech Stokowiec, Wang Ling, Lingpeng Kong, Phil Blunsom, and Chris Dyer. 2020. Better document-level machine translation with Bayes’ rule. Transactions of the Association for Computational Linguistics, 8:346–360. Jiacheng Zhang, Huanbo Luan, Maosong Sun, Feifei Zhai, Jingfang Xu, Min Zhang, and Yang Liu. 2018. Improving the transformer translation model with document-level context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 533–542, Brussels, Belgium. Association for Computational Linguistics. Meishan Zhang, Yue Zhang, and Duy-Tin Vo. 2016. Gated neural networks for targeted sentiment analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30. Pei Zhang, Boxing Chen, Niyu Ge, and Kai Fan. 2020. Long-short term masking transformer: A simple but effective baseline for document-level neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1081–1087, Online. Association for Computational Linguistics. Zaixiang Zheng, Xiang Yue, Shujian Huang, Jiajun Chen, and Alexandra Birch. 2020. Towards making the most of context in neural machine translation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 3983–3989. 3454 A Evaluation Metrics Following Liu et al. (2020), we use sentence-level BLEU score (s-BLEU) as the major metric for our evaluation. However, when document-level Transformer is compared, we use document-level BLEU score (d-BLEU) since the sentence-to-sentence alignment is not available. s-BLEU. To calculate sentence-level BLEU score on document translations, we first split the translations into sentences, mapping to the corresponding source sentences. Then we calculate the BLEU score on pairs of translation and reference of the same source sentence. d-BLEU. When the alignments between translation and source sentences are not available, we calculate the BLEU score on document-level, matching n-grams in the whole document. B Transformer B.1 Model Transformer (Vaswani et al., 2017) has an encoderdecoder structure, using multi-head attention and feed-forward network as basic modules. In this paper, we mainly concern about the attention module. Attention. An attention module works as a function, mapping a query and a set of key-value pairs to an output, that the query, keys, values, and output are all vectors. The output is computed as a weighted sum of the values, where the weight assigned to each value is computed by a matching function of the query with the corresponding key. Formally, for matrix inputs of query Q, key K, and value V , Attention(Q, K, V ) = softmax QKT √dk V, (11) where dk is the dimensions of the key vector. Multi-Head Attention. Build upon single-head attention module, multi-head attention allows the model to attend to different positions of a sequence, gathering information from different representation subspaces by heads. MultiHead(Q, K, V ) = Concat(head1, ..., headh)W O, (12) where headi = Attention(QW Q i , KW K i , V W V i ), (13) that the projections of W O, W Q i , W K i , and W V i are parameter matrices. Encoder. The encoder consists of a stack of N identical layers. Each layer has a multi-head selfattention, stacked with a feed-forward network. A residual connection is applied to each of them. Decoder. Similar as the encoder, the decoder also consists of a stack of N identical layers. For each layer, a multi-head self-attention is used to represent the target itself, and a multi-head crossattention is used to attend to the encoder outputs. The same structure of feed-forward network and residual connection as the encoder is used. B.2 Training Settings We build our experiments based on Transformer implemented by Fairseq (Ott et al., 2019). We use shared dictionary between source and target, and use a shared embedding table between the encoder and the decoder. We use the default setting proposed by Transformer (Vaswani et al., 2017), which uses Adam optimizer with β1 = 0.9 and β2 = 0.98, a learning rate of 5e−4, and an inversesquare schedule with warmup steps of 4000. We apply label-smoothing of 0.1 and dropout of 0.3 on all settings. To study the impact of input length, data scale, and model size, we take the learning rate and other settings as controlled variables that are fixed for all experiments. We determine the number of updates/steps automatically by early stop on validation set. We train base and big models on 4 GPUs of Navidia 2080ti, and large model on 4 GPUs of v100. C G-Transformer C.1 Training Settings We generate the corresponding group tag sequence dynamically in the model according to the special sentence-mark tokens <s> and </s>. Taking a document “<s> there is no public transport . </s> <s> local people struggle to commute . </s>” as an example, a group-tag sequence G = {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2} is generated according to Eq 3, where 1 starts on the first <s> and ends on the first </s>, 2 the second, and so on. The model can be trained either randomly initialized or fine-tuned. Randomly Initialized. We use the same settings as Transformer to train G-Transformer, using label-smoothing of 0.1, dropout of 0.3, Adam optimizer, and a learning rate of 5e −4 with 4000 warmup steps. To encourage inferencing the translation from the context, we apply a word-dropout 3455 Method TED News Europarl s-BLEU d-BLEU s-BLEU d-BLEU s-BLEU d-BLEU G-Transformer random initialized (Base) 23.53 25.84 23.55 25.23 32.18 33.87 G-Transformer random initialized (Big) 23.29 25.48 22.22 23.82 32.04 33.77 G-Transformer random initialized (Large) 6.23 8.95 13.68 15.33 31.51 33.21 Table 7: G-Transformer on different model size. (Bowman et al., 2016) with a probability of 0.3 on both the source and the target inputs. Fine-tuned on Sentence-Level Transformer. We use the parameters of an existing sentencelevel Transformer to initialize G-Transformer. We copy the parameters of the multi-head attention in Transformer to the group multi-head attention in G-Transformer, leaving the global multi-head attention and the gates randomly initialized. For the global multi-head attention and the gates, we use a learning rate of 5e−4, while for other components, we use a smaller learning rate of 1e −4. All the parameters are jointly trained using Adam optimizer with 4000 warmup steps. We apply a word-dropout with a probability of 0.1 on both the source and the target inputs. Fine-tuned on mBART25. Similar as the finetuning on sentence-level Transformer, we also copy parameters from mBART25 (Liu et al., 2020) to G-Transformer, leaving the global multi-head attention and the gates randomly initialized. We following the settings (Liu et al., 2020) to train the model, using Adam optimizer with a learning rate of 3e −5 and 2500 warmup steps. Here, we do not apply word-dropout, which empirically shows a damage to the performance. C.2 Results on Model Size As shown in Table 7, G-Transformer has a relatively stable performance on different model size. When increasing the model size from Base to Big, the performance drops for about 0.24, 1.33, and 0.14 s-BLEU points, respectively. Further to Large model, the performance drops further for about 17.06, 8.54, and 0.53 s-BLEU points, respectively. Although the performance drop on small dataset is large since overfitting on larger model, the drop on large dataset Europarl is relatively small, indicating a stable training on different model size.
|
2021
|
267
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3456–3468 August 1–6, 2021. ©2021 Association for Computational Linguistics 3456 Prevent the Language Model from being Overconfident in Neural Machine Translation Mengqi Miao1∗, Fandong Meng2∗, Yijin Liu2, Xiao-Hua Zhou3† , and Jie Zhou2 1Peking University, China 2Pattern Recognition Center, WeChat AI, Tencent Inc, China 3Beijing International Center for Mathematical Research, National Engineering Lab for Big Data Analysis and Applications, Department of Biostatistics, Peking University, Beijing, China [email protected], {fandongmeng, yijinliu}@tencent.com [email protected], [email protected] Abstract The Neural Machine Translation (NMT) model is essentially a joint language model conditioned on both the source sentence and partial translation. Therefore, the NMT model naturally involves the mechanism of the Language Model (LM) that predicts the next token only based on partial translation. Despite its success, NMT still suffers from the hallucination problem, generating fluent but inadequate translations. The main reason is that NMT pays excessive attention to the partial translation while neglecting the source sentence to some extent, namely overconfidence of the LM. Accordingly, we define the Margin between the NMT and the LM, calculated by subtracting the predicted probability of the LM from that of the NMT model for each token. The Margin is negatively correlated to the overconfidence degree of the LM. Based on the property, we propose a Margin-based Token-level Objective (MTO) and a Margin-based Sentencelevel Objective (MSO) to maximize the Margin for preventing the LM from being overconfident. Experiments on WMT14 Englishto-German, WMT19 Chinese-to-English, and WMT14 English-to-French translation tasks demonstrate the effectiveness of our approach, with 1.36, 1.50, and 0.63 BLEU improvements, respectively, compared to the Transformer baseline. The human evaluation further verifies that our approaches improve translation adequacy as well as fluency. 1 1 Introduction Neural Machine Translation (NMT) has achieved great success in recent years (Sutskever et al., 2014; ∗Equal contribution. This work was done when Mengqi Miao was interning at Pattern Recognition Center, WeChat AI, Tencent Inc, China. †Corresponding author. 1Code is available at https://github.com/Mlair 77/nmt adequacy Cho et al., 2014; Bahdanau et al., 2014; Luong et al., 2015; Vaswani et al., 2017; Meng and Zhang, 2019; Zhang et al., 2019a; Yan et al., 2020b), which generates accurate and fluent translation through modeling the next word conditioned on both the source sentence and partial translation. However, NMT faces the hallucination problem, i.e., translations are fluent but inadequate to the source sentences. One important reason is that the NMT model pays excessive attention to the partial translation to ensure fluency while failing to translate some segments of the source sentence (Weng et al., 2020b), which is actually the overconfidence of the Language Model (LM). In the rest of this paper, the LM mentioned refers to the LM mechanism involved in NMT. Many recent studies attempt to deal with the inadequacy problem of NMT from two main aspects. One is to improve the architecture of NMT, such as adding a coverage vector to track the attention history (Tu et al., 2016), enhancing the crossattention module (Meng et al., 2016, 2018; Weng et al., 2020b), and dividing the source sentence into past and future parts (Zheng et al., 2019). The other aims to propose a heuristic adequacy metric or objective based on the output of NMT. Tu et al. (2017) and Kong et al. (2019) enhance the model’s reconstruction ability and increase the coverage ratio of the source sentences by translations, respectively. Although some researches (Tu et al., 2017; Kong et al., 2019; Weng et al., 2020b) point out that the lack of adequacy is due to the overconfidence of the LM, unfortunately, they do not propose effective solutions to the overconfidence problem. From the perspective of preventing the overconfidence of the LM, we first define an indicator of the overconfidence degree of the LM, called the Margin between the NMT and the LM, by subtracting the predicted probability of the LM from that of the NMT model for each token. A small Mar3457 gin implies that the NMT might concentrate on the partial translation and degrade into the LM, i.e., the LM is overconfident. Accordingly, we propose a Margin-based Token-level Objective (MTO) to maximize the Margin. Furthermore, we observe a phenomenon that if target sentences in the training data contain many words with negative Margin, they always do not correspond to the source sentences. These data are harmful to model performance. Therefore, based on the MTO, we further propose a Margin-based Sentence-level Objective (MSO) by adding a dynamic weight function to alleviate the negative effect of these “dirty data”. We validate the effectiveness and superiority of our approaches on the Transformer (Vaswani et al., 2017), and conduct experiments on large-scale WMT14 English-to-German, WMT19 Chinese-toEnglish, and WMT14 English-to-French translation tasks. Our contributions are: • We explore the connection between inadequacy translation and the overconfidence of the LM in NMT, and thus propose an indicator of the overconfidence degree, i.e., the Margin between the NMT and the LM. • Furthermore, to prevent the LM from being overconfident, we propose two effective optimization objectives to maximize the Margin, i.e., the Margin-based Token-level Objective (MTO) and the Margin-based Sentence-level Objective (MSO). • Experiments on WMT14 English-to-German, WMT19 Chinese-to-English, and WMT14 English-to-French show that our approaches bring in significant improvements by +1.36, +1.50, +0.63 BLEU points, respectively. Additionally, the human evaluation verifies that our approaches can improve both translation adequacy and fluency. 2 Background Given a source sentence x = {x1, x2, ..., xN }, the NMT model predicts the probability of a target sentence y = {y1, y2, ..., yT } word by word: P(y|x) = T Y t=1 p(yt|y<t, x), (1) where y<t = {y1, y2, ..., yt−1} is the partial translation before yt. From Eq. 1, the source sentence x and partial translation y<t are considered in the meantime, suggesting that the NMT model is essentially a joint language model and the LM is instinctively involved in NMT. Based on the encoder-decoder architecture, the encoder of NMT maps the input sentence x to hidden states. At time step t, the decoder of NMT employs the output of the encoder and y<t to predict yt. The training objective of NMT is to minimize the negative log-likelihood, which is also known as the cross entropy loss function: LNMT ce = − T X t=1 log p(yt|y<t, x). (2) The LM measures the probability of a target sentence similar to NMT but without knowledge of the source sentence x: P(y) = T Y t=1 p(yt|y<t). (3) The LM can be regarded as the part of NMT decoder that is responsible for fluency, only takes y<t as input. The training objective of the LM is almost the same as NMT except for the source sentence x: LLM ce = − T X t=1 log p(yt|y<t). (4) The NMT model predicts the next word yt according to the source sentence x and meanwhile ensures that yt is fluent with the partial translation y<t. However, when NMT pays excessive attention to translation fluency, some source segments may be neglected, leading to inadequacy problem. This is exactly what we aim to address in this paper. 3 The Approach In this section, we firstly define the Margin between the NMT and the LM (Section 3.1), which reflects the overconfidence degree of the LM. Then we put forward the token-level (Section 3.2) and sentencelevel (Section 3.3) optimization objectives to maximize the Margin. Finally, we elaborate our twostage training strategy (Section 3.4). 3.1 Margin between the NMT and the LM When the NMT model excessively focuses on partial translation, i.e., the LM is overconfident, the NMT model degrades into the LM, resulting in hallucinated translations. To prevent the overconfidence problem, we expect that the NMT model outperforms the LM as much as possible in predicting golden tokens. Consequently, we define the Margin between the NMT and the LM at the 3458 t-th time step by the difference of the predicted probabilities of them: ∆(t) = pNMT (yt|y<t, x) −pLM (yt|y<t), (5) where pNMT denotes the predicted probability of the NMT model, i.e., p(yt|y<t, x), and pLM denotes that of the LM, i.e., p(yt|y<t). The Margin ∆(t) is negatively correlated to the overconfidence degree of the LM, and different values of the Margin indicate different cases: • If ∆(t) is big, the NMT model is apparently better than the LM, and yt is strongly related to the source sentence x. Hence the LM is not overconfident. • If ∆(t) is medium, the LM may be slightly overconfident and the NMT model has the potential to be enhanced. • If ∆(t) is small, the NMT model might degrade to the LM and not correctly translate the source sentence, i.e., the LM is overconfident.2 Note that sometimes, the model needs to focus more on the partial translation such as the word to be predicted is a determiner in the target language. In this case, although small ∆(t) does not indicate the LM is overconfident, enlarging the ∆(t) can still enhance the NMT model. 3.2 Margin-based Token-level Objective Based on the Margin, we firstly define the Margin loss LM and then fuse it into the cross entropy loss function to obtain the Margin-based Tokenevel Optimization Objective (MTO). Formally, we define the Margin loss LM to maximize the Margin as follow: LM = T X t=1 (1 −pNMT (t))M(∆(t)), (6) where we abbreviate pNMT (yt|y<t, x) as pNMT (t). M(∆(t)) is a function of ∆(t), namely Margin function, which is monotonically decreasing (e.g., 1 −∆(t)). Moreover, when some words have the same ∆(t) but different pNMT (t), their meanings are quite different: (1) If pNMT (t) is big, the NMT model learns the token well and does not need to focus on the Margin too much; (2) If pNMT (t) is 2In addition, if pNMT (yt|y<t, x) is large, less attention will be paid to this data because yt has been learned well, which will be described in detail in Section 3.2. 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 0.0 0.2 0.4 0.6 0.8 1.0 M( ) Linear Cube Quitic Log Figure 1: The four Margin functions M(∆). All of them are monotonically decreasing, yet with different slopes. Compared with Linear, the three non-linear functions are more stable around |∆| = 0 and steeper around |∆| = 1. We set α in Log to 10 in this figure. small, the NMT model is urgently to be optimized on the token thus the weight of M(∆(t)) should be enlarged. Therefore, as the weight of M(∆(t)), 1 −pNMT (t) enables the model treat tokens wisely. Variations of M(∆). We abbreviate Margin function M(∆(t)) as M(∆) hereafter. A simple and intuitive definition is the Linear function: M(∆) = 1 −∆, which has the same gradient for different ∆. However, as illustrated in Section 3.1, different ∆has completely various meaning and needs to be treated differently. Therefore, we propose three non-linear Margin functions M(∆) as follows: • Cube: (1 −∆3)/2. • Quintic (fifth power): (1 −∆5)/2. • Log: 1 αlog( 1−∆ 1+∆) + 0.5. where α is a hyperparamater for Log. As shown in Figure 1, the four variations3 have quite different slopes. Specifically, the three nonlinear functions are more stable around ∆= 0 (e.g., ∆∈[−0.5, 0.5]) than Linear, especially Quintic. We will report the performance of the four M(∆) concretely and analyze why the three non-linear M(∆) perform better than Linear in Section 5.4. Finally, based on LM, we propose the Marginbased Token-level Objective (MTO): LT = LNMT ce + λMLM, (7) where LNMT ce is the cross-entropy loss of the NMT model defined in Eq. 2 and λM is the hyperparameter for the Margin loss LM. 3In order to keep the range of M(∆) roughly [0,1], we set Linear function to (1 −∆)/2. 3459 Source 尽管他们是孪生儿, 但性格却截然不同. Target How did your mother succeed in keeping the peace between these two very different men? Expert Translation Although they are twins, they are quite different in character. Figure 2: The parallel sentences, i.e., the source and target sentences, are sampled from the WMT19 Chineseto-English training dataset. We also list an expert translation of the source sentence. The words in bold red have negative Margin. This target sentence has more than 50% tokens with negative Margin, and these tokens are almost irrelevant to the source sentence. Apparently, the target sentence is a hallucination and will harm the model performance. 3.3 Margin-based Sentence-level Objective Furthermore, through analyzing the Margin distribution of target sentences, we observe that the target sentences in the training data which have many tokens with negative Margin are almost “hallucinations” of the source sentences (i.e., dirty data), thus will harm the model performance. Therefore, based on MTO, we further propose the Marginbased Sentence-level Objective (MSO) to address this issue. Compared with the LM, the NMT model predicts the next word with more prior knowledge (i.e., the source sentence). Therefore, it is intuitive that when predicting yt, the NMT model should predict more accurately than the LM, as follow: pNMT (yt|y<t, x) > pLM (yt|y<t). (8) Actually, the above equation is equivalent to ∆(t) > 0. The larger ∆(t) is, the more the NMT model exceeds the LM. However, there are many tokens with negative Margin through analyzing the Margin distribution. We conjecture the reason is that the target sentence is not corresponding to the source sentence in the training corpus, i.e., the target sentence is a hallucination. Actually, we also observe that if a large proportion of tokens in a target sentence have negative Margin (e.g., 50%), the sentence is probably not corresponding to the source sentence, such as the case in Figure 2. These “dirty” data will harm the performance of the NMT model. To measure the “dirty” degree of data, we define the Sentence-level Negative Margin Ratio of parallel sentences (x, y) as follow: R(x, y) = #{yt ∈y : ∆(t) < 0} #{yt : yt ∈y} , (9) where #{yt ∈y : ∆(t) < 0} denotes the number of tokens with negative ∆(t) in y, and #{yt : yt ∈ y} is the length of the target sentence y. When R(x, y) is larger than a threshold k (e.g., k=50%), the target sentence may be desperately inadequate, or even completely unrelated to the source sentence, as shown in Figure 2. In order to eliminate the impact of these seriously inadequate sentences, we ignore their loss during training by the Margin-based Sentence-level Objective (MSO): LS = IR(x,y)<k · LT , (10) where IR(x,y)<k is a dynamic weight function in sentence level. The indicative function IR(x,y)<k equals to 1 if R(x, y) < k, else 0, where k is a hyperparameter. LT is MTO defined in Eq. 7. IR(x,y)<k is dynamic at the training stage. During training, as the model gets better, its ability to distinguish hallucinations improves thus IR(x,y)<k becomes more accurate. We will analyze the changes of IR(x,y)<k in Section 5.4. 3.4 Two-stage Training We elaborate our two-stage training in this section, 1) jointly pretraining an NMT model and an auxiliary LM, and 2) finetuning the NMT model. Jointly Pretraining. The language model mechanism in NMT cannot be directly evaluated, thus we train an auxiliary LM to represent it. We pretrain them together using a fusion loss function: Lpre = LNMT ce + λLMLLM ce , (11) where LNMT ce and LLM ce are the cross entropy loss functions of the NMT model and the LM defined in Eq. 2 and Eq. 4, respectively. λLM is a hyperparameter. Specifically, we jointly train them through sharing their decoders’ embedding layers and their pre-softmax linear transformation layers (Vaswani et al., 2017). There are two reasons for joint training: (1) making the auxiliary LM as consistent as possible with the language model mechanism in NMT; (2) avoiding abundant extra parameters. Finetuning. We finetune the NMT model by minimizing the MTO (LT in Eq. 7) and MSO (LS in Eq. 10).4 Note that the LM is not involved at the inference stage. 4The LM can be fixed or trained along with the NMT after pretraining. Our experimental results show that continuous training the LM and fixing the LM have analogous performance during the finetuning stage. Therefore, we only report the results of keeping the LM fixed in this paper. 3460 4 Experimental Settings We conduct experiments on three large-scale NMT tasks, i.e., WMT14 English-to-German (En→De), WMT14 English-to-French (En→Fr), and WMT19 Chinese-to-English (Zh→En). Datasets. For En→De, we use 4.5M training data. Following the same setting in (Vaswani et al., 2017), we use newstest2013 as validation set and newstest2014 as test set, which contain 3000 and 3003 sentences, respectively. For En→Fr, the training dataset contains about 36M sentence pairs, and we use newstest2013 with 3000 sentences as validation set and newstest2014 with 3003 sentences as test set. For Zh→En, we use 20.5M training data and use newstest2018 as validation set and newstest2019 as test set, which contain 3981 and 2000 sentences, respectively. For Zh→En, the number of merge operations in byte pair encoding (BPE) (Sennrich et al., 2016a) is set to 32K for both source and target languages. For En→De and En→Fr, we use a shared vocabulary generated by 32K BPEs. Evaluation. We measure the case-sensitive BLEU scores using multi-bleu.perl 5 for En→De and En→Fr. For Zh→En, case-sensitive BLEU scores are calculated by Moses mteval-v13a.pl script6. Moreover, we use the paired bootstrap resampling (Koehn, 2004) for significance test. We select the model which performs the best on the validation sets and report its performance on the test sets for evaluation. Model and Hyperparameters. We conduct experiments based on the Transformer (Vaswani et al., 2017) and implement our approaches with the opensource tooklit Opennmt-py (Klein et al., 2017). Following the Transformer-Base setting in (Vaswani et al., 2017), we set the hidden size to 512 and the encoder/decoder layers to 6. All three tasks are trained with 8 NVIDIA V100 GPUs, and the batch size for each GPU is 4096 tokens. The beam size is 5 and the length penalty is 0.6. Adam optimizer (Kingma and Ba, 2014) is used in all the models. The LM architecture is the decoder of the Transformer excluding the cross-attention layers, sharing the embedding layer and the pre-softmax 5https://github.com/moses-smt/mosesde coder/blob/master/scripts/generic/multibleu.perl 6https://github.com/moses-smt/mosesde coder/blob/mast-er/scripts/generic/mteva l-v13a.pl linear transformation with the NMT model. For En→De, Zh→En, and En→Fr, the number of training steps is 150K for jointly pretraining stage and 150K for finetuning7. During pretraining, we set λLM to 0.01 for all three tasks8. Experimental results shown in Appendix A indicate that the LM has converged after pretraining for all the three tasks. During finetuning, the Margin function M(∆) in Section 3.2 is set to Quintic, and we will analyze the four M(∆) in Section 5.4. λM in Eq. 7 is set to 5, 8, and 8 on En→De, En→Fr and Zh→En, respectively. For MSO, the threshold k in Eq. 10 is set to 30% for En→De and Zh→En, 40% for En→Fr. The two hyperparameters (i.e., λM and k) are searched on validation sets, and the selection details are shown in Appendix B. The baseline model (i.e., vanilla Transformer) is trained for 300k steps for En→De, En→Fr and Zh→En. Moreover, we use a joint training model as our secondary baseline, namely NMT+LM, by jointly training the NMT model and the LM throughout the training stage with 300K steps. The training steps of all the models are consistent, thus the experiment results are strictly comparable. 5 Results and Analysis We first evaluate the main performance of our approaches (Section 5.1 and 5.2). Then, the human evaluation further confirms the improvements of translation adequacy and fluency (Section 5.3). Finally, we analyze the positive impact of our models on the distribution of Margin and explore how each fragment of our method works (Section 5.4). 5.1 Results on En→De The results on WMT14 English-to-German (En→De) are summarized in Table 1. We list the results from (Vaswani et al., 2017) and several related competitive NMT systems by various methods, such as Minimum Risk Training (MRT) objective (Shen et al., 2016), Simple Fusion of NMT and LM (Stahlberg et al., 2018), optimizing adequacy metrics (Kong et al., 2019; Feng et al., 2019) and improving the Transformer architecture (Yang et al., 2018; Zheng et al., 2019; Yang et al., 2019; Weng et al., 2020b; Yan et al., 2020a). We re7The LM does not need to be state-of-the-art. The previous study of (Baziotis et al., 2020) has shown that a more powerful LM does not lead to further improvements to NMT. 8The experimental results show that the model is insensitive to λLM. Therefore we make λLM consistent for all the three tasks. 3461 System En→De ↑ Existing NMT systems Transformer (Vaswani et al., 2017) 27.3 MRT* (Shen et al., 2016) 27.71 Simple Fusion** (Stahlberg et al., 2018) 27.88 Localness (Yang et al., 2018) 28.11 Context-Aware (Yang et al., 2019) 28.26 AOL (Kong et al., 2019) 28.01 Eval. Module (Feng et al., 2019) 27.55 Past&Future (Zheng et al., 2019) 28.10 Dual (Yan et al., 2020a) 27.86 Multi-Task (Weng et al., 2020b) 28.25 Our NMT systems NMT (Transformer) 27.22 ref + LM 27.97 +0.75 + MTO 28.47†‡ +1.25 + MSO 28.58†‡ +1.36 Table 1: Case-sensitive BLEU scores (%) on the test set of WMT14 En→De. ↑denotes the improvement compared with the NMT baseline (i.e., Transformer). “†”: significantly better than NMT (p<0.01). “‡”: significantly better than the joint model NMT+LM (p<0.01). (MRT* in (Shen et al., 2016) is RNN-based, and the result reported here is implemented on Transformer by Weng et al. (2020b). **: we re-implement Simple Fusion on upon of Transformer.) implement the Transformer model (Vaswani et al., 2017) as our baseline. Similarly, we re-implement the Simple Fusion (Stahlberg et al., 2018) model. 9 Finally, the results of the joint training model NMT+LM, and models with our MTO and MSO objectives are reported. Compared with the baseline, NMT+LM yields +0.75 BLEU improvement. Based on NMT+LM, our MTO achieves further improvement with +0.50 BLEU scores, indicating that preventing the LM from being overconfident could significantly enhance model performance. Moreover, MSO performs better than MTO by +0.11 BLEU scores, which implies that the “dirty data” in the training dataset indeed harm the model performance, and the dynamic weight function IR(x,y)<k in Eq. 10 could reduce the negative impact. In conclusion, our approaches improve up to +1.36 BLEU scores on En→De compared with the Transformer baseline and substantially outperforms the existing NMT systems. The results demonstrate the effectiveness and superiority of our approaches. 9The architectures of the LM and NMT model in Simple Fusion are consistent with our MTO and MSO. System En→Fr Zh→En BLEU ↑ BLEU ↑ Vaswani et al. (2017)* 38.1 NMT (Transformer) 41.07 ref 25.75 ref + LM 41.14 +0.07 25.90 +0.15 + MTO 41.56†‡ +0.49 26.94†‡ +1.19 + MSO 41.70†‡ +0.63 27.25†‡ +1.50 Table 2: Case-sensitive BLEU scores (%) on the test set of WMT14 En→Fr and WMT19 Zh→En. ↑denotes the improvement compared with the NMT baseline (i.e., Transformer). “†”: significantly better than NMT (p<0.01). “‡”: significantly better than the joint model NMT+LM (p<0.01). * denotes the results come from the cited paper. 5.2 Results on En→Fr and Zh→En The results on WMT14 English-to-French (En→Fr) and WMT19 Chinese-to-English (Zh→En) are shown in Table 2. We also list the results of (Vaswani et al., 2017) and our reimplemented Transformer as the baselines. On En→Fr, our reimplemented result is higher than the result of (Vaswani et al., 2017), since we update 300K steps while Vaswani et al. (2017) only update 100K steps. Many studies obtain similar results to ours (e.g., 41.1 BLEU scores from (Ott et al., 2019)). Compared with the baseline, NMT+LM yields +0.07 and +0.15 BLEU improvements on En→Fr and Zh→En, respectively. The improvement of NMT+LM on En→De in Table 1 (i.e., +0.75) is greater than these two datasets. We conjecture the reason is that the amount of training data of En→De is much smaller than that of En→Fr and Zh→En, thus NMT+LM is more likely to improve the model performance on En→De. Compared with NMT+LM, our MTO achieves further improvements with +0.42 and +1.04 BLEU scores on En→Fr and Zh→En, respectively, which demonstrates the performance improvement is mainly due to our Margin-based objective rather than joint training. Moreover, based on MTO, our MSO further yields +0.14 and +0.31 BLEU improvements. In summary, our approaches improve up to +0.63 and +1.50 BLEU scores on En→Fr and Zh→En compared with the baselines, respectively, which demonstrates the effectiveness and generalizability of our approaches. 5.3 Human Evaluation We conduct the human evaluation for translations in terms of adequacy and fluency. Firstly, we ran3462 Model Adequacy Fluency Ave. NMT (Transformer) 4.04 4.66 4.35 + LM 4.12 4.86 4.49 + MTO 4.26 4.87 4.57 + MSO 4.41 4.91 4.66 Table 3: Human evaluation on adequacy and fluency. 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 0 50K 100K 150K 200K 250K 300K Frequency NMT+LM MSO Figure 3: The distribution of ∆of NMT+LM and MSO. We randomly sample 100K sentence pairs from the training dataset of Zh→En and compute the Margin of their tokens. The purple area is the overlap of the two models’ ∆distributions. The two distributions are quite different. Compared with NMT+LM, MSO reduces the distribution around ∆= 0 and meanwhile increases the distribution around ∆= 1. domly sample 100 sentences from the test set of WMT19 Zh→En. Then we invite three annotators to evaluate the translation adequacy and fluency. Five scales have been set up, i.e., 1, 2, 3, 4, 5. For adequacy, “1” means totally irrelevant to the source sentence, and “5” means equal to the source sentence semantically. For fluency, “1” represents not fluent and incomprehensible; “5” represents very “native”. Finally, we take the average of the scores from the three annotators as the final score. The results of the baseline and our approaches are shown in Table 3. Compared with the NMT baseline, NMT+LM, MTO and MSO improve adequacy with 0.08, 0.22, and 0.37 scores, respectively. Most improvements come from our Margin-based methods MTO and MSO, and MSO performs the best. For fluency, NMT+LM achieves 0.2 improvement compared with NMT. Based on NMT+LM, MTO and MSO yield further improvements with 0.01 and 0.05 scores, respectively. Human evaluation indicates that our MTO and MSO approaches remarkably improve translation adequacy and slightly enhance translation fluency. Model Percent of ∆< 0 (↓) Average ∆(↑) NMT + LM 12.45% (ref) 0.33 (ref) + MTO 10.17% (-2.28%) 0.44 (+0.11) + MSO 10.89% (-1.56%) 0.44 (+0.11) Table 4: The percent of ∆< 0 and average ∆of models computed from the 100K sentence pairs introduced in Figure 3. Compared with NMT+LM, both MTO and MSO effectively reduce the percent of ∆< 0 and improve the average ∆. 5.4 Analysis Margin between the NMT and the LM. Firstly, we analyze the distribution of the Margin between the NMT and the LM (i.e., ∆in Eq. 5). As shown in Figure 3, for the joint training model NMT+LM, although most of the Margins are positive, there are still many tokens with negative Margin and a large amount of Margins around 0. This indicates that the LM is probably overconfident for many tokens, and addressing the overconfidence problem is meaningful for NMT. By comparison, the Margin distribution of MSO is dramatically different with NMT+LM: the tokens with Margin around 0 are significantly reduced, and the tokens with Margin in [0.75, 1.0] are increased apparently. More precisely, we list the percentage of tokens with negative Margin and the average Margin for each model in Table 4. Compared with NMT+LM, MTO and MSO reduce the percentage of negative Margin by 2.28 and 1.56 points, respectively. We notice MSO performs slightly worse than MTO, because MSO neglects the hallucinations during training. As there are many tokens with negative Margin in hallucinations, the ability of MSO to reduce the proportion of ∆< 0 is weakened. We further analyze effects of MTO and MSO on the average of Margin. Both MTO and MSO improve the average of the Margin by 33% (from 0.33 to 0.44). In conclusion, MTO and MSO both indeed increase the Margin between the NMT and the LM. Variations of M(∆). We compare the performance of the four Margin functions M(∆) defined in Section 3.2. We list the BLEU scores of the Transformer baseline, NMT+LM and our MTO approach with the four M(∆) in Table 5. All the four variations bring improvements over NMT and NMT+LM. The results of Log with different α are similar to Linear, while far lower than Cube and Quintic. And Quintic performs the best among all the four variations. We speculate the reason is that 3463 Function BLEU ↑ NMT (Transformer) 25.75 ref + LM 25.90 +0.15 + Linear 26.13 +0.38 + Cube 26.45 +0.60 + Quintic 26.94 +1.19 + Log (α = 5) 26.12 +0.37 + Log (α = 10) 26.07 +0.32 + Log (α = 20) 26.24 +0.49 Table 5: Case-sensitive BLEU scores (%) on Zh→En test set of MTO with several variations of M(∆). α is the hyperparameter of Log. All four M(∆) achieve BLEU improvements compared with NMT and NMT+LM, and Quintic performs the best. Models Valid Test NMT (Transformer) 23.67 25.75 + LM 23.61 25.90 + MTO w/ Weight 24.09 26.94 + MTO w/o Weight 23.36 25.85 Table 6: Case-sensitive BLEU scores (%) on Zh→En validation set and test set of MTO with (w/) and without (w/o) the weight 1 −pNMT (t). ∆∈[−0.5, 0.5] is the main range for improvement, and Quintic updates more careful on this range (i.e., with smaller slopes) as shown in Figure 1. Effects of the Weight of M(∆). In MTO, we propose the weight 1−pNMT (t) of the Margin function M(∆) in Eq. 6. To validate the importance of it, we remove the weight and the Margin loss degrades to LM = PT t=1 M(∆(t)). The results are listed in Table 6. Compared with NMT+LM, MTO without weight performs worse with 0.25 and 0.05 BLEU decreases on the validation set and test set, respectively. Compared with MTO with weight, it decreases 0.73 and 1.09 BLEU scores on the validation set and test set, respectively. This demonstrates that the weight 1 −pNMT (t) is indispensable for our approach. Changes of IR(x,y)<k During Training. In MSO, we propose a dynamic weight function IR(x,y)<k in Eq. 10. Figure 4 shows the changes of IR(x,y)<k in MSO and the BLEU scores of MSO and MTO during finetuning. As the training continues, our model gets more competent, and the proportion of sentences judged to be “dirty data” by our model increases rapidly at first and then 23.0 23.2 23.4 23.6 23.8 24.0 24.2 24.4 24.6 24.8 25.0 0.056 0.058 0.060 0.062 0.064 0.066 0.068 0.070 0.072 155K 175K 195K 215K 235K 255K 275K 295K BLEU Proportion Training steps Propotion of I=0 MTO MSO Figure 4: Changes of the proportion of IR(x,y)<30% = 0 on Zh→En during finetuning for MSO, and BLEU scores (%) on the validation set of Zh→En for MTO and MSO. The orange line corresponds to the left yaxis, and the green and blue lines correspond to the right y-axis. We sample 100K sentence pairs in the training data and compute IR(x,y)<30%. flattens out, which is consistent with the trend of BLEU of MSO. Moreover, by adding the dynamic weight function, MSO outperforms MTO at most steps. Case Study. To better illustrate the translation quality of our approach, we show several translation examples in Appendix C. Our approach grasps more segments of the source sentences, which are mistranslated or neglected by the Transformer. 6 Related Work Translation Adequacy of NMT. NMT suffers from the hallucination and inadequacy problem for a long time (Tu et al., 2016; M¨uller et al., 2020; Wang and Sennrich, 2020; Lee et al., 2019). Many studies improve the architecture of NMT to alleviate the inadequacy issue, including tracking translation adequacy by coverage vectors (Tu et al., 2016; Mi et al., 2016), modeling a global representation of source side (Weng et al., 2020a), dividing the source sentence into past and future parts (Zheng et al., 2019), and multi-task learning to improve encoder and cross-attention modules in decoder (Meng et al., 2016, 2018; Weng et al., 2020b). They inductively increase the translation adequacy, while our approaches directly maximize the Margin between the NMT and the LM to prevent the LM from being overconfident. Other studies enhance the translation adequacy by adequacy metrics or additional optimization objectives. Tu et al. (2017) minimize the difference between the original source sentence and the reconstruction source sentence of NMT. Kong et al. (2019) pro3464 pose a coverage ratio of the source sentence by the model translation. Feng et al. (2019) evaluate the fluency and adequacy of translations with an evaluation module. However, the metrics or objectives in the above approaches may not wholly represent adequacy. On the contrary, our approaches are derived from the criteria of the NMT model and the LM, thus credible. Language Model Augmented NMT. Language Models are always used to provide more information to improve NMT. For low-resource tasks, the LM trained on extra monolingual data can rerank the translations by fusion (G¨ulc¸ehre et al., 2015; Sriram et al., 2017; Stahlberg et al., 2018), enhance NMT’s representations (Clinchant et al., 2019; Zhu et al., 2020), and provide prior knowledge for NMT (Baziotis et al., 2020). For data augmentation, LMs are used to replace words in sentences (Kobayashi, 2018; Wu et al., 2018; Gao et al., 2019). Differently, we mainly focus on the Margin between the NMT and the LM, and no additional data is required. Stahlberg et al. (2018) propose the Simple Fusion approach to model the difference between NMT and LM. Differently, it is trained to optimize the residual probability, positively correlated to pNMT /pLM which is hard to optimize and the LM is still required in inference, slowing down the inference speed largely. Data Selection in NMT. Data selection and data filter methods have been widely used in NMT. To balance data domains or enhance the data quality generated by back-translation (Sennrich et al., 2016b), many approaches have been proposed, such as utilizing language models (Moore and Lewis, 2010; van der Wees et al., 2017; Zhang et al., 2020), translation models (JunczysDowmunt, 2018; Wang et al., 2019a), and curriculum learning (Zhang et al., 2019b; Wang et al., 2019b). Different from the above methods, our MSO dynamically combines language models with translation models for data selection during training, making full use of the models. 7 Conclusion We alleviate the problem of inadequacy translation from the perspective of preventing the LM from being overconfident. Specifically, we firstly propose an indicator of the overconfidence degree of the LM in NMT, i.e., Margin between the NMT and the LM. Then we propose Margin-based Tokenlevel and Sentence-level objectives to maximize the Margin. Experimental results on three large-scale translation tasks demonstrate the effectiveness and superiority of our approaches. The human evaluation further verifies that our methods can improve translation adequacy and fluency. Acknowledgments The research work descried in this paper has been supported by the National Nature Science Foundation of China (No. 12026606). The authors would like to thank the anonymous reviewers for their valuable comments and suggestions to improve this paper. References Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. Christos Baziotis, Barry Haddow, and Alexandra Birch. 2020. Language model prior for low-resource neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7622–7634, Online. Association for Computational Linguistics. Kyunghyun Cho, Bart van Merri¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724– 1734, Doha, Qatar. Association for Computational Linguistics. Stephane Clinchant, Kweon Woo Jung, and Vassilina Nikoulina. 2019. On the use of BERT for neural machine translation. In Proceedings of the 3rd Workshop on Neural Generation and Translation, pages 108–117, Hong Kong. Association for Computational Linguistics. Yang Feng, Wanying Xie, Shuhao Gu, Chenze Shao, Wen Zhang, Zhengxin Yang, and Dong Yu. 2019. Modeling fluency and faithfulness for diverse neural machine translation. arXiv preprint arXiv:1912.00178. Fei Gao, Jinhua Zhu, Lijun Wu, Yingce Xia, Tao Qin, Xueqi Cheng, Wengang Zhou, and Tie-Yan Liu. 2019. Soft contextual data augmentation for neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5539–5544, Florence, Italy. Association for Computational Linguistics. 3465 C¸ aglar G¨ulc¸ehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Lo¨ıc Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2015. On using monolingual corpora in neural machine translation. CoRR, abs/1503.03535. Marcin Junczys-Dowmunt. 2018. Dual conditional cross-entropy filtering of noisy parallel corpora. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 888–895, Belgium, Brussels. Association for Computational Linguistics. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. OpenNMT: Opensource toolkit for neural machine translation. In Proceedings of ACL 2017, System Demonstrations, pages 67–72, Vancouver, Canada. Association for Computational Linguistics. Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 452–457, New Orleans, Louisiana. Association for Computational Linguistics. Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388– 395, Barcelona, Spain. Association for Computational Linguistics. Xiang Kong, Zhaopeng Tu, Shuming Shi, Eduard Hovy, and Tong Zhang. 2019. Neural machine translation with adequacy-oriented learning. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):6618–6625. Katherine Lee, Orhan Firat, Ashish Agarwal, Clara Fannjiang, and David Sussillo. 2019. Hallucinations in neural machine translation. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412–1421, Lisbon, Portugal. Association for Computational Linguistics. Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. 2016. Interactive attention for neural machine translation. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2174–2185, Osaka, Japan. Fandong Meng, Zhaopeng Tu, Yong Cheng, Haiyang Wu, Junjie Zhai, Yuekui Yang, and Di Wang. 2018. Neural machine translation with key-value memoryaugmented attention. In Proceedings of IJCAI. Fandong Meng and Jinchao Zhang. 2019. DTMT: A novel deep transition architecture for neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 224– 231. Haitao Mi, Baskaran Sankaran, Zhiguo Wang, and Abe Ittycheriah. 2016. Coverage embedding models for neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 955–960, Austin, Texas. Association for Computational Linguistics. Robert C. Moore and William Lewis. 2010. Intelligent selection of language model training data. In Proceedings of the ACL 2010 Conference Short Papers, pages 220–224, Uppsala, Sweden. Association for Computational Linguistics. Mathias M¨uller, Annette Rios, and Rico Sennrich. 2020. Domain robustness in neural machine translation. In Proceedings of the 14th Conference of the Association for Machine Translation in the Americas (Volume 1: Research Track), pages 151–164, Virtual. Association for Machine Translation in the Americas. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53, Minneapolis, Minnesota. Association for Computational Linguistics. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725, Berlin, Germany. Association for Computational Linguistics. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725, Berlin, Germany. Association for Computational Linguistics. Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2016. Minimum risk training for neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683–1692, Berlin, Germany. Association for Computational Linguistics. 3466 Anuroop Sriram, Heewoo Jun, Sanjeev Satheesh, and Adam Coates. 2017. Cold fusion: Training seq2seq models together with language models. CoRR, abs/1708.06426. Felix Stahlberg, James Cross, and Veselin Stoyanov. 2018. Simple fusion: Return of the language model. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 204–211, Brussels, Belgium. Association for Computational Linguistics. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27:3104–3112. Zhaopeng Tu, Yang Liu, Lifeng Shang, Xiaohua Liu, and Hang Li. 2017. Neural machine translation with reconstruction. In Proceedings of the 31st AAAI Conference on Artificial Intelligence. Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and Hang Li. 2016. Modeling coverage for neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 76–85. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008. Curran Associates, Inc. Chaojun Wang and Rico Sennrich. 2020. On exposure bias, hallucination and domain shift in neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3544–3552, Online. Association for Computational Linguistics. Shuo Wang, Yang Liu, Chao Wang, Huanbo Luan, and Maosong Sun. 2019a. Improving back-translation with uncertainty-based confidence estimation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 791– 802, Hong Kong, China. Association for Computational Linguistics. Wei Wang, Isaac Caswell, and Ciprian Chelba. 2019b. Dynamically composing domain-data selection with clean-data selection by “co-curricular learning” for neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1282–1292. Marlies van der Wees, Arianna Bisazza, and Christof Monz. 2017. Dynamic data selection for neural machine translation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1400–1410, Copenhagen, Denmark. Association for Computational Linguistics. Rongxiang Weng, Haoran Wei, Shujian Huang, Heng Yu, Lidong Bing, Weihua Luo, and Jiajun Chen. 2020a. Gret: Global representation enhanced transformer. AAAI. Rongxiang Weng, Heng Yu, Xiangpeng Wei, and Weihua Luo. 2020b. Towards enhancing faithfulness for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2675–2684, Online. Association for Computational Linguistics. Xing Wu, Shangwen Lv, Liangjun Zang, Jizhong Han, and Songlin Hu. 2018. Conditional bert contextual augmentation. Jianhao Yan, Fandong Meng, and Jie Zhou. 2020a. Dual past and future for neural machine translation. Jianhao Yan, Fandong Meng, and Jie Zhou. 2020b. Multi-unit transformers for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1047–1059, Online. Baosong Yang, Jian Li, Derek F Wong, Lidia S Chao, Xing Wang, and Zhaopeng Tu. 2019. Context-aware self-attention networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 387– 394. Baosong Yang, Zhaopeng Tu, Derek F. Wong, Fandong Meng, Lidia S. Chao, and Tong Zhang. 2018. Modeling localness for self-attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4449– 4458, Brussels, Belgium. Association for Computational Linguistics. Boliang Zhang, Ajay Nagesh, and Kevin Knight. 2020. Parallel corpus filtering via pre-trained language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8545–8554, Online. Association for Computational Linguistics. Wen Zhang, Yang Feng, Fandong Meng, Di You, and Qun Liu. 2019a. Bridging the gap between training and inference for neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4334– 4343, Florence, Italy. Xuan Zhang, Pamela Shapiro, Gaurav Kumar, Paul McNamee, Marine Carpuat, and Kevin Duh. 2019b. Curriculum learning for domain adaptation in neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1903–1915, Minneapolis, Minnesota. Association for Computational Linguistics. Zaixiang Zheng, Shujian Huang, Zhaopeng Tu, XinYu Dai, and Jiajun Chen. 2019. Dynamic past and 3467 future for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 931–941, Hong Kong, China. Association for Computational Linguistics. Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tieyan Liu. 2020. Incorporating bert into neural machine translation. In International Conference on Learning Representations. A Loss of the Language Model To validate whether the LM is converged or not after pretraining, we plot the loss of the LM as shown in Figure 5. The loss of the LM remains stable after training about 80K steps for En→De, Zh→En and En→Fr, indicating that the LM is converged during pretraining stage. 0 20K 40K 60K 80K 100K 120K 140K Training Steps 5.0 5.5 6.0 6.5 7.0 7.5 8.0 Loss En >De Zh >En En >Fr Figure 5: The loss of the LM on the validation set during pretraining for En→De, Zh→En and En→Fr. The LM converges after training nearly 80K steps for all the three tasks. B Hyperparameters Selection The results of our approaches with different λM (defined in Eq. 7) and k (defined in Eq. 10) on the validation sets of WMT14 En→De, WMT14 En→Fr and WMT19 Zh→En are shown in Figure 6. We firstly search the best λM based on MTO. All the three datasets achieve better performance for λM ∈[5, 10]. The model reaches the peak when λM =5, 8, and 8 for the three tasks, respectively. Then, fixing the best λM for each dataset, we search the best threshold k. As shown in the right of Figure 6, the best k is 30% for En→De and Zh→En, 40% for En→Fr. This is consistent with our observations. When the proportion of tokens 1 3 5 8 10 M 27.0 27.1 27.2 27.3 27.4 BLEU (a) λM (En→De) 20% 30% 40% 50% threshold in S 27.1 27.2 27.3 27.4 27.5 BLEU (b) k (En→De) 1 3 5 8 10 M 41.2 41.3 41.4 41.5 41.6 BLEU (c) λM (En→Fr) 20% 30% 40% 50% threshold in S 41.75 41.85 41.95 42.05 BLEU (d) k (En→Fr) 1 3 5 8 10 M 23.7 23.8 23.9 24.0 24.1 24.2 BLEU (e) λM (Zh→En) 20% 30% 40% 50% threshold in S 24.0 24.1 24.2 24.3 24.4 24.5 BLEU (f) k (Zh→En) Figure 6: Case-sensitive BLEU scores (%) on validation sets of WMT14 En→De, WMT14 En→Fr and WMT19 Zh→En with different hyperparameters, respectively. λM is defined in Eq. 7, and the search results are shown in Figure (a), (c) and (e). The threshold k for MSO is defined in Eq. 10 and the results of it are shown in Figure (b), (d), and (f). with negative Margin in a target sentence is greater than 30% or 40%, the sentence is most likely to be a hallucination. C Case Study As shown in Figure 7, our approach outperforms the base model (i.e., the Transformer) in translation adequacy. In case 1, the base model generates “on Tuesday”, which is unrelated to the source sentence, i.e., hallucination, and under-translates “November 5” and “the website of the Chinese embassy in Mongolia” information in the source sentence. However, our approach translates the above two segments well. In Case 2, the base model reverses the chronological order of the source sentence, thus generates a mis-translation, while our model translates perfectly. In Case 3, the base model neglects two main segments of the source sentence (the text in bold blue font) and leads to the inadequacy problem. However, our model takes them into account. According to the three examples, we conclude that our approach alleviates the inadequacy problem which is extremely harmful to NMT. 3468 Case 1 SRC 1 中新网11月5日电据中国驻蒙古国大使馆网站4日消息,近日,中国公民郭玉芹和毛润新在 蒙旅游期间失联。 REF 1 Report on November 5 of China News: the website of the Chinese embassy in Mongolia reported on November 5 that Chinese citizens Guo Yuqin and Mao Runxin had been missing when traveling in Mongolia. BASE 1 Chinese citizens Guo Yu-Qin and Mao Yunxin lost their ties during a trip to Mongolia, China said on Tuesday. OURS 1 Chinese citizens Guo Yuqin and Mao Runxin lost their ties during a trip to Mongolia, according to the website of the Chinese Embassy in Mongolia on November 5. Case 2 SRC 2 对此央视发表快评:这是我国英雄烈士保护法施行后第一个烈士纪念日。 REF 2 For this, CCTV issued a quick comment: this was the first Memorial Day after the implementation of the law for the protection of heroes and martyrs in China. BASE 2 CCTV released a quick comment on this: this is our heroic martyrs protection law after the implementation of the first martyr anniversary. OURS 2 CCTV issued a quick comment on this: this is the first martyr memorial day after the implementation of our country's heroic martyr protection law. Case 3 SRC 3 据外媒报道,南非首都比勒陀利亚郊区的一处保育中心里,两只小狮子一起嬉闹玩耍,很难 看出有任何异常之处,不过它们其实绝无仅有。 REF 3 According to foreign media reports, it was hard for people to find anything unusual in two little lions playing in a conservation center located in the suburb in Pretoria, the capital of South Africa, but they were absolutely unique. BASE 3 It's hard to see anything unusual in a nursing home in a suburb of Pretoria, South Africa's capital, where two lions play together. OURS 3 According to foreign media reports, in a care center on the outskirts of Pretoria, South Africa, two lions play together, it is difficult to see any abnormalities, but they are unique. Figure 7: Several example sentence pairs (SRC, REF) from WMT19 Zh→En test set. We list the translation of the Transformer baseline (BASE) and our MSO method (OURS). The text in bold red font is mistranslated by the base model. The text in bold blue font is mistranslated or under-translated by the base model but translated correctly by our model.
|
2021
|
268
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3469–3483 August 1–6, 2021. ©2021 Association for Computational Linguistics 3469 Towards Emotional Support Dialog Systems Siyang Liu1,2∗, Chujie Zheng1∗, Orianna Demasi3, Sahand Sabour1, Yu Li3, Zhou Yu4, Yong Jiang2, Minlie Huang1† 1The CoAI group, DCST, Institute for Artificial Intelligence, State Key Lab of Intelligent Technology and Systems, 1Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing 100084, China 2Tsinghua-Berkeley Shenzhen Institute, Tsinghua Shenzhen International Graduate School, 2Tsinghua University, Shenzhen, China 3University of California, Davis 4Columbia University [email protected], [email protected], [email protected] Abstract Emotional support is a crucial ability for many conversation scenarios, including social interactions, mental health support, and customer service chats. Following reasonable procedures and using various support skills can help to effectively provide support. However, due to the lack of a well-designed task and corpora of effective emotional support conversations, research on building emotional support into dialog systems remains untouched. In this paper, we define the Emotional Support Conversation (ESC) task and propose an ESC Framework, which is grounded on the Helping Skills Theory (Hill, 2009). We construct an Emotion Support Conversation dataset (ESConv) with rich annotation (especially support strategy) in a help-seeker and supporter mode. To ensure a corpus of high-quality conversations that provide examples of effective emotional support, we take extensive effort to design training tutorials for supporters and several mechanisms for quality control during data collection. Finally, we evaluate state-of-the-art dialog models with respect to the ability to provide emotional support. Our results show the importance of support strategies in providing effective emotional support and the utility of ESConv in training more emotional support systems 1. 1 Introduction Emotional support (ES) aims at reducing individuals’ emotional distress and helping them understand and work through the challenges that they face (Burleson, 2003; Langford et al., 1997; Heaney and Israel, 2008). It is a critical capacity to train into dialog systems that interact with users ∗Equal Contribution. †Corresponding author. 1Our data and codes are available at https://github.com/thu-coai/ Emotional-Support-Conversation. 😿 I feel so frustrated. I should first understand his/her situation... Let me explore his/her experiences 😯 (Question) May I ask why you are feeling frustrated? 😿 My school was closed without any prior warning due to the pandemic. I should comfort him/her when gradually learning about his/her situation (Providing Suggestions) Have you thought about talking to your parents or a close friend about this? 🤔 (Self-disclosure) I understand you. I would also have been really frustrated if that happened to me. 😔 😿 Yeah! I don't even know what is going to happen with our final. Mere comforting cannot solve the problem... Let me help him/her take some action and get out of the difficulty (Reflection of Feelings) That is really upsetting and stressful. Figure 1: An example chat showing effective emotional support (adapted from ESConv) being provided to the help-seeker(left) by the supporter(right). The support strategies (skills) used by the supporter are marked in the parentheses before the utterances. The red bold texts in the dashed boxes highlight the three stages of our proposed ESC Framework (Figure 3). on daily basis (Van der Zwaan et al., 2012; Zhou et al., 2020), particularly for settings that include social interactions (accompanying and cheering up the user), mental health support (comforting a frustrated help-seeker and helping identify the problem), customer service chats (appeasing an angry customer and providing solutions), etc. Recent research has also shown that people prefer dialog systems that can provide more supportive responses (Rains et al., 2020). Research has shown that providing emotional support is not intuitive (Burleson, 2003), so procedures and conversational skills have been suggested (Hill, 2009) to help provide better support through conversation. Such skills can be seen in the example conversation that we collected and is shown in Figure 1. To identify the causes of the help3470 seeker’s distress, the supporter first explores the help-seeker’s problems. Without exploration, the support is unlikely to understand the help-seeker’s experiences and feelings, and thus it may be offensive or even harmful if the supporter would give irrelevant advice, like ‘You could go for a walk to relax’. While learning about the help-seeker’s situation, the supporter may express understanding and empathy to relieve the help-seeker’s frustration by using various skills (e.g., Self-disclosure, Reflection of Feelings, etc.). After understanding the help-seeker’s problem, the supporter may offer suggestions to help the help-seeker cope with the problem. If the supporter only comforts the help-seeker without any inspiration for action to change, the supporter may not effectively help the help-seeker’s emotions improve. Finally, during the data collection of this example conversation, the help-seeker reported that their emotion intensity decreased from 5 to 2 (emotion intensity is labeled in our corpus, we give detailed annotations of this conversation example in Appendix A), which indicates the effectiveness of the ES provided by the supporter. Despite the importance and complexity of ES, research on data-driven ES dialog systems is limited due to a lack of both task design and relevant corpora of conversations that demonstrate diverse ES skills in use. First, existing research systems that relate to emotional chatting (Zhou et al., 2018) or empathetic responding (Rashkin et al., 2019) return messages that are examples of emotion or empathy and are thus limited in functionality, as they are not capable of many other skills that are often used to provide effective ES (Hill, 2009). Figure 2 illustrates the relationship between the three tasks and we provide further discussion in Section 2.1. Second, people are not naturally good at being supportive, so guidelines have been developed to train humans how to be more supportive. Without trained individuals, existing online conversation datasets(Sharma et al., 2020a; Rashkin et al., 2019; Zhong et al., 2020; Sun et al., 2021) do not naturally exhibit examples or elements of supportive conversations. As a result, data-driven models that leverage such corpora (Radford et al., 2019; Zhang et al., 2020; Roller et al., 2020) are limited in their ability to explicitly learn how to utilize support skills and thus provide effective ES. In this paper, we define the task of Emotional Support Conversation (ESC), aiming to provide Emotional Support Conversation Reduce users' emotional distress and help them work through the challenges Empathetic Responding Understand users' feelings and reply accordingly Emotional Chatting Accurately express emotions in responses Figure 2: Emotional support conversations (our work) can include elements of emotional chatting (Zhou et al., 2018) and empathetic responding(Rashkin et al., 2019). support through social interactions (like the interactions between peers, friends, or families) rather than professional counseling, and propose an ESC Framework, which is grounded on the Helping Skills Theory (Hill, 2009) and tailored to be appropriate for a dialog system setting (Figure 3). We carefully design the ESC Framework for a dialog system setting by adapting relevant components of Hill’s Helping Skills model of conversational support. The ESC Framework proposes three stages (Exploration, Comforting and Action), where each stage contains several support strategies (or skills). To facilitate the research of emotional support conversation, we then construct an Emotional Support Conversation dataset, ESConv, and take multiple efforts to ensure rich annotation and that all conversations are quality examples for this particularly complex dialog task. ESConv is collected with crowdworkers chatting in help-seeker and supporter roles. We design tutorials based on the ESC framework and train all the supporters and devise multiple manual and automatic mechanisms to ensure effectiveness of emotional support in conversations. Finally, we evaluate the state-of-the-art models and observe significant improvement in the emotional support provided when various support strategies are utilized. Further analysis of the interactive evaluation results shows the Joint model can mimic human supporters’ behaviors in strategy utilization. We believe our work will facilitate research on more data-driven approaches to build dialog systems capable of providing effective emotional support. 2 Related Work 2.1 Emotional & Empathetic Conversation Figure 2 intuitively shows the relationships among ESC, emotional conversation, and empathetic conversation. Emotion has been shown to be impor3471 tant for building more engaging dialog systems (Zhou et al., 2018; Li et al., 2017; Zhou and Wang, 2018; Huber et al., 2018; Huang et al., 2020). As a notable work of emotional conversation, Zhou et al. (2018) propose Emotional Chatting Machine (ECM) to generate emotional responses given a pre-specified emotion. This task is required to accurately express (designated or not) emotions in generated responses. While ES may include expressing emotions, such as happiness or sadness, it has a broader aim of reducing the user’s emotional distress through the utilization of proper support skills, which is fundamentally different from emotional chatting. Emotional chatting is merely a basic quality of dialog systems, while ES is a more high-level and complex ability that dialog systems are expected to be equipped with. Another related task is empathetic responding (Rashkin et al., 2019; Lin et al., 2019; Majumder et al., 2020; Zandie and Mahoor, 2020; Sharma et al., 2020a; Zhong et al., 2020; Zheng et al., 2021), which aims at understanding users’ feelings and then replying accordingly. For instance, Rashkin et al. (2019) argued that dialog models can generate more empathetic responses by recognizing the interlocutor’s feelings. Effective ES naturally requires expressing empathy according to the help-seeker’s experiences and feelings, as shown in our proposed Emotional Support Framework (Section 3.2, Figure 3). Hence, empathetic responding is only one of the necessary components of emotional support. In addition to empathetic responding, an emotional support conversation needs to explore the users’ problems and help them cope with difficulty. 2.2 Related Datasets for Emotional Support Various works have considered conversations of emotional support in a social context, such as on social media or online forums (Medeiros and Bosse, 2018; Sharma et al., 2020b; Hosseini and Caragea, 2021). Medeiros and Bosse (2018) collected stressrelated posts and response pairs from Twitter and classified replies into supportive categories. In (Sharma et al., 2020b), the post-response pairs from TalkLife and mental health subreddits are annotated with the communication mechanisms of text-based empathy expression (only the data of the Reddit part is publicly available). Hosseini and Caragea (2021) also collected such post-response pairs from online support groups, which have been annotated as needing or expressing support. The dialogues in these corpora are either single-turn interactions (post-response pair) or very short conversations, which limits the potential for effective ES, as ES often requires many turns of interaction (Hill, 2009). 2.3 Emotional Support Dialog Systems Some traditional dialog systems have applied human-crafted rules to provide emotional support responses (Van der Zwaan et al., 2012; van der Zwaan et al., 2012). A recent system has considered a rule-based algorithm that determines the supportive act used in the response and then selects proper replies from the pre-defined list of candidates (Medeiros and Bosse, 2018). Another conversational system designed to provide support for coping with COVID-19 was implemented by identifying topics that users mentioned and then responding with a reflection from a template or a message from a pre-defined lexicon (Welch et al., 2020). Few studies have focused on generating supportive responses, and those that have have been limited in scope. For example, Shen et al. (2020) explored how to generate supportive responses via reflecting on user input. 3 Emotional Support Conversation 3.1 Task Definition When a user is in a bad emotional state, perhaps due to a particular problem, they may seek help to improve their emotional state. In this setting, the user can be tagged with a negative emotion label e, a emotion intensity level l (e.g., ranging from 1 to 5), and an underlying challenge that the user is going through. The supporter (or the system) needs to comfort the user in a conversation with support skills to lower their intensity level. Note that the user’s state is unknown to the supporter prior to the conversation. During the conversation, the supporter needs to identify the problem that the user is facing, comfort the user, and then provide some suggestions or information to help the user take action to cope with their problem. An emotional support conversation is effective if the intensity level of the user is lowered at the end of the conversation, or more concretely, if the supporter can effectively identify the problem, comfort the user, and provide solutions or suggestions. The ESC task has several sub-problems: (1) Support strategy selection and strategy-constrained response generation. As shown in our later experiments (Section 6.4), the timing of applying strategies is relevant to the effectiveness of ES. It is thus important that a generated response conforms to a 3472 Strategies Stages Examples Lexical Features Question Can you talk more about your feelings at that time? do you (15.0), are you (13.8), how (13.7), what (12.3), do (11.5) Restatement or Paraphrasing It sounds that you feel like everyone is ignoring you. Is it correct? is that (8.2), so you (8.2), it sounds (7.1), correct (7.1), so (6.6) Reflection of Feelings I understand how anxious you are. can tell (7.4), understand how (5.8), are feeling (5.1), tell (5.1), understand (4.9) Self-disclosure I feel the same way! I also don't know what to say to strangers. my (15.3), was (10.5), me (10.2), had (9.7), myself (7.8) Affirmation and Reassurance You've done your best and I believe you will get it! its (5.7), thats (5.6), will (5.4), through this (5.1), you will (4.7) Providing Suggestions Deep breaths can help people calm down. Could you try to take a few deep breaths? maybe (7.3), if (6.5), have you (6.4), talk to (5.8), suggest (5.8) Information Apparently, lots of research has found that getting enough sleep before an exam can help students perform better. there are (4.4), will (3.8), available (3.7), seen (3.3), possible (3.3) Others I am glad to help you! welcome (9.6), hope (9.6), glad (7.3), thank (7.0), hope you (6.9) ③Action Help the seeker solve the problems ②Comforting Comfort the seeker through expressing empathy and understanding ①Exploration Explore to identify the problems Figure 3: Overview of our proposed ESC Framework. It contains three stages and suggested support strategies. The procedure of emotional support generally follows the order: 1⃝Exploration →2⃝Comforting →3⃝Action (as indicated by the black arrows), but it can also be adapted to the individual conversation as needed (indicated by the dashed gray arrows). The column of “Lexical Features” displays top 5 unigrams or bigrams associated with messages that use each strategy in our dataset. Each feature is ranked by the rounded z-scored log odds ratios (Monroe et al., 2008) in the parentheses. specified strategy. (2) Emotion state modeling. It is important to model and track the user’s emotion state dynamically, both for dynamic strategy selection and for measuring the effectiveness of ESC. (3) Evaluation of support effectiveness. In addition to the traditional dimension of evaluating a conversation’s relevance, coherence, and user engagement, ESC raises a new dimension of evaluating the effectiveness of ES. 3.2 ESC Framework We present an ESC Framework, which characterizes the procedure of emotional support into three stages, each with several suggested support strategies. We ground the ESC Framework on Hill’s Helping Skills Theory (Hill, 2009) and adapt it more appropriate for a dialog system setting, aiming to provide support through social interactions (like the interactions between peers, friends, or families) rather than merely professional counseling. An overview of the conversational stages and strategies in the ESC Framework is shown in Figure 3. Stages Hill (2009) proposes three stages of supporting people: exploration (exploring to help the help-seeker identify the problems), insight (helping the help-seeker move to new depths of selfunderstanding), and action (helping the help-seeker make decisions on actions to cope with the problems). However, we note that insight usually requires re-interpreting users’ behaviors and feelings, which is both difficult and risky for the supporters without sufficient support experience. We thus adapt insight to comforting (defined as providing support through empathy and understanding). While it is suggested that emotional support conversations target these three ordered stages, in practice conversations cannot follow a fixed or linear order and must adapt appropriately. As suggested in (Hill, 2009), the three stages can be flexibly adjusted to meet the help-seeker’s needs. Strategies Hill (2009) also provides several recommended conversational skills for each stage. Some of the described skills are not appropriate2 in a dialog system setting without professional supervision and experience. To adapt these skills appropriate to the dialog system setting, we extract seven methods from these skills (along with an “Others” one), which we called strategies in our task and hereafter. We provide a detailed definition of each strategy in Appendix B. 4 Data Collection To facilitate the research of emotional support skills in dialog systems, we introduce an Emotional Support Conversation Dataset, ESConv, which is collected in a help-seeker and supporter mode with crowdworkers. As high-quality conversation examples are needed for this complex task, we took tremendous effort to try to ensure the effectiveness of ES in conversations. Our efforts included the following major aspects: (1) Because providing conversational support is a skill that must be trained 2For instance, one skill named challenging refers to pointing out the discrepancies or irrational beliefs that the helpseeker is unaware of or unwilling to change. Such skills usually require professional experience, which is too difficult for an average person. 3473 for supporters to be effective (Burleson, 2003), we design a tutorial with the ESC Framework and train crowdworkers to be supporters. Only those who pass the examination are admitted to the task. (2) We require help-seekers to complete a pre-chat survey on their problems and emotions and to provide feedback during and after the conversations. (3) We devise and use multiple manual or automatic mechanisms to filter out the low-quality conversations after collecting raw dialog data. 4.1 Supporter-specific Tasks Training and Examination To teach crowdworkers how to provide effective emotional support, we designed a tutorial with the ESC Framework. Inspired by 7cups (7cups.com) (Baumel, 2015), we developed eleven sub-tasks (3 + 8) to help workers to learn the definitions of the three stages and the eight support strategies. Each sub-task includes an example conversation excerpt and a corresponding quiz question. As noted in Section 3.2, we also informed participants that following a fixed order may not be possible and that they may need to be flexible with adjusting the stage transitions. Strategy Annotation To encourage supporters to use the ESC support strategies during the conversation and to structure the resulting dataset, we ask the supporter to first select a proper strategy that they would like to use according to the dialog context. They are then able to write an utterance reflecting their selected strategy. We encourage supporters to send multiple messages if they would like to use multiple strategies to provide support. Post-chat Survey After each conversation, the supporter is asked to rate the extent that the seeker goes into detail about their problems on five-point Likert scales. 4.2 Seeker-specific Tasks Pre-chat Survey Before each conversation, the help-seeker was asked to complete the following survey: (1) Problem & emotion category: the helpseeker should select one problem from 5 options and one emotion from 7 options (the options were based on conversations collected in pilot data collection trials). (2) Emotion intensity: a score from 1 to 5 (the larger number indicates a more intense emotion). (3) Situation: open text describing the causes of the emotional problem. (4) Experience origin: whether the described situation was the current experience of the help-seeker or based on prior life circumstances. We found that 75.2% of converRoles Aspects Criteria Supporter (≥3)* Understanding the help-seeker’s experiences and feelings (rated by the helpseeker) >= 3 Relevance of the utterances to the conversation topic (rated by the help-seeker) >= 4 Average length of utterances >= 8 Improvement in the help-seeker’s emotion intensity (rated by the helpseeker)** >= 1 Seeker Describing details about the own emotional problems (rated by the supporter) not required Average length of utterances >= 6 Table 1: Criteria of high-quality conversations. * denotes that supporters must meet at least two of the three criteria. In **, the improvement of the help-seeker’s emotion intensity was calculated by subtracting the intensity after from that before the conversation. sations originated from the help-seekers’ current experiences. Feedback During the conversation, the helpseeker was asked to give feedback after every two new utterances they received from the supporter. Their feedback scored the helpfulness of the supporter messages on a 5-star scale. We divided each conversation into three phases and calculated the average feedback score for each phase. The scores in the three phases are 4.03, 4.30, and 4.44 respectively, indicating that the supporters were sufficiently trained to effectively help the help-seekers feel better. Post-chat Survey After each conversation, the help-seeker is asked to rate their emotion and the performance of the supporter on the following fivepoint Likert scales: (1) Their emotion intensity after the emotional support conversation (a decrease from the intensity before the conversation reflects emotion improvement), (2) the supporter’s empathy and understanding of the help-seeker’s experiences and feelings, and (3) the relevance of the supporter’s responses to the conversation topic. 4.3 Quality Control We use multiple methods to ensure that the corpus contains high-quality examples of effective emotional support conversations. Preliminary Filtering Mechanisms When recruiting participants for the supporter role, we initially received 5,449 applicants, but only 425 (7.8%) passed the training tutorial. From the 2,472 conversations that we initially collected, we filtered out those that were not finished by the help-seekers or that had fewer than 16 utterances. This filtering 3474 left 1,342 conversations (54.3%) for consideration. Auto-approval Program for Qualified Conversations We carefully designed the auto-approval program, which is the most important part of data quality control. This program uses criteria based on the post-chat survey responses from both roles and the length of utterances, which are summarized in Table 1. These criteria are based on initial human reviewing results. We show how to choose these auto-approval criteria in Appendix D. The computed average emotion intensity before conversations is 4.04 and 2.14 after. Such improvement demonstrates the effectiveness of the emotional support provided by the supporters. In a small number of conversations, the help-seeker did not finish the post-chat surveys, so we added another criterion for these conversations requiring that the last two feedback scores from the help-seekers are both greater than 4. Thus, among all the conversations without post-chat surveys, only those who met both (2) and (3) were qualified. Using these quality criteria, 1,053 (78.5% of 1,342) of collected conversations were qualified. Annotation Correction To further ensure data quality, we reviewed and revised incorrect annotations of support strategy and seeker’s emotion intensity. (1) For strategy annotation correction, we asked new qualified supporters to review and revise annotations on previously collected conversations as necessary, which led to 2,545 utterances (17.1%) being reviewed. We manually reviewed annotations where more than 75% of reviewers disagreed and revised 139 of them. (2) According to the auto-approval criteria (Table 7), a conversation can be qualified when the score of the seeker’s emotion improvement is less than one, but the other three criteria are satisfied. Upon review, we found this to most often result from seekers mistaking negative emotion intensity as the positiveness of their emotion. We manually re-checked and revised the emotion intensity of these conversations by using other helpful information, such as the responses to the post-chat survey open question and the seekers’ feedback scores during the chat. Of 130 such conversations, 92% were revised and included in the corpus. 5 Data Characteristics 5.1 Statistics The overall statistics of the 1,053 ESConv examples are shown in table 2. Relatively long conversations (avg. 29.8 utterances) indicate that providing Category Total Supporter Seeker # dialogues 1,053 Avg. Minutes per Chat 22.6 # Workers 854 425 532 # Utterances 31,410 14,855 16,555 Avg. length of dialogues 29.8 14.1 15.7 Avg. length of utterances 17.8 20.2 15.7 Table 2: Statistics of ESConv. Categories Num Proportion Seeker’s Problem Ongoing Depression 306 29.1% Job Crisis 233 22.1% Breakup with Partner 216 20.5% Problems with Friends 159 15.1% Academic Pressure 139 13.2% Overall 1,053 100.0% Seeker’s Emotion Anxiety 281 26.7% Depression 276 26.2% Sadness 250 23.7% Anger 96 9.1% Fear 88 8.4% Disgust 32 3.0% Shame 30 2.8% Overall 1,053 100.0% Seeker’s Feedback 1 (Very Bad) 71 1.1% 2 (Bad) 183 2.9% 3 (Average) 960 15.5% 4 (Good) 1,855 29.9% 5 (Excellent) 3,144 50.6% Overall 6,213 100.0% Support Strategy Question 3,109 20.9% Restatement or Paraphrasing 883 5.9% Reflection of Feelings 1,156 7.8% Self-disclosure 1,396 9.4% Affirmation and Reassurance 2,388 16.1% Providing Suggestions 2,323 15.6% Information 904 6.1% Others 2,696 18.1% Overall 14,855 100.0% Table 3: Statistics of all the annotations, including the help-seekers’ problems, emotions, feedback, and the support strategies. effective ES usually requires many turns of interaction and considerably more turns than typical for previous emotional chatting (Zhou et al., 2018) or empathetic dialog (Rashkin et al., 2019) datasets. We also present the statistics of other annotations in Table 3. Perhaps due to the current outbreak of COVID-19, ongoing depression and job crisis are the most commonly stated problems for the help-seekers and depression and anxiety are the most commonly noted emotions. From the helpseekers’ feedback, we found that they are usually highly satisfied with the emotional support, which further indicates that the training tutorial based on the ESC Framework indeed helps supporters learn to provide effective ES. We release all these annotations to facilitate further research. 3475 Figure 4: The distribution of strategies at different conversation progress. 5.2 Strategy Analysis Lexical Features We extracted lexical features of each strategy by calculating the log odds ratio, informative Dirichlet prior (Monroe et al., 2008) of all the unigrams and bigrams for each strategy contrasting to all other strategies. We list the top 5 phrases for each strategy in Figure 3. Those strategies are all significantly (z-score > 3) associated with certain phrases (e.g., Question with “are you”, Self-disclosure with “me”). Strategy Distribution We computed the distribution of strategies at different phases of the conversation. For a conversation with L utterances in total, the k-th (1 ≤k ≤L) utterance is from the supporter and adopts the strategy st, we say that it locates at the conversation progress k/L. Specifically, we split the conversation progress into six intervals: [0, 1] = S4 i=0[i/5, (i + 1)/5) S{1}. Then, for all the conversations in ESConv, we counted the proportions of different strategies in the six intervals. We split the conversation progress into six intervals: [0, 1] = S4 i=0[i/5, (i + 1)/5) S{1} and drew the distributions on the six intervals at six points i/5(i = 0, . . . , 5) respectively and connected them, finally obtaining Figure 4. The supporters generally follow the stage order suggested by the ESC Framework (Figure 3), but there is also flexible adjustment of stages and adoption of strategies. For instance, at the early phase of conversation, the supporters usually adopt exploratory strategies such as Question. After knowing help-seekers’ situations, the supporters tend to provide their opinions (such as Providing Suggestions). Throughout the entire conversation, the comforting strategies (such as Affirmation and Reassurance) are used and label a relatively constant proportion of messages. Strategy Transition We present the top-5 most frequent strategy transitions with 3 / 4 hops in Appendix (Table 6). These transitions indicate that, as the tutorial of ESC framework trains, supporters usually ask questions and explore the help-seekers’ situations before comforting the help-seekers. 6 Experiments Our experiments focus on two key questions: (1) How much can ESConv with strategy annotation improve state-of-the-art generative dialog models? (2) Can these models learn to provide effective emotional support from ESConv? 6.1 Backbone Models We used two state-of-the-art pre-trained models as the backbones of the compared variant models: BlenderBot BlenderBot (Roller et al., 2020) is an open-domain conversational agent trained with multiple communication skills, including empathetic responding. As such, BlenderBot should be capable of providing ES for users to some extent. We used the small version3 of BlenderBot in experiments, because the larger versions have the limitation of maximum context length 128, which we found harms the model performance and response coherence. DialoGPT We additionally evaluated DialoGPT (Zhang et al., 2020), which is a GPT-2-based model pre-trained on large-scale dialog corpora. We used the small version4. 6.2 Variant Models Taking each of the above pre-trained models as the backbone, we built the following variant models: Vanilla Directly fine-tuning the backbone model on ESConv with no access to strategy annotations. Formally, suppose the flattened dialog history is x and the response to be generated is y, we maximize the conditional probability: P(y|x) = Q|y| i=1 P (yi|x, y≤i). Variants with strategy To incorporate the strategy annotation into the backbone model, we used a special token to represent each strategy. For each utterance y from the supporters, we appended the corresponding strategy token before this utterance: ˜y = [st] ⊕y, where [st] denotes the special token of the used strategy. Then, taking the flattened dialog history x as input, the model generates the response conditioned on the first predicted (or designated) strategy token: P(˜y|x) = P([st]|x) Q|y| i=1 P (yi|x, [st], y<i). 3https://huggingface.co/facebook/ BlenderBotbot_small-90M 4https://huggingface.co/microsoft/ DialoGPT-small 3476 Backbones Variants PPL B-2 R-L Extrema DialoGPT Vanilla 15.51 5.13 15.26 49.80 Joint 5.00 15.09 49.97 Oracle 15.19 5.52 15.82 50.18 BlenderBot Vanilla 16.23 5.45 15.43 50.49 Joint 5.35 15.46 50.27 Oracle 16.03 6.31 17.90 51.65 Table 4: Results of automatic evaluation. The results in bold are significantly better than all the competitors (Student’s t-test, p-value < 0.05). We studied three variants that use strategy annotation in the later experiments. (1) Oracle: responses are generated conditioned on the gold reference strategy tokens. (2) Joint: responses are generated conditioned on predicted (sampled) strategy tokens. (3) Random: responses are generated conditioned on randomly selected strategies. Implementation details are in Appendix C. 6.3 Automatic Evaluation To investigate the impact of utilizing support strategies on the model performance with either BlenderBot or DialoGPT as the backbone, we compared the performance of the Vanilla, Joint, and Oracle variants described above. The automatic metrics we adopted include perplexity (PPL), BLEU-2 (B2) (Papineni et al., 2002), ROUGE-L (R-L) (Lin, 2004), and the BOW Embedding-based (Liu et al., 2016) Extrema matching score. The metrics except PPL were calculated with an NLG evaluation toolkit5 (Sharma et al., 2017) with responses tokenized by NLTK6 (Loper and Bird, 2002). There are three major findings from the experiments (Table 4). (1) The Oracle models are significantly superior to the Vanilla models on all the metrics, indicating the great utility of support strategies. (2) The Joint models obtain sightly lower scores than the Vanilla models, as, if the predicted strategy is different from the ground truth, the generated response will be much different from the reference response. However, learning to predict strategies is important when there are no ground truth labels provided, and we will further investigate the performance of the Joint model in human interactive evaluation (Section 6.4). (3) The BlenderBot variants consistently perform better than the DialoGPT ones, indicating that BlenderBot is more suitable for the ESC task. Thus, in the subsequent human evaluation, we will focus evaluation on the Blender5https://github.com/Maluuba/nlg-eval 6https://www.nltk.org/ Joint vs. w/o ft Vanilla Random Win Lose Win Lose Win Lose Fluency 71‡ 24 52† 35 53† 35 Identification 65‡ 25 50 34 54† 37 Comforting 75‡ 20 54‡ 34 47 39 Suggestion 72‡ 21 47 39 48† 27 Overall 73‡ 20 51† 34 56‡ 36 Table 5: Results of the human interactive evaluation. Ties are not shown. All the models use BlenderBot as the backbone. ‘w/o ft’ denotes the BlenderBot model without fine-tuning on ESConv. The Joint model outperforms all the competitors on all the metrics (sign test, †/‡ denote p-value < 0.1/0.05 respectively). Bot variants. 6.4 Human Interactive Evaluation We recruited participants from Amazon Mechanical Turk to chat with the models. The online tests were conducted on the same platform as our data collection, but with the role of supporter taken by a model. Each participant chatted with two different models that were randomly ordered to avoid exposure bias. Participants were asked to compare the two models based on the following questions: (1) Fluency: which bot’s responses were more fluent and understandable? (2) Identification: which bot explored your situation more in depth and was more helpful in identifying your problems? (3) Comforting: which bot was more skillful in comforting you? (4) Suggestion: which bot gave you more helpful suggestions for your problems? (5) Overall: generally, which bot’s emotional support do you prefer? The metrics in (2), (3), and (4) correspond to the three stages in the ESC Framework. We compare three pairs of models: (a) Joint vs. BlenderBot (without fine-tuning on ESConv), (b) Joint vs. Vanilla, and (c) Joint vs. Random (using randomly selected strategies). To better simulate the real strategy occurrence, the Random model randomly selects a strategy following the strategy distribution in ESConv (Table 3). Each pair of models was compared by 100 conversations with human participants (Table 5). The results of comparison (a) show that BlenderBot’s capability of providing ES is significantly improved on all the metrics after being fine-tuned on ESConv. From comparison (b), we found that utilizing strategies can better comfort the users. The results of comparison (c) also demonstrate that the proper timing of strategies is critical to help users identify their problems and to provide effective suggestions. In general, through being fine-tuned with the su3477 Figure 5: The Joint model’s generation distribution. The meanings of all the graphics and abbreviations are consistent with Figure 4. pervision of strategy prediction on ESConv, the pre-trained models become preferred by the users, which proves the high-quality and utility of ESConv. 6.5 Further Analysis of Human Interactive Evaluation In this section, we explore what the dialog models learned from ESConv. Firstly, we analyzed the strategy distribution based on the 300 dialogs between users and the Joint model in human interactive experiments. We can see in Figure 5 (the calculation was consistent with Figure 4), the strategies that the Joint model adopted have a very similar distribution compared with the truth distribution in ESConv (Figure 4). It provides important evidence that models mimic strategy selection and utilization as human supporters do to achieve more effective ES. Secondly, we present a case study in Figure 7. We see in cases that the Joint model provides more supportive responses and uses more skills in conversation, while BlenderBot without fine-tuning seems not to understand the user’s distress very well and prefers to talk more about itself. This may imply that having more supportive responses and a diverse set of support strategies are crucial to effective emotional support. 7 Conclusion In this work, we define the task of Emotional Support Conversation and present an ESC Framework. The ESC Framework is adapted from the Helping Skills Theory into a dialog system setting, which characterizes three stages with corresponding support strategies useful at each stage. We then construct an Emotional Support Conversation dataset, ESConv. We carefully design the process of data collection and devise multiple mechanisms to ensure the effectiveness of ES in conversations. Finally, we evaluate the ES ability with state-of-theart dialog models. Experimental results show the potential utility of ESConv in terms of improving dialog systems’ ability to provide effective ES. Our work can facilitate future research of ES dialog systems, as well as improve models for other conversation scenarios where emotional support plays an important role. Strategy selection and realization, user state modeling, and task evaluation are important directions for further research. Acknowledgments This work was supported by the NSFC projects (Key project with No. 61936010 and regular project with No. 61876096). This work was also supported by the Guoqiang Institute of Tsinghua University, with Grant No. 2019GQG1 and 2020GQG0005. Ethical Considerations There are many types and levels of support that humans can seek to provide, e.g., professional versus peer support, and some of these levels may be inappropriate, unrealistic, and too risky for systems to deliver. However, as dialog systems become more common in daily use, opportunities will arise when at least some basic level of supportive statements may be required. In developing the ESC Framework, we have carefully considered which elements of conversational support may be relevant for a dialog system and omitted elements that are clear oversteps. Considerable additional work is needed to determine what are appropriate levels of support for systems to provide or that can be expected from systems, but our work provides a cautious, yet concrete, step towards developing systems capable of reasonably modest levels of support. The corpus we construct can also provide examples to enable future work that probes the ethical extent to which systems can or should provide support. In addition to these broader ethical considerations, we have sought to ethically conduct this study, including by transparently communicating with crowdworkers about data use and study intent, compensating workers at a reasonable hourly wage, and obtaining study approval from the Institutional Review Board. References Amit Baumel. 2015. Online emotional support delivered by trained volunteers: users’ satisfaction and their perception of the service compared to psychotherapy. Journal of Mental Health, 24(5):313– 320. 3478 Brant R Burleson. 2003. Emotional support skill. HANDBOOK OF COMMUNICATION AND SOCIAL INTERACTION SKILLS, page 551. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. Catherine A Heaney and Barbara A Israel. 2008. Social networks and social support. Health behavior and health education: Theory, research, and practice, 4:189–210. Clara E Hill. 2009. Helping skills: Facilitating, exploration, insight, and action. American Psychological Association. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. In International Conference on Learning Representations. Mahshid Hosseini and Cornelia Caragea. 2021. It takes two to empathize: One to seek and one to provide. Proceedings of the 35th American Association for Artificial Intelligence (AAAI 2021). Minlie Huang, Xiaoyan Zhu, and Jianfeng Gao. 2020. Challenges in building intelligent open-domain dialog systems. ACM Transactions on Information Systems (TOIS), 38(3):1–32. Bernd Huber, Daniel McDuff, Chris Brockett, Michel Galley, and Bill Dolan. 2018. Emotional dialogue generation using image-grounded language models. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, pages 1–12. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Catherine Penny Hinson Langford, Juanita Bowsher, Joseph P Maloney, and Patricia P Lillis. 1997. Social support: a conceptual analysis. Journal of advanced nursing, 25(1):95–100. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986–995, Taipei, Taiwan. Asian Federation of Natural Language Processing. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Zhaojiang Lin, Andrea Madotto, Jamin Shin, Peng Xu, and Pascale Fung. 2019. MoEL: Mixture of empathetic listeners. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 121–132, Hong Kong, China. Association for Computational Linguistics. Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2122–2132, Austin, Texas. Association for Computational Linguistics. Edward Loper and Steven Bird. 2002. Nltk: the natural language toolkit. arXiv preprint cs/0205028. Navonil Majumder, Pengfei Hong, Shanshan Peng, Jiankun Lu, Deepanway Ghosal, Alexander Gelbukh, Rada Mihalcea, and Soujanya Poria. 2020. MIME: MIMicking emotions for empathetic response generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8968–8979, Online. Association for Computational Linguistics. Lenin Medeiros and Tibor Bosse. 2018. Using crowdsourcing for the development of online emotional support agents. In International Conference on Practical Applications of Agents and Multi-Agent Systems, pages 196–209. Springer. Burt L Monroe, Michael P Colaresi, and Kevin M Quinn. 2008. Fightin’words: Lexical feature selection and evaluation for identifying the content of political conflict. Political Analysis, 16(4):372–403. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. Stephen A Rains, Corey A Pavlich, Bethany Lutovsky, Eric Tsetsi, and Anjali Ashtaputre. 2020. Support seeker expectations, support message quality, and supportive interaction processes and outcomes: The case of the comforting computer program revisited. Journal of Social and Personal Relationships, 37(2):647–666. Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic opendomain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370–5381, Florence, Italy. Association for Computational Linguistics. 3479 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M Smith, et al. 2020. Recipes for building an open-domain chatbot. arXiv preprint arXiv:2004.13637. Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020a. A computational approach to understanding empathy expressed in text-based mental health support. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5263–5276, Online. Association for Computational Linguistics. Ashish Sharma, Adam S Miner, David C Atkins, and Tim Althoff. 2020b. A computational approach to understanding empathy expressed in text-based mental health support. arXiv preprint arXiv:2009.08441. Shikhar Sharma, Layla El Asri, Hannes Schulz, and Jeremie Zumer. 2017. Relevance of unsupervised metrics in task-oriented dialogue for evaluating natural language generation. arXiv preprint arXiv:1706.09799. Siqi Shen, Charles Welch, Rada Mihalcea, and Verónica Pérez-Rosas. 2020. Counseling-style reflection generation using generative pretrained transformers with augmented context. In Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 10–20, 1st virtual meeting. Association for Computational Linguistics. Hao Sun, Zhenru Lin, Chujie Zheng, Siyang Liu, and Minlie Huang. 2021. Psyqa: A chinese dataset for generating long counseling text for mental health support. In Findings of the Association for Computational Linguistics: ACL 2021. Janneke M van der Zwaan, Virginia Dignum, and Catholijn M Jonker. 2012. A conversation model enabling intelligent agents to give emotional support. In Modern Advances in Intelligent Systems and Tools, pages 47–52. Springer. JM Van der Zwaan, V Dignum, and CM Jonker. 2012. A bdi dialogue agent for social support: Specification and evaluation method. In AAMAS 2012: Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, Workshop on Emotional and Empathic Agents, Valencia, Spain, 4-8 June 2012; authors version. International Foundation for Autonomous Agents and Multiagent Systems (IFAAMAS). Charles Welch, Allison Lahnala, Veronica Perez-Rosas, Siqi Shen, Sarah Seraj, Larry An, Kenneth Resnicow, James Pennebaker, and Rada Mihalcea. 2020. Expressive interviewing: A conversational system for coping with COVID-19. In Proceedings of the 1st Workshop on NLP for COVID-19 (Part 2) at EMNLP 2020, Online. Association for Computational Linguistics. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Rohola Zandie and Mohammad H Mahoor. 2020. Emptransfo: A multi-head transformer architecture for creating empathetic dialog systems. arXiv preprint arXiv:2003.02958. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT : Largescale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270– 278, Online. Association for Computational Linguistics. Chujie Zheng, Yong Liu, Wei Chen, Yongcai Leng, and Minlie Huang. 2021. Comae: A multi-factor hierarchical framework for empathetic response generation. In Findings of the Association for Computational Linguistics: ACL 2021. Peixiang Zhong, Chen Zhang, Hao Wang, Yong Liu, and Chunyan Miao. 2020. Towards persona-based empathetic conversational models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6556–6566, Online. Association for Computational Linguistics. Hao Zhou, Minlie Huang, Tianyang Zhang, Xiaoyan Zhu, and Bing Liu. 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. In Proceedings of the AAAI Conference on Artificial Intelligence. Li Zhou, Jianfeng Gao, Di Li, and Heung-Yeung Shum. 2020. The design and implementation of xiaoice, an empathetic social chatbot. Computational Linguistics, 46(1):53–93. Xianda Zhou and William Yang Wang. 2018. MojiTalk: Generating emotional responses at scale. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1128–1137, Melbourne, Australia. Association for Computational Linguistics. 3480 A Data Example from ESConv Here we detail the conversation that Figure 1 demonstrates to show the annotations that our dataset contains. The detailed example can be seen in Figure 6. Each pre-chat survey of conversation is labeled its problem category, emotion category, emotion intensity, and a brief of the situation of the seeker. In the context of each conversation, the strategies used by supporters are labeled and the seeker’s feedback score per two utterances of the supporter’s responses are also given in our dataset. Note that not all conversations have the label of emotion intensity after the conversation. It is because some seekers don’t finish the post-chat survey but we still include such conversations into our dataset due to their high quality that meets our criteria. Pre-chat Survey Problem: Academic pressure Emotion: Anxiety Emotion Intensity: 5 Situation: My school was closed due to the pandemic. Conversation Seeker: I feel so frustrated. Supporter (Questions): May I ask why you are feeling frustrated? Seeker: My school was closed without any prior warning due to the pandemic. Supporter (Affirmation and Reassurance): That is really upsetting and stressful. I commend you for having to deal with that! Supporter (Self-disclosure): I know I would have been really frustrated if that happened to me. System: Do those messages help you feel better? ⭐⭐⭐⭐ Seeker: Yeah! I don't even know what is going to happen with our finals now. Supporter (Restatement or Paraphrasing): I can see how that would make you frustrated. Supporter (Providing Suggestions): Have you thought about talking to your parents or a close friend about this? System: Do those messages help you feel better? ⭐⭐⭐⭐⭐ … Seeker: I really appreciate your assistance today. I feel better and will take some action this week. Thank you! Supporter (Others): You're very welcome! Feel free to chat if you need anything else! Post-chat Survey Emotion Intensity: 2 Figure 6: Data example from ESConv. Blue text: the help-seeker’s pre-chat survey. Red text: strategies used by the supporter. Orange text: the question that the systems ask help-seeker to evaluate the helpfulness per two utterances from the supporter. Thus the stars denote the seeker’s feedback score. B Definitions of Strategies Question Asking for information related to the problem to help the help-seeker articulate the issues that they face. Open-ended questions are best, Strategy Transition Proportion 3-Hop Qu →AR →Qu 19.65 ‰ Qu →RP →Qu 14.55 ‰ Qu →RP →AR 12.37 ‰ AR →Qu →AR 11.96 ‰ Ot →Qu →RP 11.64 ‰ 4-Hop Qu →AR →Qu →AR 7.00 ‰ AR →Qu →AR →Qu 5.13 ‰ Ot →Qu →RP →Qu 4.20 ‰ PS →Ot →PS →Ot 3.85 ‰ Qu →RP →AR →Qu 3.85 ‰ Table 6: Proportions of top-5 strategy transitions in supporter utterances. Abbreviations are consistent with Figure 4. and closed questions can be used to get specific information. Restatement or Paraphrasing A simple, more concise rephrasing of the help-seeker’s statements that could help them see their situation more clearly. Reflection of Feelings Articulate and describe the help-seeker’s feelings. Self-disclosure Divulge similar experiences that you have had or emotions that you share with the help-seeker to express your empathy. Affirmation and Reassurance Affirm the helpseeker’s strengths, motivation, and capabilities and provide reassurance and encouragement. Providing Suggestions Provide suggestions about how to change, but be careful to not overstep and tell them what to do. Information Provide useful information to the help-seeker, for example with data, facts, opinions, resources, or by answering questions. Others Exchange pleasantries and use other support strategies that do not fall into the above categories. C Implementation Details The implementation of all models was based on Transformer library7 (Wolf et al., 2020). We split ESConv into the sets of training / validation / test with the proportions of 6:2:2. since the conversations in ESConv usually have long turns, we cut each dialog into conversation pieces with 5 utterances, which contain one supporter’s response and the preceding 4 utterances. During training, we trained all the models with Adam (Kingma and Ba, 2014) optimizer with learning rate 5e−5. All the models were trained for 5 epochs, and the check7https://github.com/huggingface/ transformers 3481 points with the lowest perplexity scores on the validation set were selected for evaluation. During inference, we masked other tokens and sampled a strategy token at the first position of the response. For the Random variant models, we sampled strategies randomly following the strategy distribution in ESConv, which is reported in Table 3. The response were decoded by Top-k and Top-p sampling with p = 0.9 (Holtzman et al., 2019), k = 30, temperature τ = 0.7, and the repetition penalty 1.03. D Auto-Approval Criteria To establish each criterion of the auto-approval program as shown in the main paper (Section 3.4), we searched the most suitable thresholds for each filtering rule. We recruited three well-trained human annotators, who have also received the same training procedures as the supporter applicants did. We then randomly sampled 100 conversations from our dataset and asked the three annotators to judge whether the conversations are qualified for providing effective emotional support. Next, we utilized the post-survey results and the lengths of speaker utterances to choose suitable thresholds for filtering rules. We then treated each auto-filtering rule as a rule annotator and computed the Cohen’s Kappa (Cohen, 1960) score between the rule annotator and each human annotator. The agreement scores in Table 7 are Cohen’s Kappa consistency among the agreement scores between each rule annotator and the three human annotators. We selected the thresholds that lead to the second-highest agreement score with human annotators and used these thresholds in the filtering rules. We didn’t use the set of thresholds that has the highest agreement score because the rule based on these thresholds is stricter so that many conversations would be filtered out. However, the second-highest score is only slightly lower than the highest so the rule based on the thresholds of second-highest score can remain more qualified conversations with little accepted cost. As a result, a qualified conversation requires that the supporter must meet at least three of all the four criteria, and the help-seeker must satisfy both of the two corresponding criteria. The final ’rule’ annotator combines the two conditions, and the averaged agreement score between the final rule annotator and the three human annotators is 0.576, indicating significant agreement. E Interface of Data Collection Platform To facilitate readers to have an intuitive understanding of our data collection process, we present an interface diagram of some important steps in the data collection process in Figure 8, which contains the surfaces of support strategy training, supporter’s chatting, help-seeker’s pre-chat survey, help-seeker’s chatting, and post-survey. 3482 Auto-approval Rule Consistency Supporter Seeker Human1 Human2 Human3 Average Improvement Avg. Length Empathy Relevance Detail Avg. Length 1 8 3 4 6 0.545 0.659 0.525 0.576 2 8 3 4 6 0.505 0.566 0.486 0.519 1 8 4 4 6 0.539 0.602 0.519 0.553 1 8 2 4 6 0.539 0.618 0.570 0.576 1 8 3 3 6 0.546 0.630 0.526 0.567 1 8 3 5 6 0.575 0.640 0.555 0.590 1 8 3 4 7 0.539 0.602 0.473 0.538 1 8 3 4 5 0.520 0.551 0.501 0.524 1 8 3 4 3 6 0.505 0.653 0.531 0.563 1 8 3 4 2 6 0.527 0.640 0.508 0.558 1 8 3 4 4 6 0.457 0.599 0.482 0.513 1 9 3 4 6 0.510 0.621 0.490 0.540 1 7 3 4 6 0.515 0.633 0.495 0.548 Table 7: The agreement score between each “rule” annotator and three human annotators. The aspects are consistent with Table 1. The first rule means: 1) The supporter improves the help-seeker’s emotional state as least one score (emotional improvement is calculated as the help-seeker’s emotion intensity before the conversation minus the emotion intensity after the conversation). 2) The average length of the supporter’s utterances is no less than eight. 3) The empathy score of the supporter’s performance is no less than three. 4) The topic relevance score of the supporter’s performance is no less than four. 5) The score that measures how well the help-seeker goes into detail about their problems is not required to be used in the rule (denoted as ‘-’). (6) The average length of the help-seeker’s utterances is no less than six. Problem: Problems with friends Emotion: Sadness Emotion Intensity: 5 Situation: My friend is drinking too much. Joint BlenderBot w/o Fine-tuning Seeker: Hi there. I feel sadness because my friend has a drinking problem. Supporter (Question): I am sorry to hear that. What do you think is going on? Seeker: I think he just got used to drink every day and I can't pursue him to stop at least for a week. Supporter (Question): I can see how that would be difficult for you. Does your friend drink at home or does it happen at work or home? Seeker: Well he is working from home. So he is drinking almost 24/7. I feel so exhausted as well. Supporter (Reflection of Feelings): I understand that you feel like you are exhausted, but I think you can get through this! Seeker: I just need a way to make him stop drinking. Maybe I should stop any communication with him for a while? Supporter (Affirmation and Reassurance): I think it's a great idea! I would do it if I were you! Seeker: Probably that's what I will do. I just can't take this anymore. If he wants he can buy drinks himself as I am tired walking to shop every day and all neighbors thinking that I am an alcoholic. Supporter (Self-disclosure): I completely understand where you are coming from. I was in the same situation as you. Seeker: It was so nice to chat with you, I feel better now. Supporter (Affirmation and Reassurance): You're welcome! I'm glad I could help! Seeker: Hi there. My friend has a drinking problem and I don't know what to do. Supporter: I know it can be difficult, but if you can do it, I'm sure you can find a way. I hope he gets better soon. Seeker: I feel such a sadness inside me. I am trying to help him but he won't listen. Supporter: I understand completely. it is tough, but I am sure you will get through it. Seeker: Every day he is drinking beer and he is not listening to me. Supporter: I can imagine. I feel like that's the only way I can keep up with him. I have tried so hard. I think I just have to try harder than him to get through this. Seeker: Few more months and I think I will give up on him Supporter: That is exactly how I feel... I am trying to find the right person for me and I can't do it! Seeker: I thought you are here to help me? Supporter: It's hard, but we all have to stick together. There are so many people out there. you just have time. Seeker: Maybe I should ignore my friend for a while? Maybe I will feel better? Supporter: I try to just hang out with friends. I have been trying to meet new people in the city but they just don't show up. I just want to be alone and be alone. Seeker: Thank you for your help. Supporter: You're welcome. I think that's what I'll do too. I'm trying to get used to the city life. I feel like I've been stuck in this city for so long. Figure 7: Case Study. Blue text: the help-seeker’s pre-chat survey. Red text: strategies predicted by the model. Green background : utterances that corresponding to the used strategies. Yellow background : utterances where the model talks about itself and is not user-centered. 3483 (a) (b) (c) (d) (e) Figure 8: (a) Support strategy training. (b) Pre-chat survey. (c) The help-seeker’s chatting interface. (d) The supporter’s chatting interface. (e) Post-chat survey.
|
2021
|
269
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 317–327 August 1–6, 2021. ©2021 Association for Computational Linguistics 317 Bridge-Based Active Domain Adaptation for Aspect Term Extraction Zhuang Chen, Tieyun Qian∗ School of Computer Science, Wuhan University, China {zhchen18, qty}@whu.edu.cn Abstract As a fine-grained task, the annotation cost of aspect term extraction is extremely high. Recent attempts alleviate this issue using domain adaptation that transfers common knowledge across domains. Since most aspect terms are domain-specific, they cannot be transferred directly. Existing methods solve this problem by associating aspect terms with pivot words (we call this passive domain adaptation because the transfer of aspect terms relies on the links to pivots). However, all these methods need either manually labeled pivot words or expensive computing resources to build associations. In this paper, we propose a novel active domain adaptation method. Our goal is to transfer aspect terms by actively supplementing transferable knowledge. To this end, we construct syntactic bridges by recognizing syntactic roles as pivots instead of as links to pivots. We also build semantic bridges by retrieving transferable semantic prototypes. Extensive experiments show that our method significantly outperforms previous approaches. 1 Introduction Aspect term extraction (ATE) is a fundamental task in aspect-based sentiment analysis. Given a review sentence “The pizza here is also absolutely delicious.”, ATE aims to extract the term pizza. Recent studies define ATE as a sequence tagging task and propose supervised taggers (Wang et al., 2017; Xu et al., 2018). However, due to the high cost of token-level annotation, the lack of labeled data becomes the main obstacle (Chen and Qian, 2019). To alleviate the data deficiency issue, unsupervised domain adaptation is proposed to transfer knowledge from the labeled source domain to the unlabeled target domain. Since ATE is a tokenlevel task, it is natural to conduct token-level domain adaptation. Then a problem arises: many *Corresponding author. ĺ ĺ ĺ ĺ ĺ ĺ 3HUFHQWDJH 7UDQVIHU3DLU 5HVWDXUDQW /DSWRS 'HYLFH Figure 1: The proportion of source aspect terms that appear in target data. R (Restaurant), L (Laptop), and D (Device) are three datasets from different domains. aspect terms are domain-specific and cannot be transferred directly. We present the proportion of source aspect terms that also appear in target test data in Figure 1. As can be seen, in distant transfer pairs like R→L, only less than 10% of source aspect terms have appeared in target data. Even in a close pair L→D, the proportion is no more than 40%. In other words, there is a wide discrepancy between the data from different domains, and many aspect terms have to be transferred under the guidance of proper references. To solve this problem, previous studies try to associate aspect terms with specific pivot words1. We name these methods passive domain adaptation because the transfer of aspect terms is dependent on their links to the pivots. There are two types of methods along this line. (1) Opinion terms as pivots. Since aspect and opinion terms usually appear in pairs, it is straightforward to extract aspect terms with the indication from opinion terms. Early studies (Li et al., 2012; Ding et al., 2017) use common opinion seeds (e.g., good, fancy) and pre-defined rules (e.g., good→amod→NN) to extract aspect terms across domains. However, it is hard to collect a complete set of seeds or define high-quality rules, and thus these methods often produce inferior performance. Several studies (Wang and Pan, 2018, 2019b) manually annotate all opinion terms in reviews and design neural models to capture aspectopinion relations via multi-task learning. While 1Pivot words are words which behave in the same way for discriminative learning in both domains (Blitzer et al., 2006). 318 getting improvements, these methods induce additional annotation costs. (2) Context terms as pivots. Since pre-trained language models (PLMs) like BERT represent words w.r.t their contexts, recent studies (Xu et al., 2019; Gong et al., 2020) leverage PLMs to transfer aspect terms with common context terms2. However, not all context terms qualify as pivots (e.g., eat). In addition, PLMs like BERT build word associations mainly based on semantic similarity in co-occurring contexts. For an aspect term like pizza, BERT tends to link it to hamburger via a flow like pizza→eat→hamburger. Consequently, it is hard for these methods to identify keyboard in the target domain based on the labeled term pizza in the source domain. 7KHSL]]DKHUHLVDOVRDEVROXWHO\GHOLFLRXV 7KHNH\ERDUGLVLQUHDVRQDEOHVL]H SL]]D NH\ERDUG ʒ DQRXQ ʓ WKHWDUJHWRIDGHWHUPLQHU ʔ WKHVXEMHFWRIDQDGMHFWLYH ʒ 7KHGLVNKHUHLVDOVRDEVROXWHO\ELJ ʓ 7KH26KHUHLVDOVRDEVROXWHO\IDVW ʔ 7KHPRXVHKHUHLVDOVRDEVROXWHO\WLQ\ 6\QWDFWLF%ULGJH 6HPDQWLF%ULGJH LGHQWLI\ UHWULHYH LGHQWLI\ 6RXUFH5HYLHZ 7DUJHW5HYLHZ UHFRJQL]H Figure 2: Illustration of syntactic and semantic bridges. In this paper, we propose a novel active domain adaptation method. Concretely, we construct two types of bridges for all words, which can help transfer aspect terms across domains. An example in Figure 2 shows how to identify the unseen target term keyboard based on the source term pizza. (1) The syntactic bridge aims to recognize transferable syntactic roles for the words across domains. Though pizza and keyboard have almost no semantic relatedness, they often play a similar role in parse trees. In view of this, we treat the involved syntactic roles (including POS tag and dependency relations) of a certain word as its syntactic bridge. Previous studies also utilize dependency information. However, we differ our method from existing ones in that we do not use dependency relations to associate pivot words with aspect terms. Instead, we treat syntactic roles themselves as pivot features and do not need any manually annotated pivot words. (2) The semantic bridge moves one step further by retrieving transferable prototypes. Intuitively, if we correlate pizza with some prototype target terms like {disk, OS, mouse}, the domain discrepancy between the training and testing reviews can be largely reduced. Hence we regard the proto2Context terms denote all words that are not aspect terms. Hence opinion terms form a subset of context terms. types of a certain word as its semantic bridge and design a syntax-enhanced similarity metric to retrieve them. Compared with previous opinion and context term-based methods, building a semantic bridge directly links aspect terms across domains and only requires unlabeled source and target data. Based on the syntactic/semantic bridges, we then develop an end-to-end tagger to fuse reviews with these transferable bridges. We conduct extensive experiments on three datasets. The results show that our method achieves a new state-of-the-art performance with a low computational cost. 2 Related Work Aspect Term Extraction Early researches for ATE mainly involve pre-defined rules (Hu and Liu, 2004; Popescu and Etzioni, 2005; Wu et al., 2009; Qiu et al., 2011) and hand-crafted features (Li et al., 2010; Liu et al., 2012, 2013; Chen et al., 2014). With the development of deep learning, supervised sequence taggers have become the mainstream due to their promising performance (Liu et al., 2015; Wang et al., 2016, 2017; Xu et al., 2018; Ma et al., 2019; Chen and Qian, 2020a). More recently, there emerge many studies that interact ATE with other tasks like aspect-level sentiment classification (Wang et al., 2018; He et al., 2019; Chen and Qian, 2020b). Since these methods highly depend on abundant domain-specific training data, they can hardly scale across the domains where labeled data is absent. Hence it would be more practical to develop unsupervised domain adaptation methods for ATE. Domain Adaptation Many domain adaptation methods have been proposed to solve coarsegrained tasks like text classification (Blitzer et al., 2006; Ganin and Lempitsky, 2015; Guo et al., 2020). The basic idea in coarse-grained tasks is to transfer pivot words, which does not fit ATE well since most aspect terms are domain-specific nonpivot words. There have been a few attempts to this problem, which fall into two lines. (1) One is to model aspect-opinion relations. Early researches use common opinion seeds and pre-defined dependency link rules to build manual features (Jakob and Gurevych, 2010), conduct bootstrapping (Li et al., 2012), and create pseudo target labels (Ding et al., 2017). Due to the incompleteness of seeds and the inflexibility of rules, they often produce inferior performance. Subsequent studies (Wang and Pan, 2018, 2019a,b; Li et al., 2019) manually 319 annotate all opinion terms in reviews and design trainable neural models to capture the relations via multi-task learning. However, they induce extra annotation costs. (2) The other aims to find aspectcontext relations. Xu et al. (2019) post-trains BERT on the cross-domain corpus to enhance its domain adaptation ability. Gong et al. (2020) and Pereg et al. (2020) further incorporate external syntactic information into BERT with auxiliary tasks or modified attention mechanisms, but they still rely on the prior knowledge in BERT. These methods often have more than 100M parameters and involve lots of computing power. Unlike all the aforementioned methods, we do not associate aspect terms with pivot words but actively transfer them via bridges. 3 Methodology In this section, we first introduce the cross-domain ATE task. We then illustrate how to construct syntactic and semantic bridges. Lastly, we present the bridge-based sequence tagging. 3.1 Problem Statement Given a review x = {x1, ..., xn}, we formulate ATE as a sequence tagging task that aims to predict a tag sequence y = {y1, ..., yn}, where each yi ∈ {B, I, O} denotes the beginning of, inside of, and outside of an aspect term. In this paper, we focus on the unsupervised domain adaptation for ATE, i.e., labeled training data is not available in the target domain. Specifically, given a set of labeled data DS = {(xS j , yS j )}NS j=1 from the source domain and a set of unlabeled data DU = {(xU j )}NU j=1 from the target domain, our goal is to predict labels yT for the unseen target test data DT = {(xT j )}NT j=1. 3.2 Bridge Construction Given a review sentence x from either domain, we map it with a lookup table E ∈Rde×|V |, and generate word embeddings E = {e1, ..., en} ∈ Rde×n, where |V | is the vocabulary size, and de is the embedding dimension. For cross-domain ATE, we construct bridges for reviews to help directly transfer aspect terms across two domains. Syntactic Bridge In natural language, linguistic expressions are rich and flexible. In contrast, the syntactic structures are limited and are general across domains. Based on this observation, we propose to build connections between source and target words based on their syntactic roles (POS tags and dependency relations) rather than the lexical items. For example, from the parsing results in the upper part of Figure 3, the word pizza with a POS tag NN and dependency relations {det, nsubj} might be an aspect term, while those with the RB tag and advmod relation might not. Note the sentence “The keyboard is in reasonable size.” in the target domain has similar parsing results. Hence the syntactic roles can serve as supplementary evidence for recognizing aspect terms across domains. Several prior studies (Wang and Pan, 2018, 2019b; Pereg et al., 2020) also make use of parsing results. However, they only use dependency relations to link words or to propagate word representations. For example, given a dependency great nsubj −→pizza in DS, where great is a known pivot and pizza is an aspect term, the goal is to extract keyboard as an aspect from the target review “The keyboard is great” in DT . The typical syntax based method Hier-Joint (Ding et al., 2017) first locates the pivot great, then utilizes the nsubj dependency to identify the term keyboard. Other methods like RNSCN (Wang and Pan, 2018) combine the embedding of the child node (pizza) with that of the parent node (great) according to the relation type, or reversely (depending on the specific design). It can be seen that the dependency relation nsubj here is only used as a link to the pivot. 2QHKRW3269HFWRU 0XOWLKRW'HSHQGHQF\9HFWRU '7 11 5% 9%= - GHW DGYPRG QVXEM FRS URRW SXQFW 7KHSL]]DKHUHLVDOVRDEVROXWHO\GHOLFLRXV '7 11 5% 9%= 5% 5% - ĚĞƚ ĂĚǀŵŽĚ ŶƐƵďũ ƉƵŶĐƚ ĂĚǀŵŽĚ ĂĚǀŵŽĚ ĐŽƉ 6\QWDFWLF%ULGJH 6 * 6 * * * 6 * 6 * 6 6 6 6 * * * * * * * * 6 * Figure 3: Construction of the syntactic bridge. If a POS tag or dependency relation is involved, its corresponding entry in the vector is set to 1, and otherwise 0. We start in the opposite direction, i.e., we aim to fully exploit syntactic roles by recognizing themselves as pivots instead of treating them as links to pivots. To achieve this, we present a novel data structure to encode the POS and dependency information by grounding them into involved words. As shown in the lower part of Figure 3, for a word xi, we use a one-hot vector bpos ∈RNpos and a multi-hot vector bdep ∈RNdep to represent its POS tag and dependency relation(s), where Npos and Ndep are the number of tag/relation types. For 320 bdep, we merge all relations involved with xi regardless of the direction (i.e., being the governor or dependent)3. To enlarge the learning capability, we project bpos and bdep to the same dimensionality with learnable weight matrices4 and concatenate them to form the syntactic bridge bsyn: bsyn = (Wpos × bpos) ⊕(Wdep × bdep), (1) where bsyn ∈Rde has the same dimensionality with the word embedding e. In training, Wpos and Wdep get trained by labeled samples. In testing, we fix them and obtain bsyn for DT . By doing this, our proposed method well preserves two types of syntactic information throughout the entire learning process. As a result, we can take full advantage of their transferable information. Semantic Bridge The semantic bridge takes the syntactic roles above as a basis but moves one step further to retrieve transferable prototypes. Unlike previous passive methods that construct information flows like pizza→good→keyboard via opinion terms or pizza→offer→keyboard via context terms, we aim to construct a direct flow like pizza→keyboard. For example, to transfer knowledge from pizza in DS to keyboard in DT , we aim to introduce some supplementary target terms like {disk, OS, mouse} in DU for pizza and directly improve its semantic relatedness with keyboard. We call these supplementary terms prototypes and will retrieve them to build the semantic bridges5. PLMs like BERT can find a set of semantically similar terms like {hamburger, salad} for pizza, which can also serve as prototypes. However, such prototypes are not suitable for the domain adaptation task, because aspect terms in one domain are often far away from those in another domain in the semantic space. To address this problem, we design a syntax-enhanced similarity metric to retrieve transferable semantic prototypes. Before starting, we filter the words in DU by frequency and only preserve those appearing more than τ times. We regard these words in unlabeled target data as candidate prototypes and build a prototype bank eV from DU accordingly. We then conduct retrieval following the procedure in Figure 4. For a query word v ∈V S (vocabulary of DS), 3This simplification almost has no side effects. If a word has a NN tag and det relation, it must be the governer. 4In all equations, W denotes a trainable weight matrix. 5We retrieve prototypes for all words in the review due to the existence of domain-specific context terms like eat. UHWULHYH WRS. DJJUHJDWH 3URWRW\SH %DQN 4XHU\ :RUG 5HWULHYHG 3URWRW\SHV 6HPDQWLF %ULGJH Figure 4: Construction of the semantic bridge. For a query word, the top-K prototypes are retrieved from the prototype bank and aggregated to its semantic bridge. we want to find a prototype term ev ∈eV that play a similar syntactic role in the target domain. Specifically, we first summarize the global usages of v by merging its POS and dependency embeddings in all reviews where v appear in DS: bg pos = {bpos,j=1 | bpos,j=2 |...| bpos,j=NS}, bg dep = {bdep,j=1 | bdep,j=2 |...| bdep,j=NS}, (2) where | is the dimension-wise OR operation and NS is the number of reviews in DS. Similarly, we can obtain ebg pos and ebg dep for ev. We then define the syntax-enhanced similarity between v and ev: s.sim(v, ev) = c(bg pos, ebg pos)×c(bg dep, ebg dep)×c(e, ee), (3) where e and ee are word embeddings and c(·, ·) is the cosine similarity. Here the POS and dependency similarities are used to find similar syntactic roles, while the word similarity is used to reduce the noise of prototypes6. Consequently, we can obtain a s.sim score matrix MS∈R|V S|×|eV |. After ranking, for v, we select the top-K words { evk}K k=1 with their s.sim scores { esk}K k=1 from the prototype bank. Lastly, we aggregate these prototypes into the semantic bridge bsem of v: bsem = K X k=1 esk · eek. (4) Following the way for DS, we also retrieve transferable prototypes for DU and DT using eV . In this way, source and target words with the same prototypes can be directly correlated to each other. For DU, we can generate a score matrix MU ∈R|V U|×|eV | by calculating the s.sim for all words in DU and all candidate prototypes in eV . Then we can obtain the semantic bridge bsem for each word in DU in training. In testing, DT is unseen and the global bg pos/bg dep are not available. Therefore, for a word w in DT , we obtain bsem using MU if w has appeared in DU. Otherwise, we temporarily use the local bpos/bdep of w in current tesing sample to replace the global bg pos/bg dep and calculate the s.sim. 6A domain-invariant word that appears frequently in both domains should preserve its own information. It will have a maximum similarity score with itself since c(e, ee) = 1. 321 3.3 Bridge-based Sequence Tagging Based on the syntactic and semantic bridges, we now propose a lightweight end-to-end sequence tagger for aspect term extraction. As shown in Figure 5, the tagger receives a mixture of DS and DU for training and then makes predictions for DT in testing. We then illustrate the details. %ULGJH )XVHU )HDWXUH ([WUDFWRU 7RNHQ &ODVVLILHU %,2 /RVV 6RXUFH 5HYLHZ 6RXUFH (PEHGGLQJ 6RXUFH )HDWXUH 7DUJHW 5HYLHZ 7DUJHW (PEHGGLQJ 7DUJHW )HDWXUH '20 /RVV 6\QWDFWLF %ULGJH 6HPDQWLF %ULGJH 'RPDLQ &ODVVLILHU *5/ Figure 5: Training of bridge-based sequence tagging. Bridge Fuser Our constructed bridges have two properties. (1) Bridges are domain-invariant and should be preserved. (2) Bridges can help extract domain-invariant information from ei. Therefore, we propose to enhance the embedding ei of a word xi with its transferable bridges bsyn,i and bsem,i. Specifically, we use a gating operation to fuse bridges. Take the syntactic bridge as an example, we first calculate a dimension-wise gate gsyn,i: gsyn,i = σ (Wsyn(ei ⊕bsyn,i)), (5) where Wsyn ∈R2de×2de, σ is the Sigmoid function, ⊕is concatenation. We then scale the concatenated vector ei ⊕bsyn,i with gsyn,i and obtain the syntactic bridge enhanced embedding esyn,i: esyn,i = gsyn,i ⊙(ei ⊕bsyn,i), (6) where ⊙is an element-wise multiplication. The semantic bridge enhanced embedding esem,i can be calculated similarly. We term the model with ei, esyn,i, and esem,i input as BaseTagger, SynBridge, and SemBridge, respectively. Three types of embeddings are collectively called einput,i . Feature Extractor Previous studies (Xu et al., 2018) show that low-level token features are insufficient for tagging terms. Therefore, we use a CNN encoder containing L stacked convolutional layers with ReLU activation to extract the high-level features fi ∈Rdf : f l+1 i = ReLU(f l i−c:i+c ∗Kl + bl), f 0 i = einput,i, (7) where K ∈Rdf×(dinput×ks) is the kernel group, ks = 2c + 1 is the kernel size. Token Classifier For recognizing aspect and opinion terms, we send f L i in the last layer to a token classifier: ˆyi = Softmax(WA × f L i ), (8) where ˆyi is the prediction of the word xi. Domain Classifier Besides BIO tagging, we further enhance the domain-invariance of bridgebased features via domain adversarial training. Specifically, we first aggregate f L i to a global representation fg: fg = MaxPool(f L 1:n). (9) Then we add a Gradient Reversal Layer (GRL) (Ganin and Lempitsky, 2015) to fg with the scale coefficient λ and train a domain classifier to distinguish the domain that fg belongs to: ˆyd = Softmax(WO × MLP(GRLλ(fg))), (10) where ˆyd is the domain prediction, and MLP contains LD layers with ReLU activation. Training Procedure In training, only samples from DS have corresponding BIO labels yS for token classification. The goal is to minimize the tagging loss for recognizing aspect terms: LBIO = − X DS n X i=1 ℓ(ˆyi, yi), (11) where ℓis the cross-entropy loss function. On the other hand, the samples from DS and DU are used to train the domain classifier and minimize the following domain classification loss: LDOM = − X DS∪DU ℓ(ˆyd, yd), (12) where yd = 0 for DS and yd = 1 for DU. The final loss for training the end-to-end tagger is defined as L = LBIO + LDOM. Notice that DT is only used in testing. There is no data leakage in training, and the task setting is strictly inductive. 4 Experiment 4.1 Experimental Setup Datasets We use three conventional English datasets from different domains and construct six directed transfer pairs, where R and L are from SemEval 2014 and 2015 (Pontiki et al., 2014, 2015), and D is collected by Hu and Liu (2004). Following previous studies (Wang and Pan, 2018, 2019b; Pereg et al., 2020), we use three different splits and each split has a fixed train-test ratio 3:1. The detailed statistics of datasets are presented in Table 17. Table 1: The statistics of datasets. Dataset Domain Total Train Test R Restaurant 5841 4381 1460 L Laptop 3845 2884 961 D Device 3836 2877 959 7Our code and data are available at https://github.com/ NLPWM-WHU/BRIDGE. 322 Table 2: Comparison of different methods. Baselines with △use annotated opinion terms. The best scores are in bold and the second best ones are underlined. Averaged results with † and ‡ are significantly better than BERTCross and BaseTagger (p < 0.05) based on one-tailed unpaired t-test, respectively. The upper bounds of three datasets (achieved by BaseTagger trained on in-domain labeled data) are 76.43 (R), 75.60 (L), and 57.10 (D). Type Model Embedding R→L L→R R→D D→R L→D D→L AVG. I TCRF Manual 19.72 28.19 21.07 6.59 29.96 24.22 21.63 RAP Manual 25.92 46.90 22.63 45.44 34.54 28.22 33.94 SAL Word2vec 29.03 44.57 22.82 38.89 38.82 47.25 36.90 Hier-Joint Word2vec 33.66 48.10 33.20 47.97 31.25 34.74 38.15 RNSCN△ Word2vec 40.43 52.91 35.10 48.36 40.42 51.14 44.73 TRNN△ Word2vec 40.15 53.78 37.33 51.17 41.19 51.66 45.88 TIMN△ Word2vec 43.68 54.12 35.45 53.82 38.63 52.46 46.36 II BERT-Base BERT 33.89 42.74 35.30 36.86 43.54 46.06 39.73 UDA BERT 44.24 50.52 40.04 53.39 41.48 52.33 47.00 SA-EXAL△ BERT 47.59 54.67 40.50 54.54 42.19 47.72 47.87 BERT-Cross BERT 46.30 51.60 43.68 53.15 44.22 50.04 48.17 III BaseTagger Word2vec 48.86 61.42 40.56 57.67 43.75 51.95 50.70† SynBridge Word2vec 51.53 63.90 42.76 59.40 44.97 52.44 52.50†‡ SemBridge Word2vec 51.53 65.96 43.03 60.61 45.37 53.77 53.38†‡ Settings We pre-process each dataset by lowercasing all words. We use the same word2vec vectors as previous studies (Wang and Pan, 2018, 2019a,b) to generate word embeddings, and set the dimensionality de=100. In the syntactic bridge, we use Stanford CoreNLP (Manning et al., 2014) for dependency parsing. There are 45 classes of POS tags and 40 classes of dependency relations in three datasets. In the semantic bridge, we set the frequency threshold τ=5, the number of prototypes K=10. In the end-to-end tagger, we set the number of convolution layers L=4, and the kernel size ks of each layer is 3, 5, 5, 5, respectively, the number of MLP layers LD=3, and dropout (Srivastava et al., 2014) is applied to layers’ outputs with the probability 0.5. The dimensionality of features df=256, the scale coefficient of GRL λ=0.1. We train the tagger for 100 epochs using Adam optimizer (Kingma and Ba, 2015) with the learning rate 1e-4 and batch size 8 in a 1080Ti GPU. Evaluation For each transfer pair, we use the labeled training data from the source domain and unlabeled training data from the target domain to train the tagger. Then we evaluate the tagger on unseen test data from the target domain. We use the mean F1-scores of aspect terms over three splits with three random seeds (i.e., nine runs for each transfer pair) for evaluation8. 4.2 Compared Methods We classify all models into three categories. Type-I denotes the opinion term-based methods. TCRF (Jakob and Gurevych, 2010), RAP (Li et al., 2012), and Hier-Joint (Ding et al., 2017) use manually defined dependency rules. RNSCN and 8The hyperparameter ranges are presented in Appendix A. TRNN (Wang and Pan, 2018, 2019a) model dependency trees with trainable recursive networks. SAL (Li et al., 2019) and TIMN (Wang and Pan, 2019b) replace the dependency tree with trainable memory interaction. Type-II denotes context term-based methods. BERT-Base uses vanilla base BERT (Devlin et al., 2019) for ATE. BERT-Cross (Xu et al., 2019) posttrains BERT on a combination of Yelp and Amazon corpus. UDA (Gong et al., 2020) and SA-EXAL (Pereg et al., 2020) incorporate syntactic information into BERT with auxiliary tasks and modified attention mechanisms9. Type-III denotes the proposed active domain adaptation strategy. BaseTagger is the tagger without bridges, while SynBridge and SemBridge use syntactic and semantic bridges, respectively. 4.3 Main Results The comparison results for all methods are shown in Table 2. It is clear that our proposed model achieves a new state-of-the-art performance in terms of the average F1-scores. For example, SemBridge outperforms the best TIMN in Type-I by 7.02% and BERT-Cross in Type-II by 5.21%, respectively. We also notice that our BaseTagger already outperforms all baselines. We attribute this to the design of CNN feature extractor and domain adversarial training (DAT). CNN focuses on the Ngram feature rather than a single word and reduces the side effects of non-pivot aspect terms. DAT is applied to the sentence-level features, such that they are not misled by the common N-grams that are labeled both 0 and 1. 9Since SAL and UDA use extra aspect sentiment labels, we show how to make them fair competitors in Appendix B. 323 SynBridge and SemBridge further improve BaseTagger with a 1.80% and 2.68% absolute gain, respectively. This proves the effectiveness of our proposed active domain adaptation strategy. Meanwhile, SemBridge is a bit superior to SynBridge. The reasons are two-fold. (1) The semantic bridges come from prototype words that possess prior embedding knowledge and also contain syntactic information, while the syntactic bridges are merely trained from scratch. (2) The retrieved top-K terms make the supplementary information in SemBridge more diverse and abundant than that in SynBridge. Among the baselines, early methods using common opinion seeds and pre-defined rules are inferior. Relying on annotated opinion terms, the methods like TIMN get some improvements but induce extra annotation costs. By incorporating pre-trained BERT with external dependency and cross-domain corpus, UDA, SA-EXAL, and BERTCross outperform previous methods, but they need high computational resources. In contrast, by using the static Word2vec embeddings, our model can outperform those with dynamic BERT representations. This is instructive for other researches in that there is still room for improvement by exploring the syntactic and semantic features beyond the popular BERT-based models10. 5 Analysis 5.1 What If There Is an OTE Task? With the proposed active domain adaptation strategy, we do not need any manually labeled opinion terms for ATE. However, this does not mean that our method cannot handle opinion term extraction (i.e., OTE). In contrast, if the labeled opinion terms are provided in DS, we can also conduct the OTE task for DT by simply modifying the tagger. In specific, we add an opinion term prediction layer in Eq.8 and then extract aspect and opinion terms simultaneously. The results are shown in Table 3. Obviously, our method again outperforms all baselines11. We find a small performance decrease in AVG-AS compared with that in Table 2. Similar results are also observed in BERT-Base. The reason is that the objective of ATE and OTE may interfere with each other without proper balancing and a sophisticated multi-task learning framework. 10We also make some explorations about combining SynBridge and SemBridge, please refer to Appendix C. 11Please refer to Appendix D for detailed results for all transfer pairs. Table 3: Comparison of different methods. AVG-AS and AVG-OP are F1-scores for ATE and OTE averaged on all transfer pairs. Model AVG-AS AVG-OP RNSCN 44.73 67.44 TRNN-GRU 45.88 67.12 TIMN 46.36 68.21 BERT-Base 39.52 66.22 SA-EXAL 47.87 69.15 BERT-Cross 48.35 69.47 BaseTagger 50.12 71.73 SynBridge 51.86 71.73 SemBridge 52.53 72.08 5.2 Ablation Study We conduct a series of ablation study to validate the effectiveness of our method. The results are shown in Table 4. Table 4: Ablation study. The scores denote the decrease of performance after removing(−) or replacing(→) a specific component. Index Model Variant AVG. 1 BaseTagger −LDOM 1.94 2 CNN→BiLSTM 8.47 3 SynBridge −bpos 1.68 4 −bdep 1.49 5 bdep→Tree-LSTM 3.97 6 bdep→GCN 4.21 7 SemBridge −c(e, ee) 1.82 8 −c(bpos, ebpos) 2.30 9 −c(bdep, ebdep) 2.52 Results 1∼2 conform to our previous discussion about BaseTagger that both CNN and domain adversarial training contribute to overall good performance. Results 3∼6 show the effectiveness of POS and dependency embeddings in SynBridge. Specifically, in 5∼6, we replace our proposed structure for dependency with frequently-used Tree-LSTM and GCN to model the dependency tree and find a significant drop in performance. Results 7∼9 show the importance of all three types of similarity for retrieving prototypes in SemBridge. 5.3 Parameter Study There are three key hyperparameters in our method: the scale coefficient of GRL λ, the frequency threshold τ, and the number of prototypes K. We vary λ in the range 10−4 ∼1.0 and τ/K in 1 ∼10 to investigate their impacts and present the results in Figure 6. In Figure 6(a), when increasing λ from 10−4 to 10−1, we enlarge the scale of domain adversarial training in GRL and get small improvements. However, the performance does not keep rising when 324 Table 5: Case study. The left columns present the selected target testing examples, and the words in red are aspect terms. The right columns denote the extraction results of corresponding models. Pair Example RNSCN BERT-Cross SynBridge SemBridge R→LS1.it has usb ports, 1 sd memory card reader and an sd memory car expansion. None card reader, sd memory car expansion usb ports, sd memory card reader, sd memory car expansion usb ports, sd memory card reader, sd memory car expansion L→RS2.The asparagus, truffle oil, parmesan bruschetta is a winner! None asparagus, bruschetta asparagus,truffle oil parmesan bruschetta asparagus,truffle oil parmesan bruschetta L→RS3.They showed up 15 minutes after the tuna melt.tuna melt None tuna melt tuna ) OJȜ 6\Q%ULGJH 6HP%ULGJH (a) Impact of λ. ) IJ. IJ . (b) Impact of τ/K. Figure 6: Impacts of hyperparameters λ, τ, and K. λ = 1.0. This result shows that simply forcing non-pivots to transfer knowledge is not suitable for domain adaptation. In Figure 6(b), τ is used to balance diversity and accuracy. A low τ means that prototypes are diverse, but some of them are long-tail words and contribute little to the reduction of domain discrepancy. On the contrary, a high τ only preserves frequent prototypes, and some meaningful prototypes are filtered out. Therefore, a middle τ=5 is an appropriate choice. For K, the curve is generally upward when more prototypes are introduced. This trend is reasonable since more prototypes equal to more target information. ) 38 6\Q%ULGJH 6HP%ULGJH (a) Impact of PU. ) 31 6\Q%ULGJH 6HP%ULGJH (b) Impact of PN. Figure 7: Impacts of PU and PN. In Figure 7, we further analyze the impacts of the percentage of unlabeled data PU and the percentage of parsing noise PN. For PU, the performance is generally better when more unlabeled target data is introduced. Moreover, around 20%∼40% unlabeled data is enough to achieve satisfactory performance. Notice that SemBridge without unlabeled data will degenerate into BaseTagger since no prototypes can be retrieved. For PN, we manually disturb the parsing results to observe the robustness of our method. Clearly, after introducing noises on parsing, the performance begins to degrade, but not by a large margin. Our method has the ability to resist parsing errors for two reasons. First, beyond syntactic roles, we also incorporate embedding similarity when retrieving prototypes (for SemBridge only). Second, the gating mechanism can further filter useless syntactic information and maintain the quality of word representations. 5.4 Case Study To have a close look, we select a few samples from testing target data for a case study. S1 and S2 show the positive impacts of bridges. Due to the space limit, we illustrate S1 in detail. Since most words in S1 are domain-specific terms in L, RNSCN fails to recognize any aspect terms by simply propagating word representations with dependency. BERTCross only extracts a part of aspect terms based on its prior knowledge. For our bridge-based method, SynBridge supplements syntactic roles {nummod, compound, obj, conj, NNS} for port. These syntactic roles also join the representation of usb and help to extract usb ports correctly. For SemBridge, the analysis is much straightforward. usb is the prototype of typical aspect terms in R like {garlic, thai, banana}, thus the tagger with semantic bridges can easily recognize usb as an aspect term. S3 further illustrates how SemBridge helps recover from the wrong parsing results. Such results make two syntax based methods RNSCN and SynBridge stop working. In contrast, tuna is the prototype of noun words like {nvidia, amd, blade} in L and melt has the verb prototype like {imagine, hang, relax} in R, thus SemBridge correctly extracts tuna and filters out melt in the same time. In Table 6, We further present several sample prototypes of the training data from the transfer pairs R→L (upper three) and L→R (lower three) in SemBridge, where three terms on the left are aspect term, opinion term, and context term, respectively. For a source non-pivot term like processor in L, SemBridge enhances it with typical target words like soup and burger. As a result, the domain discrepancy between the source and target data is largely reduced with the help of prototypes. 325 Table 6: Top-10 prototypes in SemBridge. Words are ranked by their s.sim scores. Term Prototypes food machine,product,keyboard,netbook,service, computer,screen,value,touchpad,processor delicious amazing,wonderful,awesome,great,good,nice, fantastic,beautiful,perfect,lightweight cook use,load,plug,work,turn,break,charge,change, help,run processorsoup,burger,meal,sauce,flavor,cheese,food, salad,seafood,fan efficient attentive,impressive,affordable,friendly,reasonable,pleasant,simple,courteous,helpful,hungry freeze eat,hang,stop,die,bring,stay,leave,start,give,keep 5.5 Analysis on Computational Cost In practice, for any transfer pairs, the one-time construction of syntactic and semantic bridges can finish within 30 seconds. Therefore, we focus on the end-to-end training costs of SynBridge/SemBridge. We run five top-performing methods on the transfer pair R→L and present the trainable parameter number and running time per epoch of each method in Table 7. We can conclude that our proposed method maintains a quite low computational cost. Table 7: Computational cost of each method. Parameter Runtime TIMN 0.8M 132s BERT-Cross 109M 84s BaseTagger 1.3M 11s SynBridge/SemBridge 1.4M 12s 6 Conclusion In this paper, we propose a novel active domain adaptation method for aspect term extraction. Unlike previous studies that conduct passive domain adaptation by associating aspect terms with pivots, we actively enhance the terms’ transferability by constructing syntactic and semantic bridges for them. We then design a lightweight end-toend tagger for bridge-based sequence tagging. Experiments on six transfer pairs demonstrate that our method achieves a new state-of-the-art performance with a quite low computational cost. Acknowledgments We thank the anonymous reviewers for their valuable comments. The work described in this paper is supported by the NSFC projects (61572376, 91646206), and the 111 project (B07037). References John Blitzer, Ryan T. McDonald, and Fernando Pereira. 2006. Domain adaptation with structural correspondence learning. In EMNLP, pages 120–128. Zhiyuan Chen, Arjun Mukherjee, and Bing Liu. 2014. Aspect extraction with automated prior knowledge learning. In ACL, pages 347–358. Zhuang Chen and Tieyun Qian. 2019. Transfer capsule network for aspect level sentiment classification. In ACL, pages 547–556. Zhuang Chen and Tieyun Qian. 2020a. Enhancing aspect term extraction with soft prototypes. In EMNLP, pages 2107–2117. Association for Computational Linguistics. Zhuang Chen and Tieyun Qian. 2020b. Relation-aware collaborative learning for unified aspect-based sentiment analysis. In ACL, pages 3685–3694. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL, pages 4171–4186. Ying Ding, Jianfei Yu, and Jing Jiang. 2017. Recurrent neural networks with auxiliary labels for crossdomain opinion target extraction. In AAAI, pages 3436–3442. Yaroslav Ganin and Victor S. Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180–1189. Chenggong Gong, Jianfei Yu, and Rui Xia. 2020. Unified feature and instance based domain adaptation for aspect-based sentiment analysis. In EMNLP, pages 7035–7045. Han Guo, Ramakanth Pasunuru, and Mohit Bansal. 2020. Multi-source domain adaptation for text classification via distancenet-bandits. In AAAI, pages 7830–7838. Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Dahlmeier. 2019. An interactive multi-task learning network for end-to-end aspect-based sentiment analysis. In ACL, pages 504–515. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In SIGKDD, pages 168– 177. Niklas Jakob and Iryna Gurevych. 2010. Extracting opinion targets in a single and cross-domain setting with conditional random fields. In EMNLP, pages 1035–1045. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR. Fangtao Li, Chao Han, Minlie Huang, Xiaoyan Zhu, Yingju Xia, Shu Zhang, and Hao Yu. 2010. Structure-aware review mining and summarization. In COLING, pages 653–661. 326 Fangtao Li, Sinno Jialin Pan, Ou Jin, Qiang Yang, and Xiaoyan Zhu. 2012. Cross-domain co-extraction of sentiment and topic lexicons. In ACL, pages 410– 419. Zheng Li, Xin Li, Ying Wei, Lidong Bing, Yu Zhang, and Qiang Yang. 2019. Transferable end-to-end aspect-based sentiment analysis with selective adversarial learning. In EMNLP-IJCNLP, pages 4589– 4599. Kang Liu, Heng Li Xu, Yang Liu, and Jun Zhao. 2013. Opinion target extraction using partially-supervised word alignment model. In IJCAI, pages 2134–2140. Kang Liu, Liheng Xu, and Jun Zhao. 2012. Opinion target extraction using word-based translation model. In EMNLP, pages 1346–1356. Pengfei Liu, Shafiq R. Joty, and Helen M. Meng. 2015. Fine-grained opinion mining with recurrent neural networks and word embeddings. In EMNLP, pages 1433–1443. Dehong Ma, Sujian Li, Fangzhao Wu, Xing Xie, and Houfeng Wang. 2019. Exploring sequence-tosequence learning in aspect term extraction. In ACL, pages 3538–3547. Christopher D. Manning, Mihai Surdeanu, John Bauer, Jenny Rose Finkel, Steven Bethard, and David McClosky. 2014. The stanford corenlp natural language processing toolkit. In ACL, pages 55–60. Oren Pereg, Daniel Korat, and Moshe Wasserblat. 2020. Syntactically aware cross-domain aspect and opinion terms extraction. In COLING, pages 1772– 1777. Maria Pontiki, Dimitris Galanis, Haris Papageorgiou, Suresh Manandhar, and Ion Androutsopoulos. 2015. Semeval-2015 task 12: Aspect based sentiment analysis. In SemEval, pages 486–495. Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. Semeval-2014 task 4: Aspect based sentiment analysis. In SemEval, pages 27–35. Ana-Maria Popescu and Oren Etzioni. 2005. Extracting product features and opinions from reviews. In EMNLP, pages 339–346. Guang Qiu, Bing Liu, Jiajun Bu, and Chun Chen. 2011. Opinion word expansion and target extraction through double propagation. Computational Linguistics, 37(1):9–27. Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 15(1):1929– 1958. Feixiang Wang, Man Lan, and Wenting Wang. 2018. Towards a one-stop solution to both aspect extraction and sentiment analysis tasks with neural multitask learning. In IJCNN, pages 1–8. Wenya Wang and Sinno Jialin Pan. 2018. Recursive neural structural correspondence network for crossdomain aspect and opinion co-extraction. In ACL, pages 2171–2181. Wenya Wang and Sinno Jialin Pan. 2019a. Syntactically meaningful and transferable recursive neural networks for aspect and opinion extraction. CL, 45(4):705–736. Wenya Wang and Sinno Jialin Pan. 2019b. Transferable interactive memory network for domain adaptation in fine-grained opinion extraction. In AAAI, pages 7192–7199. Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, and Xiaokui Xiao. 2016. Recursive neural conditional random fields for aspect-based sentiment analysis. In EMNLP, pages 616–626. Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, and Xiaokui Xiao. 2017. Coupled multi-layer attentions for co-extraction of aspect and opinion terms. In AAAI, pages 3316–3322. Yuanbin Wu, Qi Zhang, Xuanjing Huang, and Lide Wu. 2009. Phrase dependency parsing for opinion mining. In EMNLP, pages 1533–1541. Hu Xu, Bing Liu, Lei Shu, and Philip S. Yu. 2018. Double embeddings and cnn-based sequence labeling for aspect extraction. In ACL, pages 592–598. Hu Xu, Bing Liu, Lei Shu, and Philip S. Yu. 2019. BERT post-training for review reading comprehension and aspect-based sentiment analysis. In NAACL-HLT, pages 2324–2335. A Ranges of Hyperparameters We present the hyperparameter ranges in Table 8. We select all hyperparameters via manual tuning. Table 8: Ranges of Hyperparameters. Hyperparameter Range Best frequency threshold τ 1,2,3,4,5,6,7,8,9,10 5 number of prototypes K 1,2,3,4,5,6,7,8,9,10 10 number of CNN layers L 1,2,3,4,5 4 dimension of CNN features df 64, 128, 256 256 kernel size ks of CNN layer 1 3,5,7,9 3 kernel size ks of CNN layer 2 3,5,7,9 5 kernel size ks of CNN layer 3 3,5,7,9 5 kernel size ks of CNN layer 4 3,5,7,9 5 number of MLP layers LD 1,2,3,4,5 3 the scale coefficient of GRL λ 10[−4,−3,−2,−1,0] 10−1 327 Table 9: Comparison of different methods when there is an OTE task. The best scores are in bold and the second best ones are underlined. AS and OP denote aspect and opinion F1-scores. Averaged results with * are significantly better than the best baseline BERT-Cross (p < 0.01) based on one-tailed unpaired t-test. Models R→L L→R R→D D→R L→D D→L AVG. AS OP AS OP AS OP AS OP AS OP AS OP AS OP RNSCN 40.43 65.85 52.91 72.51 35.10 60.17 48.36 73.75 40.42 61.15 51.14 71.18 44.73 67.44 TRNN 40.15 65.63 53.78 73.40 37.33 60.32 51.17 74.37 41.19 60.20 51.66 68.79 45.88 67.12 TIMN 43.68 68.44 54.12 73.69 35.45 59.05 53.82 76.52 38.63 62.22 52.46 69.32 46.36 68.21 BERT-Base 34.70 73.84 37.07 80.12 37.17 64.52 40.54 60.45 43.45 59.59 44.19 58.77 39.52 66.22 SA-EXAL 47.59 75.79 54.67 80.05 40.50 63.33 54.54 71.57 42.19 60.19 47.72 63.98 47.87 69.15 BERT-Cross 44.00 75.38 54.31 81.97 43.12 66.57 51.97 70.58 44.35 58.49 50.01 63.81 48.35 69.47 BaseTagger 47.78 70.61 58.39 79.53 39.71 63.63 57.56 80.18 44.49 64.14 52.77 72.30 50.12 71.73∗ SynBridge 50.59 70.74 60.94 79.86 42.42 63.37 59.92 79.88 45.30 64.22 51.97 72.33 51.86∗ 71.73∗ SemBridge 50.67 71.51 63.04 80.48 43.34 63.46 60.19 80.21 44.91 64.15 53.02 72.63 52.53∗ 72.08∗ B Modification of SAL and UDA Since SAL and UDA are designed for end-to-end cross-domain aspect-based sentiment analysis, they have access to the aspect sentiment labels in training. As previous studies show, aspect term extraction and aspect-level sentiment classification can benefit each other. Therefore, it is unfair to directly compare our method with SAL and UDA. We choose to modify SAL and UDA and make them fair competitors. We degrade the collapsed tags {B-POS, I-POS, B-NEG, I-NEG, B-NEU, INEU, O} to {B, I, O} thus remove the aspectlevel sentiment classification task. Following other BERT-based methods, we use BERT-Base as the backbone of UDA. C Can We Combine SynBridge and SemBridge? Since SynBridge and SemBridge contain transferable syntactic and semantic information, it is intuitive to combine them for a better performance than either individual model. Here we apply a very simple operation for combination. For a word xi with embedding ei, we first obtain its syntactic and semantic bridges bsyn,i and bsem,i, and merge them into a combined bridge: bcom,i = (Wsyn × bsyn,i) + (Wsem × bsem,i), (13) Then we conduct a similar gating operation and get the combined bridge enhanced embedding ecom,i: gcom,i = σ (Wcom(ei ⊕bcom,i)) ecom,i = gcom,i ⊙(ei ⊕bcom,i), (14) Lastly, we regard ecom,i as the input of tagger and make predictions for aspect terms. We term this model ComBridge and present the results in Table 10. Table 10: Comparison of different bridge-based methods. The best scores are in bold and the second best ones are underlined. Model R→L L→R R→D D→R L→D D→L AVG. BaseTagger 48.86 61.42 40.56 57.67 43.75 51.95 50.70 SynBridge 51.53 63.90 42.76 59.40 44.97 52.44 52.50 SemBridge 51.53 65.96 43.03 60.61 45.39 53.77 53.38 ComBridge 53.32 66.20 42.56 60.99 44.74 53.32 53.52 ComBridge slightly outperforms SemBridge and achieves the optimal results in all bridge-based methods. The small improvement is explicable since SemBridge already contains most of the syntactic information in SynBridge and we do not use any sophisticated methods in combination. D Detailed Results for an Additional OTE Task When opinion terms are labeled, our method can also conduct aspect term extraction and opinion term extraction simultaneously. For recognizing aspect and opinion terms, we only need to add an opinion term prediction layer: ˆya,i = Softmax(WA × f L i ), ˆyo,i = Softmax(WO × f L i ), (15) where ˆya,i / ˆyo,i are the predictions of {B, I, O} for the aspect / opinion terms. And the resulted BIO loss is calculated as follow: LBIO = − X DS n X i=1 ℓ(ˆya,i, ya,i) + ℓ(ˆyo,i, yo,i) (16) where ℓis the cross-entropy loss function. We present the detailed results in Table 9. Obviously, our proposed SynBridge and SemBridge outperform other baselines in both aspect and opinion F1-scores.
|
2021
|
27
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3484–3494 August 1–6, 2021. ©2021 Association for Computational Linguistics 3484 Novel Slot Detection: A Benchmark for Discovering Unknown Slot Types in the Task-Oriented Dialogue System Yanan Wu1∗, Zhiyuan Zeng1∗, Keqing He2∗, Hong Xu1 Yuanmeng Yan1, Huixing Jiang2, Weiran Xu1∗ 1Pattern Recognition & Intelligent System Laboratory 1Beijing University of Posts and Telecommunications, Beijing, China 2Meituan Group, Beijing, China {yanan.wu,zengzhiyuan,xuhong,yanyuanmeng,xuweiran}@bupt.edu.cn {hekeqing,jianghuixing}@meituan.com Abstract Existing slot filling models can only recognize pre-defined in-domain slot types from a limited slot set. In the practical application, a reliable dialogue system should know what it does not know. In this paper, we introduce a new task, Novel Slot Detection (NSD), in the task-oriented dialogue system. NSD aims to discover unknown or out-of-domain slot types to strengthen the capability of a dialogue system based on in-domain training data. Besides, we construct two public NSD datasets, propose several strong NSD baselines, and establish a benchmark for future work. Finally, we conduct exhaustive experiments and qualitative analysis to comprehend key challenges and provide new guidance for future directions1. 1 Introduction Slot filling plays a vital role to understand user queries in personal assistants such as Amazon Alexa, Apple Siri, Google Assistant, etc. It aims at identifying a sequence of tokens and extracting semantic constituents from the user queries. Given a large scale pre-collected training corpus, existing neural-based models (Mesnil et al., 2015; Liu and Lane, 2015, 2016; Goo et al., 2018; Haihong et al., 2019; Chen et al., 2019; He et al., 2020b,d; Yan et al., 2020; Louvan and Magnini, 2020; He et al., 2020a) have been actively applied to slot filling and achieved promising results. Existing slot filling models can only recognize pre-defined entity types from a limited slot set, which is insufficient in the practical application scenario. A reliable slot filling model should not only predict the pre-defined slots but also detect potential unknown slot types to know what it doesn’t ∗The first three authors contribute equally. Weiran Xu is the corresponding author. 1https://github.com/ChestnutWYN/ACL20 21-Novel-Slot-Detection Agent: What can I do for you? User: Play is this my world by leo arnaud. Play is this my world by leo arnaud. playlist artist Dialogue System DM & NLG Module is this my world is an unknown slot type (denoted as NS). It’s the name of leo arnaud’s album. Dialogue System without Novel Slot Detector Agent: What can I do for you? User: Play is this my world by leo arnaud. Dialogue System DM & NLG Module Dialogue System with Novel Slot Detector NLU Module Play is this my world by leo arnaud. NS artist Novel Slot Detector Human Annotated Collected Novel Slots Update Dialogue System Agent: “is this my world” is probably a novel slot. The current system can not handle it. Agent: You don’t have a playlist called “is this my world” × √ Figure 1: An example of Novel Slot Detection in the task-oriented dialogue system. Without NSD, the dialogue system gives the wrong response since it misunderstands the unknown slot “is this my world” as the indomain playlist type. In contrast, NSD recognizes “is this my world” as NS and the system gives a fallback response. Meanwhile, with human-in-the-loop annotation, the system can increase its functions or skills. know, which we call Novel Slot Detection (NSD) in this paper. NSD is particularly crucial in deployed systems—both to avoid performing the wrong action and to discover potential new entity types for future development and improvement. We display an example as Fig 1 shows. In this paper, we define Novel Slot (NS) as new slot types that are not included in the pre-defined slot set. NSD aims to discover potential new or out-of-domain entity types to strengthen the capability of a dialogue system based on in-domain precollected training data. There are two aspects in the previous work related to NSD, out-of-vocabulary (OOV) recognition (Liang et al., 2017a; Zhao and Feng, 2018; Hu et al., 2019; He et al., 2020c,d; Yan et al., 2020; He et al., 2020e) and out-of-domain (OOD) intent detection (Lin and Xu, 2019; Larson et al., 2019; Xu et al., 2020a; Zeng et al., 2021b,a). OOV means many slot types can have a large number of new slot values while the training set only obtains a tiny part of slot values. OOV aims to recognize unseen slot values in training set 3485 Utterance play is this my world by leo arnaud Slot Filling Labels O B-album I-album I-album I-album O B-artist I-artist Novel Slot Detection Labels O NS NS NS NS O B-artist I-artist Table 1: Comparison between slot filling and novel slot detection. In the novel slot detection labels, we consider “album” as an unknown slot type that is out of the scope of the pre-defined slot set. Meanwhile, “artist” belonging to in-domain slot types still needs to be recognized as the original slot filling task. for pre-defined slot types, using character embedding (Liang et al., 2017a), copy mechanism (Zhao and Feng, 2018), few/zero-shot learning (Hu et al., 2019; He et al., 2020e; Shah et al., 2019), transfer learning (Chen and Moschitti, 2019; He et al., 2020c,b) and background knowledge (Yang and Mitchell, 2017; He et al., 2020d), etc. Compared to OOV recognition, our proposed novel slot detection task focuses on detecting unknown slot types, not just unseen values. NSD faces the challenges of both OOV and no sufficient context semantics (see analysis in Section 6.2), greatly increasing the complexity of the task. Another line of related work is OOD intent detection (Hendrycks and Gimpel, 2017; Lee et al., 2018; Lin and Xu, 2019; Ren et al., 2019; Zheng et al., 2020; Xu et al., 2020a) which aims to know when a query falls outside the range of predefined supported intents. The main difference is that NSD detects unknown slot types in the token level while OOD intent detection identifies out-of-domain intent queries. NSD requires a deep understanding of the query context and is prone to label bias of O (see analysis in Section 5.3.1), making it challenging to identify unknown slot types in the task-oriented dialog system. In this paper, we first introduce a new and important task, Novel Slot Detection (NSD), in the task-oriented dialogue system (Section 2.2). NSD plays a vital role in avoiding performing the wrong action and discovering potential new entity types for the future development of dialogue systems. Then, we construct two public NSD datasets, SnipsNSD and ATIS-NSD, based on the original slot filling datasets, Snips (Coucke et al., 2018) and ATIS (Hemphill et al., 1990) (Section 2.2). From the perspective of practical application, we consider three kinds of dataset construction strategies, Replace, Mask and Remove. Replace denotes we label the novel slot values with all O in the training set. Mask is to label with all O and mask the novel slot values. Remove is the most strict strategy where all the queries containing novel slots are removed. We dive into the details of the three different construction strategies in Section 3.2 and perform a qualitative analysis in Section 5.3.1. Besides, we propose two kinds of evaluation metrics, span-level F1 and token-level F1 in Section 3.4, following the slot filling task. Span F1 considers the exact matching of a novel slot span while Token F1 focuses on prediction accuracy on each word of a novel slot span. We discuss performance comparison between the two metrics and propose a new metric, restriction-oriented span evaluation (ROSE), to combine the advantages of both in Section 5.3.3. Then, we establish a fair benchmark and propose extensive strong baselines for NSD in Section 4. Finally, we perform exhaustive experiments and qualitative analysis to shed light on the challenges that current approaches faced with NSD in Section 5.3 and 6. Our contributions are three-fold: (1) We introduce a Novel Slot Detection (NSD) task in the task-oriented dialogue system. NSD helps avoid performing the wrong action and discovering potential new entity types for increasing functions of dialogue systems. (2) We construct two public NSD datasets and establish a benchmark for future work. (3) We conduct exhaustive experiments and qualitative analysis to comprehend key challenges and provide new guidance for future NSD work. 2 Problem Formulation 2.1 Slot Filling Given a sentence X = {x1, ..., xn} with n tokens, the slot filling task is to predict a corresponding tag sequence Y = {y1, ..., yn} in BIO format, where each yi can take three types of values: B-slot type, I-slot type and O, where “B” and “I” stand for the beginning and intermediate word of a slot and “O” means the word does not belong to any slot. Here, slot filling assumes yi ∈y, where y denotes a pre-defined slot set of size M. Current approaches typically model slot filling as a sequence labeling problem using RNN (Liu and Lane, 2015, 2016; Goo et al., 2018) or pre-trained language models (Chen et al., 2019). 2.2 Novel Slot Detection We refer to the above training data D as in-domain (IND) data. Novel slot detection aims to identify 3486 Original Utterance play is this my world by leo arnaud Original Slot Filling Labels O B-album I-album I-album I-album O B-artist I-artist Strategy Replace play is this my world by leo arnaud O O O O O O B-artist I-artist Mask play MASK MASK MASK MASK by leo arnaud O O O O O O B-artist I-artist Remove Table 2: Comparison between three processing strategies in the training set. We consider “album” as an unknown slot type and “-” denotes the sentence is removed from the training data. unknown or out-of-domain (OOD) slot types via IND data while correctly labeling in-domain data. We denote unknown slot type as NS and in-domain slot types as IND in the following sections. Note that we don’t distinguish between B-NS and I-NS and unify them as NS because we empirically find existing models hardly discriminate B and I for an unknown slot type. We provide a detailed analysis in Section 5.3.3. We show an example of NSD in Table 1. The challenges of recognizing NSD come from two aspects, O tags and in-domain slots. On the one hand, models need to learn entity information for distinguishing NS from O tags. On the other hand, they require discriminating NS from other slot types in the pre-defined slot set. We provide a detailed error analysis in Section 6.1. 3 Dataset Since there are not existing NSD datasets, we construct two new datasets based on the two widely used slot filling datasets, Snips (Coucke et al., 2018) and ATIS (Hemphill et al., 1990). We first briefly introduce Snips and ATIS, then elaborate on data construction and processing in detail, and display the statistic of our NSD datasets, Snips-NSD and ATIS-NSD. Finally, we define two evaluation metrics for the NSD task, Span F1 and Token F1. 3.1 Original Slot Filling Datasets Snips2 is a custom intent engine dataset. It originally has 13,084 train utterances, 700 and 700 test utterances. ATIS3 contains audio recordings of people making flight reservations. It originally has 4,478 train utterances, 500 dev and 893 test utterances. The full statistic is shown in Table 3. Note that the vocabulary only contains words in the training set, and test set words that do not exist in the vocabulary are referred to OOV words. The percentage of OOV words represents the portion of OOV words in the test set. 2https://github.com/sonos/nlubenchmark/tree/master/2017-06-custom-intent-engines 3https://github.com/yvchen/JointSLU/tree/master/data Snips ATIS Vocabulary Size 11,241 722 Percentage of OOV words 5.95% 0.77% Number of Slots 39 79 Training Set Size 13,084 4,478 Development Set Size 700 500 Testing Set Size 700 893 Table 3: Statistics of ATIS and Snips datasets. 3.2 Data Construction and Processing For Snips and ATIS datasets, we keep some slot classes in training as unknown and integrate them back during testing, following (Fei and Liu, 2016; Shu et al., 2017; Lin and Xu, 2019). We randomly select part of slot types in Snips and ATIS as unknown slots(5%, 15%, and 30% in this paper). Note that the original train/val/test split is fixed. Considering class imbalance, we perform weighted sampling where the chosen probability is relevant to the number of class examples similar to (Lin and Xu, 2019). To avoid randomness of experiment results, we report the average result over 10 runs. After we choose the unknown slot types, a critical problem is how to handle sentences including these unknown slot types in training set. For OOD intent detection, we just need to remove these sentences in training and validation set. However, for Novel Slot Detection, a sentence perhaps contains both in-domain slots and unknown slots, which is nontrivial for tackling unknown slots at the token level. We need to balance the performance of recognizing unknown slots and in-domain slots. Therefore, we propose three different processing strategies as follows: (1) Replace: We label the unknown slot values with all O in the training set while the original values remain unchanged. (2) Mask: We label the unknown slot values with all O and mask these slot values with a special token MASK. (3) Remove: All the sentences containing unknown slots are directly removed. We display examples of the above three strategies in Table 2. For the val and test set, we just label the unknown slot values with all NS while keeping the in-domain labeling fixed. Note that NS 3487 Snips-NSD-15% Train Val Test number of in-domain slots 33 33 33 number of unknown slots 6 6 6 percentage of OOV words 8.51% number of queries 9,329 700 700 number of queries including unknown slots 0 192 202 number of slot values 23,176 1,794 1,790 number of unknown slot values 0 210 220 Table 4: The detailed statistics of Snips-NSD-15%. tags only exist in the val and test set, not in the training set. Besides, we keep original in-domain slots fixed to evaluate the performance of both NS and in-domain slots. We aim to simulate the practical scenario where we can hardly know what unknown slots are. These three strategies all have its practical significance. Compared with others, Remove is the most suitable strategies for real-world scenarios. In practical scenario, dialog systems first train in the data set labeled by human annotators, and then applied to the actual application. In the process of interaction with the real users, novel slot types appear gradually. Therefore, we consider that the training set doesn’t contain potential novel slots sentences. In other words, Remove is the most suitable strategy for NSD in real applications. What’s more, Section 5.3.1 demonstrates Remove performs best while the others suffer from severe model bias by O tags. Therefore, we adopt Remove as the main strategy in this paper. 3.3 Statistic of New NSD Datasets Table 4 shows the detailed statistics of Snips-NSD15% constructed by Remove strategy, where we choose 15% classes in the training data as unknown slots. 4 Combining Table 3 and Table 4, we can find Remove strategy removes 28.70% of queries in the original Snips training set, hence increases the percentage of OOV word from 5.95% to 8.51%. And unknown slot values account for 12.29% of total slot values in the test set. 3.4 Metrics The traditional slot filling task uses Span F1 5 for evaluation. Span F1 considers the exact span matching of an unknown slot span. However, we find in Section 5.3.3 that this metric is too strict to NSD 4Since different proportions of unknown slots have different statistics, here we only display the results of Snips-NSD15% for brevity. 5https://www.clips.uantwerpen.be/conl l2000/chunking/conlleval.txt Contextual Encoder Embedding Layer play is this my world by leo arnaud ...... Softmax Layer MSP/GDA In-domain Slot Types Novel Slot Types Training Test Figure 2: The overall architecture of our approach. models. In the practical application, we only need to coarsely mine parts of words of unknown slots, then send these queries containing potential unknown slot tokens to human annotators, which has effectively reduced extensive labor and improved efficiency. Therefore, we define a more reasonable metric, Token F1 which focuses on the word-level matching of a novel slot span. We also propose a new metric, Restriction-Oriented Span Evaluation (ROSE), for a fair comparison in Section 5.3.3. 4 Methodology In this section, we introduce the NSD models proposed in this paper and illustrate the differences between the various parallel approaches during the training and test stage. 4.1 Overall Framework The overall structure of model is shown in Fig 2. In the training stage, we either train a multiple-class classifier or binary classifier using different training objectives. We use public BERT-large (Devlin et al., 2019) embedding layer and BiLSTM-CRF (Huang et al., 2015) for token level feature extraction. Then, in the test stage, we use the typical neural multiple classifier to predict the in-domain slot labels. Meanwhile, we use the detection algorithm, MSP or GDA to figure out novel slot tokens. Finally, we override the slot token labels which are detected as NS. In terms of training objectives, detection algorithms, and distance strategies, we compare different variants as follows. Training objective. For in-domain slots, we propose two training objectives. Multiple classifier refers to the traditional slot filling objective setting, which performs token-level multiple classifications on the BIO tags (Ratinov and Roth, 2009) combined with different slots. Binary classifier unifies all non-O tags into one class, and the model makes 3488 Models 5% 15% 30% IND NSD IND NSD IND NSD detection method objective distance strategy Span F1 Span F1 Token F1 Span F1 Span F1 Token F1 Span F1 Span F1 Token F1 MSP binary 87.21 12.34 25.16 71.44 12.31 39.50 58.88 8.73 40.38 multiple 88.05 14.04 30.50 79.71 20.97 40.02 78.52 25.26 46.91 binary+multiple 89.59 23.58 37.55 83.72 24.70 45.32 79.08 30.66 52.10 GDA binary difference 87.95 23.83 35.83 83.65 22.06 43.99 78.72 32.50 44.13 binary minimum 61.29 10.36 17.08 49.11 16.91 31.10 48.07 15.56 33.78 multiple difference 93.14 29.73 45.99 90.07 31.96 53.02 85.56 36.16 54.55 multiple minimum 93.10 31.67* 46.97* 90.18 32.19 53.75* 86.26* 38.64* 55.24* Table 5: IND and NSD results with different proportions (5%, 15% and 30%) of classes are treated as unknown slots on Snips-NSD. * indicates the significant improvement over all baselines (p < 0.05). Models 5% 15% 30% IND NSD IND NSD IND NSD detection method objective distance strategy Span F1 Span F1 Token F1 Span F1 Span F1 Token F1 Span F1 Span F1 Token F1 MSP binary 92.04 19.73 29.63 91.74 23.40 33.89 80.49 21.88 39.17 multiple 94.33 27.15 31.16 92.54 39.88 42.29 87.63 40.42 47.64 binary+multiple 94.41 32.49 43.48 93.29 41.23 43.13 90.14 41.76 51.87 GDA binary difference 93.69 27.02 34.21 92.13 30.51 36.30 88.73 30.91 45.64 binary minimum 93.57 15.90 20.96 90.98 24.53 27.26 88.21 26.40 39.83 multiple difference 95.20 47.78* 51.54* 93.92 50.92* 52.24* 92.02 51.26* 56.59* multiple minimum 95.31* 41.74 45.91 93.88 43.78 46.18 91.67 45.44 52.37 Table 6: IND and NSD results with different proportions (5%, 15% and 30%) of classes are treated as unknown slots on ATIS-NSD. * indicates the significant improvement over all baselines (p < 0.05). a token-level binary classification of O or non-O on the sequence. Note that in the test stage, for indomain prediction, we both use the multiple classifier. While, for novel slot detection, we use the multiple classifier, or the binary classifier, or both of them. In Table 5 and Table 6, binary+multiple means the token will be labeled as NS only if both classifiers predict it as NS. Detection algorithm. MSP and GDA are detection algorithms in the test stage. MSP (Maximum Softmax Probability) (Hendrycks and Gimpel, 2017) applies a threshold on the maximum softmax probability, if the maximum falls below the threshold, the token will be predicted to be a novel slot token. GDA (Gaussian Discriminant Analysis) (Xu et al., 2020a) is a generative distancebased classifier for out-of-domain detection with Euclidean space. We treat tokens not belonging to any in-domain slots (including O) as novel slot tokens for both methods. For example, with a binary classifier, if the softmax probabilities belonging to O or non-O are both lower than an MSP threshold, then the token is labeled as NS. Distance strategy. The GDA detection is based on the distances between a target and each slot representation cluster. In original GDA, when the minimum distance is greater than a certain threshold, it is predicted to be novel slots. We propose a novel strategy named Difference, which uses the maximum distance minus the minimum distance, when the difference value of a target is less than a threshold, it is predicted as novel slots. Both of their thresholds are obtained by optimizing the NSD metrics on the validation set. 5 Experiment and Analysis 5.1 Implementation Details We use the public pre-trained Bert-large-uncased model to embed tokens which has 24 layers, 1024 hidden states, 16 heads and 336M parameters. The hidden size for the BiLSTM layer is set to 128. Adam is used for optimization with an initial learning rate of 2e-5. The dropout value is fixed as 0.5, and the batch size is 64. We train the model only on in-domain labeled data. The training stage has an early stopping setting with patience equal to 10. We use the best F1 scores on the validation set to calculate the MSP and GDA thresholds adaptively. Each result of the experiments is tested for 10 times under the same setting and reports the average value. The training stage of our model lasts about 28 minutes on single Tesla T4 GPU(16 GB of memory). 5.2 Main Results Table 5 and 6 show the experiment results with seven different models on two benchmark slot filling datasets Snips-NSD and ATIS-NSD constructed by Remove strategy. We both report NSD and IND results using Span F1 and Token F1. We compare these models from three perspectives, detection method, objective and distance strategy in the following. The analysis of effect of the propor3489 Strategy 5% 15% 30% IND NSD IND NSD IND NSD Span Span Token Span Span Token Span Span Token Replace 94.52 1.93 5.27 94.33 0.66 2.29 94.02 0.27 0.82 Mask 90.08 23.10 37.91 86.52 25.07 45.92 83.37 32.14 50.68 Remove 93.10 31.67 46.97 90.18 32.19 53.75 86.26 38.64 55.24 Table 7: Comparison between different data processing strategies on Snips-NSD using GDA+Multiple+Minimum. 5% 15% 30% 32.5 35.0 37.5 40.0 42.5 45.0 47.5 50.0 F1-score (macro) your title name Snips NSD span F1 ATIS NSD span F1 5% 15% 30% 86 88 90 92 94 Snips IND span F1 ATIS IND span F1 Proportion of Unknown Slot Types Figure 3: Effect of the proportion of unknown slot types. tion of unknown slot types is described in 5.3.2. Detection Method: MSP vs GDA. Under the same setting of objective, GDA performs better than MSP in both IND and NSD, especially in NSD. We argue that GDA models the posterior distribution on representation spaces of the feature extractor and avoids the issue of overconfident predictions (Guo et al., 2017; Liang et al., 2017b, 2018). Besides, comparing Snips-NSD and ATISNSD, NSD Token F1 scores on ATIS-NSD are much higher than Snips-NSD but no significant difference exists for NSD Span F1 scores. The reason is that Snips-NSD has a higher average entity length (1.83) than ATIS-NSD (1.29), making it harder to detect the exact NS span. Objective: Binary vs Multiple. Under all settings, Multiple outperforms Binary with a large margin on two datasets in both IND and NSD metrics. For MSP, combining Multiple and Binary get higher F1 scores. Specifically, the Binary classifier is used to calculate the confidence of a token belonging to non-O type, which can judge whether the token belongs to entities and distinguish NS from type O. On the other hand, we use the Multiple classifier to calculate the confidence for tokens that are of type NS, to distinguish NS from all predefined non-O slot types. For GDA, we do not combine Multiple and Binary because of poor performance. Multiple achieves the best results for all the IND and NSD F1 scores. We suppose multi-class classification can better capture semantic features than binary classification. Distance Strategy: Minimum vs Difference. We find under the same setting of Binary, Difference strategy outperforms Minimum on both datasets for NSD metrics. But under the same setting of Multiple, there is no consistent superiority between the two distance strategies. For example, Difference outperforms Minimum for NSD metrics on ATIS-NSD, opposite to the results on Snips-NSD. We argue different distance strategies are closely related to objective settings and dataset complexity. We will leave the theoretical analysis to the future. 5.3 Qualitative Analysis 5.3.1 Effect of Different Data Processing Strategies Table 7 displays IND and NSD metrics of three different dataset processing strategies on Snips-NSD using the same model GDA+Multiple+Minimum. In this section, we will dive into the analysis of the effects of different data processing strategies. Results show the Replace strategy gets poor performance in NSD, which proves labeling unknown slots as O tags will severely mislead the model. The Mask and Remove strategies are more reasonable since they remove unknown slots from the training data. Their main difference is that Mask only deletes token-level information, while Remove even eliminates the contextual information. For NSD in all datasets, Remove gains significantly better performance on both Token F1 and Span F1 than Mask by 9.06%(5%), 7.83%(15%) and 4.56%(30%) on Token F1, and 8.57%(5%), 7.12%(15%) and 6.5%(30%) on Span F1. We argue the remaining context is still misleading even if the novel slot tokens are not directly trained in the Mask strategy. Besides, Mask does not conform to the real NSD scenario. Generally, Remove is the most suitable strategy for NSD in real applications and can achieve the best performance. 5.3.2 Effect of the Proportion of Unknown Slot Types Fig 3 displays the effect of the proportion of unknown slot types using the Remove strategy in GDA+Multiple+Minimum. Results show that with the increase of the proportion of unknown slot types, the NSD F1 scores get improvements while IND F1 scores decrease. We suppose fewer indomain slot types help the model distinguish unknown slots from IND slots, thus NSD F1 scores get improvements. However, for in-domain slot detection, since Remove deletes all the sentences containing unknown slots in the training data, our 3490 ROSE-25% ROSE-50% ROSE-75% ROSE-100% Span F1 Metrics 15 20 25 30 35 40 F1-score (macro) MSP+bin. MSP+mul. MSP+bin.+mul. GDA+bin.+min. GDA+bin.+diff. GDA+mul.+min. GDA+mul.+diff. Figure 4: Effect of varying degrees of restrictions GDA+mul.+min. MSP+bin.+mul. ROSE-mean 40.73 34.71 ROSE-100% 40.39 33.74 ROSE-50% 41.00 35.46 Table 8: ROSE metrics on Snips-NSD using GDA+Multiple+Minimum and MSP+Binary+Multiple models suffer from the lack of sufficient context to recognize IND slots so IND F1 scores decrease. 5.3.3 New Metric: ROSE The previous results have shown Span F1 is much lower than the token F1. The reason is that Span F1 is a strict metric, where the model needs to correctly predict all NS tokens and the correct boundary. This is difficult for NSD models due to the lack of supervised information. In fact, NSD models only need to mark some tokens in the span of novel slots and send the total sequence containing the NS tokens back to the humans. A small number of token omissions or misjudgments are acceptable. Therefore, to meet a reasonable NSD scenario, we propose a new metric, restriction-oriented span evaluation (ROSE), to evaluate the span prediction performance under different restrictions. First, we do not punish the situation where tokens prediction exceeds the span. Then, we consider a span is correct when the number of correctly predicted tokens is greater than a settable proportion p of the span length. We take the average of the ROSE score and the original span F1 to avoid the model obtaining an outstanding result through over-long prediction. The results using Snips with 15% of novel slots are shown in Figure 4. As the degree of restriction increases, the metrics tend to decline. It indicates that the model can mostly identify more than half Type Proportion(%) Span Length Token F1 Span F1 top 5 Object name 21.42 3.71 55.64 20.82 TimeRange 15.29 2.35 53.65 30.15 Entity name 23.14 3.09 48.56 22.83 Music item 14.86 1.05 46.23 34.59 Artist 15.29 2.05 45.26 26.36 bottom 5 City 8.57 1.32 18.72 15.85 Country 6.29 1.57 14.19 11.11 State 5.54 1.10 13.55 10.83 Best rating 6.14 1.00 11.04 11.04 Year 3.43 1.00 10.24 10.24 Table 9: Results of single unknown slot. Type 1 Type 2 Token F1 Span F1 Object name 55.64 20.82 TimeRange 53.65 30.15 Party size number 33.44 28.57 City 18.72 15.85 State 13.55 10.83 Object name TimeRange 53.88 23.37 Object name Party size number 52.81 22.35 Object name City 57.92 21.42 Object name State 56.32 19.27 TimeRange Party size number 71.27∗ 51.03∗ City State 29.33∗ 27.14∗ Table 10: Results of combining multiple unknown slots. * denotes that NSD performance of the combination of two unknown slots is significantly better than each single slot. of the tokens in spans. To make a comprehensive evaluation, we defined the ROSE-mean, namely the mean of ROSE-25%, ROSE-50%, ROSE-75%, and ROSE-100%. We present results on part of proposed models in Table 8. 5.3.4 Analysis of Single Unknown Slot To analyze the relationship between NSD performance and a single specific slot, we calculate the token and span metrics treating each single slot type as an unknown slot and show the results of the top five and bottom five for Token F1 scores in Table 9. We find that the slots with better performance often account for a larger percentage of the data set, such as Object name or Entity name. They also tend to have a larger value space, such as TimeRange, Music item, or Artist. These characteristics allow the semantic representation of these slots to be distributed over a large area rather than clustered tightly together. We consider that this distribution is more reasonable because in a real application scenario, novel slots are diverse and its distribution tends to be diffuse. Performance on these types also proves that the NSD models we propose can be better generalized to a reasonable data setting. 3491 NSD error proportion(%) O Open vocabulary slots Other slots Sum Prediction is NS 17.79 18.84 9.07 45.70 Target is NS 18.47 7.54 28.29 54.30 Sum 36.26 26.38 37.36 100.00 Table 11: Relative proportions of several types of errors. Error type NS Example NS to O movie name (m name) text: when will paris by night aired true: O O B-m name I-m name I-m name O predict: O O NS O NS O NS to open slot album text: play the insoc ep true: O B-album I-album I-album predict: O B-object name I-object name NS NS to other slot artist text: play kurt cobain ballad tunes true: O B-artist I-artist B-music item O predict: O B-genre I-genre B-music item O O to NS artist text: the workout playlist needs more chris cross true: O B-playlist O O O B-artist I-artist predict: O B-playlist O O NS NS NS open slots to NS object type text: tell me the actors of the saga awards true: O O O B-object name O O B-object type O predict: O O O NS O O NS O other slots to NS city text: what is the weather of east portal ks true: O O O O O B-city I-city B-state predict: O O O O O NS NS NS Table 12: Error case from NSD prediction. 5.3.5 Analysis for Relationship of Multiple Unknown Slots In order to explore the effect of inter-slot relationships on NSD, we conducted experiments in which two types are mixed as novel slots. Some of the results are shown in Table 10. In the five types shown in the table, Object name is an open vocabulary slot with a wide range of values and contains many OOV tokens, TimeRange and Party size number often contain numbers, City and State are usually similar in semantics and context. We found that when the other types combined with Object name, NSD performance is often maintained close to treat Object name as a novel slot alone. The reason, on the one hand, is that the proportion of other types in the dataset is relatively small, so the overall impact on the metrics is smaller. On the other hand, due to the large semantic distribution range of the open vocabulary slot, there is a latent inclusion relationship for other types, so the mixing of a single type tends to have a slight impact on the NSD performance. We also found that the appropriate combination can significantly improve the efficiency of NSD. Such as TimeRange with Party size number, or City with State. This indicates that when the novel slot is similar to the in-domain slot, the model tends to predict the novel slot as a similar slot, which leads to errors. When both are treated as novel slots, these errors can be mitigated. 6 Discussion In this section, we empirically divide all the error samples into three categories. Each type of problem contains two aspects, corresponding to NSD precision and recall, respectively. We present the relative proportions of several types of errors in Table 11, which using Snips dataset with 5% novel slots on GDA+multiple+minimum model. For each error type, we present an example in Table 12 to describe the characteristics and analyze the causes. Then, we dive into identifying the key challenges and finally proposed possible solutions for future work. 6.1 Error Analysis Tag O. Tag O is the largest and most widely distributed type in the dataset, and it generally refers to the independent function tokens. Therefore, when identifying, it is easy to be confused with other types, and the confusion is more serious for novel slots without supervised learning. We observed that tokens with O label detected as novel slots usually exist near spans, and the function words in the span labeled as a novel slot have a probability of being predicted as O. We consider that this kind of problem is related to the context. Although the processing strategy of Remove can effectively reduce the misleading of O for the novel slots, tag O will still be affected by context information of other in-domain slots. Open Vocabulary Slots. We observe that a large number of novel slot tokens are mispredicted as open vocabulary slots, while the reverse situation is much less likely to happen. This indicates that in Snips, open vocabulary slots tend to overlap or contain most other slots semantically. Even in traditional slot filling tasks, open vocabulary slots are often confused with other slots. We demonstrate this hypothesis in the analysis. Section 5.3.5 shows that NSD performs better when open vocabulary slots are treated as novel slots, and Section 5.3.4 shows that there is no significant performance change when open vocabulary slots are mixed with some semantically concentrated slots. The reason for this problem is that the definition of the dataset is not reasonable. Slots with a large value range can hardly help the personal assistant to give an appropriate reply, and the supervised information of these slots is usually incomplete. Similar Slots. Except for the two cases mentioned above, predicting novel slots as other in-domain 3492 slots is the most common type of error, in which similar slots account for a large part of it. Due to the overlap between vocabulary or shared similar context, the model often tend to be overconfident to predict similar slot labels, we analyze the phenomenon in Table 10, when similar types is treated as a new slot at the same time, NSD efficiency will rise significantly. We employ a generative classification method GDA, compared with the traditional MSP method, to make full use of data features and alleviate the problem. 6.2 Challenges Based on the above analysis, we summarize the current challenges faced by the NSD task: Function tokens. Articles, prepositions, and so on that act as connective words in a sequence. It is usually labeled with type O, but also found in some long-span slots, such as Movie name. It can lead to confusion between O and novel slot when this kind of slot is the target of NSD. Insufficient context. Correct slot detection often depends on the context, and this supervised information is missing for novel slots. Models can only conduct NSD to tokens using the original embeddings or representations trained in other contexts, which can lead to bias in the semantic modeling of the novel slot. Dependencies between slots. There are some semantic overlaps or inclusion relationships in the slot definition of the current benchmark slot filling datasets. As a result, the semantic features are not sufficiently discriminative, and thus some outliers tokens in in-domain slots are easily confused with the novel slots. Open vocabulary slots. Open vocabulary slots is a special kind of slot, its definition is usually macroscopic and can be further divided, the value range is broad. The representation distribution for Open vocabulary slots tends to be diffuse and uneven, which can be misleading to NSD. 6.3 Future Directions For tag O, a possible solution is to use a binary model to assist identification between O and non-O function tokens, we provide a simple method in this paper and leave further optimizing to future work. Then, to decouple the dependencies between slots, it is critical to learn more discriminative features for in-domain data, using contrastive learning or prototypical network is expected to help. Besides, in the traditional slot filling task, the open vocabulary slot problem has been researched for a long time, and accumulate many achievements. Adaptive combination and improvement of relevant methods with NSD tasks is also an important direction of our future research. 7 Related Work OOV Recognition OOV aims to recognize unseen slot values in training set for pre-defined slot types, using character embedding (Liang et al., 2017a), copy mechanism (Zhao and Feng, 2018), few/zeroshot learning (Hu et al., 2019; Shah et al., 2019), transfer learning (Chen and Moschitti, 2019; He et al., 2020c) and background knowledge (Yang and Mitchell, 2017; He et al., 2020d), etc. Our proposed NSD task focuses on detecting unknown slot types, not just unseen values. OOD Intent Detection Lee et al. (2018); Lin and Xu (2019); Xu et al. (2020a) aim to know when a query falls outside the range of predefined supported intents. Generally, they first learn discriminative intent representations via in-domain (IND) data, then employs detecting algorithms, such as Maximum Softmax Probability (MSP) (Hendrycks and Gimpel, 2017), Local Outlier Factor (LOF) (Lin and Xu, 2019), Gaussian Discriminant Analysis (GDA) (Xu et al., 2020b) to compute the similarity of features between OOD samples and IND samples. Compared to our proposed NSD, the main difference is that NSD detects unknown slot types in the token level while OOD intent detection identifies sentence-level OOD intent queries. 8 Conclusion In this paper, we defined a new task, Novel Slot Detection(NSD), then provide two public datasets and establish a benchmark for it. Further, we analyze the problems of NSD through multi-angle experiments and extract the key challenges of the task. We provide some strong models for these problems and offer possible solutions for future work. Acknowledgements This work was partially supported by National Key R&D Program of China No. 2019YFF0303300 and Subject II No. 2019YFF0303302, DOCOMO Beijing Communications Laboratories Co., Ltd, MoE-CMCC ”Artifical Intelligence” Project No. MCM20190701. 3493 Broader Impact Dialog systems have demonstrated remarkable performance across a wide range of applications, with the promise of a significant positive impact on human production mode and lifeway. The first step of the dialog system is to identify users’ key points. In practical industrial scenario, users may make unreasonable queries which fall outside of the scope of the system-supported slot types. Previous dialogue systems will ignore this problem, which will lead to wrong operations and limit the system’s development. In this paper, we firstly propose to detect not only pre-defined slot types but also potential unknown or out-of-domain slot types using MSP and GDA methods. According to exhaustive experiments and qualitative analysis, we also discuss several major challenges in Novel Slot Detection for future work. The effectiveness and robustness of the model are significantly improved by adding Novel Slot Detection, which takes a step towards the ultimate goal of enabling the safe real-world deployment of dialog systems in safety-critical domains. The experimental results have been reported on standard benchmark datasets for considerations of reproducible research. References Lingzhen Chen and Alessandro Moschitti. 2019. Transfer learning for sequence labeling using source model and target data. ArXiv, abs/1902.05309. Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. arXiv preprint arXiv:1902.10909. A. Coucke, A. Saade, Adrien Ball, Th´eodore Bluche, A. Caulier, D. Leroy, Cl´ement Doumouro, Thibault Gisselbrecht, F. Caltagirone, Thibaut Lavril, Ma¨el Primet, and J. Dureau. 2018. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. ArXiv, abs/1805.10190. J. Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT. Geli Fei and B. Liu. 2016. Breaking the closed world assumption in text classification. In HLT-NAACL. Chih-Wen Goo, Guang Gao, Yun-Kai Hsu, Chih-Li Huo, Tsung-Chieh Chen, Keng-Wei Hsu, and YunNung Chen. 2018. Slot-gated modeling for joint slot filling and intent prediction. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 753–757. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. 2017. On calibration of modern neural networks. In ICML. E Haihong, Peiqing Niu, Zhongfu Chen, and Meina Song. 2019. A novel bi-directional interrelated model for joint intent detection and slot filling. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5467– 5471. Keqing He, Shuyu Lei, Yushu Yang, Huixing Jiang, and Zhongyuan Wang. 2020a. Syntactic graph convolutional network for spoken language understanding. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2728– 2738, Barcelona, Spain (Online). International Committee on Computational Linguistics. Keqing He, Weiran Xu, and Yuanmeng Yan. 2020b. Multi-level cross-lingual transfer learning with language shared and specific knowledge for spoken language understanding. IEEE Access, 8:29407– 29416. Keqing He, Yuanmeng Yan, Si hong Liu, Z. Liu, and Weiran Xu. 2020c. Learning label-relational output structure for adaptive sequence labeling. 2020 International Joint Conference on Neural Networks (IJCNN), pages 1–8. Keqing He, Yuanmeng Yan, and Weiran Xu. 2020d. Learning to tag OOV tokens by integrating contextual representation and background knowledge. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 619– 624, Online. Association for Computational Linguistics. Keqing He, Jinchao Zhang, Yuanmeng Yan, Weiran Xu, Cheng Niu, and Jie Zhou. 2020e. Contrastive zeroshot learning for cross-domain slot filling with adversarial attack. In COLING. C. T. Hemphill, J. J. Godfrey, and G. Doddington. 1990. The atis spoken language systems pilot corpus. In HLT. Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-ofdistribution examples in neural networks. ArXiv, abs/1610.02136. Ziniu Hu, Ting Chen, Kai-Wei Chang, and Yizhou Sun. 2019. Few-shot representation learning for out-ofvocabulary words. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4102–4112, Florence, Italy. Association for Computational Linguistics. Zhiheng Huang, W. Xu, and Kai Yu. 2015. Bidirectional lstm-crf models for sequence tagging. ArXiv, abs/1508.01991. 3494 Stefan Larson, Anish Mahendran, Joseph Peper, Christopher Clarke, Andrew Lee, P. Hill, Jonathan K. Kummerfeld, Kevin Leach, M. Laurenzano, L. Tang, and J. Mars. 2019. An evaluation dataset for intent classification and out-of-scope prediction. ArXiv, abs/1909.02027. Kimin Lee, Kibok Lee, H. Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting outof-distribution samples and adversarial attacks. In NeurIPS. Dongyun Liang, Weiran Xu, and Yinge Zhao. 2017a. Combining word-level and character-level representations for relation classification of informal text. In Proceedings of the 2nd Workshop on Representation Learning for NLP, pages 43–47, Vancouver, Canada. Association for Computational Linguistics. Shiyu Liang, Yixuan Li, and R. Srikant. 2017b. Principled detection of out-of-distribution examples in neural networks. ArXiv, abs/1706.02690. Shiyu Liang, Yixuan Li, and R. Srikant. 2018. Enhancing the reliability of out-of-distribution image detection in neural networks. arXiv: Learning. Ting-En Lin and H. Xu. 2019. Deep unknown intent detection with margin loss. ArXiv, abs/1906.00434. Bing Liu and Ian Lane. 2015. Recurrent neural network structured output prediction for spoken language understanding. In Proc. NIPS Workshop on Machine Learning for Spoken Language Understanding and Interactions. Bing Liu and Ian Lane. 2016. Attention-based recurrent neural network models for joint intent detection and slot filling. arXiv preprint arXiv:1609.01454. Samuel Louvan and B. Magnini. 2020. Recent neural methods on slot filling and intent classification for task-oriented dialogue systems: A survey. In COLING. Gr´egoire Mesnil, Yann Dauphin, Kaisheng Yao, Yoshua Bengio, Li Deng, Dilek Z. Hakkani-Tur, Xiaodong He, Larry Heck, Gokhan Tur, Dong Yu, and Geoffrey Zweig. 2015. Using recurrent neural networks for slot filling in spoken language understanding. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23:530–539. Lev Ratinov and Dan Roth. 2009. Design challenges and misconceptions in named entity recognition. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009), pages 147–155. J. Ren, Peter J. Liu, E. Fertig, Jasper Snoek, Ryan Poplin, Mark A. DePristo, Joshua V. Dillon, and Balaji Lakshminarayanan. 2019. Likelihood ratios for out-of-distribution detection. In NeurIPS. Darsh J. Shah, Raghav Gupta, A. Fayazi, and Dilek Z. Hakkani-T¨ur. 2019. Robust zero-shot crossdomain slot filling with example values. ArXiv, abs/1906.06870. Lei Shu, Hu Xu, and Bing Liu. 2017. Doc: Deep open classification of text documents. ArXiv, abs/1709.08716. H. Xu, Keqing He, Yuanmeng Yan, Si hong Liu, Z. Liu, and Weiran Xu. 2020a. A deep generative distancebased classifier for out-of-domain detection with mahalanobis space. In COLING. Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2020b. A deep generative distance-based classifier for out-of-domain detection with mahalanobis space. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1452–1460, Barcelona, Spain (Online). International Committee on Computational Linguistics. Yuanmeng Yan, Keqing He, Hong Xu, Sihong Liu, Fanyu Meng, Min Hu, and Weiran Xu. 2020. Adversarial semantic decoupling for recognizing openvocabulary slots. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6070–6075, Online. Association for Computational Linguistics. B. Yang and Tom Michael Mitchell. 2017. Leveraging knowledge bases in lstms for improving machine reading. In ACL. Zhiyuan Zeng, Keqing He, Yuanmeng Yan, Hong Xu, and Weiran Xu. 2021a. Adversarial self-supervised learning for out-of-domain detection. In NAACL. Zhiyuan Zeng, Hong Xu, Keqing He, Yuanmeng Yan, Sihong Liu, Zijun Liu, and Weiran Xu. 2021b. Adversarial generative distance-based classifier for robust out-of-domain detection. In ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7658–7662. Lin Zhao and Zhe Feng. 2018. Improving slot filling in spoken language understanding with joint pointer and attention. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 426– 431, Melbourne, Australia. Association for Computational Linguistics. Yinhe Zheng, Guanyi Chen, and Minlie Huang. 2020. Out-of-domain detection for natural language understanding in dialog systems. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:1198–1209.
|
2021
|
270
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3495–3506 August 1–6, 2021. ©2021 Association for Computational Linguistics 3495 GTM: A Generative Triple-Wise Model for Conversational Question Generation Lei Shen1,2 Fandong Meng3 Jinchao Zhang3 Yang Feng1,2∗ Jie Zhou3 1Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China 2University of Chinese Academy of Sciences, Beijing, China 3Pattern Recognition Center, WeChat AI, Tencent Inc, China [email protected], {fandongmeng,dayerzhang}@tencent.com [email protected], [email protected] Abstract Generating some appealing questions in opendomain conversations is an effective way to improve human-machine interactions and lead the topic to a broader or deeper direction. To avoid dull or deviated questions, some researchers tried to utilize answer, the “future” information, to guide question generation. However, they separate a post-questionanswer (PQA) triple into two parts: postquestion (PQ) and question-answer (QA) pairs, which may hurt the overall coherence. Besides, the QA relationship is modeled as a one-to-one mapping that is not reasonable in open-domain conversations. To tackle these problems, we propose a generative triple-wise model with hierarchical variations for open-domain conversational question generation (CQG). Latent variables in three hierarchies are used to represent the shared background of a triple and one-to-many semantic mappings in both PQ and QA pairs. Experimental results on a largescale CQG dataset show that our method significantly improves the quality of questions in terms of fluency, coherence and diversity over competitive baselines. 1 Introduction Questioning in open-domain dialogue systems is indispensable since a good system should have the ability to well interact with users by not only responding but also asking (Li et al., 2017). Besides, raising questions is a proactive way to guide users to go deeper and further into conversations (Yu et al., 2016). Therefore, the ultimate goal of opendomain conversational question generation (CQG) is to enhance the interactiveness and maintain the continuity of a conversation (Wang et al., 2018). Joint work with Pattern Recognition Center, WeChat AI, Tencent Inc, China. ∗Yang Feng is the corresponding author. Post: I ate out with my friends this evening. Question Candidates: Q1.1: Which restaurant did you go? Q1.2: Where did you eat? Q2.1: What food did you eat? Q2.2: Did you eat something special? Q3: What do you mean? Q4: How about drinking together? Answer Candidates: A1: We went to an Insta-famous cafeteria. A2: We ate steak and pasta. Table 1: An example of CQG task which is talking about a person’s eating activity. There are one-to-many mappings in both PQ and QA pairs. The content of each meaningful and relevant question (Q1.1 to Q2.2) is decided by its post and answer. Q3 (dull) and Q4 (deviated) are generated given only the post. CQG differs fundamentally from traditional question generation (TQG) (Zhou et al., 2019; Kim et al., 2019; Li et al., 2019) that generates a question given a sentence/paragraph/passage and a specified answer within it. While in CQG, an answer always follows the to-be-generated question, and is unavailable during inference (Wang et al., 2019). At the same time, each utterance in open-domain scenario is casual and can be followed by several appropriate sentences, i.e., one-to-many mapping (Gao et al., 2019; Chen et al., 2019). At first, the input information of CQG was mainly a given post (Wang et al., 2018; Hu et al., 2018), and the generated questions were usually dull or deviated (Q3 and Q4 in Table 1). Based on the observation that an answer has strong relevance to its question and post, Wang et al. (2019) tried to integrate answer into the question generation process. They applied a reinforcement learning framework that firstly generated a question given the post, and then used a pre-trained matching model to estimate the relevance score (reward) between 3496 answer and generated question. This method separates a post-question-answer (PQA) triple into post-question (PQ) and question-answer (QA) pairs rather than considering the triple as a whole and modeling the overall coherence. Furthermore, the training process of the matching model only utilizes one-to-one relation of each QA pair and neglects the one-to-many mapping feature. An open-domain PQA often takes place under a background that can be inferred from all utterances in the triple and help enhance the overall coherence. When it comes to the semantic relationship in each triple, the content of a specific question is under the control of its post and answer (Lee et al., 2020). Meanwhile, either a post or an answer could correspond to several meaningful questions. As shown in Table 1, the triple is about a person’s eating activity (the background of the entire conversation). There are one-to-many mappings in both PQ and QA pairs that construct different meaningful combinations, such as P-Q1.1-A1, P-Q1.2-A1, P-Q2.1-A2 and P-Q2.2-A2. An answer connects tightly to both its post and question, and in turn helps decide the expression of a question. On these grounds, we propose a generative triplewise model (GTM) for CQG. Specifically, we firstly introduce a triple-level variable to capture the shared background among PQA. Then, two separate variables conditioned on the triple-level variable are used to represent the latent space for question and answer, and the question variable is also dependent on the answer one. During training, the latent variables are constrained to reconstruct both the original question and answer according to the hierarchical structure we define, making sure the triple-wise relationship flows through the latent variables without any loss. For the question generation process, we sample the triple-level and answer variable given a post, then obtain the question variable conditioned on them, and finally generate a question based on the post, triple-level and question variables. Experimental results on a largescale CQG dataset show that GTM can generate more fluent, coherent, and intriguing questions for open-domain conversations. The main contribution is threefold: • To generate coherent and informative questions in the CQG task, we propose a generative triple-wise model that models the semantic relationship of a triple in three levels: PQA, PQ, and QA. Figure 1: The graphical representation of GTM for training process. zt is used to capture the shared background among PQA, while zq and za are used to model the diversity in PQ and QA pairs. Solid arrows illustrate the generation of q, a (not used in inference), and qt, while dashed arrows are for posterior distributions of latent variables. • Our variational hierarchical structure can not only utilize the “future” information (answer), but also capture one-to-many mappings in PQ and QA, which matches the open-domain scenario well. • Experimental results on a large-scale CQG corpus show that our method significantly outperforms the state-of-the-art baselines in both automatic and human evaluations. 2 Proposed Model Given a post as the input, the goal of CQG is to generate the corresponding question. Following the work of Zhao et al. (2017) and Wang et al. (2019), we leverage the question type qt to control the generated question, and take advantage of the answer information a to improve coherence. In training set, each conversation is represented as {p, q, qt, a}, consisting of post p = {pi}|p| i=1, question q = {qi}|q| i=1 with its question type qt, and answer a = {ai}|a| i=1. 2.1 Overview The graphical model of GTM for training process is shown in Figure 1. θ, ϕ, and φ are used to denote parameters of generation, prior, and recognition network, respectively. We integrate answer generation to assist question generation with hierarchical latent variables. Firstly, a triple-level variable zt is imported to capture the shared background and 3497 Prior Network(a) Recognition Network (a) Recognition Network (q) Prior Network (q) MLP𝑡𝑟1 MLP𝑡𝑟2 𝐳𝑡 𝐳𝑞 𝐳𝑞 𝐳𝑎 𝐳𝑎 𝐳𝑡 𝜇′𝑞 𝜎′𝑞 𝜇𝑞 𝜎𝑞 𝜇𝑎 𝜎′𝑎 𝜇′𝑎 𝜎𝑎 𝑞𝑡 𝐡𝑒𝑛𝑐 𝑞 𝐡𝑒𝑛𝑐 𝑝 𝐡𝑒𝑛𝑐 𝑎 𝐡𝑐𝑡𝑥 𝑝 𝐡𝑐𝑡𝑥 𝑞 𝑞𝑡 𝑞𝑡′ 𝐡𝑑𝑒𝑐 𝑎,0 𝐡𝑑𝑒𝑐 𝑞,0 Encoder & Prior/Recognition Network Answer Decoder Question Decoder KL KL 𝐡𝑐𝑡𝑥 𝑎 Question Post Answer MLP𝑞𝑡 𝑞𝑡′ Question Type Prediction Bi-GRU Encoder Bi-GRU Encoder Bi-GRU Encoder GRU Decoder GRU Decoder I ate out with my friends this evening. What did you eat? We ate steak and pasta. What did you eat? We ate steak and pasta. [What] 𝐳𝑎 Figure 2: The architecture of GTM. ⊕denotes the concatenation operation. In training process, latent variables obtained from recognition networks and the real question type qt are used for decoding. Red dashed arrows refer to inference process, in which we get latent variables from prior networks, and the predicted question type qt′ is fed into the question decoder. The answer decoder is only utilized during training to assist the triple-wise modeling. is inferred from PQA utterances. Then answer latent variable za and question latent variable zq are sampled from Gaussian distributions conditioned on both post and zt. To ensure that the question is controlled by answer, zq is also dependent on za. 2.2 Input Representation We use a bidirectional GRU (Cho et al., 2014) as encoder to capture the semantic representation of each utterance. Take post p as an example. Each word in p is firstly encoded into its embedding vector. The GRU then computes forward hidden states {−→h i}|p| i=1 and backward hidden states {←−h i}|p| i=1: −→h i = −−→ GRU(epi, −→h i−1), ←−h i = ←−− GRU(epi, ←−h i+1), where epi is employed to represent the embedding vector of word pi. We finally get the post representation by concatenating the last hidden states of two directions henc p = [−→h |p|; ←−h 1]. Similarly, we can obtain representations of question q and answer a, denoted as henc q and henc a , respectively. The question type qt is represented by a realvalued, low dimensional vector vqt which is updated during training and is regarded as a linguistic feature that benefits the training of latent variables (Zhao et al., 2017). We use the actual question type qt during training to provide the information of interrogative words that is the most important feature to distinguish question types. 2.3 Triple-level Latent Variable To capture the shared background of entire triple, we introduce a triple-level latent variable zt that is inferred from PQA utterances and is in turn responsible for generating the whole triple. Inspired by Park et al. (2018), we use a standard Gaussian distribution as the prior distribution of zt: pϕ(zt) = N(z|0, I), where I represents the identity matrix. For the inference of zt in training set, we consider three utterance representations henc p , henc q and henc a as a sequence, and use a bidirectional GRU to take individual representation as the input of each time step. The triple representation ht is obtained by concatenating the last hidden states of both directions. Then, zt is sampled from: qφ(zt|p, q, a) = N(z|µt, σtI), µt = MLPt φ(ht), σt = softplus(MLPt φ(ht)), where MLP(·) is a feed-forward network, and softplus function is a smooth approximation to ReLU and can be used to ensure positiveness (Park et al., 2018; Serban et al., 2017). 2.4 One-to-many Mappings After obtaining zt, we use a GRU f to get a vector hctx p for connecting p and q/a. hctx p is then transformed to hctx q and hctx a that are used in prior and recognition networks for zq and za: hctx p = f(zt, henc p ), hctx q = MLPtr1 θ (hctx p ), hctx a = MLPtr2 θ (hctx p ). 3498 To model one-to-many mappings in PQ and QA pairs under the control of zt, we design two utterance-level variables, zq and za, to represent latent spaces of question and answer. We define the prior and posterior distributions of za as follows: pϕ(za|p, zt) = N(z|µa, σaI), qφ(za|p, zt, a) = N(z|µ ′ a, σ ′ aI), where µa, σa, µ ′ a, and σ ′ a, the parameters of two Gaussian distributions, are calculated as: µa = MLPa ϕ([hctx a ; zt]), σa = softplus(MLPa ϕ([hctx a ; zt])), µ ′ a = MLPa φ([hctx a ; zt; henc a ]), σ ′ a = softplus(MLPa φ([hctx a ; zt; henc a ])). To make sure the content of question is also decided by answer and improve their relatedness, we import za into zq space. The prior and posterior distributions of zq are computed as follows: pϕ(zq|p, zt, za) = N(z|µq, σqI), qφ(zq|p, zt, q, qt, za) = N(z|µ ′ q, σ ′ qI), where µq, σq, µ ′ q, and σ ′ q are calculated as: µq = MLPq ϕ([hctx q ; zt; za]), σq = softplus(MLPq ϕ([hctx q ; zt; za])), µ ′ q = MLPq φ([hctx q ; zt; henc q ; vqt; za]), σ ′ q = softplus(MLPq φ([hctx q ; zt; henc q ; vqt, za])). 2.5 Question Generation Network Following the work of Zhao et al. (2017) and Wang et al. (2019), a question type prediction network MLPqt is introduced to approximate pθ(qt|zq, zt, p) in training process and produces question type qt′ during inference. As shown in Figure 2, there are two decoders in our model, one is for answer generation that is an auxiliary task and only exists in the training process, and the other is for desired question generation. The question decoder employs a variant of GRU that takes the concatenation result of zq, zt, hctx q , and qt as initial state s0, i.e., s0 = [zq; zt, hctx q , qt]. For each time step j, it calculates the context vector cj following Bahdanau et al. (2015), and computes the probability distribution pθ(q|zq, zt, p, qt) over all words in the vocabulary: sj = GRU(ej−1, sj−1, cj) ˜sj = MLP([ej−1; cj; sj]), pθ(qj|q<j, zq, zt, p, qt) = softmax(Wo˜sj), where ej−1 represents the embedding vector of the (j −1)-th question word. Similarly, the answer decoder receives the concatenation result of za, zt, and hctx a as initial state to approximate the probability pθ(a|za, zt, p). 2.6 Training and Inference Importantly, our model GTM is trained to maximize the log-likelihood of the joint probability p(p, q, a, qt): logp(p, q, a, qt) = log Z zt p(p, q, a, qt, zt). However, the optimization function is not directly tractable. Inspired by Serban et al. (2017) and Park et al. (2018), we convert it to the following objective that is based on the evidence lower bound and needs to be maximized in training process: LGTM = −KL(qφ(zt|p, q, a)||pϕ(zt)) −KL(qφ(za|p, zt, a)||pϕ(za|p, zt)) −KL(qφ(zq|p, zt, q, qt, za)||pϕ(zq|p, zt, za)) + Eza,zt∼qφ[log pθ(a|za, zt, p)] + Ezq,zt∼qφ[log pθ(q|zq, zt, p, qt)] + Ezq,zt∼qφ[log pθ(qt|zq, zt, p)]. The objective consists of two parts: the variational lower bound (the first five lines) and question type prediction accuracy (the last line). Meanwhile, the variational lower bound includes the reconstruction terms and KL divergence terms based on three hierarchical latent variables. The gradients to the prior and recognition networks can be estimated using the reparameterization trick (Kingma and Welling, 2014). During inference, latent variables obtained via prior networks and predicted question type qt′ are fed to the question decoder, which corresponds to red dashed arrows in Figure 2. The inference process is as follows: 3499 (1) Sample triple-level LV: zt ∼qφ(zt|p)1. (2) Sample answer LV: za ∼pϕ(za|p, zt). (3) Sample question LV: zq ∼pϕ(zq|p, zt, za). (4) Predict question type: qt ∼pθ(qt|zq, zt, p). (5) Generate question: q ∼pθ(zq, zt, p, qt). 3 Experiments In this section, we conduct experiments to evaluate our proposed method. We first introduce some empirical settings, including dataset, hyperparameters, baselines, and evaluation measures. Then we illustrate our results under both automatic and human evaluations. Finally, we give out some cases generated by different models and do further analyses over our method. 3.1 Dataset We apply our model on a large-scale CQG corpus2 extracted from Reddit3 by Wang et al. (2019). There are over 1.2 million PQA triples which have been divided into training/validation/test set with the number of 1,164,345/30,000/30,000. The dataset has been tokenized into words using the NLTK tokenizer (Bird et al., 2009). The average number of words in post/question/answer is 18.84/19.03/19.30, respectively. Following Fan et al. (2018) and Wang et al. (2019), we categorize questions in training and validation set into 9 types based on interrogative words, i.e., “what”, “when”, “where”, “who”, “why”, “how”, “can (could)”, “do (did, does)”, “is (am, are, was, were)” 3.2 Hyper-parameter Settings We keep the top 40,000 frequent words as the vocabulary and the sentence padding length is set to 30. The dimension of GRU layer, word embedding and latent variables is 300, 300, and 100. The prior networks and MLPs have one hidden layer with size 300 and tanh non-linearity, while the number of hidden layers in recognition networks for both triple-level and utterance-level variables is 2. We apply dropout ratio of 0.2 during training. The mini-batch size is 64. For optimization, we use Adam (Kingma and Ba, 2015) with a learning rate of 1e-4. In order to alleviate degeneration problem of variational framework (Park et al., 2018), we 1Inspired by Park et al. (2018), using zt inferred from post with the posterior distribution is better than sampling it from the prior one, i.e., a standard Gaussian distribution. 2https://drive.google.com/drive/ folder/1wNG30YPHiMc_ZNyE3BH5wa1uVtR8l1pG 3http://www.reddit.com apply KL annealing, word drop (Bowman et al., 2016) and bag-of-word (BOW) loss (Zhao et al., 2017)4. The KL multiplier λ gradually increases from 0 to 1, and the word drop probability is 0.25. We use Pytorch to implement our model, and the model is trained on Titan Xp GPUs. 3.3 Baselines We compare our methods with four groups of representative models: (1) S2S-Attn: A simple Seq2Seq model with attention mechanism (Shang et al., 2015). (2) CVAE&kgCVAE: The CVAE model integrates an extra BOW loss to generate diverse questions. The kgCVAE is a knowledge-guided CVAE that utilizes some linguistic cues (question types in our experiments) to learn meaningful latent variables (Zhao et al., 2017). (3) STD&HTD: The STD uses soft typed decoder that estimates a type distribution over word types, and the HTD uses hard typed decoder that specifies the type of each word explicitly with Gumbel-softmax (Wang et al., 2018). (4) RL-CVAE: A reinforcement learning method that regards the coherence score (computed by a one-to-one matching network) of a pair of generated question and answer as the reward function (Wang et al., 2019). RL-CVAE is the first work to utilize the future information, i.e., answer, and is also the state-of-the-art model for CQG5. Additionally, we also conduct ablation study to better analyze our method as follows: (5) GTMzt: GTM without the triple-level latent variable, which means zt is not included in the prior and posterior distributions of both zp and za. (6) GTMa: the variant of GTM that does not take answer into account. That is, answer decoder and za are removed from the loss function and the prior and posterior distributions of zq. Besides, zt here does not capture the semantics from answer. (7) GTMzq/za: GTM variant in which distributions of zq are not conditioned on za, i.e., the fact that the content of question is also controlled by answer is not modelled explicitly by latent variables. In our model, we use an MLP to predict question types during inference, which is different from the conditional training (CT) methods (Li et al., 2016b; Zhou et al., 2018; Shen and Feng, 2020) 4The total BOW loss is calculated as the sum of all BOW losses between each latent variable and q/a. Please refer to Park et al. (2018) for more details. 5For those methods with open-source codes, we run the original codes; otherwise, we re-implement them based on the corresponding paper. 3500 Model Embedding Metrics Diversity BLEU Scores RUBER Scores Average Extrema Greedy Dist-1 Dist-2 BLEU-1 BLEU-2 RubG RubA S2S-Attn 0.634 0.322 0.413 0.0132 0.0830 0.0936 0.0298 0.584 0.622 CVAE 0.646 0.337 0.421 0.0160 0.1599 0.1422 0.0306 0.649 0.687 kgCVAE 0.647 0.332 0.425 0.0153 0.1587 0.1491 0.0310 0.650 0.682 STD 0.637 0.326 0.418 0.0144 0.1325 0.1327 0.0302 0.633 0.663 HTD 0.648 0.330 0.423 0.0154 0.1582 0.1475 0.0314 0.653 0.689 RL-CVAE 0.662 0.343 0.437 0.0161 0.1785 0.1503 0.0320 0.660 0.701 GTM-zt 0.672 0.351 0.448 0.0165 0.1872 0.1521 0.0332 0.661 0.710 GTM-a 0.653 0.338 0.428 0.0158 0.1679 0.1482 0.0317 0.657 0.692 GTM-zq/za 0.687 0.360 0.449 0.0170 0.1934 0.1528 0.0329 0.669 0.713 GTM 0.697 0.365 0.454 0.0176 0.2028 0.1537 0.0331 0.671 0.720 Table 2: Automatic evaluation results for different models based on four types of metrics. that provide the controllable feature, i.e., question types, in advance for inference. Therefore, we do not consider CT-based models as comparable ones. 3.4 Evaluation Measures To better evaluate our results, we use both quantitative metrics and human judgements in our experiments. Automatic Metrics For automatic evaluation, we mainly choose four kinds of metrics: (1) BLEU Scores: BLEU (Papineni et al., 2002) calculates the n-gram overlap score of generated questions against ground-truth questions. We use BLEU-1 and BLEU-2 here and normalize them to 0 to 1 scale. (2) Embedding Metrics: Average, Greedy and Extrema metrics are embedding-based and measure the semantic similarity between the words in generated questions and ground-truth questions (Serban et al., 2017; Liu et al., 2016). We use word2vec embeddings trained on the Google News Corpus6 in this part. Please refer to Serban et al. (2017) for more details. (3) Dist-1& Dist-2: Following the work of Li et al. (2016a), we apply Distinct to report the degree of diversity. Dist-1/2 is defined as the ratio of unique uni/bi-grams over all uni/bi-grams in generated questions. (4) RUBER Scores: Referenced metric and Unreferenced metric Blended Evaluation Routine (Tao et al., 2018) has shown a high correlation with human annotation in open-domain conversation evaluation. There are two versions, one is RubG based on geometric averaging and the other is RubA based on arithmetic averaging. Embedding metrics and BLEU scores are used to measure the similarity between generated and ground-truth questions. RubG/A reflects the se6https://code.google.com/archive/p/ word2vec/ mantic coherence of PQ pairs (Wang et al., 2019), while Dist-1/2 evaluates the diversity of questions. Human Evaluation Settings Inspired by Wang et al. (2019), Shen et al. (2019), and Wang et al. (2018), we use following three criteria for human evaluation: (1) Fluency measures whether the generated question is reasonable in logic and grammatically correct. (2) Coherence denotes whether the generated question is semantically consistent with the given post. Incoherent questions include dull cases. (3) Willingness measures whether a user is willing to answer the question. This criterion is to justify how likely the generated questions can elicit further interactions. We randomly sample 500 examples from test set, and generate questions using models mentioned above. Then, we send each post and corresponding 10 generated responses to three human annotators without order, and require them to evaluate whether each question satisfies criteria defined above. All annotators are postgraduate students and not involved in other parts of our experiments. 3.5 Experimental Results Now we demonstrate our experimental results on both automatic evaluation and human evaluation. Automatic Evaluation Results Now we demonstrate our experimental results on both automatic evaluation and human evaluation. The automatic results are shown in Table 2. The top part is the results of all baseline models, and we can see that GTM outperforms other methods on all metrics (significance tests (Koehn, 2004), p-value < 0.05), which indicates that our proposed model can improve the overall quality of generated questions. Specifically, Dist-2 and RubA have been improved by 2.43% and 1.90%, respectively, compared to the state-of-the-art RL-CVAE model. 3501 First, higher embedding metrics and BLEU scores show that questions generated by our model are similar to ground truths in both topics and contents. Second, taking answer into account and using it to decide the expression of question can improve the consistency of PQ pairs evaluated by RUBER scores. Third, higher distinct values illustrate that one-to-many mappings in PQ and QA pairs make the generated responses more diverse. The bottom part of Table 2 shows the results of our ablation study, which demonstrates that taking advantage of answer information, modeling the shared background in entire triple, and considering one-to-many mappings in both PQ and QA pairs can help enhance the performance of our hierarchical variational model in terms of relevance, coherence and diversity. Human Evaluation Results As shown in Table 3, GTM can alleviate the problem of generating dull and deviated questions compared with other models (significance tests (Koehn, 2004), p-value < 0.05). Both our proposed model and the state-of-the-art model RL-CVAE utilize the answer information and the results of them could prove that answers assist the question generation process. Besides, GTM can produce more relevant and intriguing questions, which indicates the effectiveness of modeling the shared background and one-to-many mappings in CQG task. The interannotator agreement is calculated with the Fleiss’ kappa (Fleiss and Cohen, 1973). Fleiss’ kappa for Fluency, Coherence and Willingness is 0.493, 0.446 and 0.512, respectively, indicating “Moderate Agreement” for all three criteria. 3.6 Question-Answer Coherence Evaluation Automatic metrics in Section “Automatic Metrics” are designed to compare generated questions with ground-truth ones (RUBER also takes the post information into consideration), but ignore answers in the evaluation process. To measure the semantic coherence between generated questions and answers, we apply two methods (Wang et al., 2019): (1) Cosine Similarity: We use the pre-trained Infersent model7 (Conneau et al., 2017) to obtain sentence embeddings and calculate cosine similarity between the embeddings of generated responses 7The Infersent model is trained to predict the meaning of sentences based on natural language inference, and the cosine similarity computed with it is more consistent with human’s judgements, which performs better than the pre-trained Transformer/BERT model in our experiments. and answers. (2) Matching Score: We use the GRUMatchPyramid (Wang et al., 2019) model that adds the MatchPyramid network (Pang et al., 2016) on top of a bidirectional GRU to calculate the semantic coherence. As shown in Table 4, questions generated by GTM are more coherent to answers. Attributing to the design of triple-level latent variable that captures the shared background, one-to-many Model Fluency Coherence Willingness S2S-Attn 0.482 0.216 0.186 CVAE 0.462 0.484 0.428 kgCVAE 0.474 0.536 0.476 STD 0.488 0.356 0.286 HTD 0.526 0.504 0.414 RL-CVAE 0.534 0.578 0.508 GTM-zt 0.538 0.580 0.516 GTM-a 0.532 0.570 0.512 GTM-zq/za 0.542 0.586 0.520 GTM 0.548 0.608 0.526 Table 3: Results for human evaluation. Model Cosine Similarity Matching Score S2S-Attn 0.498 5.306 CVAE 0.564 8.047 kgCVAE 0.578 8.054 STD 0.542 6.879 HTD 0.583 8.059 RL-CVAE 0.607 8.423 GTM-zt 0.613 8.427 GTM-a 0.605 8.424 GTM-zq/za 0.618 8.472 GTM 0.629 8.517 Table 4: Evaluation results for QA coherence. mappings in PQ and QA pairs, and relationship modeling for zq and za, GTM can improve the relevance in QA pairs. 3.7 Case Study In Table 5, we list the generated results of two posts from the test set to compare the performance of different models. In the first case, both the post and answer mention two topics, “donation” and “song”, so the question is better to consider their relations. Besides, the answer here begins with “because”, then “why” and “what (reason)” questions are reasonable. For the second case, the post only talks about “pen”, while the answer refers to “ink”, which means there is a topic transition the question needs to cover. The second case shows the effectiveness of an answer that not only decides the expression of question but also improves the entire coherence of a tripe. Questions generated by GTM are more relevant to 3502 Post Question Answer S2S-Attn: what does that mean? CVAE: what’s the best way to do that? because i like the aspect of and it’s a donation to food kgCVAE: is there anything else to buy? song criticising may banks, even if you dislike STD: is it a good donation? reaching to the charts as the song. HTD: where are those food banks? well as the fact it goes to RL-CVAE: why do we need to do the stupid thing? charity. GTM: why not just donate money to food banks? S2S-Attn: what colors do you have? CVAE: are you colorblind? i have some coming, it’s a very blue pen. i’m kgCVAE: it has nothing to complain. where’s the pen? hopefully it will be here more into brighter colours, STD: what color are you using? early next week. right now but this is still pretty nice. HTD: what’s the colour scheme? it’s got green ink and it’s RL-CVAE: what kind of ink should i buy? really messing me up. GTM: i’ll take it though. do you also sell the ink? Table 5: Two cases comparison among GTM and other baselines. both posts and answers, and could attract people to give an answer to them. However, other baselines may generate dull or deviated responses, even the RL-CVAE model that considers the answer information would only contain the topic words in answers (e.g., the question in case two), but fail to ensure the PQA coherence. Figure 3: Total KL divergence (per word) of all latent variables in GTM and GTM-a model (first 30 epochs of validation set). 3.8 Further Analysis of GTM Variational models suffer from the notorious degeneration problem, where the decoders ignore latent variables and reduce to vanilla Seq2Seq models (Zhao et al., 2017; Park et al., 2018; Wang et al., 2019). Generally, KL divergence measures the amount of information encoded in a latent variable. In the extreme case where the KL divergence of latent variable z equals to zero, the model completely ignores z, i.e., it degenerates. Figure 3 shows that the total KL divergence of GTM model maintains around 2 after 18 epochs indicating that the degeneration problem does not exist in our model and latent variables can play their corresponding roles. 4 Related Work The researches on open-domain dialogue systems have developed rapidly (Majumder et al., 2020; Zhan et al., 2021; Shen et al., 2021), and our work mainly touches two fields: open-domain conversational question generation (CQG), and context modeling in dialogue systems. We introduce these two fields as follows and point out the main differences between our method and previous ones. 4.1 CQG Traditional question generation (TQG) has been widely studied and can be seen in reading comprehension (Zhou et al., 2019; Kim et al., 2019), sentence transformation (Vanderwende, 2008), question answering (Li et al., 2019; Nema et al., 2019), visual question generation (Fan et al., 2018) and task-oriented dialogues (Li et al., 2017). In such tasks, finding information via a generated question is the major goal and the answer is usually part of the input. Different from TQG, CQG aims to enhance the interactiveness and persistence of conversations (Wang et al., 2018). Meanwhile, the answer is the “future” information which means it is unavailable in the inference process. Wang et al. (2018) first studied on CQG, and they used soft and hard typed decoders to capture the distribution of different word types in a question. Hu et al. (2018) added a target aspect in the input and proposed an extended Seq2Seq model to generate aspect-specific questions. Wang et al. (2019) devised two methods based on either reinforcement learning or generative adversarial network (GAN) 3503 to further enhance semantic coherence between posts and questions under the guidance of answers. 4.2 Context Modeling in Dialogue Systems Existing methods mainly focus on the historical context in multi-turn conversations, and hierarchical models occupy a vital position in this field. Serban et al. (2016) proposed the hierarchical recurrent encoder-decoder (HRED) model with a context RNN to integrate historical information from utterance RNNs. To capture utterance-level variations, Serban et al. (2017) raised a new model Variational HRED (VHRED) that augments HRED with CVAEs. After that, VHCR (Park et al., 2018) added a conversation-level latent variable on top of the VHRED, while CSRR (Shen et al., 2019) used three-hierarchy latent variables to model the complex dependency among utterances. In order to detect relative utterances in context, Tian et al. (2017) and Zhang et al. (2018) applied cosine similarity and attention mechanism, respectively. HRAN (Xing et al., 2018) combined the attention results on both word-level and utterance-level. Besides, the future information has also been considered for context modeling. Shen et al. (2018) separated the context into history and future parts, and assumed that each of them conditioned on a latent variable is under a Gaussian distribution. Feng et al. (2020) used future utterances in the discriminator of a GAN, which is similar to Wang et al. (2019). The differences between our method and aforementioned ones in Section 4.1 and 4.2 are: (1) Rather than dividing PQA triples into two parts, i.e., PQ (history and current utterances) and QA (current and future utterances) pairs, we model the entire coherence by utilizing a latent variable to capture the share background in a triple. (2) Instead of regarding the relationship between question and answer as a text matching task that lacks the consideration of diversity, we incorporate utterance-level latent variables to help model one-to-many mappings in both PQ and QA pairs. 5 Conclusion We propose a generative triple-wise model for generating appropriate questions in open-domain conversations, named GTM. GTM models the entire background in a triple and one-to-many mappings in PQ and QA pairs simultaneously with latent variables in three hierarchies. It is trained in a onestage end-to-end framework without pre-training like the previous state-of-the-art model that also takes answer into consideration. Experimental results on a large-scale CQG dataset show that GTM can generate fluent, coherent, informative as well as intriguing questions. Acknowledgements We would like to thank all the reviewers for their insightful and valuable comments and suggestions. References Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015. Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural language processing with Python: analyzing text with the natural language toolkit. ” O’Reilly Media, Inc.”. Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew Dai, Rafal Jozefowicz, and Samy Bengio. 2016. Generating sentences from a continuous space. In Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning, pages 10–21. Chaotao Chen, Jinhua Peng, Fan Wang, Jun Xu, and Hua Wu. 2019. Generating multiple diverse responses with multi-mapping and posterior mapping selection. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 4918–4924. AAAI Press. Kyunghyun Cho, Bart van Merri¨enboer Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, pages 1724–1734. Alexis Conneau, Douwe Kiela, Holger Schwenk, Lo¨ıc Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 670–680, Copenhagen, Denmark. Association for Computational Linguistics. Zhihao Fan, Zhongyu Wei, Piji Li, Yanyan Lan, and Xuanjing Huang. 2018. A question type driven framework to diversify visual question generation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4048–4054. AAAI Press. 3504 Shaoxiong Feng, Hongshen Chen, Kan Li, and Dawei Yin. 2020. Posterior-gan: Towards informative and coherent response generation with posterior generative adversarial network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7708–7715. Joseph L Fleiss and Jacob Cohen. 1973. The equivalence of weighted kappa and the intraclass correlation coefficient as measures of reliability. Educational and psychological measurement, 33(3):613– 619. Xiang Gao, Sungjin Lee, Yizhe Zhang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2019. Jointly optimizing diversity and relevance in neural response generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1229–1238, Minneapolis, Minnesota. Association for Computational Linguistics. Wenpeng Hu, Bing Liu, Jinwen Ma, Dongyan Zhao, and Rui Yan. 2018. Aspect-based question generation. Yanghoon Kim, Hwanhee Lee, Joongbo Shin, and Kyomin Jung. 2019. Improving neural question generation using answer separation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6602–6609. Diederick P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In the 3rd International Conference on Learning Representations. Diederik P Kingma and Max Welling. 2014. Autoencoding variational bayes. In the 2nd International Conference on Learning Representations. Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388– 395, Barcelona, Spain. Association for Computational Linguistics. Dong Bok Lee, Seanie Lee, Woo Tae Jeong, Donghwan Kim, and Sung Ju Hwang. 2020. Generating diverse and consistent QA pairs from contexts with information-maximizing hierarchical conditional VAEs. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 208–224, Online. Association for Computational Linguistics. Jingjing Li, Yifan Gao, Lidong Bing, Irwin King, and Michael R. Lyu. 2019. Improving question generation with to the point context. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3216–3226, Hong Kong, China. Association for Computational Linguistics. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016a. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California. Association for Computational Linguistics. Jiwei Li, Michel Galley, Chris Brockett, Georgios Spithourakis, Jianfeng Gao, and Bill Dolan. 2016b. A persona-based neural conversation model. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 994–1003, Berlin, Germany. Association for Computational Linguistics. Jiwei Li, Alexander H Miller, Sumit Chopra, Marc’Aurelio Ranzato, and Jason Weston. 2017. Learning through dialogue interactions by asking questions. ICLR. Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2122–2132, Austin, Texas. Association for Computational Linguistics. Bodhisattwa Prasad Majumder, Harsh Jhamtani, Taylor Berg-Kirkpatrick, and Julian McAuley. 2020. Like hiking? you probably enjoy nature: Personagrounded dialog with commonsense expansions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9194–9206, Online. Association for Computational Linguistics. Preksha Nema, Akash Kumar Mohankumar, Mitesh M. Khapra, Balaji Vasan Srinivasan, and Balaraman Ravindran. 2019. Let’s ask again: Refine network for automatic question generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3314–3323, Hong Kong, China. Association for Computational Linguistics. Liang Pang, Yanyan Lan, Jiafeng Guo, Jun Xu, Shengxian Wan, and Xueqi Cheng. 2016. Text matching as image recognition. In Thirtieth AAAI Conference on Artificial Intelligence. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. 3505 Yookoon Park, Jaemin Cho, and Gunhee Kim. 2018. A hierarchical latent structure for variational conversation modeling. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1792–1801, New Orleans, Louisiana. Association for Computational Linguistics. Iulian V Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pages 3776–3783. Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron Courville, and Yoshua Bengio. 2017. A hierarchical latent variable encoder-decoder model for generating dialogues. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, pages 3295–3301. Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. Neural responding machine for short-text conversation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1577–1586, Beijing, China. Association for Computational Linguistics. Lei Shen and Yang Feng. 2020. CDL: Curriculum dual learning for emotion-controllable response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 556–566, Online. Association for Computational Linguistics. Lei Shen, Yang Feng, and Haolan Zhan. 2019. Modeling semantic relationship in multi-turn conversations with hierarchical latent variables. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5497–5502, Florence, Italy. Association for Computational Linguistics. Lei Shen, Haolan Zhan, Xin Shen, and Yang Feng. 2021. Learning to select context in a hierarchical and global perspective for open-domain dialogue generation. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7438–7442. IEEE. Xiaoyu Shen, Hui Su, Wenjie Li, and Dietrich Klakow. 2018. NEXUS network: Connecting the preceding and the following in dialogue generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4316– 4327, Brussels, Belgium. Association for Computational Linguistics. Chongyang Tao, Lili Mou, Dongyan Zhao, and Rui Yan. 2018. Ruber: An unsupervised method for automatic evaluation of open-domain dialog systems. In Thirty-Second AAAI Conference on Artificial Intelligence. Zhiliang Tian, Rui Yan, Lili Mou, Yiping Song, Yansong Feng, and Dongyan Zhao. 2017. How to make context more useful? an empirical study on contextaware neural conversational models. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 231–236, Vancouver, Canada. Association for Computational Linguistics. Lucy Vanderwende. 2008. The importance of being important: Question generation. In Proceedings of the 1st Workshop on the Question Generation Shared Task Evaluation Challenge, Arlington, VA. Weichao Wang, Shi Feng, Daling Wang, and Yifei Zhang. 2019. Answer-guided and semantic coherent question generation in open-domain conversation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5066– 5076, Hong Kong, China. Association for Computational Linguistics. Yansen Wang, Chenyi Liu, Minlie Huang, and Liqiang Nie. 2018. Learning to ask questions in opendomain conversational systems with typed decoders. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2193–2203, Melbourne, Australia. Association for Computational Linguistics. Chen Xing, Yu Wu, Wei Wu, Yalou Huang, and Ming Zhou. 2018. Hierarchical recurrent attention network for response generation. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, pages 5610–5617. Zhou Yu, Ziyu Xu, Alan W Black, and Alexander Rudnicky. 2016. Strategy and policy learning for nontask-oriented conversational systems. In Proceedings of the 17th annual meeting of the special interest group on discourse and dialogue, pages 404– 412. Haolan Zhan, Hainan Zhang, Hongshen Chen, Zhuoye Ding, Yongjun Bao, and Yanyan Lan. 2021. Augmenting knowledge-grounded conversations with sequential knowledge transition. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5621–5630, Online. Association for Computational Linguistics. Weinan Zhang, Yiming Cui, Yifa Wang, Qingfu Zhu, Lingzhi Li, Lianqiang Zhou, and Ting Liu. 2018. Context-sensitive generation of open-domain conversational responses. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2437–2447. Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics 3506 (Volume 1: Long Papers), pages 654–664, Vancouver, Canada. Association for Computational Linguistics. Hao Zhou, Minlie Huang, Tianyang Zhang, Xiaoyan Zhu, and Bing Liu. 2018. Emotional chatting machine: Emotional conversation generation with internal and external memory. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32. Wenjie Zhou, Minghua Zhang, and Yunfang Wu. 2019. Question-type driven question generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6032–6037, Hong Kong, China. Association for Computational Linguistics.
|
2021
|
271
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3507–3520 August 1–6, 2021. ©2021 Association for Computational Linguistics 3507 Diversifying Dialog Generation via Adaptive Label Smoothing Yida Wang1,2∗, Yinhe Zheng1,3∗, Yong Jiang2, Minlie Huang1 † 1 The CoAI group, DCST, Institute for Artificial Intelligence, State Key Lab of Intelligent Technology and Systems, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China 2 Tsinghua-Berkeley Shenzhen Institute, Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China 3 Samsung Research China - Beijing (SRC-B) [email protected], [email protected], [email protected], [email protected] Abstract Neural dialogue generation models trained with the one-hot target distribution suffer from the over-confidence issue, which leads to poor generation diversity as widely reported in the literature. Although existing approaches such as label smoothing can alleviate this issue, they fail to adapt to diverse dialog contexts. In this paper, we propose an Adaptive Label Smoothing (AdaLabel) approach that can adaptively estimate a target label distribution at each time step for different contexts. The maximum probability in the predicted distribution is used to modify the soft target distribution produced by a novel light-weight bi-directional decoder module. The resulting target distribution is aware of both previous and future contexts and is adjusted to avoid over-training the dialogue model. Our model can be trained in an endto-end manner. Extensive experiments on two benchmark datasets show that our approach outperforms various competitive baselines in producing diverse responses. 1 Introduction The success of neural models has greatly advanced the research of dialog generation (Huang et al., 2020; Wang et al., 2020; Zhang et al., 2020). However, most of these models suffer from a lowdiversity issue where models tend to generate bland and generic responses such as I don’t know or I’m OK (Li et al., 2016). Although various approaches have been proposed to tackle this issue (Li et al., 2016; Zhao et al., 2017; Du et al., 2018; Zhou et al., 2018; Welleck et al., 2020; Zheng et al., 2020b), there are still remarkable gaps between responses generated by neural models and those from humans (Holtzman et al., 2020). Further, some existing methods may even harm the fluency or coherence when improving the diversity of generated ∗Equal contribution † Corresponding Author: [email protected] So, what exactly do you do around here ? I make the robots seem more ___ Post: Response: human 0.9 1.0 0.61 0.01 0.0 0.01 bank 0.01 0.0 0.01 fights 0.01 0.0 0.10 ugly 0.01 0.0 0.08 dull 0.01 0.0 0.11 fun … Hard Target (One hot) Label Smoothing AdaLabel (Ours) Figure 1: A dialogue sampled from the OpenSubtitles dataset. We demonstrate the hard target, label smoothing, and Adaptive Label Smoothing approach when learning to predict the next word (“human”). responses. (Ippolito et al., 2019; Massarelli et al., 2020; Zheng et al., 2020a). Recently, Jiang and de Rijke (2018); Jiang et al. (2019) show that there is a strong connection between the low-diversity problem and the overconfidence issue. i.e., over-confident dialogue models tend to produce low-diversity responses. One of the reasons can be attributed to the supervision target. Specifically, training a dialogue generation model with the Maximum Likelihood Estimation (MLE) objective under the hard target (i.e., one-hot distribution as ground truth) makes the model favor high-frequency tokens and produce over-confident probability estimation (Gowda and May, 2020), which ultimately leads to poor calibration (Mukhoti et al., 2020), and thus low diversity (Jiang et al., 2019). Hinton et al. (2015) and Yang et al. (2018) suggest that the ideal training target should be a soft target that assigns probability mass on multiple valid candidates (see Figure 1). With such a soft target, the over-confidence issue can be alleviated (M¨uller et al., 2019), and thus the diversity of the output responses can be improved. Unfortunately, the ideal soft target is challenging to obtain. Early works try to tackle this issue 3508 using label smoothing (Szegedy et al., 2016), i.e., a small probability is uniformly assigned to nontarget words. However, the target distribution constructed in this way is far from ideal: First, the probability of the target word is chosen manually and fixed, which cannot adapt to different contexts. However, as Holtzman et al. (2020) demonstrated, human text distribution exhibits remarkable fluctuations in the per-token perplexity. We argue that different target probabilities should be used for different contexts. Second, the uniform assignment of the probability mass on non-target words ignores the semantic relationship between the context and each word. Ideally, a word should receive more probability mass if it is more relevant to the context. For the example shown in Figure 1, word “fun” is more likely to appear behind the context “I make the robots seem more ” than word “bank”. To address the above issue, we propose an Adaptive Label smoothing (AdaLabel) method that can dynamically estimate a soft target distribution at each time step for different contexts. Specifically, for each target word yt in the training data, the probability distribution predicted by the current model is first obtained. The maximum probability pmax in this distribution measures the confidence of the current prediction, i.e., a higher pmax means higher confidence for the current prediction. To avoid over-confidence, we use pmax as the supervision signal for the target word yt in the training process so that the model will not be optimized towards yt when it correctly predicts yt. A word-level factor is also introduced to facilitate the learning of low-frequency words. Moreover, we introduce a novel auxiliary decoder module Da to produce the supervision signals for these non-target words in each training step. Da only contains one transformer block, and it is optimized to predict words based on bi-directional contexts. A novel Target-Mask attention scheme is devised to prevent Da from seeing the target word in the training process. This scheme also enables parallel training and inference of Da. We perform extensive experiments on two benchmark datasets: DailyDialog and OpenSubtitles. Our method outperforms various competitive baselines and significantly improves the diversity of generated responses while ensuring fluency and coherency. Our major contributions are summarized: 1. We propose AdaLabel, a method that can produce a soft target distribution considering the current context and the model’s confidence. Specifically, AdaLabel ensures that the dialogue model will not be optimized toward the target word yt if yt has been correctly predicted. This prevents our model from being over-confident. 2. We introduce a light-weight bi-directional decoder that can produce context-aware supervision signals for non-target words. A novel Target-Mask attention scheme is devised to facilitate the parallel training and inference of this decoder. 3. Extensive experiments on two benchmark dialogue datasets with both automatic and human evaluation results show that our method helps to alleviate the model over-confident issue and significantly improves the model’s diversity. 2 Related work Diversity Promotion: Existing approaches for solving the low diversity issue of neural dialogue models generally involve two categories: The first category is training-based, where new training objectives are designed (Li et al., 2016; Zhang et al., 2018; Gao et al., 2019) or latent variables are introduced (Zhao et al., 2017; Zhou et al., 2018) in the dialogue model. Some methods also try to refine the training target used in the MLE loss (Choi et al., 2020; Jiang et al., 2019; Li et al., 2019), or directly penalize the trivial responses with auxiliary loss terms (Welleck et al., 2020; Li et al., 2020). Unlike these existing approaches, our method tries to adaptively adjust the training target by utilizing the current predictions. The second category is decoding-based, in which different heuristic decoding rules are designed (Holtzman et al., 2020; Kulikov et al., 2019). Note that these decoding techniques are independent of the model setting, and our method can be used in combination with these techniques. Confidence Calibration: Modern deep neural networks suffer from the over-confidence issue (Guo et al., 2017; Kumar and Sarawagi, 2019), and various remedies are proposed (Pereyra et al., 2017; Mukhoti et al., 2020; Lin et al., 2017). Following the work of Jiang and de Rijke (2018); Jiang et al. (2019), our method is proposed to tackle the overconfidence issue to improve the diversity of the generated responses. However, different from existing approaches, our method enables more flexible controls over the target distribution. Knowledge Distillation: Another important technique similar to our work is knowledge distilla3509 Encoder Decoder Context 𝑋 Auxiliary Decoder 𝒟 Training Response 𝜖 ൈሺ1 െ𝜖ሻ ൈ𝜖 𝐵𝑂𝑆 𝒗𝑦ଵ 𝑦ଵ 𝒗𝑦ଶ 𝑦ଶ 𝒗𝑦ଷ 𝑦் 𝒗ሾ𝐸𝑂𝑆ሿ 𝑦ଷ 𝒗𝑦ସ … … 𝐵𝑂𝑆 𝑦ଵ 𝑦ଵ 𝑦ଶ 𝑦ଶ 𝑦ଷ 𝛼 𝑝ሺ𝑦ଷሻ Auxiliary Distribution 𝒗 𝑝ሺ𝑦ଷሻ Hard Target 𝒒 𝑝ሺ𝑦ଷሻ Adaptive Soft Target 𝒒ᇱ ℒሺ𝒒ᇱ, 𝒑ሻ Partial Response Predicted Distribution 𝒑 𝑝௫ 𝑝ሺ𝑦ଷሻ Figure 2: Overview of constructing the adaptive soft target q′ using AdaLabel: The maximum probability pmax in the predicted distribution p is used to obtain an adaption factor ϵ, which is further used to combine the hard target q and the auxiliary distribution v to obtain q′. A bi-directional auxiliary decoder Da is used to produce v. tion, in which a learned teacher model is distilled to a student model by minimizing a KL term (Hinton et al., 2015; Kim and Rush, 2016). The most related work comparing to ours is the C-MLM approach (Chen et al., 2020), in which a BERT model is fine-tuned to be a teacher. Our approach and C-MLM’s primary difference is that our auxiliary decoder Da is a one layer module that is jointly trained with the dialogue model. However, the BERT teacher in C-MLM contains much more parameters, and it is trained using an expensive pretrained and then fine-tuned process. Moreover, the target-masked attention scheme in Da enables parallel inferences of v for each training sequence Y . In contrast, multiple independent forward passes are required for the BERT teacher. 3 Method 3.1 Background: MLE with Hard Target The goal of generative dialogue modeling is to learn a conditional probability distribution p(Y |X), where X is the dialogue context, Y = y1, ..., yT is a response word sequence, and yi ∈V is a word from the vocabulary V. In an auto-regressive manner, p(Y |X) is factorized as Q t p(yt|y<t, X). For each target word yt in the training sequence Y , a conventional MLE training approach try to optimize the following cross entropy loss: L(q, p) = − X wk∈V qklog [p(wk|y<t, X)] , (1) where q is a one-hot distribution (i.e., a hard target) that assigns a probability of 1 for the target word yt and 0 otherwise, i.e., qk = 1 only when wk = yt. For simplicity of notation, we abbreviate the dependency of yt in the notation of each distribution in our paper, i.e., different target word yt in Y corresponds to different values of q and p. 3.2 Method Overview We propose to adaptively construct a soft target distribution q′ to replace q in Eq. 1. Specifically, q′ = ε · q + (1 −ε) · v, (2) where ε ∈[0, 1] is an adaption factor, and v is an auxiliary distribution vector that depends on the current time step. (see Figure 2 for an overview). In this study, we constrain v to assign zero probability for the target word yt and non-zero probabilities for these non-target words V̸=yt = {yi|yi ∈ V, yi ̸= yt}. This constraint allows us to explicitly control the supervisions assigned to yt. Specifically, the first term ε · q and the second term (1 −ε) · v in Eq. 2 respectively determines how much probability q′ assigns to yt and V̸=yt. This setting differs from conventional knowledge distillation (Kim and Rush, 2016) because it facilitates more flexible controls over q′, so that we can use the factor ε to determine the supervision signal provided for the target word yt. The following sections detail how to compute ε and v. 3.3 Target Word Probability We control the probability of the target word yt in p′ by manipulating the adaption factor ε in Eq. 2. Specifically, for a training dialogue pair ⟨X, Y ⟩and each target word yt ∈Y , the current distribution p(·|y<t, X) is first calculated, and the maximum probability in this distribution is obtained: pmax = max wk∈V p(wk|y<t, X). (3) 3510 ε is then obtained: ε = max(pmax, λ), (4) where λ serves as a lower-bound of ε (i.e., ε ≥λ). The basic intuition behind Eq. 4 is to set ε = pmax when pmax is reasonably large. This design prevents our model from receiving supervisions sharper than pmax, when the current prediction is confidence enough. Further, to ensure that the target word yt always receives the largest probability in q′, i.e., to ensure ε > (1 −ε) · max(v) (see Eq. 2), in which max(v) is the maximum probabilities for non-target words V̸=yt, we have to enforce ε > max(v) 1+max(v). Thus we propose to calculate the lower-bound λ of ε as: λ = max(v) 1 + max(v) + η, (5) where η > 0 is a hyper-parameter that controls the margin between the probability of the target word and non-target words in p′. To facilitate faster converge and better learning of low-probability words, an empirical factor α ∈ [0, 1] is further introduced to adjust the calculation of ε on the basis of Eq. 4: ε = 1 −α · (1 −max(pmax, λ)), (6) where α is calculated as the relative ratio to pmax: α = p(yt|y<t, X) pmax 2 , (7) where p(yt|y<t, X) is the probability for the target word yt. Note that Eq. 6 and Eq. 4 is equivalent if α = 1. Intuitively, α accelerates the training of lowfrequency words because if yt is of low-frequency in the corpus, then yt is usually under-trained and thus p(yt|y<t, X) is generally small. This leads to a small α and thus increases the probability for yt in p′. Note that ε, λ and α are all time-step specific variables, whereas η is a fixed hyper-parameter. This allows the values adapt to dynamic contexts. In our experiments, Eq. 6 is used to calculate ε. 3.4 Non-target Words Probabilities The auxiliary distribution v in Eq. 2 is calculated using an auxiliary decoder Da, which is a singlelayer transformer-based decoder that is jointly optimized with the generation model. Figure 3 shows the structure of Da, in which a novel target-masked 𝐵𝑂𝑆 𝒗𝑦ଵ 𝑦ଵ 𝒗𝑦ଶ 𝑦ଶ 𝒗𝑦ଷ 𝒗[𝐸𝑂𝑆] 𝑦ଷ 𝒗𝑦ସ 𝑄 𝐾, 𝑉 𝑦ସ Target-Masked Attention (b) (c) Feed Forward Target-Masked Multi-head Attention Add & Norm Multi-head Attention Add & Norm Encoder Outputs Add & Norm 1× (a) 𝑄 𝐾, 𝑉 Figure 3: (a) The auxiliary decoder Da; (b) The targetmasked attention scheme used to compute the auxiliary distribution v for the target word y3, specifically, y2 is used as the query and y3 is masked; (c) The attention pattern used in the target-masked attention scheme, white dots represent masked positions. attention scheme is devised to mask each target word yt in the self attention module of the decoder when calculating the corresponding v (see Figure 3b and 3c). In this way, bi-directional contexts can be utilized when predicting the auxiliary distribution v for yt. Moreover, it is important to use only one decoder layer in Da because stacking multiple layers in Da leaks the information of yt to v. Note that using one layer in Da does not necessarily downgrade its performance (Kasai et al., 2021). Our experiment results in Section 5.1 indicate that with the help of bi-directional contexts, the accuracy of Da largely outperforms the unidirectional dialogue decoder that is much deeper than Da. Moreover, for a training response Y , the structure of Da enables us infer the auxiliary distribution in parallel for all the target words in Y within a single forward pass. This differs from the BERT teacher used by Chen et al. (2020), in which multiple independent forward passes are needed to get the teacher distributions for all the words in Y . When training Da, the following standard MLE loss is optimized for each target word yt: L(q, v) = − |V| X k=1 qklogvk, (8) in which the notation of qk follows Eq. 1. The outputs of Da are used as the logits to infer v to be further used in Eq. 2. Specifically, the logit of the target word yt is masked to −∞before Softmax to ensure yt always receives zero probability in v. Moreover, we also follow the approach used by Tang et al. (2020) to truncate the head and tail of the remaining logits before inferring v in Eq. 3511 Train Valid Test DailyDialog 65.8K 6.13K 5.80K OpenSubtitles 1.14M 20.0K 10.0K Table 1: Dataset statistics. 2, i.e., all the logits are ranked in a descending order and only the logits ranked from n to m are kept while the rest logits are masked to −∞. This masks the head and tail probabilities in v to zero. We argue that truncating the tail probabilities of v filters noises, and truncating the head probabilities of v encourages the dialogue model to focus more on low-probability words. In our experiments, we set n = 2 and m = 500. An extensive hyperparameter search indicates that our method is not sensitive to the value of n and m. There are two major differences between our auxiliary decoder Da and the teacher model used in conventional knowledge distillation approaches: First, conventional teacher models usually carry more parameters than their students, whereas Da is rather light-weight. Second, conventional teacher models are typically pre-trained before being utilized in the distillation process, whereas Da is trained jointly with our dialogue model. 4 Experiments 4.1 Dataset We use two benchmark datasets for open-domain dialogue generation: DailyDialog (Li et al., 2017) is a high-quality multi-turn dialogue dataset that is collected from daily conversations. OpenSubtitles 1 contains dialogues collected from movie subtitles. Moreover, we follow Li et al. (2016) and Jiang et al. (2019) to focus on short conversations, i.e., dialogues with posts or responses longer than 100 tokens are removed. See Table 1 for more details. 4.2 Implementation Details The backbone of our model is the transformerbased sequence to sequence model (Vaswani et al., 2017), and most hyper-parameters follow Cai et al. (2020). Specifically, the encoder and decoder each contains 6 layers. Each layer has 8 attention heads, and the hidden size is set to 512. The auxiliary decoder Da follows the same hyper-parameter setting as the dialogue decoder, but it only contains one layer. The WordPiece tokenizer provided by 1http://opus.nlpl.eu/OpenSubtitles.php BERT (Devlin et al., 2019) is used, and the Adam optimizer (Kingma and Ba, 2015) is employed to train our model from random initializations with a learning rate of 1e-4. η in Eq. 5 is set to 0.2 for all datasets. See Appendix A for more details. 2 4.3 Baselines We compared our method with two groups of baselines that try to tackle the over-confidence issue. The first group modifies the training target used to compute the loss function: 1) LS (Szegedy et al., 2016): uses the label smoothing approach to construct a target distribution by adding the onehot target and a uniform distribution; 2) FL (Lin et al., 2017): uses the focal loss to down-weigh well-classified tokens in each time step. 3) FACE (Jiang et al., 2019): uses the frequency-aware crossentropy loss to balance per-token training losses. Specifically, relative low losses are assigned to high-frequency words to explicitly tackle the overconfidence issue. We used the best performing “Pre-weigh” version in our experiments. 4) F2 (Choi et al., 2020): factorizes the target distribution based on the token frequencies. The second group of baselines add some penalty term to the standard MLE loss: 5) CP (Pereyra et al., 2017): a confidence penalty term is added to regularize the entropy of the model, so that over-confident predictions are penalized; 6) UL (Welleck et al., 2020): an unlikelihood loss term is added to penalize the frequently generated words. 7) NL (He and Glass, 2020): works similarly with baseline UL except a negative loss term is used instead of the unlikelihood loss term. 8) D2GPo (Li et al., 2019): augments the MLE loss with a data-dependent gaussian prior objective to assign different losses for different non-target words. We also compared to: 9) CE: a vanilla Seq2Seq model trained with the cross-entropy loss. For fair comparisons, the C-MLM model proposed by Chen et al. (2020) is not used as our baseline since the BERT teacher in C-MLM requires a large amount of extra data to pre-train. Nevertheless, AdaLabel still surpasses C-MLM on various metrics (see Appendix F for more analysis). All our baselines are adapted from the authors’ official codes with the same backbone architecture and hyper-parameters as our model (see details in Appendix B). Following the original setting, a train2Our code is available at: https://github.com/ lemon234071/AdaLabel 3512 Model DailyDialog OpenSubtitles Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 CE 1.67 9.43 4.53 6.59 2.99 7.56 4.38 2.61 2.55 9.87 4.13 5.58 0.84 7.60 4.30 2.57 LS 1.48 8.78 4.48 6.55 2.44 7.98 4.68 2.86 2.77 13.08 4.45 6.57 0.51 8.91 5.57 3.84 FL 2.38 13.42 4.7 7.04 5.05 9.74 6.12 4.11 3.19 13.16 4.42 6.50 1.04 8.06 4.79 3.08 FACE 1.62 11.04 4.96 7.27 4.11 8.78 5.06 3.06 3.31 14.06 4.77 7.05 1.33 7.69 4.40 2.70 F2 1.40 7.91 4.35 6.28 2.32 7.78 4.45 2.60 2.89 11.40 4.24 6.14 0.99 7.52 4.30 2.62 CP 2.35 12.91 4.64 6.89 4.07 9.06 5.68 3.79 3.11 12.72 4.36 6.35 0.98 8.06 4.82 3.12 UL 2.35 12.99 4.68 6.98 4.96 10.83 6.87 4.61 2.84 11.64 4.31 6.32 0.76 7.73 4.59 2.96 NL 1.66 9.18 4.47 6.58 4.30 9.83 5.83 3.60 3.24 12.98 4.42 6.49 1.08 7.56 4.38 2.71 D2GPo 1.26 8.06 4.43 6.48 2.20 8.30 4.82 2.93 2.07 11.01 4.32 6.36 0.19 8.41 5.08 3.35 AdaLabel 3.96 23.53 5.17 8.00 8.49 17.42 13.38 11.01 4.78 22.88 4.96 7.66 1.47 9.80 6.48 4.75 Human 6.59 37.74 5.67 8.91 13.7 N/A N/A N/A 8.62 43.16 5.89 9.36 4.75 N/A N/A N/A Table 2: Automatic evaluation results (%). Best results among all the models are in bold. and-refine strategy is used in baseline 3, 6, and 7, i.e., these baselines are refined based on CE. We follow the setting of Jiang et al. (2019) to use deterministic decoding scheme (particularly, greedy decoding) for our model and all baselines. Note that our method can be adapted to other decoding schemes such as beam-search or top-K sampling. See Appendix C for more detailed analysis. 4.4 Automatic Evaluation Metrics: We first used automatic metrics to evaluate our method: 1) Distinct (Dist) (Li et al., 2016) calculates the proportion of unique n-grams (n=1, 2) in the generated responses, which is widely used to measure the response diversity. 2) Entropy (Ent) (Zhang et al., 2018) evaluates how evenly the empirical n-gram (n=1, 2) distribution is. Higher sores mean more diverse of the response. 3) LowFrequency Token Ratio (LF) (Li et al., 2019) further measures the model diversity by counting the ratio of low-frequency words in the generated responses. We chose words with a frequency less than 100 in each corpus as low-frequency words. Over-confident models tend to omit low-frequency words (i.e., get low LF scores) and yield less diversified responses. 4) BLEU (Papineni et al., 2002) measures n-gram (n=2, 3, 4) overlap between the generated responses and references. Results: As shown in Table 2, our method AdaLabel outperforms all the baselines by large margins on all the datasets. We can further observe that: 1) AdaLabel achieves the best diversity scores (Dist-1,2, Ent-1,2, and LF). This indicates that our method yields better training targets that help to produce more diverse responses; 2). The models that explicitly tackle the over-confidence issue (i.e., AdaLabel and FACE) generally outperform other baselines in diversity-related metrics. For example, FACE obtains the second-best diversity scores (i.e., Dist, Ent, and LF) on the OpenSubtitles dataset. This verifies our motivation that alleviating the over-confidence issue helps to produce more diverse responses. Note that our method also outperforms all the baselines using the stochastic decoding scheme. Please refer to Appendix C for more details. 4.5 Manual Evaluation Metrics: Pairwise manual evaluations are conducted to further validate our method. Specifically, for a given dialogue post, our model’s response is paired with the one from a baseline. Three individual annotators were employed to rank each response pair from three aspects: 1) Fluency (Flu.): which response is more fluent; 2) Coherency (Coh.): which response is more coherent to the context; 3) Informativeness (Info.): which response contains more informative content. We also asked the annotator to choose an overall preferred response (Pref.). Ties were allowed. Results: 200 posts were randomly sampled from each of these two datasets, respectively, and totally 3.6K response pairs were generated. The inter-rater annotation agreement was measured using Fleiss’s kappa κ (Fleiss, 1971). Particularly, the κ value on DailyDialog, OpenSubtitles dataset was 0.59 and 0.55, respectively, indicating moderate agreement. As shown in Table 3, AdaLabel outperforms all the baselines on the informativeness measure. This means that our method can respond with more informative content. We can further observe that: 1). All models achieve competitive fluency because it is easy for neural models to produce fluent responses by yielding trivial responses like “I 3513 Comparison DailyDialog OpenSubtitles Pref. Flu. Coh. Info. Pref. Flu. Coh. Info. AdaLabel vs CE 17.00‡ 1.33 12.5‡ 28.33‡ 6.33 1.17 7.33† 13.67‡ AdaLabel vs LS 2.67 0.17 3.33 24.83‡ 5.3 -0.67 3.17 8.50‡ AdaLabel vs FL 4.50 1.67 7.00† 22.0‡ 8.00† 1.00 6.00 5.50 AdaLabel vs FACE 6.67† 3.50† 7.17† 8.50† 4.50 0.50 1.83 2.50 AdaLabel vs F2 7.67† 0.33 6.83† 8.67‡ 4.33 -0.50 1.67 9.50‡ AdaLabel vs CP 10.50‡ -0.17 8.00† 23.83‡ 8.00† 1.50 6.17 16.83‡ AdaLabel vs UL 7.83† 0.83 6.67† 17.33‡ 6.83† 2.00 5.83 15.00‡ AdaLabel vs NL 9.17† 2.67† 9.17† 7.67† 5.17 0.17 2.17 15.5‡ AdaLabel vs D2GPo 0.83 0.00 3.33 15.17‡ 3.17 7.33‡ 1.00 6.33† Table 3: Pairwise human evaluation results (%). The absolute gains of AdaLabel (i.e., Win rate −Lose rate) are reported. †, ‡ indicates significant improvement with p-value < 0.05 and < 0.005, respectively (sign test). Model BLEU-3,4 Dist-1,2 Ent-1,2 LF 1.w/o ε 5.46 3.57 2.52 13.21 4.64 6.89 4.85 2.w/o α 11.35 8.70 3.62 20.56 5.02 7.70 7.30 3.Orig. v 8.15 5.77 3.71 19.53 5.00 7.58 8.25 4.Uniform 5.66 3.61 2.24 14.96 4.84 7.33 4.98 5.Rand 6.27 4.07 2.03 13.47 4.7 7.08 4.56 6.BERT 11.6 9.34 3.67 20.97 5.02 7.71 7.28 AdaLabel 13.38 11.01 3.96 23.53 5.17 8.00 8.49 Table 4: Ablation study results on DailyDialog (%). don’t know”. However, our model surpasses most baselines in terms of fluency while ensuring high diversity scores. This demonstrates the superiority of our method in producing high quality responses. 2). AdaLabel produces more coherent responses comparing to most baselines. This verifies that our model does not sacrifice the response quality when achieving high diversity scores. In fact, by controlling the model’s confidence, more lowfrequency words are encouraged, and thus AdaLabel can produce more relevant and coherent responses. This claim is further verified by observing that our model achieves the best overall preference score among all the baselines. 4.6 Ablation study Ablation studies were performed to verify the effect of each component in our method. Specifically, two groups of variants were tested: The first group validates the effectiveness of the calculated target word probability, i.e., ε: 1). w/o ε directly sets a fixed value for ε in Eq. 2. The specific value of ε is searched from 0.1 to 0.7 with a stride of 0.1; 2). w/o α omits the empirical factor α in calculating ε, i.e., the value of ε in Eq. 2 is calculated using Eq. 4 in instead of Eq. 6. The second group validates the effectiveness of the non-target word probabilities produced by Da, i.e., v: 3). Orig. v does not truncate the head of v when inferring from Da. Note that the truncation for the tail of v is still applied since its effectiveness has already been proved in previous studies (Tang et al., 2020; Tan et al., 2019); 4). Uniform uses an uniform distribution as v in Eq. 2. Note that different from the baseline LS, the value of ε is calculated using Eq. 6 in this ablation model, whereas the value of ε in the baseline LS is fixed ; 5). Rand use a random distributions as v in Eq. 2; 6). BERT follows the work of Chen et al. (2020) to fine-tune a pre-trained BERT model to produce v. Note that our dialogue model may benefit from the multi-task training of Da since Da shares the same encoder with our dialogue model. Optimizing Eq. 8 may help the encoder to capture better features. For fair comparison, we kept the task of optimizing Da in ablation models 4-6 although it is not used to infer v. Table 4 shows the results of ablation models on the DailyDialog dataset. As can be seen from the first two rows, our method to adaptively calculate ε helps to improve the performance of our model by a large margin, and the empirical adjustment factor α helps to further improve our performance by facilitating the learning of low-probability words. The performance of ablation models 3-6 in Table 4 proves that v captures reliable distribution and helps our model produce more diverse responses. Moreover, truncating the head distribution of v enables the dialogue model to focus more on the low-frequency words and thus facilitates more informative responses. It is also interesting to note that our auxiliary decoder Da surpasses the BERT teacher used by Chen et al. (2020) in helping the dialogue model 3514 DailyDialog OpenSubtitles Auxiliary Decoder Da 64.03 64.92 Dialog Decoder in AdaLabel 44.16 43.90 Dialog Decoder in CE 38.58 41.57 Table 5: Prediction accuracy of decoders on test sets. 0.00 0.25 0.50 0.75 Confidence Score 0 1 2 3 Density (%) (a) AdaLabel 0.00 0.25 0.50 0.75 1.00 Confidence Score 0 1 2 Density (%) (b) CE 0.00 0.25 0.50 0.75 1.00 Confidence Score 0 1 2 Density (%) (c) LS 0.00 0.25 0.50 0.75 1.00 Confidence Score 0 1 2 Density (%) (d) FACE Figure 4: Empirical distribution of confidence scores for high-frequency words on the OpenSubtitles dataset. Words occupying the top 40% of the frequency mass in the training set are regarded as high-frequency words. to produce more diverse responses. This further proves the effectiveness of Da considering that BERT contains 6 times parameters than Da and consumes much more computation resources. 5 Discussion 5.1 Auxiliary Decoder To further test the performance of Da, we evaluated the averaged accuracy score of Da when predicting each target word in the test set (first row in Table 5). Specifically, a target word yt in the reference response is determined to be correctly predicted if it is top-ranked in the predicted distribution p(·|y<t, X). A better decoder is generally believed to obtain a higher accuracy. Table 5 also reports the uni-directional dialogue decoders’ accuracy in AdaLabel and CE. It can be seen that Da can make substantially more accurate predictions with the help of modeling bi-directional contexts using only one layer. Moreover, the dialogue model’s decoder in AdaLabel, which is guided by Da, achieves better accuracies than the CE. This further proves that our light-weight Da is capable of producing effective v. 5.2 Prediction Confidence We also visualized the distribution of confidence scores assigned by each dialogue model to highfrequency words. Figure 4 shows the results of [1, 200] [201, 400] [401, 600] [601, 800] [801, 1000] Token Frequency 0.0 0.5 1.0 1.5 2.0 2.5 % of Generated Tokens AdaLabel FACE NL FL F2 CP CE UL LS D2GPo Figure 5: Ratios of low-frequency tokens in the generated responses on the OpenSubtitles dataset. Tokens in each group are determined based on the frequency on the training set. four best performing models on the OpenSubtitles dataset. The spikes of high confidence score observed in Figure 4b and 4d indicate that CE and FACE assign extremely high confidence scores to a large number of high-frequency words. Although the smoothed labels in LS manage to alleviate these high-confidence-spikes (Figure 4c), a considerable amount of words still receives high confidence scores in LS. Our model outperforms all the baselines to avoid assigning over-confidence scores, thus alleviating the over-confidence issue. A similar trend is also observed on the DailyDialog dataset (see Appendix D for results of all models on both datasets). 5.3 Predicted Rare Word Distribution Over-confident models produce less diversified responses because they usually under-estimate rare words. To evaluate the effectiveness of AdaLabel, we tested whether AdaLabel encourages more “rare words” in its generations. Specifically, the ratio of generated tokens corresponding to different token frequency bins is calculated, and the results on the OpenSubtitles dataset are shown in Figure 5. It can be seen that AdaLabel produces more rare words in the generated responses than other baselines. Similar results are also observed on the DailyDialog dataset (see Appendix E). 6 Conclusion We address the low-diversity issue of neural dialogue models by introducing an adaptive label smoothing approach, AdaLabel. In our method, the probability of each target word is estimated based on the current dialogue model’s prediction, and the probabilities for these non-target words are calculated using a novel auxiliary decoder Da. A target-masked attention scheme is introduced in Da 3515 to help capture forward and backward contexts. We evaluate our method on two benchmark datasets: DailyDialog and OpenSubtitles. Extensive experiments show that our method effectively alleviates the over-confidence issue and improves the diversity of the generated responses. As future work, we believe this method is extensible to other text generation tasks. Acknowledgments This work was partly supported by the NSFC projects (Key project with No. 61936010 and regular project with No. 61876096). This work was also supported by the Guoqiang Institute of Tsinghua University, with Grant No. 2019GQG1 and 2020GQG0005. We thank Jinchao Zhang and Yao Qiu for early discussions and insightful comments of this work. References Hengyi Cai, Hongshen Chen, Yonghao Song, Cheng Zhang, Xiaofang Zhao, and Dawei Yin. 2020. Data manipulation: Towards effective instance learning for neural dialogue generation via learning to augment and reweight. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6334–6343, Online. Association for Computational Linguistics. Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2020. Distilling knowledge learned in BERT for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7893–7905, Online. Association for Computational Linguistics. Byung-Ju Choi, Jimin Hong, David Park, and Sang Wan Lee. 2020. Fˆ2-softmax: Diversifying neural text generation via frequency factorized softmax. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9167–9182, Online. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Jiachen Du, Wenjie Li, Yulan He, Ruifeng Xu, Lidong Bing, and Xuan Wang. 2018. Variational autoregressive decoder for neural response generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3154– 3163, Brussels, Belgium. Association for Computational Linguistics. Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378. Xiang Gao, Sungjin Lee, Yizhe Zhang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2019. Jointly optimizing diversity and relevance in neural response generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1229–1238, Minneapolis, Minnesota. Association for Computational Linguistics. Thamme Gowda and Jonathan May. 2020. Finding the optimal vocabulary size for neural machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3955–3964, Online. Association for Computational Linguistics. Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330. PMLR. Tianxing He and James Glass. 2020. Negative training for neural dialogue response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2044– 2058, Online. Association for Computational Linguistics. Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. CoRR, abs/1503.02531. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Minlie Huang, Xiaoyan Zhu, and Jianfeng Gao. 2020. Challenges in building intelligent open-domain dialog systems. ACM Transactions on Information Systems (TOIS), 38(3):1–32. Daphne Ippolito, Reno Kriz, Jo˜ao Sedoc, Maria Kustikova, and Chris Callison-Burch. 2019. Comparison of diverse decoding methods from conditional language models. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3752–3762, Florence, Italy. Association for Computational Linguistics. Shaojie Jiang, Pengjie Ren, Christof Monz, and Maarten de Rijke. 2019. Improving neural response diversity with frequency-aware cross-entropy loss. In The World Wide Web Conference, pages 2879– 2885. 3516 Shaojie Jiang and Maarten de Rijke. 2018. Why are sequence-to-sequence models so dull? understanding the low-diversity problem of chatbots. In Proceedings of the 2018 EMNLP Workshop SCAI: The 2nd International Workshop on Search-Oriented Conversational AI, pages 81–86, Brussels, Belgium. Association for Computational Linguistics. Jungo Kasai, Nikolaos Pappas, Hao Peng, James Cross, and Noah Smith. 2021. Deep encoder, shallow decoder: Reevaluating non-autoregressive machine translation. In International Conference on Learning Representations. Yoon Kim and Alexander M. Rush. 2016. Sequencelevel knowledge distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317–1327, Austin, Texas. Association for Computational Linguistics. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. OpenNMT: Opensource toolkit for neural machine translation. In Proceedings of ACL 2017, System Demonstrations, pages 67–72, Vancouver, Canada. Association for Computational Linguistics. Ilia Kulikov, Alexander Miller, Kyunghyun Cho, and Jason Weston. 2019. Importance of search and evaluation strategies in neural dialogue modeling. In Proceedings of the 12th International Conference on Natural Language Generation, pages 76–87, Tokyo, Japan. Association for Computational Linguistics. Aviral Kumar and Sunita Sarawagi. 2019. Calibration of encoder decoder models for neural machine translation. CoRR, abs/1903.00802. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California. Association for Computational Linguistics. Margaret Li, Stephen Roller, Ilia Kulikov, Sean Welleck, Y-Lan Boureau, Kyunghyun Cho, and Jason Weston. 2020. Don’t say that! making inconsistent dialogue unlikely with unlikelihood training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4715– 4728, Online. Association for Computational Linguistics. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986–995, Taipei, Taiwan. Asian Federation of Natural Language Processing. Zuchao Li, Rui Wang, Kehai Chen, Masso Utiyama, Eiichiro Sumita, Zhuosheng Zhang, and Hai Zhao. 2019. Data-dependent gaussian prior objective for language generation. In International Conference on Learning Representations. Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980– 2988. Luca Massarelli, Fabio Petroni, Aleksandra Piktus, Myle Ott, Tim Rockt¨aschel, Vassilis Plachouras, Fabrizio Silvestri, and Sebastian Riedel. 2020. How decoding strategies affect the verifiability of generated text. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 223–235, Online. Association for Computational Linguistics. Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip H. S. Torr, and Puneet K. Dokania. 2020. Calibrating deep neural networks using focal loss. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Rafael M¨uller, Simon Kornblith, and Geoffrey E Hinton. 2019. When does label smoothing help? In Advances in Neural Information Processing Systems, pages 4694–4703. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Gabriel Pereyra, George Tucker, Jan Chorowski, Lukasz Kaiser, and Geoffrey E. Hinton. 2017. Regularizing neural networks by penalizing confident output distributions. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net. Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826. Xu Tan, Yi Ren, Di He, Tao Qin, Zhou Zhao, and Tie-Yan Liu. 2019. Multilingual neural machine translation with knowledge distillation. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. 3517 Jiaxi Tang, Rakesh Shivanna, Zhe Zhao, Dong Lin, Anima Singh, Ed H. Chi, and Sagar Jain. 2020. Understanding and improving knowledge distillation. CoRR, abs/2002.03532. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 49, 2017, Long Beach, CA, USA, pages 5998–6008. Yida Wang, Pei Ke, Yinhe Zheng, Kaili Huang, Yong Jiang, Xiaoyan Zhu, and Minlie Huang. 2020. A large-scale chinese short-text conversation dataset. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 91–103. Springer. Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2020. Neural text generation with unlikelihood training. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Chenglin Yang, Lingxi Xie, Siyuan Qiao, and Alan L. Yuille. 2018. Knowledge distillation in generations: More tolerant teachers educate better students. CoRR, abs/1805.05551. Rongsheng Zhang, Yinhe Zheng, Jianzhi Shao, Xiaoxi Mao, Yadong Xi, and Minlie Huang. 2020. Dialogue distillation: Open-domain dialogue augmentation using unpaired data. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3449–3460. Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, and Bill Dolan. 2018. Generating informative and diverse conversational responses via adversarial information maximization. In Advances in Neural Information Processing Systems, pages 1810–1820. Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 654–664, Vancouver, Canada. Association for Computational Linguistics. Yinhe Zheng, Zikai Chen, Rongsheng Zhang, Shilei Huang, Xiaoxi Mao, and Minlie Huang. 2020a. Stylized dialogue response generation using stylized unpaired texts. In AAAI. Yinhe Zheng, Rongsheng Zhang, Minlie Huang, and Xiaoxi Mao. 2020b. A pre-training based personalized dialogue generation model with persona-sparse data. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9693–9700. Ganbin Zhou, Ping Luo, Yijun Xiao, Fen Lin, Bo Chen, and Qing He. 2018. Elastic responding machine for dialog generation with dynamically mechanism selecting. In Thirty-Second AAAI Conference on Artificial Intelligence. A Implementation Details This appendix describes the implementation details of our model. All our experiments are implemented with python 3.7.4, PyTorch 1.7.1, and the OpenNMT package (Klein et al., 2017). Training is performed on one TITAN Xp GPU. Our model’s backbone is the transformer-based sequence to sequence model, the encoder and decoder each contains 6 transformer layers with 8 attention heads, and the hidden size is set to 512. The dimension of the feedforward layer is also 512. The WordPiece tokenizer provided by BERT-base-uncased is used (the vocabulary contains 30522 tokens). The total number of parameters in our model is about 90M. The Adam optimizer is employed to train our model from random initializations with β1 = 0.9, β2 = 0.999, ϵ = 1e −9 and a learning rate of 1e-4. The batch size is set to 64 with 2 gradient accumulation so that 2 * 64 samples are used for each parameter update. The model is evaluated every 1000 steps on the validation set. We use early-stopping with patience 10, 30 for DailyDialog and OpenSubtitles, respectively. Specifically, the model stops training when the evaluation perplexity and accuracy are not increased for “patience” steps. The model training takes 4 hours and 3 days on DailyDialog and OpenSubtitles, respectively. The auxiliary distribution produced by the auxiliary decoder is smoothed with the temperature scaling approach. The temperature used in this process is searched in [1, 1.5, 2]. The temperature value of 1.5 and 1.0 is used for DailyDialog, and OpenSubtitles, respectively. The hyper-parameter value of η is set to 0.2 for all datasets. The fixed value of epsilon in our ablation model w/o ϵ is searched in [0.1, 0.2, 0.3, 0.4, 0.5, 0.6], and we find the value of 0.1 works best. B Baseline Implementation Details This appendix contains more implementation details of our baselines. All the baselines utilize the same backbone architecture and basic hyperparameter settings as our model (see Appendix A). The hyper-parameters specialized for each baseline is determined with the grid search based on the Dist 3518 Model DailyDialog OpenSubtitles Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 CE 1.79 8.21 4.19 5.90 2.57 4.06 2.49 1.58 2.48 9.21 4.07 5.74 0.76 7.03 4.26 2.82 LS 1.71 8.01 4.16 5.89 2.17 4.13 2.55 1.65 2.89 12.79 4.27 6.24 0.47 8.24 5.57 4.20 FL 2.40 11.37 4.39 6.35 4.46 6.01 3.95 2.75 3.10 12.37 4.25 6.13 0.82 7.13 4.56 3.25 FACE 1.80 9.47 4.54 6.40 3.48 5.65 3.43 2.17 3.12 12.62 4.47 6.40 1.02 5.97 3.63 2.43 F2 1.61 7.22 4.04 5.70 2.11 4.32 2.55 1.52 2.89 10.63 4.03 5.72 0.89 6.92 4.27 2.91 CP 2.30 10.39 4.28 6.16 3.25 5.31 3.39 2.30 3.14 11.87 4.17 5.97 0.85 7.28 4.60 3.21 UL 2.42 11.0 4.40 6.42 4.55 7.94 5.26 3.69 2.77 10.43 3.98 5.62 0.62 6.89 4.36 3.03 NL 1.61 7.53 4.19 6.05 4.02 7.09 4.41 2.91 2.65 10.14 4.21 6.05 0.75 7.16 4.32 2.85 D2GPo 1.57 7.83 4.14 5.91 2.26 4.47 2.71 1.71 2.06 10.43 4.15 6.00 0.12 7.32 4.69 3.33 AdaLabel 4.25 21.47 4.95 7.51 7.68 14.71 11.63 9.80 4.91 21.53 4.71 7.08 1.35 8.68 6.08 4.68 AdaLabel (Greedy) 3.96 23.53 5.17 8.00 8.49 17.42 13.38 11.01 4.78 22.88 4.96 7.66 1.47 9.80 6.48 4.75 Human 6.59 37.74 5.67 8.91 13.7 N/A N/A N/A 8.62 43.16 5.89 9.36 4.75 N/A N/A N/A Table 6: Automatic evaluation results (%) using the beam search decoding scheme (beam size is 5). The best results among all these beam-search-decoded models are in bold. measures on the validation set: For Label smoothing (LS), we searched the smoothing parameter in [0.05, 0.1, 0.2, 0.3, 0.4, 0.5], and found 0.1 works best on all the datasets; For Confidence penalty (CP), we searched the weight of penalty in [0.0005, 0.001, 0.01, 0.05, 0.1] and found 0.05 works best on all the datasets while ensuring the loss to be positive; For Focal loss (FL), we searched the hyperparameter γ in [0.1, 0.5, 1, 2, 3], and found 2 works best on all the datasets. For Unlikelihood loss (UL), we searched the weight of penalty in [1, 10, 100, 1000], and select 1000 on all the datasets. For FACE, we experiment with the Output token frequency & PRe-weigh version, which is reported to be the best version of FACE. For Negative loss (NL), F2-softmax (F2) and Datadependent Gaussian Prior objective (D2GPo), the selection of hyper-parameters follows the author’s suggestion. C Automatic Evaluation Results with Other Decoding Schemes This appendix reports our model’s automatic evaluation results and all the baselines when different decoding schemes are used. Specifically, Table 6 shows the results for the beam search decoding scheme (beam size of 5), and Table 7 shows the results when the top-K decoding scheme (k = 10) is used. Note that for the F2-softmax, we use the decoupled top-k sampling as the authors suggested. As can be seen from Table 6 and 7, our method outperforms all the baselines on the diversityrelated scores (i.e., Dist, Ent, and LF) by a large margin. This indicates that our method can produce more diverse responses even with the stochastic based decoding scheme. We also include the results of AdaLabel when the greedy decoding scheme is used in Table 6 and Table 7 (the second line from the bottom). It is interesting to see that the greedily decoded responses from AdaLabel are more diverse than some baselines that are decoded using the sampling scheme (see Table 7). Moreover, our model AdaLabel with the greedy decoding scheme achieves the best BLEU among all the baselines on both datasets. D Prediction Confidence This appendix reports the prediction confidence scores assigned by each model to high-frequency words. Specifically, words occupying the top 40% of the frequency mass in the training set of each dataset are regarded as high-frequency words. Figure 6 shows the results of our model and all the baselines on the DailyDialog dataset. Figure 7 shows the results of our model and all the baselines on the OpenSubtitles dataset. It can be seen that most of our baselines assign extremely high confidence scores (nearly 1.0) to these high-frequency words, and thus resulting in a spike of high confidence scores in the plotted distribution. Our model outperforms all the baselines in avoiding assigning extremely high confidence scores to these highfrequency words. E Predicted Rare Word Distribution on DailyDialog This appendix shows the distribution of rare words in the generated responses on the DailyDialog 3519 Model DailyDialog OpenSubtitles Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 Dist-1, 2 Ent-1, 2 LF BLEU-2,3,4 CE 2.22 19.05 5.07 7.87 4.09 6.78 3.29 1.61 3.78 20.58 5.07 7.97 1.23 5.94 2.84 1.46 LS 1.95 17.74 5.02 7.82 3.69 7.08 3.50 1.77 3.46 21.27 5.10 8.12 0.78 6.15 3.16 1.85 FL 2.71 20.98 5.19 8.17 6.44 8.09 4.13 2.24 3.82 22.14 5.15 8.25 1.27 5.34 2.54 1.34 FACE 2.29 21.14 5.36 8.3 5.73 7.07 3.47 1.82 4.25 23.95 5.30 8.37 1.51 5.34 2.54 1.33 F2 2.16 19.33 5.04 7.85 3.97 6.31 3.12 1.58 4.10 22.53 5.13 8.11 1.32 5.27 2.51 1.31 CP 3.16 22.38 5.11 7.96 6.01 8.11 4.38 2.50 4.06 22.62 5.13 8.14 1.33 6.00 2.94 1.52 UL 2.92 20.81 5.12 7.99 6.44 9.36 5.13 3.00 3.74 20.97 5.01 7.94 1.00 6.01 2.99 1.64 NL 2.39 18.35 4.99 7.79 5.72 8.71 4.64 2.63 3.57 20.36 5.05 7.97 1.06 5.84 2.86 1.46 D2GPo 1.75 17.09 5.00 7.81 3.40 7.45 3.73 1.97 2.74 19.21 5.00 7.97 0.36 6.32 3.15 1.72 AdaLabel 4.11 32.65 5.58 8.93 10.99 8.87 4.84 2.90 4.78 29.58 5.43 8.78 1.53 5.12 2.32 1.19 AdaLabel (Greedy) 3.96 23.53 5.17 8.00 8.49 17.42 13.38 11.01 4.78 22.88 4.96 7.66 1.47 9.80 6.48 4.75 Human 6.59 37.74 5.67 8.91 13.7 N/A N/A N/A 8.62 43.16 5.89 9.36 4.75 N/A N/A N/A Table 7: Automatic evaluation results (%) using the top-k sampling decoding scheme (k = 10). The best results among all these top-k-decoded models are in bold. 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (a) AdaLabel 0.0 0.5 1.0 Confidence Score 0 1 2 3 Density (%) (b) CE 0.0 0.5 1.0 Confidence Score 0.0 0.5 1.0 1.5 Density (%) (c) LS 0.0 0.5 1.0 Confidence Score 0 1 2 3 Density (%) (d) FACE 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (e) F^2 0.0 0.5 1.0 Confidence Score 0 1 2 3 Density (%) (f) NL 0.0 0.5 1.0 Confidence Score 0 2 4 Density (%) (g) UL 0.0 0.5 1.0 Confidence Score 0 2 4 Density (%) (h) CP 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (i) D2GPo 0.0 0.5 1.0 Confidence Score 0.0 0.5 1.0 1.5 Density (%) (j) FL Figure 6: Confidence score distributions for high-frequency words on the DailyDialog dataset. Words occupying the top 40% of the frequency mass in the training set of DailyDialog are regarded as high-frequency words. 0.0 0.5 Confidence Score 0 1 2 3 Density (%) (a) AdaLabel 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (b) CE 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (c) LS 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (d) FACE 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (e) F^2 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (f) NL 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (g) UL 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (h) CP 0.0 0.5 1.0 Confidence Score 0 1 2 3 Density (%) (i) D2GPo 0.0 0.5 1.0 Confidence Score 0 1 2 Density (%) (j) FL Figure 7: Confidence score distributions for high-frequency words on the OpenSubtitles dataset. Words occupying the top 40% of the frequency mass in the training set of OpenSubtitles are regarded as high-frequency words. dataset (see Figure 8). It can be seen that more “rare words” are predicted by our method on the DailyDialog dataset. This observation is in line with the results on the OpenSubtitles dataset as reported in Section 5.3. F Use BERT Model to Obtain v This appendix provides more experiment results comparing to the CMLM model (Chen et al., 2020): 1). CMLM exactly follows the setting of Chen et al. (2020), i.e., the teacher distribution produced by 3520 [0, 20] [21, 40] [41, 60] [61, 80] [81, 100] Token Frequency 0.0 0.5 1.0 1.5 2.0 % of Generated Tokens AdaLabel UL FL NL CP FACE CE LS F2 D2GPo Figure 8: Ratios of low-frequency tokens in the generated responses on the DailyDialog dataset. Tokens in each group are determined based on the frequency on the training set. Model BLEU-3,4 Dist-1,2 Ent-1,2 LF 1. CMLM 6.18 4.09 2.20 11.83 4.59 6.79 4.62 2. CMLM+ε 9.36 7.31 3.78 21.05 4.96 7.61 6.88 3. CMLM+ε+Da 11.6 9.34 3.67 20.97 5.02 7.71 7.28 AdaLabel 13.38 11.01 3.96 23.53 5.17 8.00 8.49 Table 8: Ablation study results based on BERT on DailyDialog (%). the BERT model is merged with the one-hot distribution using a fixed ε. 2). CMLM+ε adaptively adjust the value of ε using Eq. 6 in our paper. 3). CMLM+ε+Da add an additional training task to optimize the auxiliary decoder Da on the basis of CMLM+ε. It is expected that optimizing Da help our dialogue encoder to capture better representations. The trained Da is not used in the training and inference phase of our dialogue model. Note that the last model CMLM+ε+Da is the same with our ablation model 6. BERT as reported in our paper. As can be seen Table 8, our approach to adaptively change ε helps to produce better dialogue responses, and the training of Da helps our dialogue encoder to learn better representations. G Case study We sampled some generated cases on the DailyDialog and OpenSubtitles dataset. The results of our model and some competitive baselines are shown in Table 9 and Table 10. It can be seen that the responses generated by our method are coherent to the context and contain richer contents. Moreover, our model also produces more rare words that make our response more diverse. Dialogue Context1 Tell me a little bit about yourself , please . UL I am sorry , sir . I have been in a foreign currency company for the company . D2GPo I ’ m afraid i ca n ’ t . I ’ m a little nervous about it . FACE Yes , i am in the mood for a few days ago . AdaLabel I am studying in the international trade company . I am studying english in my company . Context1 Oh , they look nice . may i try that black one on ? Context2 Of course . the fitting room is right over there . UL I think you ’ re right . i think you ’ ll have a try . D2GPo I think it ’ s a good idea . FACE How much is it ? AdaLabel I bought this one . it ’ s the latest fashion . Context1 Well , we don ’ t have any pre - existing illnesses or conditions , so we should be fine . UL And most people are n ’ t so quiet and interviewing . D2GPo And the company company are a good job in the company . FACE And your company insurance , or workshops are booked in a private company ? AdaLabel I wish our company or country provided us with healthcare . Table 9: Cases on DailyDialog. Dialogue Context You ’ re going to love this baby . LS I ’ m going to get you . D2GPo I ’ m going to go to the bathroom . FACE You ’ re going to love this baby . AdaLabel You ’ re going to be a good father . Context Mr. Matlack can ’ t offend . LS This is our official business . D2GPo This is the most important thing . FACE It ’ s a pleasure to meet you , mr . President . AdaLabel Timothy Matlack is the official scribe of the continental congress . Context Brave men discover they are really cowards . LS I ’ m a samurai . D2GPo I ’ ve never seen such a beast in my life . FACE They are not afraid of you . AdaLabel Father , you ’ re a coward ! Context I found this whole operation strange . LS I ’ m not going to let you go . D2GPo I ’ m not sure i ’ m going to do it . FACE I ’ m sorry , but i ’ m not sure i can handle it . AdaLabel It ’ s an incredibly delicate system . Context If they make it , they ’ re clear into a safe zone where they can get medical supplies and food . LS We ’ il get them to the safe . D2GPo We ’ il have to get back to the hotel . FACE They ’ re gon na get us out of here . AdaLabel So we can use it as a safe field . Table 10: Cases on OpenSubtitles.
|
2021
|
272
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3521–3532 August 1–6, 2021. ©2021 Association for Computational Linguistics 3521 Out-of-Scope Intent Detection with Self-Supervision and Discriminative Training Li-Ming Zhan1 Haowen Liang1∗ Bo Liu1∗ Lu Fan1 Xiao-Ming Wu1† Albert Y.S. Lam2 Department of Computing, The Hong Kong Polytechnic University, Hong Kong S.A.R.1 Fano Labs, Hong Kong S.A.R.2 {lmzhan.zhan, michael.liang, doc-bo.liu}@connect.polyu.edu.hk {cslfan, csxmwu}@comp.polyu.edu.hk, [email protected] Abstract Out-of-scope intent detection is of practical importance in task-oriented dialogue systems. Since the distribution of outlier utterances is arbitrary and unknown in the training stage, existing methods commonly rely on strong assumptions on data distribution such as mixture of Gaussians to make inference, resulting in either complex multi-step training procedures or hand-crafted rules such as confidence threshold selection for outlier detection. In this paper, we propose a simple yet effective method to train an out-of-scope intent classifier in a fully end-to-end manner by simulating the test scenario in training, which requires no assumption on data distribution and no additional postprocessing or threshold setting. Specifically, we construct a set of pseudo outliers in the training stage, by generating synthetic outliers using inliner features via self-supervision and sampling out-of-scope sentences from easily available open-domain datasets. The pseudo outliers are used to train a discriminative classifier that can be directly applied to and generalize well on the test task. We evaluate our method extensively on four benchmark dialogue datasets and observe significant improvements over state-of-the-art approaches. Our code has been released at https:// github.com/liam0949/DCLOOS. 1 Introduction Conversational system is becoming an indispensable component in a variety of AI applications and acts as an interactive interface provided to users to improve user experience. Language understanding is essential for conversational systems to provide appropriate responses to users, and intent detection is usually the first step of language understanding. The primary goal is to identify diverse intentions ∗Equal contribution. † Corresponding author. Figure 1: t-SNE visualization of the learned embeddings of the test samples of CLINC150. Top: Previous K-way training; Bottom: Our proposed (K + 1)-way training. Better view in color and enlarged. behind user utterances, which is often formalized as a classification task. However, intent classes defined during training are inevitably inadequate to cover all possible user intents at the test stage due to the diversity and randomness of user utterances. Hence, out-of-scope (or unknown) intent detection is essential, which aims to develop a model that can accurately identify known (seen in training) intent classes while detecting the out-of-scope classes that are not encountered during training. Due to the practical importance of out-of-scope intent detection, recent efforts have attempted to solve this problem by developing effective intent classification models. In general, previous works approach this problem by learning decision boundaries for known intents and then using some confidence measure to distinguish known and unknown intents. For examples, LMCL (Lin and Xu, 2019) 3522 learns the decision boundaries with a margin-based optimization objective, and SEG (Yan et al., 2020b) assumes the known intent classes follow the distribution of mixture of Gaussians. After learning the decision boundaries, an off-the-shell outlier detection algorithm such as LOF (Breunig et al., 2000) is commonly employed to derive confidence scores (Yan et al., 2020b; Shu et al., 2017; Lin and Xu, 2019; Hendrycks and Gimpel, 2017). If the confidence score of a test sample is lower than a predefined threshold, it is identified as an outlier. However, it may be problematic to learn decision boundaries solely based on the training examples of known intent classes. First, if there are sufficient training examples, the learned decision boundaries can be expected to generalize well on known intent classes, but not on the unknown. Therefore, extra steps are required in previous methods, such as using an additional outlier detection algorithm at the test stage or adjusting the confidence threshold by cross-validation. On the other hand, if there are not sufficient training examples, the learned boundaries may not generalize well on both known and unknown intents. As a result, these methods often underperform when not enough training data is given. Hence, it is important to provide learning signals of unknown intents at the training stage to overcome these limitations. In contrast to previous works, we adopt a different approach by explicitly modeling the distribution of unknown intents. Particularly, we construct a set of pseudo out-of-scope examples to aid the training process. We hypothesize that in the semantic feature space, real-world outliers can be well represented in two types: “hard” outliers that are geometrically close to the inliers and “easy” outliers that are distant from the inliners. For the “hard” ones, we construct them in a self-supervised manner by forming convex combination of the features of inliers from different classes. For the “easy” ones, the assumption is that they are very unrelated to the known intent classes, so they can be used to simulate the randomness and diversity of user utterances. They can be easily constructed using public datasets. For example, in our experiments, we randomly collect sentences from datasets of other NLP tasks such as question answering and sentiment analysis as open-domain outliers. In effect, by constructing pseudo outliers for the unknown class during training, we form a consistent (K + 1) classification task (K known classes + 1 unknown class) for both training and test. Our model can be trained with a cross-entropy loss and directly applied to test data for intent classification and outlier detection without requiring any further steps. As shown in Figure 1 (better view in color and enlarged), our method can learn better utterance representations, which make each known intent class more compact and push the outliers away from the inliers. Our main contributions are summarized as follows. • We propose a novel out-of-scope intent detection approach by matching training and test tasks to bridge the gap between fitting to training data and generalizing to test data. • We propose to efficiently construct two types of pseudo outliers by using a simple selfsupervised method and leveraging publicly available auxiliary datasets. • We conduct extensive experiments on four real-world dialogue datasets to demonstrate the effectiveness of our method and perform a detailed ablation study. 2 Related Work 2.1 Out-of-Distribution Detection Early studies on outlier detection often adopt unsupervised clustering methods to detect malformed data (Hodge and Austin, 2004; Chandola et al., 2009; Zimek et al., 2012). In recent years, a substantial body of work has been directed towards improving the generalization capacity of machine learning models on out-of-distribution (OOD) data (Ruff et al., 2021; Hendrycks et al., 2020a). Hendrycks and Gimpel (2017) find that simple statistics derived from the outputting softmax probabilities of deep neural networks can be helpful for detecting OOD samples. Following this work, Liang et al. (2018) propose to use temperature scaling and add small perturbation to input images to enlarge the gap between in-scope and OOD samples. Lee et al. (2017) propose to add a Kullback-Leibler divergence term in the loss function to encourage assigning lower maximum scores to OOD data. Recently, there is a line of work that employs synthetic or real-world auxiliary datasets to provide learning signals for improving model robustness under various forms of distribution shift (Goodfellow et al., 2015; Orhan, 2019; Hendrycks et al., 3523 2019; Lee et al., 2017). Particularly, Hendrycks et al. (2018) propose to leverage large-scale public datasets to represent outliers during training time and form a regularization term based on that. This idea is similar to our proposal of constructing opendomain outliers, but we use a simpler, end-to-end, (K+1)-way discriminative training procedure without any regularization term or threshold parameter. 2.2 Out-of-Scope Intent Detection While Hendrycks et al. (2020b) find that pretrained transformer-based models like BERT are intrinsically more robust to OOD data, they suggest that there are still margins for improvement. Therefore, we build our model on top of BERT to improve intent detection under significant distribution shift. Previous methods for out-of-scope (or out-of-distribution) intent detection are commonly threshold-based, where models output a decision score and then compare it with a threshold that is predefined or selected by cross-validation. There are mainly three branches of related work. The first group uses a confidence score which determines the likelihood of an utterance being outof-scope. For example, Shu et al. (2017) build m binary Sigmoid classifiers for m known classes respectively and select a threshold to reject OOD inputs that may have lower probabilities than the threshold across all m classifiers. Similar to the OOD data generation method used in Lee et al. (2017), Ryu et al. (2018) employ GAN (Goodfellow et al., 2014) to generate simulated OOD examples with the generator and learn to reject simulated OOD examples with the discriminator. The second group identifies out-of-scope sentences through reconstruction loss. For example, Ryu et al. (2017) build an autoencoder to encode and decode in-scope utterances and obtain reconstruction loss by comparing input embeddings with decoded ones. Out-of-scope utterances result in higher reconstruction loss. The third group leverages off-the-shell outlier detection algorithms such as local outlier factor (LOF) (Breunig et al., 2000), one-class SVM (Sch¨olkopf et al., 2001), robust covariance estimators (Rousseeuw and Driessen, 1999), and isolation forest (Liu et al., 2008) to detect out-ofscope examples. Utterance embeddings belonging to a specific class will be mapped to the corresponding cluster (usually modeled by a Gaussian distribution) while out-of-scope samples will be pushed away from all in-scope clusters. Examples of this kind include SEG (Yan et al., 2020a) and LMCL (Lin and Xu, 2019). Very recently, Zhang et al. (2021) propose to learn adaptive decision boundaries after pre-training instead of using offthe-shell outlier detection algorithms. In addition, some other work focuses on outof-scope detection in few-shot scenarios. Tan et al. (2019) leverage independent source datasets as simulated OOD examples to form a hinge loss term. Zhang et al. (2020) propose to pretrain BERT by a natual language understanding task with largescale training data to transfer useful information for few-shot intent detection. Finally, for our proposal of constructing synthetic outliers, the most similar method is Mixup proposed by Zhang et al. (2018). However, their method is designed for data augmentation to enhance in-distribution performance and requires corresponding combinations in the label space (Thulasidasan et al., 2019). 3 Methodology Problem Statement In a dialogue system, given K predefined intent classes Sknown = {Ci}K i=1, an unknown intent detection model aims at predicting the category of an utterance u, which may be one of the known intents or an out-of-scope intent Coos. Essentially, it is a K + 1 classification problem at the test stage. At the training stage, a set of N labeled utterances Dl = {(xi, ci) | ci ∈Sknown)}N i=1 is provided for training. Previous methods typically train a K-way classifier for the known intents. Overview of Our Approach The mismatch between the training and test tasks, i.e., K-way classification vs. (K + 1)-way classification, leads to the use of strong assumptions and additional complexity in previous methods. Inspired by recent practice in meta learning to simulate test conditions in training (Vinyals et al., 2016), we propose to match the training and test settings. In essence, as shown in Figure 2, we formalize a (K + 1)-way classification task in the training stage by constructing out-of-scope samples via self-supervision and from open-domain data. Our method simply trains a (K + 1)-way classifier without making any assumption on the data distribution. After training, the classifier can be readily applied to the test task without any adaptation or post-processing. In the following, we elaborate on the details of our proposed method, including representation learning, 3524 Figure 2: An illustration of our proposed method. We use BERT as the utterance encoder. At training stage, we train a (K+1)-way classifier by constructing two types of pseudo outliers. The open-domain outliers are collected from an auxiliary dataset disjoint from both the training and test data. The synthetic self-supervised outliers are generated during training by random convex combinations of features of inliers from different known classes. construction of pseudo outliers, and discriminative training. 3.1 Representation Learning We employ BERT (Devlin et al., 2019) – a deep Transformer network as text encoder. Specifically, we take the d-dimensional output vector of the special classification token [CLS] as the representation of an utterance u, i.e., h = BERT(u) ∈Rd, where d = 768 by default. The training set Dl is then mapped to Dtr l = {(hi, ci) | hi = BERT(ui), (ui, ci) ∈Dl}N i=1 in the feature space. 3.2 Construction of Outliers We construct two different types of pseudo outliers to be used in the training stage: synthetic outliers that are generated by self-supervision, and opendomain outliers that can be easily acquired. Synthetic Outliers by Self-Supervision To improve the generalization ability of the unknown intent detection model, we propose to generate “hard” outliers in the feature space, which may have similar representations to the inliers of known intent classes. We hypothesize that those outliers may be geometrically close to the inliers in the feature space. Based on this assumption, we propose a selfsupervised method to generate the “hard” outliers using the training set Dtr l . Specifically, in the feature space, we generate synthetic outliers by using convex combinations of the features of inliers from different intent classes: hoos = θ ∗hβ + (1 −θ) ∗hα, (1) where hβ and hα are the representations of two utterances which are randomly sampled from different intent classes in Dtr l , i.e., cβ ̸= cα, and hoos is the synthetic outlier. For example, θ can be sampled from a uniform distribution U(0, 1). In this case, when θ is close to 0 or 1, it will generate “harder” outliers that only contain a small proportion of mix-up from different classes. In essence, “hard” outliers act like support vectors in SVM (Cortes and Vapnik, 1995), and “harder” outliers could help to train a more discriminative classifier. The generated outliers hoos are assigned to the class of Coos, the (K + 1)-th class in the feature space, forming a training set Dtr co = {(hoos i , ci = Coos)}M i=1. (2) Notice that since the outliers are generated in the feature space, it is very efficient to construct a large outlier set Dtr co. Open-Domain Outliers In practical dialogue systems, user input can be arbitrary free-form sentences. To simulate real-world outliers and provide learning signals representing them in training, we propose to construct a set of open-domain outliers, 3525 which can be easily obtained. Specifically, the set of free-form outliers Dfo can be constructed by collecting sentences from various public datasets that are disjoint from the training and test tasks. There are many datasets available, including the question answering dataset SQuaD 2.0 (Rajpurkar et al., 2018), the sentiment analysis datasets Yelp (Meng et al., 2018) and IMDB (Maas et al., 2011), and dialogue datasets from different domains. In the feature space, Dfo is mapped to Dtr fo = {(hoos i , ci = Coos) | hoos i = BERT(ui), ui ∈ Dfo}H i=1. Both synthetic outliers and open-domain outliers are easy to construct. As will be demonstrated in Section 4, both of them are useful, but synthetic outliers are much more effective than open-domain outliers in improving the generalization ability of the trained (K + 1)-way intent classifier. 3.3 Discriminative Training After constructing the pseudo outliers, in the feature space, our training set Dtr now consists of a set of inliers Dtr l and two sets of outliers Dtr co and Dtr fo, i.e., Dtr = Dtr l ∪Dtr co ∪Dtr fo and |Dtr| = N + M + H. Therefore, in the training stage, we can train a (K + 1)-way classifier with the intent label set S = Sknown ∪{Coos}, which can be directly applied in the test stage to identify unknown intent and classify known ones. In particular, we use a multilayer perceptron network, Φ(·), as the classifier in the feature space. The selection of the classifier is flexible, and the only requirement is that it is differentiable. Then, we train our model using a cross-entropy loss: L = − 1 |Dtr| X Dtr log exp(Φ(hi)ci/τ) P j∈S exp(Φ(hi)j/τ), where Φ(hi)ci refers to the output logit of Φ(·) for the ground-truth class ci, and τ ∈R+ is an adjustable scalar temperature parameter. 4 Experiments In this section, we present the experimental results of our proposed method on the targeted task of unknown intent detection. Given a test set comprised of known and unknown intent classes, the primary goal of an unknown intent detection model is to assign correct intent labels to utterances in the test set. Notice that the unknown intent label Coos is also included as a special class for prediction. 4.1 Datasets and Baselines We evaluate our proposed method on four benchmark datasets as follows, three of which are newly released dialogue datasets designed for intent detection. The statistics of the datasets are summarized in Table 2. CLINC150 (Larson et al., 2019) is a dataset specially designed for out-of-scope intent detection, which consists of 150 known intent classes from 10 domains. The dataset includes 22, 500 in-scope queries and 1, 200 out-of-scope queries. For the in-scope ones, we follow the original splitting, i.e., 15, 000, 3, 000 and 4, 500 for training, validation, and testing respectively. For the out-of-scope ones, we group all of the 1, 200 queries into the test set. StackOverflow (Xu et al., 2015) consists of 20 classes with 1, 000 examples in each class. We follow the original splitting, i.e., 12, 000 for training, 2, 000 for validation, and 6, 000 for test. Banking (Casanueva et al., 2020) is a finegrained intent detection dataset in the banking domain. It consists of 9, 003, 1, 000, and 3, 080 user queries in the training, validation, and test sets respectively. M-CID (Arora et al., 2020) is a recently released dataset related to Covid-19. We use the English subset of this dataset referred to as M-CID-EN in our experiments, which covers 16 intent classes. The splitting of M-CID-EN is 1, 258 for training, 148 for validation, and 339 for test. We extensively compare our method with the following unknown intent detection methods. • Maximum Softmax Probability (MSP) (Hendrycks and Gimpel, 2017) employs the confidence score derived from the maximum softmax probability to predict the class of a sample. The idea under the hood is that the lower the confidence score is, the more likely the sample is of an unknown intent class. • DOC (Shu et al., 2017) considers to construct m 1-vs-rest sigmoid classifiers for m seen classes respectively. It uses the maximum probability from these classifiers as the confidence score to conduct classification. • SEG (Yan et al., 2020a) models the intent distribution as a margin-constrained Gaussian mixture distribution and uses an additional outlier detector – local outlier factor 3526 CLINC150 StackOverflow Banking M-CID-EN Methods Accuracy Macro-F1 Accuracy Macro-F1 Accuracy Macro-F1 Accuracy Macro-F1 25% MSP 66.60 51.20 33.94 45.68 48.15 48.47 52.05 43.14 DOC 64.43 44.60 60.68 60.51 37.78 46.35 49.32 46.59 SEG 72.86 65.44 47.00 52.83 51.11 55.68 44.51 50.14 LMCL 68.57 62.42 41.60 48.21 52.77 56.73 41.44 46.99 Softmax 76.50 67.74 46.17 50.78 57.88 58.32 41.95 45.46 Ours 88.44 80.73 68.74 65.64 74.11 69.93 87.08 79.67 50% MSP 68.61 51.20 56.33 62.92 53.83 65.33 61.21 54.33 DOC 62.46 70.01 61.62 68.97 58.29 57.30 59.97 62.28 SEG 77.05 79.42 68.50 74.18 68.44 76.48 67.91 72.37 LMCL 78.63 80.42 64.34 71.80 63.59 73.99 63.42 69.04 Softmax 82.47 82.86 65.96 71.94 67.44 74.19 64.72 69.35 Ours 88.33 86.67 75.08 78.55 72.69 79.21 81.05 79.73 75% MSP 73.41 81.81 76.73 77.63 71.92 80.77 72.89 77.34 DOC 74.63 78.63 63.98 62.07 72.02 78.04 69.79 71.18 SEG 81.92 86.57 80.83 84.78 78.87 85.66 75.73 79.97 LMCL 84.59 88.21 80.02 84.47 78.66 85.33 77.11 80.96 Softmax 86.26 89.01 77.41 82.28 78.20 84.31 76.99 80.82 Ours 88.08 89.43 81.71 85.85 81.07 86.98 80.24 82.75 Table 1: Overall accuracy and macro f1-score for unknown intent detection with different proportion of seen classes. For each setting, the best result is marked in bold. Dataset Vocab Avg. Length Samples Classes CLINC150 8,376 8.31 23,700 150 StackOverflow 17,182 9.18 20,000 20 Banking 5028 11.9 13,083 77 M-CID-EN 1,254 6.74 1,745 16 Table 2: Dataset statistics. (LOF) (Breunig et al., 2000) to achieve unknown intent detection. • LMCL (Lin and Xu, 2019) considers to learn discriminative embeddings with a large margin cosine loss. It also uses LOF as the outlier detection algorithm. • Softmax (Yan et al., 2020a) uses a softmax loss to learn discriminative features based on the training dataset, which also requires an additional outlier detector such as LOF for detecting the unknown intents. 4.2 Experimental Setup and Evaluation Metrics To compare with existing methods, we follow the setting in LMCL (Lin and Xu, 2019). Specifically, for each dataset, we randomly sample 75%, 50%, and 25% of the intent classes from the training set as the known classes to conduct training, and we set aside the rest as the unknown classes for test. Notice that for training and validation, we only use data within the chosen known classes and do not expose our model to any of test-time outliers. Unless otherwise specified, in each training batch, we keep the ratio of inliers, open-domain outliers and self-supervised outliers roughly as 1 : 1 : 4. This setting is empirically chosen and affected by the memory limit of NVIDIA 2080TI GPU, which we use for conducting the experiments. The number of pseudo outliers can be adjusted according to different environments, and a larger number of self-supervised outliers typically takes more time to converge. We use Pytorch (Paszke et al., 2019) as the backend to conduct the experiments. We use the pretrained BERT mdoel (bert-base-uncased) provided by Wolf et al. (2019) as the encoder for utterances. We use the output vector of the special classification token [CLS] as the utterance embedding and fix its dimension as 768 by default throughout all of our experiments. To ensure a fair comparison, all baselines and our model use the same encoder. For model optimization, we use AdamW provided by Wolf et al. (2019) to fine-tune BERT and Adam proposed by Kingma and Ba (2015) to train the MLP clasisfier Φ(·). We set the learning rate for BERT as 1e−5 as suggested by Devlin et al. (2019). For the MLP clasisfier, the learning rate is fixed as 1e−4. Notice that the fine-tuning of BERT 3527 CLINC150 StackOverflow Banking M-CID-EN Methods Unknown Known Unknown Known Unknown Known Unknown Known 25% MSP 73.20 50.62 22.59 50.30 49.98 48.39 56.27 37.86 DOC 71.08 43.91 66.11 59.39 31.41 47.14 53.08 44.92 SEG 79.90 65.06 46.17 54.16 53.22 55.81 42.73 51.99 LMCL 75.61 62.01 38.85 50.15 55.29 56.81 36.99 49.50 Softmax 83.04 67.34 45.52 51.83 62.52 58.10 35.39 46.22 Ours 92.35 80.43 74.86 63.80 80.12 69.39 91.15 76.80 50% MSP 57.78 68.03 35.18 70.09 29.31 66.28 58.55 53.80 DOC 57.62 70.17 47.96 71.07 49.88 57.50 47.22 64.16 SEG 78.02 79.43 60.89 75.51 60.42 76.90 61.04 73.80 LMCL 79.89 80.42 53.12 71.80 50.30 74.62 51.11 71.29 Softmax 84.19 82.84 56.80 73.45 60.28 74.56 56.30 70.98 Ours 90.30 86.54 71.88 79.22 67.26 79.52 82.44 79.39 75% MSP 57.83 82.02 41.73 80.03 23.86 81.75 39.56 80.50 DOC 64.62 78.76 49.50 62.91 39.47 78.72 49.41 72.99 SEG 76.12 86.67 62.30 86.28 54.43 86.20 51.51 82.34 LMCL 80.42 88.28 61.40 84.47 53.26 85.89 54.61 83.16 Softmax 83.12 89.61 54.07 84.11 56.90 84.78 58.73 82.66 Ours 86.28 89.46 65.44 87.22 60.71 87.47 69.00 83.89 Table 3: Macro f1-score of the known classes and f1-score of the unknown class with different proportion of seen classes. For each setting, the best result is marked in bold. is conducted simultaneously with the training of the classifier Φ(·) with the same cross-entropy loss. The MLP classifier Φ(·) has a two-layer architecture with [1024, 1024] as hidden units. The temperature parameter τ is selected by cross-validation and set as 0.1 in all experiments. Following LMCL (Lin and Xu, 2019), we use overall accuracy and macro f1-score as evaluation metrics. All results reported in this section are the average of 10 runs with different random seeds, and each run is stopped until reaching a plateau on the validation set. For baselines, we follow their original training settings except using the aforementioned BERT as text encoder. 4.3 Result Analysis We present our main results in Table 1 and Table 3. Specifically, Table 1 gives results in overall accuracy and macro f1-score for all classes including the outlier class, while Table 3 shows results in macro f1-score for the known classes and f1-score for the outlier class respectively. It can be seen that, on all benchmarks and in almost every setting, our model significantly outperforms the baselines. As shown in Table 3, our method achieves favorable performance on both unknown and known intent classes simultaneously. It is worth mentioning that the large improvements of our method in scenarios with small labeled training sets (25% and 50% settings) indicate its great potential in real-life applications, since a practical dialogue system often needs to deal with a larger proportion of outliers than inliers due to different user demographic, ignorance/unfamiliarity of/with the platform, and limited intent classes recognized by the system (especially at the early development stage). More importantly, referring to Table 3, as the proportion of known intents increases, it can be seen that the performance gains of the baselines mainly lie in the known classes. In contrast, our method can strike a better balance between the known and unknown classes without relying on additional outlier detector, margin tuning, and threshold selection, demonstrating its high effectiveness and generality. Take the Softmax baseline for example, in the 75% case of CLINC150, it achieves a slightly higher result than our model on the known classes but a substantially lower result on the unknown ones. 4.4 Effect of Pseudo Outliers We conduct an ablation study on the effectiveness of the two kinds of pseudo outliers and summarize the results in Table 4. The first row of the three settings (25%, 50%, and 75%) stands for training solely with the labeled examples of CLINC150 3528 (a) (b) (c) (d) (e) (f) Figure 3: Effect of the number of pseudo outliers on CLINC150. (a), (b), and (c) display overall accuracy, f1-score on the unknown class and overall macro f1-score with varying number of self-supervised outliers respectively. (d), (e), and (f) display the corresponding results with varying number of open-domain outliers. Figure 4: Effect of the number of self-supervised outliers on overall intent detection accuracy under the 75% setting of Banking. without using any pseudo outliers. In general, selfsupervised synthetic outliers and open-domain outliers both lead to positive effects on classification performance. For each setting, comparing the second row with the third, we can observe that the synthetic outliers produced by convex combinations lead to a much larger performance gain than that of pre-collected open-domain outliers. Finally, combining them for training leads to the best results, as shown in the fourth row of each setting. Next, we conduct experiments to study the impact of varying the number of the two kinds of pseudo outliers separately, as shown in Figure 3. We first fix the number of open-domain outliers as zero and then increase the number of selfsupervised outliers. The results are displayed in Figure 3 (a), (b) and (c). In particular, as the number of self-supervised outliers grows, the performance first increases quickly and then grows slowly. On the other hand, we fix the number of self-supervised outliers as zero and then increases the number of open-domain outliers. The results are shown in Figure 3 (d), (e) and (f), where it can be seen that dozens of open-domain outliers already can bring significant improvements, though the gain is much smaller compared to that of the self-supervised outliers. Finally, we investigate the impact of the number of self-supervised outliers on overall intent detection accuracy with both the number of inliers and the number of open-domain outliers fixed as 100 per training batch. As shown in Figure 4, we increase the number of self-supervised outliers from 0 to 5000. Note that 400 is the default setting used in Table 1 and Table 3. We can see that comparable results can be obtained for a wide range of numbers. However, when the number grows to 5000, the performance exhibits a significant drop. We hypothesize that as the number increases, the 3529 Dtr co Dtr fo Acc Macro-F1 F1 Unknown 25% 19.79 41.05 ✓ 81.96 71.15 87.8 ✓ 37.55 45.14 36.91 ✓ ✓ 88.44 80.73 92.35 50% 38.78 60.35 ✓ 83.12 82.62 85.03 ✓ 48.62 63.19 28.82 ✓ ✓ 88.33 86.67 90.30 75% 57.43 73.6 ✓ 84.16 86.9 80.36 ✓ 69.61 79.42 48.29 ✓ ✓ 88.08 89.43 86.28 Table 4: An ablation study on the effectiveness of pseudo outliers. Dtr fo Acc Macro-F1 25% Open-bank 89.36 81.22 Open-stack 88.38 80.42 Open-big 88.44 80.73 50% Open-bank 87.35 86.41 Open-stack 88.23 86.37 Open-big 88.33 86.67 75% Open-bank 87.19 89.33 Open-stack 87.52 89.17 Open-big 88.08 89.43 Table 5: Results on CLINC150 with different sets of open-domain outliers. generated synthetic outliers may be less accurate, because some convex combinations may fall within the scope of known classes. To summarize, self-supervised outliers play a much more important role than open-domain outliers for unknown intent classification. Selfsupervised outliers not only provide better learning signals for the unknown intents, but also impose an important positive effect on the known ones. For the open-domain outliers, if used alone, they can only provide limited benefit. But in combination with the self-supervised ones, they can further enhance the performance. 4.5 Selection of Open-Domain Outliers To demonstrate the flexibility of our method in selecting open-domain outliers as described in Section 3.2, we train our model on CLINC150 using open-domain outliers from different sources. The results are summarized in Table 5. Specifically, Open-bank and Open-stack stand for using Figure 5: Comparison of training time (per epoch) and test time with baselines. the training set of Banking and StackOverflow as the source of open-domain outliers respectively. Open-big stands for the source of open-domain outliers used in other experiments, which consists of ∼0.5 million sentences randomly selected from SQuaD 2.0 (Rajpurkar et al., 2018), Yelp (Meng et al., 2018), and IMDB (Maas et al., 2011). It can be seen that the performance of our model is insensitive to the selection of open-domain outliers. 4.6 Efficiency We provide a quantitative comparison on the training and test efficiency for our method and the baselines, by calculating the average time (in seconds) for training per epoch and the total time for testing under the 75% setting. Here, we only compare with the strongest baselines. As shown in Figure 5, even with the pseudo outliers, the training time of our method is comparable to that of the baselines. Importantly, in the test stage, our method demonstrates significant advantages in efficiency, which needs much less time to predict intent classes for all samples in the test set. 5 Conclusion We have proposed a simple, effective, and efficient approach for out-of-scope intent detection by overcoming the limitation of previous methods via matching train-test conditions. Particularly, at the training stage, we construct self-supervised and open-domain outliers to improve model generalization and simulate real outliers in the test stage. Extensive experiments on four dialogue datasets show that our approach significantly outperforms state-of-the-art methods. In the future, we plan to investigate the theoretical underpinnings of our approach and apply it to more applications. Acknowledgments We would like to thank the anonymous reviewers for their helpful comments. This research was supported by the grant HK ITF UIM/377. 3530 References Abhinav Arora, Akshat Shrivastava, Mrinal Mohit, Lorena Sainz-Maza Lecanda, and Ahmed Aly. 2020. Cross-lingual transfer learning for intent detection of covid-19 utterances. Markus M. Breunig, Hans-Peter Kriegel, Raymond T. Ng, and J¨org Sander. 2000. Lof: Identifying densitybased local outliers. SIGMOD Rec., 29(2):93–104. I˜nigo Casanueva, Tadas Temcinas, Daniela Gerz, Matthew Henderson, and Ivan Vulic. 2020. Efficient intent detection with dual sentence encoders. CoRR, abs/2003.04807. Varun Chandola, Arindam Banerjee, and Vipin Kumar. 2009. Anomaly detection: A survey. ACM computing surveys (CSUR), 41(3):1–58. Corinna Cortes and Vladimir Vapnik. 1995. Supportvector networks. Machine learning, 20(3):273–297. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics. Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial networks. arXiv preprint arXiv:1406.2661. Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2020a. The many faces of robustness: A critical analysis of out-of-distribution generalization. arXiv preprint arXiv:2006.16241. Dan Hendrycks and Kevin Gimpel. 2017. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. Dan Hendrycks, Kimin Lee, and Mantas Mazeika. 2019. Using pre-training can improve model robustness and uncertainty. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2712–2721. PMLR. Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song. 2020b. Pretrained transformers improve out-ofdistribution robustness. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 2744–2751. Association for Computational Linguistics. Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. 2018. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606. Victoria J. Hodge and Jim Austin. 2004. A survey of outlier detection methodologies. Artif. Intell. Rev., 22(2):85–126. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Stefan Larson, Anish Mahendran, Joseph J Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K Kummerfeld, Kevin Leach, Michael A Laurenzano, Lingjia Tang, et al. 2019. An evaluation dataset for intent classification and out-ofscope prediction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 1311–1316. Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. 2017. Training confidence-calibrated classifiers for detecting out-of-distribution samples. arXiv preprint arXiv:1711.09325. Shiyu Liang, Yixuan Li, and R. Srikant. 2018. Enhancing the reliability of out-of-distribution image detection in neural networks. 6th International Conference on Learning Representations, ICLR 2018 ; Conference date: 30-04-2018 Through 03-05-2018. Ting-En Lin and Hua Xu. 2019. Deep unknown intent detection with margin loss. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 5491–5496. Association for Computational Linguistics. Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. 2008. Isolation forest. In Proceedings of the 2008 Eighth IEEE International Conference on Data Mining, ICDM ’08, page 413–422, USA. IEEE Computer Society. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In The 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Proceedings of the Conference, 19-24 June, 3531 2011, Portland, Oregon, USA, pages 142–150. The Association for Computer Linguistics. Yu Meng, Jiaming Shen, Chao Zhang, and Jiawei Han. 2018. Weakly-supervised neural text classification. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM 2018, Torino, Italy, October 2226, 2018, pages 983–992. ACM. A. Emin Orhan. 2019. Robustness properties of facebook’s resnext WSL models. CoRR, abs/1907.07640. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K¨opf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 8024–8035. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for squad. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 1520, 2018, Volume 2: Short Papers, pages 784–789. Association for Computational Linguistics. Peter Rousseeuw and Katrien Driessen. 1999. A fast algorithm for the minimum covariance determinant estimator. Technometrics, 41:212–223. Lukas Ruff, Jacob R. Kauffmann, Robert A. Vandermeulen, Gr´egoire Montavon, Wojciech Samek, Marius Kloft, Thomas G. Dietterich, and Klaus-Robert M¨uller. 2021. A unifying review of deep and shallow anomaly detection. Proc. IEEE, 109(5):756– 795. Seonghan Ryu, Seokhwan Kim, Junhwi Choi, Hwanjo Yu, and Gary Geunbae Lee. 2017. Neural sentence embedding using only in-domain sentences for outof-domain sentence detection in dialog systems. Pattern Recogn. Lett., 88(C):26–32. Seonghan Ryu, Sangjun Koo, Hwanjo Yu, and Gary Geunbae Lee. 2018. Out-of-domain detection based on generative adversarial network. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 714– 718, Brussels, Belgium. Association for Computational Linguistics. B. Sch¨olkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson. 2001. Estimating the support of a high-dimensional distribution. Neural Computation, 13(7):1443–1471. Lei Shu, Hu Xu, and Bing Liu. 2017. DOC: deep open classification of text documents. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP 2017, Copenhagen, Denmark, September 9-11, 2017, pages 2911–2916. Association for Computational Linguistics. Ming Tan, Yang Yu, Haoyu Wang, Dakuo Wang, Saloni Potdar, Shiyu Chang, and Mo Yu. 2019. Out-ofdomain detection for low-resource text classification tasks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3566–3572, Hong Kong, China. Association for Computational Linguistics. Sunil Thulasidasan, Gopinath Chennupati, Jeff A. Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. 2019. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 814, 2019, Vancouver, BC, Canada, pages 13888– 13899. Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. 2016. Matching networks for one shot learning. In Proceedings of the 30th International Conference on Neural Information Processing Systems, pages 3637–3645. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, et al. 2019. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. Jiaming Xu, Peng Wang, Guanhua Tian, Bo Xu, Jun Zhao, Fangyuan Wang, and Hongwei Hao. 2015. Short text clustering via convolutional neural networks. In Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing, VS@NAACL-HLT 2015, June 5, 2015, Denver, Colorado, USA, pages 62–69. The Association for Computational Linguistics. Guangfeng Yan, Lu Fan, Qimai Li, Han Liu, Xiaotong Zhang, Xiao-Ming Wu, and Albert Y. S. Lam. 2020a. Unknown intent detection using gaussian mixture model with an application to zero-shot intent classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 1050– 1060. Association for Computational Linguistics. Guangfeng Yan, Lu Fan, Qimai Li, Han Liu, Xiaotong Zhang, Xiao-Ming Wu, and Albert Y.S. Lam. 2020b. Unknown intent detection using Gaussian mixture model with an application to zero-shot intent classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 3532 pages 1050–1060, Online. Association for Computational Linguistics. Hanlei Zhang, Hua Xu, and Ting-En Lin. 2021. Deep open intent classification with adaptive decision boundary. Proceedings of the AAAI Conference on Artificial Intelligence, 35(16):14374–14382. Hongyi Zhang, Moustapha Ciss´e, Yann N. Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. Jianguo Zhang, Kazuma Hashimoto, Wenhao Liu, Chien-Sheng Wu, Yao Wan, Philip Yu, Richard Socher, and Caiming Xiong. 2020. Discriminative nearest neighbor few-shot intent detection by transferring natural language inference. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5064–5082, Online. Association for Computational Linguistics. Arthur Zimek, Erich Schubert, and Hans-Peter Kriegel. 2012. A survey on unsupervised outlier detection in high-dimensional numerical data. Statistical Analysis and Data Mining: The ASA Data Science Journal, 5(5):363–387.
|
2021
|
273
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3533–3546 August 1–6, 2021. ©2021 Association for Computational Linguistics 3533 Document-level Event Extraction via Heterogeneous Graph-based Interaction Model with a Tracker Runxin Xu1, Tianyu Liu1, Lei Li3 and Baobao Chang1,2∗ 1Key Laboratory of Computational Linguistics, Peking University, MOE, China 2Peng Cheng Laboratory, Shenzhen, China 3ByteDance AI Lab [email protected],[email protected] {tianyu0421,chbb}@pku.edu.cn Abstract Document-level event extraction aims to recognize event information from a whole piece of article. Existing methods are not effective due to two challenges of this task: a) the target event arguments are scattered across sentences; b) the correlation among events in a document is non-trivial to model. In this paper, we propose Heterogeneous Graph-based Interaction Model with a Tracker (GIT) to solve the aforementioned two challenges. For the first challenge, GIT constructs a heterogeneous graph interaction network to capture global interactions among different sentences and entity mentions. For the second, GIT introduces a Tracker module to track the extracted events and hence capture the interdependency among the events. Experiments on a large-scale dataset (Zheng et al., 2019) show GIT outperforms the existing best methods by 2.8 F1. Further analysis reveals GIT is effective in extracting multiple correlated events and event arguments that scatter across the document. Our code is available at https: //github.com/RunxinXu/GIT. 1 Introduction Event Extraction (EE) is one of the key and challenging tasks in Information Extraction (IE), which aims to detect events and extract their arguments from the text. Most previous methods (Chen et al., 2015; Nguyen et al., 2016; Liu et al., 2018; Yang et al., 2019; Du and Cardie, 2020b) focus on sentence-level EE, extracting events from a single sentence. The sentence-level model, however, fails to extract events whose arguments spread in multiple sentences, which is much more common in real-world scenarios. Hence, extracting events at the document-level is critical. It has attracted much attention recently (Yang et al., 2018; Zheng et al., 2019; Du and Cardie, 2020a; Du et al., 2020). *Corresponding author. [1] On Nov 6, 2014, the company received a letter of share reduction from Mingting Wu, the shareholder of the company. [2] Mingting Wu decreased his holding of 7.2 million shares of the company on the Shenzhen Stock Exchange on Nov 6, 2014. [3] The 7.2 million shares of the company Mingting Wu reduced this time were transferred to Xiaoting Wu. [4] Xiaoting Wu is the daughter of Mingting Wu, and they were identified as persons acting in concert according to relevant regulations. EventTypeEquityHolderTradedShares StartDate 7.2 million Nov 6, 2014 Xiaoting Wu EO EU Mingting Wu 7.2 million Nov 6, 2014 … … … Figure 1: An example document from a Chinese dataset proposed by Zheng et al. (2019) in the financial domain, and we translate it into English for illustration. Entity mentions are colored. Due to space limitation, we only show four associated sentences and three argument roles of each event type. The complete original document can be found in Appendix C. EU: Equity Underweight, EO: Equity Overweight. Though promising, document-level EE still faces two critical challenges. Firstly, the arguments of an event record may scatter across sentences, which requires a comprehensive understanding of the cross-sentence context. Figure 1 illustrates an example that one Equity Underweight (EU) and one Equity Overweight (EO) event records are extracted from a financial document. It is less challenging to extract the EU event because all the related arguments appear in the same sentence (Sentence 2). However, for the arguments of EO record, Nov 6, 2014 appears in Sentence 1 and 2 while Xiaoting Wu in Sentence 3 and 4. It would be quite challenging to identify such events without considering global interactions among sentences and entity mentions. Secondly, a document may express several correlated events simultaneously, and recognizing the interdependency among them is 3534 fundamental to successful extraction. As shown in Figure 1, the two events are interdependent because they correspond to exactly the same transaction and therefore share the same StartDate. Effective modeling on such interdependency among the correlated events remains a key challenge in this task. Yang et al. (2018) extracts events from a central sentence and query the neighboring sentences for missing arguments, which ignores the cross-sentence correspondence between augments. Though Zheng et al. (2019) takes a first step to fuse the sentences and entities information via Transformer, they neglect the interdependency among events. Focusing on single event extraction, Du and Cardie (2020a) and Du et al. (2020) concatenate multiple sentences and only consider a single event, which lacks the ability to model multiple events scattered in a long document. To tackle the aforementioned two challenges, in this paper, we propose a Heterogeneous Graphbased Interaction Model with a Tracker (GIT) for document-level EE. To deal with scattered arguments across sentences, we focus on the Global Interactions among sentences and entity mentions. Specifically, we construct a heterogeneous graph interaction network with mention nodes and sentence nodes, and model the interactions among them by four types of edges (i.e., sentence-sentence edge, sentence-mention edge, intra-mention-mention edge, and inter-mentionmention edge) in the graph neural network. In this way, GIT jointly models the entities and sentences in the document from a global perspective. To facilitate the multi-event extraction, we target on the Global Interdependency among correlated events. Concretely we propose a Tracker module to continually tracks the extracted event records with a global memory. In this way, the model is encouraged to incorporate the interdependency with other correlated event records while predicting. We summarize our contributions as follows: • We construct a heterogeneous graph interaction network for document-level EE. With different heterogeneous edges, the model could capture the global context for the scattered event arguments across different sentences. • We introduce a novel Tracker module to track the extracted event records. The Tracker eases the difficulty of extracting correlated events, as interdependency among events would be taken into consideration. • Experiments show GIT outperforms the previous state-of-the-art model by 2.8 F1 on the large-scale public dataset (Zheng et al., 2019) with 32, 040 documents, especially on crosssentence events and multiple events scenarios (with 3.7 and 4.9 absolute increase on F1). 2 Preliminaries We first clarify some important notions. a) entity mention: a text span within document that refers to an entity object; b) event argument: an entity playing a specific event role. Event roles are predefined for each event type; c) event record: an entry of a specific event type containing arguments for different roles in the event. For simplicity, we use record for short in the following sections. Following Zheng et al. (2019), given a document composed of sentences D = {si}|D| i=1 and a sentence containing a sequence of words si = {wj}|si| j=1, the task aims to handle three sub-tasks : 1) entity extraction: extracting entities E = {ei}|E| i=1 from the document to serve as argument candidates. An entity may have multiple mentions across the document. 2) event types detection: detecting specific event types that are expressed by the document. 3) event records extraction: finding appropriate arguments for the expressed events from entities, which is the most challenging and also the focus of our paper. The task does not require to identify event triggers (Zeng et al., 2018; Liu et al., 2019b), which reduces manual effort of annotation and the application scenarios becomes more extensive. 3 Methodology As shows in Figure 2, GIT first extracts candidate entities through sentence-level neural extractor (Sec 3.1). Then we construct a heterogeneous graph to model the interactions among sentences and entity mentions (Sec 3.2), and detect event types expressed by the document (Sec 3.3). Finally we introduce a Tracker module to continuously track all the records with global memory, in which we utilize the global interdependency among records for multi-event extraction (Sec 3.4). 3.1 Entity Extraction Given a sentence s = {wj}|s| j=1 ∈D, we encode s into a sequence of vectors {gj}|si| j=1 using Trans3535 Global Memory Encoder CRF layer Classifier EntityPledge EntityFreeze EntityOverweight … … Decoding Module [1] Mingting Wu decreased ... 7.2 million shares ... on Nov 6, 2014. [2] The 7.2 million shares ... Mingting Wu … to Xiaoting Wu. [3] Xiaoting Wu is the daughter of Mingting Wu … 1 2 3 N M M M X X S S M X N S 1 2 3 Record 1 Record 2 Representations Document Sentences Heterogeneous Interaction Graph Network Types Detection and Records Extraction M S X N 1 2 3 Mention Node Sentence Node Sentence-Sentence Edge Sentence-Mention Edge Intra-Mention-Mention Edge Inter-Mention-Mention Edge Tracker 𝓛𝐧𝐞𝐫 𝓛𝐫𝐞𝐜𝐨𝐫𝐝 𝓛𝐝𝐞𝐭𝐞𝐜𝐭 Figure 2: Overview of our GIT. Firstly, sentences of the document are fed into the encoder to obtain contextualized representation, followed by a CRF layer to extract entities. Then GIT constructs a heterogeneous graph interaction network with mention nodes and sentence nodes, which captures the global interactions among them based on GCNs. After obtaining document-aware representations of entities and sentences, GIT detects event types and extracts records through the decoding module with a Tracker. The Tracker tracks extracted records with global memory, based on which the decoding module incorporates global interdependency among correlated event records. Different entities are marked by different colors. M: Mingting Wu. X: Xiaoting Wu. N: Nov 6, 2014. S: 7.2 million. former (Vaswani et al., 2017): {g1, . . . , g|s|} = Transformer({w1, . . . , w|s|}) The word representation of wj is a sum of the corresponding token and position embeddings. We extract entities at the sentence level and formulate it as a sequence tagging task with BIO (Begin, Inside, Other) schema. We leverage a conditional random field (CRF) layer to identify entities. For training, we minimize the following loss: Lner = − X s∈D log P(ys|s) (1) where ys is the golden label sequence of s. For inference, we use Viterbi algorithm to decode the label sequence with the maximum probability. 3.2 Heterogeneous Graph Interaction Network An event may span multiple sentences in the document, which means its corresponding entity mentions may also scatter across different sentences. Identifying and modeling these entity mentions in the cross-sentence context is fundamental in document EE. Thus we build a heterogeneous graph G which contains entity mention nodes and sentence nodes in the document D. In the graph G, interactions among multiple entity mentions and sentences can be explicitly modeled. For each entity mention node e, we initialize node embedding h(0) e = Mean({gj}j∈e) by averaging the representation of the contained words. For each sentence node s, we initialize node embedding h(0) s = Max({gj}j∈s) + SentPos(s) by maxpooling all the representation of words within the sentence plus sentence position embedding. To capture the interactions among sentences and mentions, we introduce four types of edges. Sentence-Sentence Edge (S-S) Sentence nodes are fully connected to each other with S-S edges. In this way, we can easily capture the global properties in the document with sentence-level interactions, e.g., the long range dependency between any two separate sentences in the document would be modeled efficiently with S-S edges. Sentence-Mention Edge (S-M) We model the local context of an entity mention in a specific sentence with S-M edge, specifically the edge connecting the mention node and the sentence node it belongs to. Intra-Mention-Mention Edge (M-Mintra) We connect distinct entity mentions in the same sentences with M-Mintra edges. The co-occurrence of mentions in a sentence indicates those mentions are likely to be involved in the same event. We 3536 explicitly model this indication by M-Mintra edges. Inter-Mention-Mention Edge (M-Minter) The entity mentions that corresponds to the same entity are fully connected with each other by M-Minter edges. As in document EE, an entity usually corresponds to multiple mentions across sentences, we thus use M-Minter edge to track all the appearances of a specific entity, which facilitates the long distance event extraction from a global perspective. In Section. 4.5, experiments show that all of these four kinds of edges play an important role in event detection, and the performance would decrease without any of them. After heterogeneous graph construction *, we apply multi-layer Graph Convolution Network (Kipf and Welling, 2017) to model the global interactions inspired by Zeng et al. (2020). Given node u at the l-th layer, the graph convolutional operation is defined as follows: h(l+1) u = ReLU X k∈K X v∈Nk(u) S{u} 1 cu,k W (l) k h(l) v where K represents different types of edges, W (l) k ∈Rdm×dm is trainable parameters. Nk(u) denotes the neighbors for node u connected in k-th type edge and cu,k is a normalization constant. We then derive the final hidden state hu for node u, hu = Wa[h(0) u ; h(1) u ; . . . ; h(L) u ] where h(0) u is the initial node embedding of node u, and L is the number of GCN layers. Finally, we obtain the sentence embedding matrix S = [h⊤ 1 h⊤ 2 . . . h⊤ |D|] ∈Rdm×|D| and entity embedding matrix E ∈Rdm×|E|. The i-th entity may have many mentions, where we simply use string matching to detect entity coreference following Zheng et al. (2019) , and the entity embedding Ei is computed by the average of its mention node embedding, Ei = Mean({hj}j∈Mention(i)). In this way, the sentences and entities are interactively represented in a context-aware way. 3.3 Event Types Detection Since a document can express events of different types, we formulate the task as a multi-label classification and leverage sentences feature matrix S to *Traditional methods in sentence-level EE also utilize graph to extract events (Liu et al., 2018; Yan et al., 2019), based on the dependency tree. However, our interaction graph is heterogeneous and have no demands for dependency tree. EquityFreeze EquityHolder FrozeShares StartDate LegalInstitution A C D E F H I EquityPledge Pledger A B K Pledgee StartDate A D F I Global Memory A B Completed Uncompleted A … B C G J A K … Tracker Virtual Node Virtual Node B J C G A C E H E F Figure 3: The decoding module of GIT. Three Equity Freeze records have been extracted completely, and GIT is predicting the StartDate role for the Equity Pledge records (in the dashed frame ), based on the global memory where Tracker tracks the records on-the-fly. Both entity E and F are predicted as the legal StartDate role while A is not. Pre-defined argument roles are shown in the blue box, and GIT extracts records in this order. Capital letters (A-K) refer to different entities. A path from root to leaf node represents one unique event record. detect event types: A = MultiHead(Q, S, S) ∈Rdm×T R = Sigmoid(A⊤Wt) ∈RT where Q ∈Rdm×T and Wt ∈Rdm are trainable parameters, and T denotes the number of possible event types. MultiHead refers to the standard multi-head attention mechanism with Query/Key/Value. Therefore, we derive the event types detection loss with golden label bR ∈RT : Ldetect = − T X t=1 I bRt = 1 log P (Rt|D) + I bRt = 0 log (1 −P (Rt|D)) (2) 3.4 Event Records Extraction Since a document is likely to express multiple event records and the number of records cannot be known in advance, we decode records by expanding a tree orderly as previous methods did (Zheng et al., 2019). However, they treat each record independently. Instead, to incorporate the interdependency among event records, we propose a Tracker module, which improves the model performance. To be self-contained, we introduce the ordered tree expanding in this paragraph. In each step, 3537 we extract event records of a specific event type. The arguments extraction order is predefined so that the extraction is modeled as a constrained tree expanding task†. Taking Equity Freeze records as an example, as shown in Figure 3, we firstly extract EquityHolder, followed by FrozeShares and others. Starting from a virtual root node, the tree expands by predicting arguments in a sequential order. As there may exist multiple eligible entities for the event argument role, the current node will expand several branches during extraction, with different entities assigned to the current role. This branching operation is formulated as multi-label classification task. In this way, each path from the root node to the leaf node is identified as a unique event record. Interdependency exists extensively among different event records. For example, as shown in Figure 1, an Equity Underweight event record is closely related to an Equity Overweight event record, and they may share some key arguments or provide useful reasoning information. To take advantage of such interdependency, we propose a novel Tracker module inspired by memory network (Weston et al., 2015). Intuitively, the Tracker continually tracks the extracted records on-the-fly and store the information into a global memory. When predicting arguments for current record, the model will query the global memory and therefore make use of useful interdependency information of other records. In detail, for the i-th record path consisting of a sequence of entities, the Tracker encodes the corresponding entity representation sequence Ui = [Ei1, Ei2, ...] into an vector Gi with an LSTM (last hidden state) and add event type embedding. Then the compressed record information is stored in the global memory G, which is shared across different event types as shown in Figure 3. For extraction, given a record path Ui ∈Rdm×(J−1) with the first J −1 arguments roles, we predict the J-th role by injecting role-specific information into entity representations, E = E + RoleJ, where RoleJ is the role embedding for the J-th role. Then we concatenate E, sentences feature S, current entities path Ui, and the global memory G, followed by a transformer to obtain new entity feature matrix eE ∈Rdm×|E|, which contains global role-specific †We simply adopt the order used by Zheng et al. (2019). information for all entity candidates.‡ [ eE, eS, eUi, eG] = Transformer([E; S; Ui; G]) We treat the path expansion as a multi-label classification problem with a binary classifier over eEi, i.e., predicts whether the i-th entity is the next argument role for the current record and expand the path accordingly as shown in Figure 3. During training, we minimize the following loss: Lrecord = − X n∈ND |E| X t=1 log P(yn t |n) (3) where ND denotes the nodes set in the event records tree, and yn t is the golden label. If the t-th entity is validate for the next argument in node n, then yn t = 1, otherwise yn t = 0. 3.5 Training We sum the losses coming from three sub-tasks with different weight respectively in Eq. (1), (2) and (3) as follows: Lall = λ1Lner + λ2Ldetect + λ3Lrecord More training details are shown in Appendix A. 4 Experiments 4.1 Dataset We evaluate our model on a public dataset proposed by Zheng et al. (2019)§, which is constructed from Chinese financial documents. It consists of up to 32, 040 documents which is the largest documentlevel EE dataset by far. It focuses on five event types: Equity Freeze (EF), Equity Repurchase (ER), Equity Underweight (EU), Equity Overweight (EO) and Equity Pledge (EP), with 35 different kinds of argument roles in total. We follow the standard split of the dataset, 25, 632/3, 204/3, 204 documents for training/dev/test set. The dataset is quite challenging, as a document has 20 sentences and consists of 912 tokens on average. Besides, there are roughly 6 sentences involved for an event record, and 29% documents express multiple events. ‡To distinguish different parts in the concatenated vector, we also add segment embedding, which is omitted in Eq. 3.4. §https://github.com/dolphin-zs/ Doc2EDAG/blob/master/Data.zip 3538 Model EF ER EU EO EP Overall DCFEE-S 46.7 80.0 47.5 46.7 56.1 60.3 DCFEE-M 42.7 73.3 45.8 44.6 53.8 56.6 Greedy-Dec 57.7 79.4 51.2 50.0 54.2 61.0 Doc2EDAG 71.0 88.4 69.8 73.5 74.8 77.5 GIT (ours) 73.4 90.8 74.3 76.3 77.7 80.3 Table 1: F1 scores on test set. GIT achieves the best performance. We also list the results reported in Zheng et al. (2019) in Appendix B, and GIT consistently outperforms other baselines. EF/ER/EU/EO/EP refer to specific event types, and Overall denotes micro F1. 4.2 Experiments Setting In our implementation of GIT, we use 8 and 4 layers Transformer (Vaswani et al., 2017) in encoding and decoding module respectively. The dimensions in hidden layers and feed-forward layers are the same as previous work (Zheng et al., 2019), i.e., 768 and 1, 024. We also use L = 3 layers of GCN, and set dropout rate to 0.1, batch size to 64. GIT is trained using Adam (Kingma and Ba, 2015) as optimizer with 1e −4 learning rate for 100 epochs. We set λ1 = 0.05, λ2 = λ3 = 1 for the loss function. 4.3 Baselines and Metrics Yang et al. (2018) proposes DCFEE that extracts arguments from the identified central sentence and queries surrounding sentences for missing arguments. The model has two variants, DCFEE-S and DCFEE-M. DCFEE-S produces one record at a time, while DCFEE-M produces multiple possible argument combinations by the closest distance from the central sentence. Besides, Doc2EDAG (Zheng et al., 2019) uses transformer encoder to obtain sentence and entity embeddings, followed by another transformer to fuse cross-sentence context. Then multiple events are extracted simultaneously. Greedy-Dec is a variant of Doc2EDAG, which produces only one record greedily. Three sub-tasks of the document-level EE are all evaluated by F1 score. Due to limited space, we leave the results of entity extraction and event types detection in Appendix B, which shows GIT only slightly outperform Doc2EDAG, because we mainly focus on event record extraction and the methods are similar to Doc2EDAG for these two sub-tasks. In the following, we mainly report and analyze the results of event record extraction. Model I II III IV DCFEE-S 64.6 70.0 57.7 52.3 DCFEE-M 54.8 54.1 51.5 47.1 Greedy-Dec 67.4 68.0 60.8 50.2 Doc2EDAG 79.6 82.4 78.4 72.0 GIT (ours) 81.9 85.7 80.0 75.7 Table 2: F1 scores on four sets with growing average number of involved sentences for records (increases from I to IV). The highest improvement of GIT comes from event records involving the most sentences (Set IV) by 3.7 F1 score compared with Doc2EDAG. 4.4 Main Results Overall performance. The results of the overall performance on the document-level EE dataset is illustrated in Table 1. As Table 1 shows, our GIT consistently outperforms other baselines, thanks to better modelling of global interactions and interdependency. Specifically, GIT improves 2.8 micro F1 compared with the previous state-of-the-art, Doc2EDAG, especially 4.5 improvement in Equity Underweight (EU) event type. Cross-sentence records scenario. There are more than 99.5% records of the test set are crosssentence event records, and the extraction becomes gradually more difficult as the number of their involved sentences grows. To verifies the effectiveness of GIT to capture cross-sentence information, we first calculate the average number of sentences that the records involve for each document, and sort them in ascending order. Then we divide them into four sets I/II/III/IV with equal size. Documents in Set. IV is considered to be the most challenging as it requires the most number of sentences to successfully extract records. As Table 2 shows, GIT consistently outperforms Doc2EDAG, especially on the most challenging Set. IV that involves the most sentences, by 3.7 F1 score. It suggests that GIT can well capture global context and mitigate the arguments-scattering challenge, with the help of the heterogeneous graph interaction network. Multiple records scenario. GIT introduces the tracker to make use of global interdependency among event records, which is important in multiple records scenario. To illustrate its effectiveness, we divide the test set into single-record set (S.) containing documents with one record, and multi-record set (M.) containing those with multiple records. As shown in Table. 3, F1 score on M. 3539 Model EF ER EU EO EP Overall S. M. S. M. S. M. S. M. S. M. S. M. DCFEE-S 55.7 38.1 83.0 55.5 52.3 41.4 49.2 43.6 62.4 52.2 69.0 50.3 DCFEE-M 45.3 40.5 76.1 50.6 48.3 43.1 45.7 43.3 58.1 51.2 63.2 49.4 Greedy-Dec 74.0 40.7 82.2 50.0 61.5 35.6 63.4 29.4 78.6 36.5 77.8 37.0 Doc2EDAG 79.7 63.3 90.4 70.7 74.7 63.3 76.1 70.2 84.3 69.3 81.0 67.4 GIT (ours) 81.9 65.9 93.0 71.7 82.0 64.1 80.9 70.6 85.0 73.5 87.6 72.3 Table 3: F1 scores on single-record (S.) and multi-record (M.) sets. Model F1 I II III IV GIT 80.3 81.9 85.7 80.0 75.7 - S-S -1.4 -0.9 -0.1 -1.9 -2.3 - S-M -1.0 -1.6 -1.7 -0.7 -0.7 - M-Mintra -1.3 -0.5 -1.4 -2.4 -1.5 - M-Minter -1.1 -0.5 -1.6 -1.4 -1.7 - Graph -2.0 -1.8 -1.5 -2.0 -2.5 Table 4: The decrease of F1 scores on ablation study for GIT’s heterogeneous graph interaction network. Removing the heterogeneous graph leads to significant drop on F1, especially for records involving the most sentences (i.e., −2.5 F1 on Set IV). Model P R F1 S. M. GIT 82.3 78.4 80.3 87.6 72.3 GIT-OT -0.6 -0.4 -0.5 -0.8 -0.7 GIT-OP -1.0 -1.6 -1.2 -1.0 -1.5 GIT-NT -2.8 +0.1 -1.3 -1.3 -1.5 Table 5: Performance of GIT on ablation study for the Tracker module. The removal of the Tracker (GITNT) brings about higher F1 decrease on M. than that on S.. S.: Single-record set, M.: Multi-record set. is much lower than that on S., indicating it is challenging to extract multiple records. However, GIT still surpasses other strong baselines by 4.9 ∼35.3 on multi-record set (M.). This is because GIT is aware of other records through the Tracker module, and leverage the interdependency information to improve the performance¶. ¶Nguyen et al. (2016) maintain three binary matrices to memorize entities and events states. Although they aim at sentence-level EE that contains fewer entities and event records, it would be also interesting to compare with them and we leave it as future work. 63 65 67 69 71 73 2 - 3 4 - 5 >=6 F1 score The number of records of documents GIT GIT-OT GIT-NT Doc2EDAG Figure 4: F1 scores on documents with different number of event records. The F1 gap between w/ (GIT) and w/o Tracker (GIT-NT) becomes wider as the number of event records of documents increases. 4.5 Analysis We conduct further experiments to analyze the key modules in GIT more deeply. On the effect of heterogeneous graph interaction network. The heterogeneous graph we constructed contains four types of edges. To explore their functions, we remove one type of edges at a time, and remove the whole graph network finally. Results are shown in Table 4, including micro F1 and F1 on the four sets, which are divided by the number of involved sentences for records as we did before. The micro F1 would decreases 1.0 ∼1.4 without a certainty type of edge. Besides, removing the whole graph causes an significant drop by 2.0 F1, especially for Set IV by 2.5, which requires the most number of sentences to extract the event record. It demonstrates that the graph interaction network helps improve the performance, especially on records involving many sentences, and all kinds of edges play an important role for extraction. On the effect of Tracker module. GIT can leverage interdependency among records based on the information of other event records tracked by Tracker. To explore its effect, firstly, we remove the global interdependency information between records of different event types, by clearing the global memory whenever we extract events for an3540 … [5] The shareholder of the company, Quanlie Chen, pledged 52.4 million to GDZQ Co., Ltd. in 2018, and supplemented the pledge recently because of the decline of the share price. … [7] Since the borrowings have been paid off, Quanlie Chen completed the pledge cancellation procedures of 35.5 million that were pledged to GTJA Co., Ltd. on Nov 7, 2018. [8] As of today, Quanlie Chen holds a total of 325.4 million of the company, and there are still 218.6 million in pledge status. … Quanlie Chen Quanlie Chen Pledger PledgedShares Pledgee TotalHoldingShares TotalPledgedShares 35.5 million GTJA Co., Ltd. 325.4 million 218.6 million 52.4 million GDZQ Co., Ltd. NULL NULL … … … Doc2EDAG Quanlie Chen Quanlie Chen Pledger PledgedShares Pledgee TotalHoldingShares TotalPledgedShares 35.5 million GTJA Co., Ltd. 325.4 million 218.6 million 52.4 million GDZQ Co., Ltd. 325.4 million 218.6 million … … … GIT No. 1 2 No. 1 2 Figure 5: The case study of our proposed GIT and Doc2EDAG, with their key prediction difference colored in red. Related entities are colored in blue. GIT successfully extract TotalHoldingShares and TotalPledgedShares for Record 2, while Doc2EDAG fails. The complete content are provided in Appendix C. other new event type (GIT-Own Type). Next, we remove all the tracking information except the own path for a record, to explore whether the tracking of other records makes effect indeed (GIT-Own Path). Finally, we remove the whole Tracker module (GIT-No Tracker). As Table 5 shows, the F1 in GIT-OT/GIT-OP decreases by 0.5/1.2, suggesting the interdependency among records of both the same and different event types do play an essential role. Besides, their F1 decrease in M. by 0.7/1.5 are more than those in S. by 0.8/1.0, verifying the effectiveness of the Tracker in multi-event scenarios. Moreover, the performances are similar between GIT-OP and GIT-NT, which also provides evidence that other records do help. We also reveal F1 on documents with different number of records in Figure 4. The gap between models with or without Tracker raises as the number of records increases, which validates the effectiveness of our Tracker. 4.6 Case Study Figure 5 demonstrates a case of the predictions of Doc2EDAG and GIT for Equity Pledge (EP) event types. The TotalHoldingShares and TotalPledgedShares information lies in Sentence 8, while the PledgedShares and Pledgee information for Record 2 lies in Sentence 5. Though Doc2EDAG fails to extract these arguments in Record 2 (colored in red), GIT succeeds because it can capture interactions between long-distance sentences, and utilize the information of Record 1 (325.4 million and 218.6 million) thanks to the Tracker model. 5 Related Work Sentence-level Event Extraction. Previous approaches mainly focus on sentence-level event extraction. Chen et al. (2015) propose a neural pipeline model that identifies triggers first and then extracts argument roles. Nguyen et al. (2016) use a joint model to extract triggers and argument roles simultaneously. Some studies also utilize dependency tree information (Liu et al., 2018; Yan et al., 2019). To utilize more knowledge, some studies leverage document context (Chen et al., 2018; Zhao et al., 2018), pre-trained language model (Yang et al., 2019), and explicit external knowledge (Liu et al., 2019a; Tong et al., 2020) such as WordNet (Miller, 1995). Du and Cardie (2020b) also try to extract events in a Question-Answer way. These studies usually conduct experiments on sentencelevel event extraction dataset, ACE05 (Walker et al., 2006). However, it is hard for the sentence-level models to extract multiple qualified events spanning across sentences, which is more common in real-world scenarios. Document-level Event Extraction. Documentlevel EE has attracted more and more attention recently. Yang and Mitchell (2016) use well-defined features to handle the event-argument relations across sentences, which is, unfortunately, quite nontrivial. Yang et al. (2018) extract events from a central sentence and find other arguments from neighboring sentences separately. Although Zheng et al. (2019) use Transformer to fuse sentences and entities, interdependency among events is neglected. Du and Cardie (2020a) try to encode the sentences in a multi-granularity way and Du et al. (2020) leverage a seq2seq model. They conduct experiments on MUC-4 (Sundheim, 1992) dataset with 1, 700 documents and 5 kinds of entity-based arguments, and it is formulated as a table-filling task, coping with single event record of single event 3541 type. However, our work is different from these studies in that a) we utilize heterogeneous graph to model the global interactions among sentences and mentions to capture cross-sentence context, b) and we leverage the global interdependency through Tracker to extract multiple event records of multiple event types. 6 Conclusion Although promising in practical application, document-level EE still faces some challenges such as arguments-scattering phenomenon and multiple correlated events expressed by a single document. To tackle the challenges, we introduce Heterogeneous Graph-based Interaction Model with a Tracker (GIT). GIT uses a heterogeneous graph interaction network to model global interactions among sentences and entity mentions. GIT also uses a Tracker to track the extracted records to consider global interdependency during extraction. Experiments on large-scale public dataset (Zheng et al., 2019) show GIT outperforms previous stateof-the-art by 2.8 F1. Further analysis verifies the effectiveness of GIT especially in cross-sentence events extraction and multi-event scenarios. Acknowledgments The authors would like to thank Changzhi Sun, Mingxuan Wang, and the anonymous reviewers for their thoughtful and constructive comments. This paper is supported in part by the National Key R&D Program of China under Grand No.2018AAA0102003, the National Science Foundation of China under Grant No.61936012 and 61876004. References Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. 2015. Scheduled sampling for sequence prediction with recurrent neural networks. In Proceedings of the 28th International Conference on Neural Information Processing Systems (NeurIPS). Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multipooling convolutional neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (ACL-IJCNLP). Yubo Chen, Hang Yang, Kang Liu, Jun Zhao, and Yantao Jia. 2018. Collective event detection via a hierarchical and bias tagging networks with gated multilevel attention mechanisms. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP). Xinya Du and Claire Cardie. 2020a. Document-level event role filler extraction using multi-granularity contextualized encoding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL). Xinya Du and Claire Cardie. 2020b. Event extraction by answering (almost) natural questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Xinya Du, Alexander M. Rush, and Claire Cardie. 2020. Document-level event-based extraction using generative template-filling transformers. arXiv preprint arXiv:2008.09249. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations (ICLR). Thomas N. Kipf and Max Welling. 2017. Semisupervised classification with graph convolutional networks. In 5th International Conference on Learning Representations (ICLR). Jian Liu, Yubo Chen, and Kang Liu. 2019a. Exploiting the ground-truth: An adversarial imitation based knowledge distillation approach for event detection. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI). Shulin Liu, Yang Li, Feng Zhang, Tao Yang, and Xinpeng Zhou. 2019b. Event detection without triggers. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT). Xiao Liu, Zhunchen Luo, and Heyan Huang. 2018. Jointly multiple events extraction via attentionbased graph information aggregation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP). George A. Miller. 1995. Wordnet: A lexical database for english. Commun. ACM. Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. In NIPS-W. 3542 Beth M. Sundheim. 1992. Overview of the fourth Message Understanding Evaluation and Conference. In Fourth Message Uunderstanding Conference (MUC4). Meihan Tong, Bin Xu, Shuai Wang, Yixin Cao, Lei Hou, Juanzi Li, and Jun Xie. 2020. Improving event detection via open-domain trigger knowledge. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL). Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS). Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. In Philadelphia: Linguistic Data Consortium. Minjie Wang, Lingfan Yu, Da Zheng, Quan Gan, Yu Gai, Zihao Ye, Mufei Li, Jinjing Zhou, Qi Huang, Chao Ma, Ziyue Huang, Qipeng Guo, Hao Zhang, Haibin Lin, Junbo Zhao, Jinyang Li, Alexander J Smola, and Zheng Zhang. 2019. Deep graph library: Towards efficient and scalable deep learning on graphs. ICLR Workshop on Representation Learning on Graphs and Manifolds. Jason Weston, Sumit Chopra, and Antoine Bordes. 2015. Memory networks. In 3rd International Conference on Learning Representations (ICLR). Haoran Yan, Xiaolong Jin, Xiangbin Meng, Jiafeng Guo, and Xueqi Cheng. 2019. Event detection with multi-order graph convolution and aggregated attention. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). Bishan Yang and Tom M. Mitchell. 2016. Joint extraction of events and entities within a document context. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL). Hang Yang, Yubo Chen, Kang Liu, Yang Xiao, and Jun Zhao. 2018. DCFEE: A document-level Chinese financial event extraction system based on automatically labeled training data. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL). Sen Yang, Dawei Feng, Linbo Qiao, Zhigang Kan, and Dongsheng Li. 2019. Exploring pre-trained language models for event extraction and generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL). Shuang Zeng, Runxin Xu, Baobao Chang, and Lei Li. 2020. Double graph based reasoning for documentlevel relation extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics. Ying Zeng, Yansong Feng, Rong Ma, Zheng Wang, Rui Yan, Chongde Shi, and Dongyan Zhao. 2018. Scale up event extraction learning via automatic training data generation. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence (AAAI). Yue Zhao, Xiaolong Jin, Yuanzhuo Wang, and Xueqi Cheng. 2018. Document embedding enhanced event detection with hierarchical and supervised attention. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL). Shun Zheng, Wei Cao, Wei Xu, and Jiang Bian. 2019. Doc2EDAG: An end-to-end document-level framework for Chinese financial event extraction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). A Training Details To mitigate the error propagation due to the gap between training and inference phrase (i.e., the extracted entities are ground truth during training but predicted results during inference), we adopt scheduled sampling strategy (Bengio et al., 2015) as Zheng et al. (2019) did. We gradually switch the entity extraction results from golden label to what the model predicts on its own. Specifically, from epoch 10 to epoch 20, we linearly increase the proportion of predicted entity results from 0% to 100%. We implement GIT under PyTorch (Paszke et al., 2017) and DGL (Wang et al., 2019) based on codes provided by Zheng et al. (2019). All the experiments (including the baselines) are run with the same 8 Tesla-V100 GPUs and the same version of python dependencies to ensure the fairness. Hyperparameters trials are listed in Table 6. The value of hyperparameters we finally adopted are in bold. Note that we do not tune all the hyperparameters, and make little effort to select the best hyperparameters for our GIT. We choose the final checkpoints for test according to the Micro F1 performance on the dev set. Table 9 illustrates the best epoch in which the model achieves the highest Micro F1 on the dev set and their according F1 score. B Additional Evaluation Results We have showed the evaluation results of event records extraction in the paper for document-level 3543 [1] 证券代码：002102证券简称：冠福股份编号：2018-112。 [2] 冠福控股股份有限公司关于大股东陈烈权先生部分股份补充质押及解除质押的公告。 [3] 本公司及董事会全体成员保证信息披露的内容真实、准确、完整，没有虚假记载、误导性陈述或者重大遗漏。 [4] 冠福控股股份有限公司（以下简称“公司”）近日接到公司大股东陈烈权先生函告，获悉其将持有的公司部分股份 办理了补充质押及解押，具体情况如下。 [5] 一、本次股份补充质押情况。公司大股东陈烈权先生原于2017年10月24日质押给国泰君安证券股份有限公司（以下 简称“国泰君安”）的公司股份69200000股、2018年2月8日质押给中信建投证券股份有限公司（以下简称“中信建投”） 的公司股份52000000股、2018年2月26日质押给国都证券股份有限公司（以下简称“国都证券”）的公司股份52369050股， 因公司近日股价下跌，分别对国君证券、中信建投及国都证券进行补充质押。 [6] 上述原有质押情况详见公司分别于2017年10月27日、2018年2月12日、3月1日在《证券时报》、《中国证券报》、 《上海证券报》和《证券日报》及巨潮资讯网上披露的《冠福控股股份有限公司关于大股东陈烈权先生部分股份质押及 解除质押的公告》（公告编号：2017-108）、《冠福控股股份有限公司关于大股东陈烈权先生部分股份解除质押及再质 押的公告》（公告编号：2018-010、2018-013）。 [7] 二、本次股份解除质押情况。陈烈权先生原质押给国泰君安的公司股份35500000股（占公司总股本的1.35%），因已 还清国泰君安的借款，分别于2018年9月7日、9月10日在国泰君安证券股份有限公司荆州便河东路营业部办理完成质押解 除手续。 [8] 三、累计质押情况。截止本公告日，陈烈权先生共持有公司股份325363822股，占公司总股本的12.35%，其中处于质 押状态的股份累计数为218569050股，占公司总股本的8.30%。 [9] 四、备查文件 [10] 1、中信建投证券股份有限公司股票质押式回购交易申请书（补充交易）； [11] 2、国都证券股份有限公司股票质押式回购交易补充质押已达成通知； [12] 3、国泰君安证券股份有限公司股票质押式回购交易协议书。 [13] 特此公告。 [14] 冠福控股股份有限公司董事会 [15] 二○一八年九月十二日 Figure 6: The original complete document corresponding to the case study in Figure 5. Sentences in red color are presented in Figure 5. Hyperparameters Value Batch Size 32, 64 Learning Rate 0.0001 Dropout 0.1 Layers of GCN 1, 2, 3, 4, 5 Number of Epochs 100 λ1 0.05 λ2 1.00 λ3 1.00 Gradient Accumulation Steps 8 Layers of Transformer in Entity Extractor 8 Layers of Transformer in Decoder Module 4 Hyperparameter Search Trials 10 Table 6: Hyperparameters for our proposed GIT. Model P R F1 DCFEE-S 86.5 88.6 87.6 DCFEE-M 86.6 89.0 87.8 Greedy-Dec 87.5 89.8 88.6 Doc2EDAG 88.0 90.0 89.0 GIT (ours) 85.8 92.6 89.1 Table 7: Results of entity extraction sub-task on the test set. The performance of different models are similar, for the reason that they all utilize the same structure and methods to extract entities. event extraction. In this section, we also illustate the results of entity extraction in Table. 7 and event types detection in Table. 8. Moreover, the comprehensive results of event record extraction is shown in Table. 10, including results reported in Zheng et al. (2019) with precison, recall and F1 score. C Complete Document for the Examples We show an example document in Figure 1 in the paper. To better illustrate, we translate it from Chinese into English and make some simplication. Here we present the original complete document example in Figure 7. For the specific meanings of argument roles, we recommend readers to refer to (Zheng et al., 2019). We also demonstrate an case study in Figure 5 in the paper. Now we also show its original Chinese version in Figure 6. 3544 Model EF ER EU EO EP Overall DCFEE-S 81.5 94.0 82.3 85.7 93.8 91.4 DCFEE-M 79.8 92.4 78.9 84.2 92.9 90.0 Greedy-Dec 99.3 99.9 96.8 95.4 99.6 99.0 Doc2EDAG 99.0 99.8 96.8 94.1 99.5 98.9 GIT (ours) 98.8 99.8 97.9 96.6 99.6 99.2 Table 8: F1 scores results of event types detection sub-task on the test set. All the models obtains more than 90.0 micro F1 score. GIT slightly outperform Doc2EDAG. Model Best Epoch EF ER EU EO EP Overall DCFEE-S 86 51.3 73.0 44.1 51.4 58.6 58.7 DCFEE-M 87 52.5 69.1 43.9 47.2 55.9 55.8 Greedy-Dec 90 57.5 76.0 55.1 49.3 57.0 59.1 Doc2EDAG 89 75.2 85.2 71.6 80.0 77.9 78.7 GIT (ours) 89 78.3 87.6 74.7 80.9 79.8 80.7 Table 9: The best epoch in which the models achieve the highest micro F1 score on the dev set and the corresponding performance. 3545 Model EF ER EU EO EP Overall P R F1 P R F1 P R F1 P R F1 P R F1 P R F1 DCFEE-S♦ 66.0 41.6 51.1 84.5 81.8 83.1 62.7 35.4 45.3 51.4 42.6 46.6 64.3 63.6 63.9 DCFEE-M♦ 51.8 40.7 45.6 83.7 78.0 80.8 49.5 39.9 44.2 42.5 47.5 44.9 59.8 66.4 62.9 Greedy-Dec♦ 79.5 46.8 58.9 83.3 74.9 78.9 68.7 40.8 51.2 69.7 40.6 51.3 85.7 48.7 62.1 Doc2EDAG♦ 77.1 64.5 70.2 91.3 83.6 87.3 80.2 65.0 71.8 82.1 69.0 75.0 80.0 74.8 77.3 DCFEE-S♠ 61.1 37.8 46.7 84.5 76.0 80.0 60.8 39.0 47.5 46.9 46.5 46.7 64.2 49.8 56.1 67.7 54.4 60.3 DCFEE-M♠ 44.6 40.9 42.7 75.2 71.5 73.3 51.4 41.4 45.8 42.8 46.7 44.6 55.3 52.4 53.8 58.1 55.2 56.6 Greedy-Dec♠ 78.5 45.6 57.7 83.9 75.3 79.4 69.0 40.7 51.2 64.8 40.6 50.0 82.1 40.4 54.2 80.4 49.1 61.0 Doc2EDAG♠ 78.7 64.7 71.0 90.0 86.8 88.4 80.4 61.6 69.8 77.2 70.1 73.5 76.7 73.0 74.8 80.3 75.0 77.5 GIT (ours)♠ 78.9 68.5 73.4 92.3 89.2 90.8 83.9 66.6 74.3 80.7 72.3 76.3 78.6 76.9 77.7 82.3 78.4 80.3 Table 10: Comprehensive results of event record extraction. Results with ♦are results reported in Zheng et al. (2019). Results with are ♠results we implement on our own. Our GIT consistently outperform other baselines. 3546 [1] 证券代码：300126 证券简称：锐奇股份公告编号: 2014-075。 [2] 上海锐奇工具股份有限公司关于控股股东股份减持计划实施进展的公告。 [3] 本公司及董事会全体成员保证信息披露的内容真实、准确、完整，没有虚假记载、误导性陈述或者 重大遗漏。 [4] 上海锐奇工具股份有限公司(以下简称”公司”)于2014年11月1日在中国证券监督管理委员会指定 的创业板信息披露网站披露了《关于控股股东股份减持计划的公告》(公告编号2014-074)。 [5] 公司于2014年11月6日接到公司控股股东吴明厅先生的《股份减持告知函》。 [6] 吴明厅先生于2014年11月6日通过深圳证券交易所大宗交易方式减持了其直接持有的公司无限售条 件流通股7200000股，占公司目前总股本的2.34%。 [7] 一、股东减持情况。吴明厅先生本次减持的公司股份7200000股为其直接持有的公司无限售条件流 通股，占公司总股本的2.34%，本次减持的公司股份全部转让给吴晓婷女士。 [8] 吴晓婷女士为吴明厅先生的女儿，两人为父女关系，根据相关规定被认定为一致行动人。 [9] 二、其他相关说明。1、本次减持没有违反《深圳证券交易所创业板股票上市规则》、《上市公司 解除限售存量股份转让指导意见》等有关法律法规及公司规章制度。 [10] 2、本次减持不存在违反《证券法》、《上市公司收购管理办法》等法律、行政法规、部门规章、 规范性文件和深圳证券交易所《创业板信息披露业务备忘录第18号：控股股东、实际控制人股份减持信 息披露》等规定的情况。 [11] 3、本次减持后，吴明厅先生直接持有公司总股本的比例下降为32.08%，通过上海瑞浦投资有限公 司持有公司总股本的14.02%，合计持有公司总股本的46.82%，仍为公司控股股东。 [12] 4、本次减持后，吴明厅、上海瑞浦投资有限公司、应媛琳、吴晓依、吴晓婷作为一致行动人，其 所合计持有的公司股份权益并未减少，仍为公司总股本的56.22%。 [13] 三、备查文件。 [14] 1、吴明厅先生的《股份减持告知函》。 [15] 2．深交所要求的其他文件。 [16] 上海锐奇工具股份有限公司董事会。 [17] 2014年11月6日。 EquityHolder TradedShares StartDate EndDate LaterHolding Shares AveragePrice 吴明厅 720000股 2014年11月6日 2014年11月6日 NULL NULL EquityHolder TradedShares StartDate EndDate LaterHolding Shares AveragePrice 吴晓婷 720000股 2014年11月6日 2014年11月6日 720000股 NULL EquityUnderweight EquityOverweight Figure 7: The original complete document corresponding to the running example in Figure 1. Sentences in red color are presented in Figure 1.
|
2021
|
274
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3547–3557 August 1–6, 2021. ©2021 Association for Computational Linguistics 3547 Nested Named Entity Recognition via Explicitly Excluding the Influence of the Best Path Yiran Wang1∗, Hiroyuki Shindo2, Yuji Matsumoto3, Taro Watanabe2 1National Institute of Information and Communications Technology (NICT), Kyoto, Japan 2Nara Institute of Science and Technology (NAIST), Nara, Japan 3RIKEN Center for Advanced Intelligence Project (AIP), Tokyo, Japan [email protected], [email protected], [email protected], [email protected] Abstract This paper presents a novel method for nested named entity recognition. As a layered method, our method extends the prior secondbest path recognition method by explicitly excluding the influence of the best path. Our method maintains a set of hidden states at each time step and selectively leverages them to build a different potential function for recognition at each level. In addition, we demonstrate that recognizing innermost entities first results in better performance than the conventional outermost entities first scheme. We provide extensive experimental results on ACE2004, ACE2005, and GENIA datasets to show the effectiveness and efficiency of our proposed method. 1 Introduction Named entity recognition (NER), as a key technique in natural language processing, aims at detecting entities and assigning semantic category labels to them. Early research (Huang et al., 2015; Ma and Hovy, 2016; Lample et al., 2016) proposed to employ deep learning methods and obtained significant performance improvements. However, most of them assume that the entities are not nested within other entities, so-called flat NER. Inherently, these methods do not work satisfactorily when nested entities exist. Figure 1 displays an example of the nested NER task. Recently, a large number of papers proposed novel methods (Fisher and Vlachos, 2019; Wang et al., 2020) for the nested NER task. Among them, layered methods solve this task through multi-level sequential labeling, in which entities are divided into several levels, where the term level indicates the depth of entity nesting, and sequential labeling is performed repeatedly. As a special case of layered method, Shibuya and Hovy (2020) force the ∗This work was done when the first author was at NAIST. Former Hogwarts headmaster Dumbledore Albus ROLE ROLE ORG ROLE PER PER Figure 1: An example of nested NER. next level entities to locate on the second-best path of the current level search space. Hence, their algorithm can repeatedly detect inner entities through applying a conventional conditional random field (CRF) (Lafferty et al., 2001) and then exclude the obtained best paths from the search space. To accelerate computation, they also designed an algorithm to efficiently compute the partition function with the best path excluded. Moreover, because they search the outermost entities first, performing the second-best path search only on the spans of extracted entities is sufficient, since inner entities can only exist within outer entities. However, we claim that the target path at the next level is neither necessary nor likely to be the second-best path at the current level. Instead, those paths sharing many overlapping labels with the current best path are likely to be the second-best path. Besides, Shibuya and Hovy (2020) reuse the same potential function at all higher levels. Thus, even though they exclude the best path, the influence of the best path is still preserved, since the emission scores of labels on the best path are used in the next level recognition. Moreover, these best path labels are treated as the target labels at the current level. However, if they are not on the best path of the next level, they will be treated as non-target labels at the next level, hence these adversarial optimization goals eventually hurt performance. In this paper, we use a different potential function at each level to solve this issue. We propose to achieve this by introducing an encoder that pro3548 duces a set of hidden states at each time step. At each level, we select some hidden states for entity recognition, then, remove these hidden states which have interaction with the best path labels before moving to the next level. In this way, the emission scores of these best path labels are completely different, so we can explicitly exclude the influence of the best path. Furthermore, we also propose three different selection strategies for fully leveraging information among hidden states. Besides, Shibuya and Hovy (2020) proposed to recognize entities from outermost to inner. We empirically demonstrate that extracting the innermost entities first results in better performance. This may due to the fact that some long entities do not contain any inner entity, so using outermostfirst encoding mixes these entities with other short entities at the same levels, therefore leading encoder representations to be dislocated. In this paper, we convert entities to the IOBES encoding scheme (Ramshaw and Marcus, 1995), and solve nested NER through applying CRF level by level. Our contributions are considered as fourfold, (a) we design a novel nested NER algorithm to explicitly exclude the influence of the best path through using a different potential function at each level, (b) we propose three different selection strategies for fully utilizing information among hidden states, (c) we empirically demonstrate that recognizing entities from innermost to outer results in better performance, (d) and we provide extensive experimental results to demonstrate the effectiveness and efficiency of our proposed method on the ACE2004, ACE2005, and GENIA datasets. 2 Proposed Method Named entities recognition task aims to recognize entities in a given sequence {xt}n t=1. For nested NER some shorter entities may be nested within longer entities, while for flat NER there is no such case. Existing algorithms solve flat NER by applying a sequential labeling method, which assigns each token a label yt ∈Y to determine the span and category of each entity and non-entity simultaneously. To solve nested NER, we follow the previous layered method and extend this sequential labeling method with a multi-level encoding scheme. In this encoding scheme, entities are divided into several levels according to their depths, we apply the sequential labeling method level by level to recognize all entities. 2.1 Encoding Schemes Shibuya and Hovy (2020) proposed to recognize the outermost entities first and recursively detect the nested inner entities. However, we find that detecting from the innermost entities results in better performance. We take the sentence in Figure 1 as an example to illustrate the details of these two encoding schemes. The results of the outermost-first encoding scheme look as follows. (level 1) B-PER I-PER I-PER I-PER E-PER (level 2) B-ROLE I-ROLE E-ROLE B-PER E-PER (level 3) O B-ROLE E-ROLE O O (level 4) O S-ORG S-ROLE O O (level 5) O O O O O (level 6) O O O O O Labels B-, I-, E- indicate the current word is the beginning, the intermediate, and the end of an entity, respectively. Label S- means this is a single word entity, and label O stands for nonentity word. For example, the outermost entity “Former Hogwarts headmaster Albus Dumbledore” appears at the first level, while innermost entities “Hogwarts” and “headmaster” appear at the fourth level. Since there exists no deeper nested entity, the remaining levels contain only label O. In contrast, the innermost-first encoding scheme converts the same example to the following label sequences. (level 1) O S-ORG S-ROLE B-PER E-PER (level 2) O B-ROLE E-ROLE O O (level 3) B-ROLE I-ROLE E-ROLE O O (level 4) B-PER I-PER I-PER I-PER E-PER (level 5) O O O O O (level 6) O O O O O In this encoding scheme, innermost entities “Hogwarts”, “headmaster”, and “Albus Dumbledore” appear at the first level. Note that the innermost-first encoding scheme is not the simple reverse of the outermost-first encoding scheme. For example, the entity “Former Hogwarts headmaster” and the entity “Albus Dumbledore” appear at the same level in the outermost-first scheme but they appear at different levels in the innermost-first scheme. 2.2 Influence of the Best Path Although the second-best path searching algorithm is proposed as the main contribution of Shibuya and Hovy (2020), we claim that forcing the target path at the next level to be the second-best path at 3549 Former Hogwarts headmaster Albus Dumbledore Emb Emb Emb Emb Emb BiLSTM O S-ORG S-ROLE B-PER E-PER O B-ROLE E-ROLE O O x L CRF CRF Figure 2: The architecture of our model. The dotted lines mean these components are shared across levels. the current level is not optimal. As the innermostfirst encoding example above, the best path at level 3 is B-ROLE,I-ROLE,E-ROLE,O,O. Therefore the second-best path is more likely to be one of those paths that share as many as possible labels with the best path, e.g., B-ROLE,I-ROLE,E-ROLE,O,S-ORG, rather than the actual target label sequence at level 4, i.e., B-PER,I-PER,I-PER,I-PER,E-PER, which does not overlap with the best path at all. In addition, Shibuya and Hovy (2020) reuse the same potential function at all higher levels. This indicates that, for instance, at level 3 and time step 1, their model encourages the dot product of the hidden state and the label embedding h⊤ 1 vB-ROLE to be larger than h⊤ 1 vB-PER, while at level 4, the remaining influence of the best path reversely forces h⊤ 1 vB-PER to be larger than h⊤ 1 vB-ROLE. These adversarial optimization goals eventually hurt performance and result in sub-optimal performance. Therefore, the crux of the matter is to introduce different emission scores for different levels. For example, encouraging h3⊤ 1 vB-ROLE > h3⊤ 1 vB-PER at level 3 and encouraging h4⊤ 1 vB-PER > h4⊤ 1 vB-ROLE at level 4 will not lead to adversarial optimization directions anymore, where h3 1 and h4 1 are two distinctive hidden states to be used at levels 3 and 4, respectively. To achieve this goal, we introduce a novel encoder which outputs m hidden states {hl t}m l=1, where m is the number of levels, as an alternative to the conventional encoder which can only output a single hidden state ht ∈Rdh at each time step. To make a distinction between our m hidden states and the conventional single hidden state, we use the term chunk from now on to refer to these hidden states hl t ∈Rdh/m. We restrict chunk dimension to be dh/m, so the total number of parameters remain unchanged. 2.3 Chunk Selection As we mentioned above, our algorithm maintains a chunk set for each time step, through selecting and removing chunks, to exclude the influence of the best path. Naturally, how to select chunk becomes the next detail to be finalized. For clarity, we use notation Hl t to denote the chunk set at level l, and use Hl to refer to all of these chunk sets at level m across time steps, i.e., {Hl t}n t=1. Because we remove one and only one chunk at each time step, |Hl t| + l = m + 1 always holds. An intuitive idea is to follow the original chunk order and simply to select the l-th chunk for level l. At level l, no matter to which label, the emission score is calculated by using hl t. In this way, this naive potential function can be defined as follow, φ (yl t−1, yl t, Hl t) = Ayl t−1,yl t + hl⊤ t vyl t (1) where A ∈R|Y|×|Y| is the transition matrix, Y is the label set, Ayl t−1,yl t indicates the transition score from label yl t−1 to label yl t, and vyl t ∈Rdh/m is the embedding of label yl t. In this case, the l-th chunk hl t ∈Hl t is just the chunk which have an interaction with target label, thus should be removed from Hl t. Hl+1 t = Hl t \ {hl t} (2) One concern of the naive potential function is that it implicitly assumes the outputs of the encoder are automatically arranged in the level order instead of other particular syntactic or semantic order, e.g., the encoder may encodes all LOC related information at the first hd/m dimensions while remaining 3550 Algorithm 1: Training input :first level chunk sets H1 input :target label sequences y1, · · · , ym output :negative log-likelihood L L ←0 for l = 1 to m do L ←L −log p (yl | Hl) for t = 1 to n do Hl+1 t ←Hl t \ {arg max h∈Hl t h⊤vyl t} end end ORG relevant information to the final hd/m dimension. For instance, at level 3 time step 1, naive potential function forces h3⊤ 1 vB-ROLE > h3⊤ 1 vB-PER. But if there exists another chunk, say h5 1, which is more similar to vB-PER, then directly selecting h5 1 and forcing h3⊤ 1 vB-ROLE > h5⊤ 1 vB-PER is more reasonable. Because it makes training harder than the former one, due to h5⊤ 1 vB-PER > h3⊤ 1 vB-PER. In other words, this selection strategy leads to hσ1⊤ t vy1 t > hσ2⊤ t vy2 t > . . . > hσm⊤ t vym t , where σl is the index of selected chunk at level l, but for naive potential function, the inequation above does not always hold. From this aspect, our method can also be considered as selecting the best path in the second-best search space. Therefore, instead of following the original chunk orders, we propose to let each label yj select the most similar chunk to it to obtain an emission score. We denote this definition as max potential function, φ (yl t−1, yl t, Hl t) = Ayl t−1,yl t + max h∈Hl t h⊤vyl t (3) In this case, we update chunk sets by removing these chunks which are selected by the target labels. Hl+1 t = Hl t \ {arg max h∈Hl t h⊤vyl t} (4) Furthermore, since the log-sum-exp operation is a well known differentiable approximation of the max operation, we also introduce it as the third potential function, φ (yl t−1, yl t, Hl t) = Ayl t−1,yl t + log X h∈Hl t exp h⊤vyl t (5) Algorithm 2: Decoding input :first level chunk sets H1 output :recognized entity set E E ←∅ for l = 1 to m do ˆyl ←arg max y′∈Yn p (y′ | Hl) for t = 1 to n do Hl+1 t ←Hl t \ {arg max h∈Hl t h⊤vˆyl t} end E ←E S label-to-entity (ˆyl) end The chunk set is updated in the same way as Equation 4. We refer to this potential function definition as logsumexp in the rest of this paper. 2.4 Embedding Layer Following previous work (Shibuya and Hovy, 2020), we convert words to word embeddings wt ∈Rdw and employ a character-level bidirectional LSTM to obtain character-based word embeddings ct ∈Rdc. The concatenation of them is fed into the encoding layer as the token representation xt = [wt, ct] ∈Rdx. 2.5 Encoding Layer We employ a three-layered bidirectional LSTM to encode sentences and leverage contextual information, {ht}n t=1 = LSTM ({xt}n t=1) (6) where ht ∈Rdh is the hidden state. In contrast to the encoders of previous work, which can only output single hidden states at each time step, we split ht into m chunks, [h1 t , . . . , hm t ] = ht (7) where hj t ∈Rdh/m, and use them as the first level chunk set, i.e., H1 t = {hj t}m j=1, to start recognition. 2.6 Decoding Layer At each level, we run a shared conventional CRF with its corresponding potential function φ (yl t−1, yl t, Hl t) and update the chunk sets until finishing all m levels. On the training stage, we remove chunks according to the selections of the target labels, while on the decoding stage, it depends on the selections of the predicted labels. 3551 2.7 Training and Decoding Following the definition of CRF, the conditional probabilistic function of a given label sequence at l-th level, i.e., yl = {yl t}n t=1, can be defined as, p (yl | Hl) = 1 Z(Hl) exp n X t=1 φ (yl t−1, yl t, Hl t) (8) Z(Hl) = X y′∈Yn exp n X t=1 φ (y′l t−1, y′l t , Hl t) (9) where Z(Hl) is the sum of all paths’ scores and is commonly known as the partition function. We optimize our model by minimizing the sum of the negative log-likelihoods of all levels. L = − m X l=1 log p (yl | Hl) (10) On the decoding stage, we iteratively apply the Viterbi algorithm (Forney, 1973) at each level to search the most probable label sequences. ˆyl = arg max y′∈Yn p (y′ | Hl) (11) The pseudocodes of the training and the decoding algorithms with max or logsumexp potential function can be found in Algorithms 1 and 2, respectively. 3 Experiments 3.1 Datasets We conduct experiments on three nested named entity recognition datasets in English, i.e., ACE2004 (Doddington et al., 2004), ACE2005 (Walker et al., 2006) and GENIA (Kim et al., 2003). We divide all these datasets into tran/dev/test split by following Shibuya and Hovy (2020) and Wang et al. (2020). The dataset statistics can be found in Table 1. Dataset Sentences Mentions |Y| m ACE2004 6,198 / 742 / 809 22,195 / 2,514 / 3,034 29 6 ACE2005 7,285 / 968 / 1,058 24,700 / 3,218 / 3,029 29 6 GENIA 15,022 / 1,669 / 1,855 47,006 / 4,461 / 5,596 21 4 Table 1: Sizes of the dataset shown in the train/dev/test split. |Y| is the size of the label set, m is the maximal depth of entity nesting. 3.2 Hyper-parameters Settings For word embeddings initialization, we utilize 100dimensional pre-trained GloVe (Pennington et al., 2014) for the ACE2004 and the ACE2005 datasets, and use 200-dimensional biomedical domain word embeddings1 (Chiu et al., 2016) for the GENIA dataset. Moreover, we randomly initialize 30dimensional vectors for character embeddings. The hidden state dimension of character-level LSTM dc is 100, i.e., 50 in each direction, thus the dimension of token representation dx is 200. We apply dropout (Srivastava et al., 2014) on token representations before feeding it into the encoder. The hidden state dimension of the three-layered LSTM is 600 for ACE2004 and ACE2005, i.e., 300 in each direction, and 400 for GENIA. Choosing a different dimension is because the maximal depth of entity nesting m is different. We apply layer normalization (Ba et al., 2016) and dropout with 0.5 ratio after each bidirectional LSTM layer. Different from Shibuya and Hovy (2020), we use only one CRF instead of employing different CRFs for different entity types. Besides, our CRF is also shared across levels, which means we learn and decode entities at all levels with the same CRF. Our model is optimized by using stochastic gradient descent (SGD), with a decaying learning rate ητ = η0/(1 + γ · τ), where τ is the index of the current epoch. For ACE2004, ACE2005, and GENIA, the initial learning rates η0 are 0.2, 0.2, and 0.1, and the decay rates γ are 0.01, 0.02, and 0.02 respectively. We set the weight decay rate, the momentum, the batch size, and the number of epochs to be 10−8, 0.5, 32, and 100 respectively, especially we use batch size 64 on the GENIA dataset. We clip the gradient exceeding 5. Besides, we also conduct experiments to evaluate the performance of our model with contextual word representations. BERT (Devlin et al., 2019) and Flair (Akbik et al., 2018) are the most commonly used contextual word representations in previous work, and have also been proved that they can substantially improve the model performance. In these settings, contextual word representations are concatenated with word and character representations to form the token representations, i.e., xt = [wt, ct, et], where et is the contextual word representation and it is not fine-tuned in any of our experiments. 1https://github.com/cambridgeltl/ BioNLP-2016 3552 Methods ACE2004 ACE2005 GENIA P R F1 P R F1 P R F1 Ju et al. (2018) 74.2 70.3 72.2 78.5 71.3 74.7 Wang et al. (2018) 74.9 71.8 73.3 74.5 71.5 73.0 78.0 70.2 73.9 Wang and Lu (2018) 78.0 72.4 75.1 76.8 72.3 74.5 77.0 73.3 75.1 Luo and Zhao (2020) 75.0 75.2 75.1 77.4 74.6 76.0 Lin et al. (2019) 76.2 73.6 74.9 75.8 73.9 74.8 Strakov´a et al. (2019) 78.92 75.33 77.08 76.35 74.39 75.36 79.60 73.53 76.44 Shibuya and Hovy (2020) 79.93 75.10 77.44 78.27 75.44 76.83 78.70 75.74 77.19 Wang et al. (2020) 80.83 78.86 79.83 79.27 79.37 79.32 77.91 77.20 77.55 Our Method (naive) 81.12 77.71 79.38 (0.31) 79.45 77.22 78.32 (0.26) 78.83 75.32 77.03 (0.13) Our Method (max) 81.90 78.05 79.92 (0.10) 80.68 77.03 78.81 (0.04) 78.80 75.71 77.22 (0.10) Our Method (logsumexp) 81.24 78.96 80.08 (0.22) 79.49 77.65 78.55 (0.12) 78.58 76.21 77.37 (0.15) Strakov´a et al. (2019) [B] 84.71 83.96 84.33 82.58 84.29 83.42 79.92 76.55 78.20 Shibuya and Hovy (2020) [B] 85.23 84.72 84.97 83.30 84.69 83.99 77.46 76.65 77.05 Wang et al. (2020) [B] 86.08 86.48 86.28 83.95 85.39 84.66 79.45 78.94 79.19 Our Method (naive)[B] 86.19 85.28 85.73 (0.24) 84.23 84.17 84.20 (0.30) 78.83 78.07 78.45 (0.32) Our Method (max)[B] 86.27 85.09 85.68 (0.09) 85.28 84.15 84.71 (0.09) 79.20 78.16 78.67 (0.18) Our Method (logsumexp)[B] 86.42 85.71 86.06 (0.10) 83.95 84.67 84.30 (0.13) 78.83 78.27 78.54 (0.02) Strakov´a et al. (2019) [B+F] 84.51 84.29 84.40 83.48 85.21 84.33 80.11 76.60 78.31 Shibuya and Hovy (2020) [B+F] 85.94 85.69 85.82 83.83 84.87 84.34 77.81 76.94 77.36 Wang et al. (2020) [B+F] 87.01 86.55 86.78 84.90 86.08 85.49 79.98 78.51 79.24 Our Method (naive)[B+F] 86.56 85.65 86.11 (0.24) 84.17 84.88 84.52 (0.21) 79.28 78.31 78.79 (0.17) Our Method (max)[B+F] 86.96 85.45 86.19 (0.17) 84.70 84.76 84.73 (0.21) 79.51 78.25 78.87 (0.04) Our Method (logsumexp)[B+F] 86.74 86.11 86.42 (0.31) 84.81 85.06 84.93 (0.24) 79.20 78.67 78.93 (0.26) Table 2: Experimental results on the ACE2004, ACE2005 and GENIA datasets. Labels [B] and [F] stand for BERT and Flair contextual word representations respectively. Bold and underlined numbers indicates the best and the second-best results respectively. naive, max, and logsumexp refer to the three potential function definitions, i.e., Equations 1, 3, and 5, respectively. These numbers in parentheses are standard deviations. BERT is a transformer-based (Vaswani et al., 2017) pre-trained contextual word representation. In our experiments, for the ACE2004 and ACE2005 datasets we use the general domain checkpoint bert-large-uncased, and for the GENIA dataset we use the biomedical domain checkpoint BioBERT large v1.1 2 (Lee et al., 2019). We average all BERT subword embeddings in the last four layers to build 1024-dimensional vectors. Flair is a character-level BiLSTM-based pretrained contextual word representation. We concatenate these vectors obtained from the news-forward and news-backward checkpoints for ACE2004 and ACE2005, and use the pubmed-forward and pubmed-backward checkpoints for GENIA, to build 4096-dimensional vectors. 3.3 Evaluation Experiments are all evaluated by precision, recall, and F1. All of our experiments were run 4 times 2https://github.com/naver/ biobert-pretrained with different random seeds and averaged scores are reported in the following tables. Our model 3 is implemented with PyTorch (Paszke et al., 2019) and we run experiments on GeForce GTX 1080Ti with 11 GB memory. 3.4 Experimental Results Table 2 shows the performance of previous work and our model on the ACE2004, ACE2005, and GENIA datasets. Our model substantially outperforms most of the previous work, especially when comparing with our baseline Shibuya and Hovy (2020). When using only word embeddings and character-based word embeddings our method exceeds theirs by 2.64 F1 score, and also achieves comparable results with the recent competitive method (Wang et al., 2020). In the case of utilizing BERT and further employing Flair, our method consistently outperforms Shibuya and Hovy (2020) by 1.09 and 0.60 by F1 scores, respectively. On the ACE2005 dataset, our method improves the F1 scores by 1.98, 0.72, and 0.59 respectively, comparing with Shibuya and Hovy (2020). Although our model performance is inferior to Wang 3https://github.com/speedcell4/nersted 3553 et al. (2020) at general, our max potential function method is slightly superior to them by 0.05 in F1 score when employing BERT. Furthermore, on the biomedical domain dataset GENIA, our method constantly outperforms Shibuya and Hovy (2020) by 0.18, 1.62, and 1.57 in F1 score, respectively. Although the low scores of Shibuya and Hovy (2020) are due to their usage of the general domain checkpoint bert-large-uncased, instead of our biomedical domain checkpoint, our model is still superior to Strakov´a et al. (2019) by 0.47 and 0.62 in F1 scores, who used the same checkpoint as us. As for these three potential functions, we notice the max and logsumexp potential functions generally works better than the naive potential function. These results demonstrate that the chunk selection strategy of the max and logsumexp can leverage information from all remaining chunks and constrains hidden states of LSTM to be more semantically ordered. When we use BERT and Flair, the advantage of the max and the logsumexp potential function is less obvious compared with the case when we only use word embeddings and characterbased word embeddings, especially on the GENIA dataset. We hypothesize that BERT and Flair can provide rich contextual information, then selecting chunks in the original order is sufficient, thus our dynamic selecting mechanism can only slightly improve the model performance. 3.5 Influence of the Encoding Scheme We also conduct experiments on the ACE2004 dataset to measure the influence of the outermostfirst and innermost-first encoding schemes. As shown in Table 3, the innermost-first encoding scheme consistently works better than the outermost-first encoding scheme with all potential functions. We hypothesize that outermost entities do not necessarily contain inner entities especially for longer ones, and that putting those diversely Encoding Scheme φ P R F1 Outermost First naive 79.08 76.57 77.80 (0.26) max 79.07 75.11 77.04 (0.20) logsumexp 79.05 76.39 77.70 (0.32) Innermost First naive 81.12 77.71 79.38 (0.31) max 81.90 78.05 79.92 (0.10) logsumexp 81.24 78.96 80.08 (0.22) Table 3: Influence of the two encoding schemes and the three potential functions. nested outermost entities at the same level would dislocate the encoding representation. Furthermore, even if we use the outermost-first encoding scheme, our method is superior to Shibuya and Hovy (2020), which further demonstrates the effectiveness of excluding the influence of the best path. 3.6 Time Complexity and Speed The time complexity of encoder is O (n), and because we employ the same tree reduction acceleration trick4 as Rush (2020), the time complexity of CRF is reduced to O (log n), therefore the overall time complexity is O (n + m · log n). Even our model outperforms slightly worse than Wang et al. (2020), the training and inference speed of our model is much faster than them, as shown in Table 4, since we do not need to stack the decoding component to 16 layers. Especially, when we increase the batch size to 64, the decoding speed is more than two times faster than their model. Method Batch Size Training Decoding Wang et al. (2020) 16 1,937.16 3,626.53 32 3,632.64 4,652.05 64 6,298.85 5,113.85 Our Method 16 4,106.03 3,761.03 32 7,219.57 6,893.03 64 10,584.80 11,652.92 Table 4: Speed comparison on the ACE2005 dataset. Numbers indicate how many words can be processed per second on average. 3.7 Level-wise Performance We display the performance on the dataset ACE2005 at each level, as in Table 5. The max potential function at the first three levels achieves constantly higher precision scores than the naive and logsumexp potential functions, while at the same time obtains the lowest recall scores. The logsumexp potential function on the contrary achieves the highest recall scores but fails to obtain satisfactory precision scores. Because most entities are located at the first two levels, the max and logsumexp achieves the best overall precision and recall scores, respectively. 3.8 Chunk Distribution We analyze the chunk distribution on the test split of the dataset ACE2005 by plotting the heat maps 4https://github.com/speedcell4/ torchlatent 3554 1 2 3 4 5 6 level 100 0 0 0 0 0 0 100 0 0 0 0 0 0 100 0 0 0 0 0 0 100 0 0 0 0 0 0 100 0 0 0 0 0 0 100 59 20 6 7 2 7 4 29 22 15 14 16 3 22 30 19 6 20 2 17 27 25 8 21 10 10 12 18 28 22 22 2 3 16 43 14 44 6 4 9 3 35 16 30 19 15 13 6 13 24 34 11 15 4 11 17 20 27 21 4 8 15 14 26 29 8 8 8 9 12 19 43 O S B I E syntactic 13 16 17 18 18 18 84 16 0 0 0 0 81 16 3 0 0 0 13 62 23 2 0 0 81 15 3 0 0 0 14 17 17 18 17 17 84 0 0 0 16 1 82 0 0 0 15 3 2 28 13 0 15 42 45 23 7 1 5 20 16 17 18 16 18 15 0 0 0 70 0 30 0 0 0 19 3 78 32 13 7 34 8 6 40 17 2 19 2 20 1 2 3 4 5 6 chunk (naive) O FAC GPE LOC ORG PER VEH WEA semantic 13 16 17 18 18 18 55 24 20 0 0 0 88 12 0 0 0 0 56 44 0 0 0 0 75 22 3 0 0 0 59 30 10 1 0 0 48 37 15 0 0 0 78 22 0 0 0 0 1 2 3 4 5 6 chunk (max) 14 17 17 18 17 17 43 13 0 0 13 30 76 6 2 0 9 7 39 20 1 0 14 26 57 9 2 0 9 23 47 17 7 0 12 16 35 16 0 0 18 31 71 1 5 0 9 14 1 2 3 4 5 6 chunk (logsumexp) 16 17 18 16 18 15 21 5 5 44 5 20 12 5 0 19 2 61 21 5 0 47 5 22 18 4 1 30 2 45 24 11 3 25 4 33 20 7 0 46 6 21 13 0 0 21 14 53 Figure 3: Chunk distributions of the naive, max, and logsumexp potential functions, respectively. Each row displays the chunk selection preferences with respect to levels, syntactic and semantic labels, respectively. Level Naive Max LogSumExp P R P R P R 1 80.83 80.12 82.14 79.51 80.98 80.12 2 73.91 68.67 74.76 70.76 73.85 70.76 3 60.09 48.80 65.26 49.10 60.17 53.01 4 100.00 16.67 37.50 10.42 66.67 14.58 5 0.00 0.00 0.00 0.00 0.00 0.00 6 0.00 0.00 0.00 0.00 0.00 0.00 Overall 79.45 77.22 80.68 77.03 79.49 77.65 Table 5: Precision and recall scores at each level with each potential functions. in Figure 3, in which these numbers indicate the percentages of each chunk being selected by a particular level or label. For example, the 35 at the upper-right corner means when using logsumexp potential function, 35% of predictions at the first level are made by choosing the sixth chunk, while the 78 at the lower-left corner shows 78% of WEA are related to the first chunk with naive. To make it easier to compare with the naive, we arranged the chunk orders of max and logsumexp, without losing generality, to make the level-chunk distribution mainly concentrate on the diagonal. The naive potential function simply selects the l-th chunk at l-th level, therefore the heat map is just diagonal. At the first level, the logsumexp potential function also prefers to select the sixth and the fourth chunks rather than the first chunk, we hypothesis this is due to most of B- and S- labels are located on the first level, and this can be confirmed according to the syntactic-chunk heat map of logsumexp where 78% B- and 70% S- labels go to the sixth and fourth chunks. Similarly, max also has a high probability to select the second chunk. Generally, the chunk distribution of logsumexp is more smooth than max. Besides, we find label O almost uniformly select chunks, in both the syntactic and semantic heat maps, while other meaningful labels have their distinguished preferences. Syntactic labels S- and B- mainly represent the beginning of an entity, while I- and E- stands for the continuation and ending of an entity. In the syntactic-chunk heat map of naive, they are indiscriminately distributed to the first chunk, because most of the entities are located on the first level. However, max and logsumexp utilize different chunks to represents these different syntactic categories. Likewise, the semantic label GPE, when using logsumexp, also has a 61% probability to select the sixth chunks other than concentrating on the first 3555 chunk as naive. These observations further demonstrate our dynamic chunk selection strategies are capable of learning more meaningful representations. 4 Related Work Existing NER algorithms commonly employ various neural networks to leverage more morphological and contextual information to improve performance. For example, to handle the out-ofvocabulary issue through introducing morphological features, Huang et al. (2015) proposed to employ manual spelling feature, while Ma and Hovy (2016) and Lample et al. (2016) suggested introducing CNN and LSTM to build word representations from character-level. Zhang et al. (2018) and Chen et al. (2019) introduced global representation to enhance encoder capability of encoding contextual information. Layered Model As a layered model, Ju et al. (2018) dynamically update span-level representations for next layer recognition according to recognized inner entities. Fisher and Vlachos (2019) proposed a merge and label method to enhance this idea further. Recently, Shibuya and Hovy (2020) designed a novel algorithm to efficiently learn and decode the second-best path on the span of detected entities. Luo and Zhao (2020) build two different graphs, one is the original token sequence, and the other is the tokens in recognized entities, to model the interaction among them. Wang et al. (2020) proposed to learn the l-gram representations at layer l through applying a decoder component to reduce a sentence layer by layer and to directly classify these l-gram spans. Region-based Model Lin et al. (2019) proposed an anchor-region network to recognize nested entities through detecting anchor words and entity boundaries first, and then classify each detected span. Exhaustive models simply enumerate all possible spans and utilize a maximum entropy tagger (Byrne, 2007) and neural networks (Xu et al., 2017; Sohrab and Miwa, 2018; Zheng et al., 2019) for classification. Luan et al. (2019) additionally aims to consider the relationship among entities and proposed a novel method to jointly learn both entities and relations. Hypergraph-based Model Lu and Roth (2015) proposed a hyper-graph structure, in which edges are connected to multiple nodes to represents nested entities. Muis and Lu (2017) and Wang and Lu (2018) resolved spurious structures and ambiguous issue of hyper-graph structure. And Katiyar and Cardie (2018) proposed another kind of hyper-graph structure. Parsing-based Model Finkel and Manning (2009) indicated all these nested entities are located in some non-terminal nodes of the constituency parses of the original sentences, thus they proposed to use a CRF-based constituency parser to obtain them. However, the cubic time complexity limits its applicability. Wang et al. (2018) instead proposed to use a transition-based constituency parser to incrementally build constituency forest, its linear time complexity ensures it can handle longer sentences. 5 Conclusion In this paper, we proposed a simple and effective method for nested named entity recognition by explicitly excluding the influence of the best path through selecting and removing chunks at each level to build different potential functions. We also proposed three different selection strategies to leverage information from all remaining chunks. Besides, we found the innermost-first encoding scheme works better than the conventional outermost-first encoding scheme. Extensive experimental results demonstrate the effectiveness and efficiency of our method. However, one of the demerits of our method is the number of chunks, i.e., the maximal depth of entity nesting, must be chosen in advance as a hyper-parameter. We will extend it to arbitrary depths as future work. Acknowledgements This work was partly supported by JST CREST Grant Number JPMJCR1513. The authors would like to thank the anonymous reviewers for their instructive comments. References Alan Akbik, Duncan Blythe, and Roland Vollgraf. 2018. Contextual string embeddings for sequence labeling. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1638–1649, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. arXiv preprint arXiv:1607.06450. 3556 K. Byrne. 2007. Nested named entity recognition in historical archive text. In International Conference on Semantic Computing (ICSC 2007), pages 589– 596. Hui Chen, Zijia Lin, Guiguang Ding, Jianguang Lou, Yusen Zhang, and Borje Karlsson. 2019. GRN: Gated relation network to enhance convolutional neural network for named entity recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6236–6243. Billy Chiu, Gamal Crichton, Anna Korhonen, and Sampo Pyysalo. 2016. How to train good word embeddings for biomedical NLP. In Proceedings of the 15th Workshop on Biomedical Natural Language Processing, pages 166–174, Berlin, Germany. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. George Doddington, Alexis Mitchell, Mark Przybocki, Lance Ramshaw, Stephanie Strassel, and Ralph Weischedel. 2004. The automatic content extraction (ACE) program – tasks, data, and evaluation. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC’04), Lisbon, Portugal. European Language Resources Association (ELRA). Jenny Rose Finkel and Christopher D. Manning. 2009. Nested named entity recognition. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 141–150, Singapore. Association for Computational Linguistics. Joseph Fisher and Andreas Vlachos. 2019. Merge and label: A novel neural network architecture for nested NER. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5840–5850, Florence, Italy. Association for Computational Linguistics. G David Forney. 1973. The viterbi algorithm. Proceedings of the IEEE, 61(3):268–278. Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF models for sequence tagging. CoRR, abs/1508.01991. Meizhi Ju, Makoto Miwa, and Sophia Ananiadou. 2018. A neural layered model for nested named entity recognition. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1446–1459, New Orleans, Louisiana. Association for Computational Linguistics. Arzoo Katiyar and Claire Cardie. 2018. Nested named entity recognition revisited. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 861–871, New Orleans, Louisiana. Association for Computational Linguistics. J.-D. Kim, T. Ohta, Y. Tateisi, and J. Tsujii. 2003. GENIA corpus—a semantically annotated corpus for bio-textmining. Bioinformatics, 19:i180–i182. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML ’01, page 282–289, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016. Neural architectures for named entity recognition. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 260–270, San Diego, California. Association for Computational Linguistics. Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2019. Biobert: a pre-trained biomedical language representation model for biomedical text mining. CoRR, abs/1901.08746. Hongyu Lin, Yaojie Lu, Xianpei Han, and Le Sun. 2019. Sequence-to-nuggets: Nested entity mention detection via anchor-region networks. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5182–5192, Florence, Italy. Association for Computational Linguistics. Wei Lu and Dan Roth. 2015. Joint mention extraction and classification with mention hypergraphs. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 857–867, Lisbon, Portugal. Association for Computational Linguistics. Yi Luan, Dave Wadden, Luheng He, Amy Shah, Mari Ostendorf, and Hannaneh Hajishirzi. 2019. A general framework for information extraction using dynamic span graphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3036–3046, Minneapolis, Minnesota. Association for Computational Linguistics. Ying Luo and Hai Zhao. 2020. Bipartite flat-graph network for nested named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6408– 6418, Online. Association for Computational Linguistics. 3557 Xuezhe Ma and Eduard Hovy. 2016. End-to-end sequence labeling via bi-directional LSTM-CNNsCRF. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1064–1074, Berlin, Germany. Association for Computational Linguistics. Aldrian Obaja Muis and Wei Lu. 2017. Labeling gaps between words: Recognizing overlapping mentions with mention separators. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2608–2618, Copenhagen, Denmark. Association for Computational Linguistics. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch´e-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8026–8037. Curran Associates, Inc. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, Doha, Qatar. Association for Computational Linguistics. Lance Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In Third Workshop on Very Large Corpora. Alexander Rush. 2020. Torch-struct: Deep structured prediction library. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 335– 342, Online. Association for Computational Linguistics. Takashi Shibuya and Eduard Hovy. 2020. Nested named entity recognition via second-best sequence learning and decoding. Transactions of the Association for Computational Linguistics, 8:605–620. Mohammad Golam Sohrab and Makoto Miwa. 2018. Deep exhaustive model for nested named entity recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2843–2849, Brussels, Belgium. Association for Computational Linguistics. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958. Jana Strakov´a, Milan Straka, and Jan Hajic. 2019. Neural architectures for nested NER through linearization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5326–5331, Florence, Italy. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc. Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. Linguistic Data Consortium, Philadelphia, 57:45. Bailin Wang and Wei Lu. 2018. Neural segmental hypergraphs for overlapping mention recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 204–214, Brussels, Belgium. Association for Computational Linguistics. Bailin Wang, Wei Lu, Yu Wang, and Hongxia Jin. 2018. A neural transition-based model for nested mention recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1011–1017, Brussels, Belgium. Association for Computational Linguistics. Jue Wang, Lidan Shou, Ke Chen, and Gang Chen. 2020. Pyramid: A layered model for nested named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5918–5928, Online. Association for Computational Linguistics. Mingbin Xu, Hui Jiang, and Sedtawut Watcharawittayakul. 2017. A local detection approach for named entity recognition and mention detection. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1237–1247, Vancouver, Canada. Association for Computational Linguistics. Yue Zhang, Qi Liu, and Linfeng Song. 2018. Sentencestate LSTM for text representation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 317–327, Melbourne, Australia. Association for Computational Linguistics. Changmeng Zheng, Yi Cai, Jingyun Xu, Ho-fung Leung, and Guandong Xu. 2019. A boundary-aware neural model for nested named entity recognition. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 357– 366, Hong Kong, China. Association for Computational Linguistics.
|
2021
|
275
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3558–3571 August 1–6, 2021. ©2021 Association for Computational Linguistics 3558 LearnDA: Learnable Knowledge-Guided Data Augmentation for Event Causality Identification Xinyu Zuo1,2, Pengfei Cao1,2, Yubo Chen1,2, Kang Liu1,2, Jun Zhao1,2, Weihua Peng3 and Yuguang Chen3 1National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 2School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 3Beijing Baidu Netcom Science Technology Co., Ltd {xinyu.zuo,pengfei.cao,yubo.chen,kliu,jzhao}@nlpr.ia.ac.cn {pengweihua,chenyuguang}@baidu.com Abstract Modern models for event causality identification (ECI) are mainly based on supervised learning, which are prone to the data lacking problem. Unfortunately, the existing NLPrelated augmentation methods cannot directly produce available data required for this task. To solve the data lacking problem, we introduce a new approach to augment training data for event causality identification, by iteratively generating new examples and classifying event causality in a dual learning framework. On the one hand, our approach is knowledge guided, which can leverage existing knowledge bases to generate well-formed new sentences. On the other hand, our approach employs a dual mechanism, which is a learnable augmentation framework, and can interactively adjust the generation process to generate task-related sentences. Experimental results on two benchmarks EventStoryLine and Causal-TimeBank show that 1) our method can augment suitable task-related training data for ECI; 2) our method outperforms previous methods on EventStoryLine and Causal-TimeBank (+2.5 and +2.1 points on F1 value respectively). 1 Introduction Event causality identification (ECI) aims to identify causal relations between events in texts, which can provide crucial clues for NLP tasks, such as logical reasoning and question answering (Girju, 2003; Oh et al., 2013, 2017). This task is usually modeled as a classification problem, i.e. determining whether there is a causal relation between two events in a sentence. For example in Figure 1, an ECI system should identify two causal relations in two sentences: (1) attack cause −→killed in S1; (2) statement cause −→protests in S2. Most existing methods for ECI heavily rely on annotated training data (Mirza and Tonelli, 2016; Kimani Gray, a young man who likes football, was killed in a police attack shortly after a tight match. In the week following the fatal violence, several protests have erupted because of the official statement. S1: S2: Kimani Gray, a young man who likes football, was killed in a police attack shortly after a tight match. EDA deletion S3: Figure 1: S1 and S2 are causal sentences that contain causal events. S3 is produced by EDA based on S1. The dotted line indicates the causal relation. Riaz and Girju, 2014b; Hashimoto et al., 2014; Hu and Walker, 2017; Gao et al., 2019). However, existing datasets are relatively small, which impede the training of the high-performance event causality reasoning model. According to our statistics, the largest widely used dataset EventStoryLine Corpus (Caselli and Vossen, 2017) only contains 258 documents, 4316 sentences, and 1770 causal event pairs. Therefore, data lacking is an essential problem that urgently needs to be addressed for ECI. Up to now, data augmentation is one of the most effective methods to solve the data lacking problem. However, most of the NLP-related augmentation methods are a task-independent framework that produces new data at one time (Zhang et al., 2015; Guo et al., 2019; Xie et al., 2019b). In these frameworks, data augmentation and target task are modeled independently. This often leads to a lack of task-related characteristics in the generated data, such as taskrelated linguistic expression and knowledge. For example, easy data augmentation (EDA) (Wei and Zou, 2019) is the most representative method that relies on lexical substitution, deletion, swapping, and insertion to produce new data. However, solely relying on such word operations often generates new data that dissatisfies task-related qualities. As shown in Figure 1, S3 is produced by EDA, it lacks a linguistic expression that expresses the causal semantics between kill and attack. Therefore, how to 3559 interactively model data augmentation and target task to generate new data with task-related characteristics is a challenging problem on ECI. Specific to ECI, we argue that an ideal taskrelated generated causal sentence needs to possess two characteristics as follows. (1) The two events in the causal sentence need to have a causal relation. We call such property as Causality. For example, there is usually a causal relation between an attack event and a kill event, while nearly no causal relation between an attack event and a born event. (2) The linguistic expressions of the causal sentence need to be well-formed to express the causal semantic of events. We call such property as Well-formedness, which consists of a) canonical sentence grammar, b) event-related entities with semantic roles (e.g. the attack was carried out by a police in S1), and c) cohesive words that express complete causal semantics (e.g. in a and other words except for events and entities in S1). To this end, we propose a learnable data augmentation framework for ECI, dubbed as Learnable Knowledge-Guided Data Augmentation (LearnDA). This framework regards sentence-torelation mapping (the target task, ECI) and relationto-sentence mapping (the augmentation task, sentence generation) as dual tasks and models the mutual relation between them via dual learning. Specifically, LearnDA can use the duality to generate task-related new sentences learning from identification and makes it more accurate to understand the causal semantic learning from generation. On the one hand, LearnDA is knowledge guided. It introduces diverse causal event pairs from KBs to initialize the dual generation which could ensure the causality of generated causal sentences. For example, the knowledge of judgment cause −→demonstration from KBs can be used to construct a novel causal sentence, which is also helpful to understand the causal semantic of statement cause −→protests. On the other hand, LearnDA is learnable. It employs a constrained generative architecture to generate well-formed linguistic expressions via iteratively learning in the dual interaction, which expresses the causal semantic between given events. Methodologically, it gradually fills the remaining missing cohesive words of the complete sentences under the constraint of given events and related entities. In experiments, we evaluate our model on two benchmarks. We first concern the standard evaluation and show that our model achieves the state-ofthe-art performance on ECI. Then we estimate the main components of LearnDA. Finally, our learnable augmentation framework demonstrates definite advantages over other augmentation methods in generating task-related data for ECI. In summary, the contributions as follows: • We propose a new learnable data augmentation framework to solve the data lacking problem of ECI. Our framework can leverage the duality between identification and generation via dual learning which can learn to generate task-related sentences for ECI. • Our framework is knowledge guided and learnable. Specifically, we introduce causal event pairs from KBs to initialize the dual generation, which could ensure the causality of generated causal sentences. We also employ a constrained generative architecture to gradually generate well-formed causal linguistic expressions of generated causal sentences via iteratively learning in the dual interaction. • Experimental results on two benchmarks show that our model achieves the best performance on ECI. Moreover, it also shows definite advantages over previous data augmentation methods. 2 Related Work To date, many researches attempt to identify the causality with linguistic patterns or statistical features. For example, some methods rely on syntactic and lexical features (Riaz and Girju, 2013, 2014b). Some focus on explicit causal textual patterns (Hashimoto et al., 2014; Riaz and Girju, 2014a, 2010; Do et al., 2011; Hidey and McKeown, 2016). And some others pay attention on statistical causal association and cues (Beamer and Girju, 2009; Hu et al., 2017; Hu and Walker, 2017). Recently, more attention is paid to the causality between events. Mirza and Tonelli (2014) annotated Causal-TimeBank of event-causal relations based on the TempEval-3 corpus. Mirza et al. (2014), Mirza and Tonelli (2016) extracted eventcausal relation with a rule-based multi-sieve approach and improved the performance incorporating with event temporal relation. Mostafazadeh et al. (2016) annotated both temporal and causal relations in 320 short stories. Caselli and Vossen (2017) annotated the EventStoryLine Corpus for 3560 Dual Cycle Primal Cycle causal/ non-causal relaiton causal/ non-causal sentence event pair Annotated Data Pre-trained Identifier Pre-trained Generator Learnable Dual Augmentation Architecture Dual-trained Identifier Dual Augmented Data Full-trained Identifier Pre-training Further training Knowledge relation->sentence->relation sentence->relation->sentence Figure 2: Overview of the learnable knowledge-guided dual data augmentation for ECI. event causality identification. Dunietz et al. (2017) presented BECauSE 2.0, a new version of the BECauSE corpus (Dunietz et al., 2015) of causal relation and other seven relations. Gao et al. (2019) modeled document-level structures to identify causality. Liu et al. (2020) identified event causality with the mention masking generalization. Unlike computer vision, the augmentation of text data in NLP is pretty rare (Chaudhary, 2020). Zuo et al. (2020) solved the data lacking problem of ECI with the distantly supervised labeled training data. However, including the distant supervision, most of the existing data augmentation methods for NLP tasks are task-independent frameworks (Related work of data augmentation and dual learning are detailed in Appendix B). Inspired by some generative methods which try to generate additional training data while preserving the class label (AnabyTavor et al., 2019; Yang et al., 2019; Papanikolaou and Pierleoni, 2020), we introduce a new learnable framework for augmenting task-related training data for ECI via dual learning enhanced with external knowledge. 3 Methodology As shown in Figure 2, LearnDA jointly models a knowledge guided sentence generator (input: event pair and its causal/non-causal relation, output: causal/non-causal sentence) and an event causality identifier (input: event pair and its sentence, output: causal/non-causal relation) with dual learning. LearnDA iteratively optimizes identifier and generator to generate task-related training data, and then utilize new data to further train the identifier. Therefore, we first present the main idea of dual learning, which is the architecture of learnable dual augmentation, including the states, actions, policies, and Identifier Relation→Sentence NCausal-Generator Causal-Generator Sentence→Relation event pair (ep) causal/non-causal relation (c) ep, s' event pair (ep) sentence (s) ep, c' Rs Rc Rc Rs Primal Cycle Dual Cycle R R I G Figure 3: The architecture of learnable dual augmentation. Causal and NCausal represent the causal and non-causal sentence generator respectively. Red parts are the process of <event pair, relation> →sentence →relation (primal cycle), while blue parts are the process of <event pair, sentence> →relation →sentence (dual cycle). Solid and dashed lines denote the main process and reward feedback direction respectively. rewards. Then, we briefly introduce the knowledge guided sentence generator, especially the processes of knowledge guiding and constrained sentence generation. Finally, we describe the event causality identifier and training processes of LearnDA. 3.1 Architecture of Learnable Dual Augmentation The architecture of learnable dual augmentation is shown in Figure 3. Specifically, I denotes the event causality identifier, and G denotes the sentence generator which consists of two independent generators. They produce causal and non-causal sentences on the relation c of input event pair ep. Generally, G generates a sentence s′ which expresses the causal or non-causal relation c of the input event pair ep. Then it receives the reward R that consists of a semantic alignment reward Rs from itself and a causality reward Rc from I (primal cycle). Similarly, I identifies the causal or non-causal relation c′ of the input event pair ep with its sentence s. Then it receives the reward R consists of a causality reward Rc from itself and a semantic alignment reward Rs from G (dual cycle). I and G are optimized interactively with dual reinforcement learning. Specifically, for G, an action is the generation from relation to sentence, a state is denoted by the representation of input event pair and its relation, a policy is defined by the parameters of generator. For I, an action is the identification from sentence to relation, a state is denoted by the representation of input event pair and its 3561 sentence, a policy is defined by the parameters of identifier. Inspired by Shen and Feng (2020), we utilize a probability distribution over actions given states to represent the policys, i.e., the probability distribution of the generation of G and identification of I. As aforementioned, we introduce two rewards, causality (Rc) and semantic alignment (Rs) rewards, which encourage G to generate taskrelated sentences with the feedback from identifier, while further optimize I with the feedback from generator. Definitions are as following: Causality Reward (Rc) If the relation of input event pair can be clearly expressed by the generated sentence, it will be easier to be understood by identifier. Therefore, we use the causal relation classification accuracy as the causality reward to evaluate the causality of generated sentences, while tune and optimize the identifier itself: Rc(ep, s) = ( p(c′|s; θI) Correct classification −p(c′|s; θI) Otherwise, (1) where θI is the parameter of I, p(c′|s; θI) denotes the probability of relation classification, s denotes the input sentence and c′ is the classified relation. Semantic Alignment Reward (Rs) We hope that the semantic of the generated sentence can be consistent with the relation of the input event pair. Additionally, if the relation of the input event pair can be more accurately classified, the semantic of the new generated sentence can be considered more consistent with it. Therefore, we measure the semantic alignment by means of the probability of constructing a sentence with similar semantic to the input relation, and the reward is: Rs(ep, c) = p(s′|c; θG) = 1 |Ts| X t∈Ts p(t|c; θG), (2) where θG is the parameter of G, c is the input relation, t is one of the generated tokens Ts of the generated sentence s′, and p(t|c; θG) is the generated probability of t. Specifically, there are two independent G with different θG. In detail, θc G is employed to generated causal sentence when the input c is causal relation, and non-causal sentence is generated via θnc G when c is non-causal relation. 3.2 Knowledge Guided Sentence Generator As shown in Figure 4, knowledge guided sentence generator (KSG) first introduces diverse causal and non-causal event pairs from KBs for causality. Then, given an event pair and its causal or non-causal relation, it employs a constrained genNcausal-Generator Causal-Generator event pair: <hurt,onrush> relation: causal Knowledge Kimani Gray, a young man who likes football, was killed in a police attack shortly after a tight match. event pair: <killed,attack> relation: causal John Henderson who is a baseball fanatic, was hurt in a gang onrush before Friday’s game. Generated sentence: Original sentence: words:events words:entities words:cohesive words Figure 4: Flow diagram of the knowledge guided sentence generator (KSG). We take causal sentence generation via lexical knowledge expanding as an example. erative architecture to generate new well-formed causal/non-causal sentences that contain them. Knowledge Guiding KSG introduces event pairs that are probabilistic causal or non-causal from multiple knowledge bases in two ways. (1) Lexical knowledge expanding: expanding annotated event pairs via external dictionaries, such as WordNet (Miller, 1995) and VerbNet (Schuler, 2005). (2) Connective knowledge introducing: introducing event pairs from external event-annotated documents (KBP corpus) assisted with FrameNet (Baker et al., 1998) and Penn Discourse Treebank (PDTB2) (Group et al., 2008). As shown in Table 1, we illustrate how to extract event pairs from multiple knowledge bases. Then, inspired by Bordes et al. (2013), we filter the extracted event pairs by converting them into triples <ei, causal/noncausal, ej> and calculating the causal-distance by maximizing L in a causal representation space: L = X (ei,ej)∈T X (e′ i,e′ j)∈T ′ [λ + d(e′ i, e′ j) −d(ei, ej)]+, (3) where T and T ′ are the causal and non-causal triples set respectively, and e is the representation of event. After that, the higher probability of causal relation, the shorter distance between two events, and we sort event pairs in ascending order by their distances. Finally, we keep the top and bottom α% sorted event pairs to obtain the causal and noncausal event pairs sets for generation. Constrained Sentence Generator Given an event pair, constrained sentence generator produces a well-formed sentence that expresses its causal or non-causal relation in three stages: (1) assigning event-related entities ensures the logic of the semantic roles of events, (2) completing sentences ensures the completeness of causal or non-causal 3562 Knowledge How to extract event pair Why causal or non-causal Lexical knowledge expanding WordNet 1) Extracting the synonyms and hypernyms from WordNet of each event in ep. 2) Assembling the items from the two groups of two events to generate causal/non-causal event pairs. Items in each group are the synonyms and hypernyms of the annotated causal/noncausal event pairs. VerbNet 1) Extracting the words from VerbNet under the same class as each event in ep. 2) Assembling the items from the two groups of two events to generate causal/non-causal event pairs. Items in each group are in the same class of the annotated causal/non-causal event pairs. e.g. < (killed, attack), causal >=⇒kill Synonyms −→ hurt, attack Synonyms −→ onrush =⇒< (hurt, onrush), causal > Original sentence: Kimani Gray, a young man who likes football, was killed in a police attack shortly after a tight match. Connective knowledge introducing FrameNet PDTB2 1) Extracting causal/non-causal connectives from FrameNet1 and PDTB2. 2) Extracting any two events connected by causal/non-causal connectives on KBP corpus to obtain causal/non-causal event pairs and original sentences respectively. Introduced event pairs are connected by causal/non-causal connectives. e.g. Looting because someone beat up someone, like the Travon Martin case. because =⇒< (loot, beat up), causal > Original sentence: Looting because someone beat up someone, like the Travon Martin case. Table 1: Extracting causal and non-causal event pairs from multiple knowledge bases. semantic expression, (3) filtering sentences ensures the quality and diversity of generated sentences. Assigning Event-related Entities. Event related entities play different semantic roles of events in sentences, which is an important part of eventsemantic expression. Hence, as shown in Figure 4, given an event pair, we firstly assign logical entities for input events to guarantee the logic of semantic roles in the new sentences, such as gang is a logical entity as the body of the event onrush. Logically, entities of the same type play the same semantic roles in similar events. Moreover, as shown in Table 1, there is a corresponding original sentence for each extracted event pair. Therefore, in new sentence, we assign the most similar entity in the same type from candidate set2 for each entity in the original sentence. For example, we assign gang for onrush in new sentence which is similar with the police related to attack in the original sentence. Specifically, we put the candidate entities in the same position in the original sentence to obtain their BERT embeddings. Then we select entities via the cosine similarity between their embeddings: E(ent) = 1 |ent| P w∈ent E(w), where ent is the entity and E(w) is the BERT embedding of ent. Completing Sentences. A well-formed sentence requires a complete linguistic expression to express the causal or non-causal semantics. Therefore, we complete sentences by filling the cohesive words between given events and assigned entities with masked BERT (Devlin et al., 2019). All words except events and entities are regarded as cohesive words. Specifically, we insert a certain number of the special token [MASK] between events and 2We collect entities from annotated data and KBP corpus. entities, and then predict the [MASK]3 tokens as new words. As shown in Figure 4, we fill cohesive tokens via two independent generators to express causal and non-causal semantic according to the relation of given events. For example, in a guiding a causal semantic filled by the causal generator. Filtering Sentences. Inspired by Yang et al. (2019), we design a filter to select new sentences that are balanced between high quality and high diversity with two key factors: 1) Perplexity (PPL): we take the average probability of the filled cohesive words in the new sentence s′ as its perplexity: PPL(s′) = 1 |T(s′)| P t∈T(s′) P(t), where T is the set of filled cohesive words. 2) Distance (DIS): we calculate the cosine similarity between generated sentence s′ and annotated data Dm as its distance: DIS(s′, Dm) = 1 |Dm| P s∈Dm E(s′)·E(s) E(s′)×E(s), where Dm is m random selected annotated sentences and E is the BERT sentence representation of the [CLS] token. A new sentence should have both appropriate high PPL which indicates the quality of generation, and appropriate high DIS which indicates the difference from the original sentences. Therefore, we select the top β% of the newly generated sentences according to Score for the further training of identifier as following: Score(s′) = µPPL(s′) + (1 −µ)DIS(s′, Dm)), where the µ is an hyper-parameter. 3.3 Training of LearnDA for ECI We briefly describe the training processes of LearnDA for ECI, including the pre-training of generator and identifier, the dual reinforcement training, and the further training of identifier. 3The inserted [MASK] is 1.2 times the number of words between events and entities in the original sentence. 3563 Algorithm 1 Dual Reinforcement Training of G I. Require: A set of knowledge guided event pairs {(ep,s,c)} A pre-trained generator G and identifier I Repeat: Early stop on the development set according to I. 1: Loop: PRIMAL CYCLE 2: for event pair (epi, si, ci) in batch do 3: Generator generates the sentence s′ i of epi; 4: Identifier re-predicts the causality c∗ i of epi; 5: Computing the reward as: 6: Rs primal = λRs(epi, ci)+(1−λ)Rc(epi, s′ i). 7: Computing the stochastic gradient of θG: 8: ∇G+ = Rs primal · ∇θGLG(epi, ci). 9: end for 10: Model batch updates: θG ←θG + η · ∇G 11: end Loop: 12: 13: Loop: DUAL CYCLE 14: for event pair (epi, si, ci) in batch do 15: Identifier predicts the causality c′ i of epi; 16: Generator re-generates the sentence s∗ i of epi; 17: Computing the reward as: 18: Rs dual = γRc(epi, si) + (1 −γ)Rs(epi, c′ i). 19: Computing the stochastic gradient of θI: 20: ∇I+ = Rs dual · ∇θILI(epi, si). 21: end for 22: Model batch updates: θI ←θI + η · ∇I 23: end Loop: Event Causality Identifier First of all, we formulate event causality identification as a sentencelevel binary classification problem. Specifically, we design a classifier based on BERT (Devlin et al., 2019) to build our identifier. The input of the identifier is the event pair ep and its sentence s. Next, we take the stitching of manually designed features (same lexical, causal potential, and syntactic features as Gao et al. (2019)) and two event representations as the input of top MLP classifier. Finally, the output is a binary vector to predict the causal/noncausal relation of the input event pair ep. Pre-training We pre-train the identifier and generator on labeled data before dual reinforcement training. On the one hand, we train identifier via the cross-entropy objective function of the relation classification. On the other hand, for generators, we keep the events and entities in the input sentences, replace the remaining tokens with a special token [MASK], and then train it via the cross-entropy objective function to re-predict the masked tokens. Specifically, causal generator and non-causal generator are pre-trained on causal and non-causal labeled sentences respectively. Dual Reinforcement Training As shown in Algorithm 1, we interactively optimize the generator and identifier by dual reinforcement learning. Specifically, we maximize the following objective functions: LG(ep, c) = ( p(s′|c; θG) = 1 |Ts| P t∈Ts p(t|c; θG) p(s′|c; θNG) = 1 |Ts| P t∈Ts p(t|c; θNG), (4) LI(ep, s) = p(c′|s; θI), (5) where θG and θNG is the parameters of causal and non-causal sentence generators respectively, Ts is the masked tokens. Finally, after dual data augmentation, we utilize generated sentences to further train the dual-trained identifier via the crossentropy objective function of relation classification. 4 Experiments 4.1 Experimental Setup Dataset and Evaluation Metrics Our experiments are conducted on two main benchmark datasets, including: EventStoryLine v0.9 (ESC) (Caselli and Vossen, 2017) described above; and (2) Causal-TimeBank (Causal-TB) (Mirza and Tonelli, 2014) which contains 184 documents, 6813 events, and 318 causal event pairs. Same as previous methods, we use the last two topics of ESC as the development set for two datasets. For evaluation, we adopt Precision (P), Recall (R), and F1-score (F1) as evaluation metrics. We conduct 5-fold and 10-fold cross-validation on ESC and Causal-TB respectively, same as previous methods to ensure comparability. All the results are the average of three independent experiments. Parameters Settings In implementations, both the identifier and generators are implemented on BERT-Base architecture4, which has 12-layers, 768-hiddens, and 12-heads. We set the learning rate of generator pre-training, identifier pretraining/further training, and dual reinforcement training as 1e-5, 1e-5, and 1e-7 respectively. We set the ratio of the augmented data used for training to the labeled data, α, β, µ, λ and γ as 1:2, 30%, 50%, 0.2, 0.5 and 0.5 respectively tuned on the development set. And we apply early stop and SGD gradient strategy to optimize all models. We also adopt a negative sampling rate of 0.5 for training the identifier, owing to the sparseness of positive examples. (See Appendix D for more details.) Compared Methods Same as previous state-ofthe-art work. For ESC, we prefer 1) LSTM (Cheng and Miyao, 2017), a dependency path based 4https://github.com/google-research/ bert 3564 sequential model that models the context between events to identify causality; 2) Seq (Choubey and Huang, 2017), a sequence model explores complex human designed features for ECI; 3) LR+ and ILP (Gao et al., 2019), document-level models adopt document structures for ECI. For Causal-TB, we prefer 1) RB, a rule-based system; 2) DD, a data driven machine learning based system; 3) VR-C, a verb rule based model with data filtering and gold causal signals enhancement. These models are designed by Mirza and Tonelli (2014); Mirza (2014) for ECI. Owing to our methods are constructed on BERT, we build BERT-based methods: 1) BERT, a BERTbased baseline, our basic proposed event causality identifier. 2) MM (Liu et al., 2020), the BERTbased SOTA method with mention masking generalization. 3) MM+Aug, the further re-trained MM with our dual augmented data. 4) KnowDis (Zuo et al., 2020) improved the performance of ECI with the distantly labeled training data. We compare with it to illustrate the quality of our generated ECI-related training data. 5) MM+ConceptAug, to make a fair comparison, we introduce causalrelated events from ConceptNet that employed by MM, and generate new sentences via KonwDis and LearnDA to further re-train MM (see Appendix C for details). Finally, we use LearnDAFull indicates our full model, which is the dual-trained identifier further trained via dual augmented data. 4.2 Our Method vs. State-of-the-art Methods Table 2 shows the results of ECI on EventStoryLine and Causal-TimeBank. From the results: 1) Our LearnDAFull outperforms all baselines and achieves the best performance (52.6%/51.9% on F1 value), outperforming the no-bert (ILP/VRC) and bert (MM/KnowDis) state-of-the-art methods by a margin of 7.9%/8.7% and 2.5%/2.1% respectively, which justifies its effectiveness. Moreover, BERT-based methods demonstrate high recall value, which is benefited from more training data and their event-related guided knowledge. 2) Comparing KnowDis with LearnDAFull, we note that training data generated by LearnDA is more helpful to ECI than distant supervision with external knowledge (+2.9%/+2.1%). This shows that LearnDA can generate more ECI-related data. 3) Comparing MM+ConceptNet with MM, with the same knowledge base, our dual augmented data can further improve the performance Methods P R F1 ESC LSTM (Cheng and Miyao, 2017) 34.0 41.5 37.4 Seq (Choubey and Huang, 2017) 32.7 44.9 37.8 LR+ (Gao et al., 2019) 37.0 45.2 40.7 ILP (Gao et al., 2019) 37.4 55.8 44.7 BERT 36.1 56.0 43.9 KnowDis (Zuo et al., 2020) 39.7 66.5 49.7 MM (Liu et al., 2020) 41.9 62.5 50.1 MM+ConceptAug (Ours) 41.2 66.5 50.9* MM+Aug (Ours) 41.0 69.3 51.5* LearnDAF ull (Ours) 42.2 69.8 52.6* Causal-TB RB (Mirza and Tonelli, 2014) 36.8 12.3 18.4 DD (Mirza and Tonelli, 2014) 67.3 22.6 33.9 VR-C (Mirza, 2014) 69.0 31.5 43.2 BERT 38.5 43.9 41.0 MM (Liu et al., 2020) 36.6 55.6 44.1 KnowDis (Zuo et al., 2020) 42.3 60.5 49.8 MM+ConceptAug (Ours) 38.8 59.2 46.9* MM+Aug (Ours) 39.2 61.9 48.0* LearnDAF ull (Ours) 41.9 68.0 51.9* Table 2: Results on event causality identification. * denotes a significant test at the level of 0.05. (+0.8%/+2.8%), which illustrates that LearnDA can make more effective use of external knowledge by generating task-related training data. 4) Comparing MM+Aug with MM, we note that training with our dual augmented data can improve the performance by 1.4%/3.9%, even though MM is designed on BERT-Large (LearnDA is constructed on BERT-Base) and also introduces external knowledge. This indicates that the augmented data generated by our LearnDA can effectively alleviate the problem of data lacking on the ECI. 4.3 Effect of Learnable Dual Augmentation We analyze the effect of the learnable dual augmentation for event causality identification. 1) For identifier. Comparing LearnDADual with BERT in Table 3, we note that the performance of the proposed identifier is improved (+2.6%) after the dual training only with the same labeled data. This indicates that the identifier can learn more informative expressions of causal semantic from generation with dual learning. 2) For generator. Comparing BERTDualAug with BERTAug in Table 3, we note that the dual augmented data is high quality and more helpful to ECI (+2.6%). This indicates generator can generate more ECI task-related data learned from identifier with dual learning. Figure 5 illustrates the learnability of our LearnDA. Specifically, as the number of training rounds of dual learning increases, the generated data gradually learns task-related information, fur3565 Method P R F BERT (Our basic identifier) 36.1 56.0 43.9 BERTOrgAug 36.6 59.7 45.4* BERTDualAug 37.8 65.6 48.0* LearnDADual 36.8 63.0 46.5* LearnDADualAug−w/o.KB 37.5 67.0 48.1* −LearnDADualAug−w/.intro 39.0 66.0 49.0* −LearnDADualAug−w/.verbnet 39.4 66.7 49.5* −LearnDADualAug−w/.wordnet 39.6 67.6 49.9* LearnDAF ull 42.2 69.8 52.6* Table 3: Ablation results on event causality identification on ESC. * denotes a significant test at the level of 0.05. BERTOrgAug and BERTDualAug denote the BERT is further trained on no-dual and dual augmented data respectively; LearnDADual denotes our identifier is only trained by dual learning without further training; LearnDADualAug−w/o.KB denotes the LearnDADual is further trained by dual augmented data without knowledge guiding; LearnDADualAug−w/.<kb> denotes LearnDADual is further trained by dual augmented data guided with knowledge base kb. Figure 5: The impact of the training rounds of dual learning on event causality identification on ESC. In each round, we generate new training data by the generator at the current round. The performance is achieved by further training the identifier at the current round with the aforementioned newly generated data. ther improving the performance accordingly. 4.4 Effect of Knowledge Guiding Table 3 also illustrates the effect of knowledge guiding on ECI depending on different knowledge bases. 1) Comparing LearnDAFull with LearnDADualAug−w/o.KB, we note that the augmented data guided by external knowledge can further improve the performance of ECI. 2) Specifically, lexical expanding and connective introducing (Sec 3.2) can both make the representation of causal relation more generalized, further making it easier for the identifier to understand the causality. 3) Moreover, the expanding is more effective than the introducing, because the former brings a wider range of effective knowledge, thus the guidance of Method P R F BERT (Our identifier) 36.1 56.0 43.9 TextSurfaceBERT 37.0 57.5 45.0* BackTranslationBERT 36.8 61.0 45.9* EDABERT 36.6 62.4 46.1* LearnDABERT 37.8 65.6 48.0* Table 4: Results of different data augmentation methods on event causality identification on ESC dataset. * denotes a significant test at the level of 0.05. Gold EDA BackTrans LearnDA Causality 3.80 3.20 3.70 3.60 Well-formedness 3.95 2.75 3.83 3.64 Diversity (Man/Auto) 0.0/1.0 3.08/0.70 2.80/0.85 3.51/0.66 Table 5: Manual (4-score rating (0, 1, 2, 3)) and automatic (BLEU score) evaluation of the generated sentences via different methods from causality, well-formedness and diversity. Causality and wellformedness are assessed manually, while diversity is assessed manually and automatically. causal-related knowledge is better. 4.5 Our Augmentation vs. Other NLP Augmentations In this section, we conduct a comparison between our augmentation framework and other NLPrelated augmentation methods to further illustrate the effectiveness of LearnDA. Effectiveness of Our Augmentation We train our identifier with augmented data produced by different NLP-related augmentation methods. As shown in Table 4, the augmented data generated by our LearnDA is more efficient for ECI, which is consistent with the previous analysis. The LearnDA can generate well-formed task-related new sentences that contain more event causal knowledge. Specifically, 1) text surface transformation brings a slight change to the labeled data, thus it has relatively little impact on ECI; 2) Back translation introduces limited new causal expressions by translation, thus it slightly increases the recall value on ECI; 3) EDA can introduce new expressions via substitution, but the augmented data is not canonical and cannot accurately express the causality, therefore, its impact on ECI is also limited. Quantitative Evaluation of Task-relevance We select five Ph.D. students majoring in NLP to manual score the 100 randomly selected augmented sentences given their corresponding original sentences as reference (Cohen’s kappa = 0.85). Furthermore, we calculate the BLEU (Papineni et al., 2002) value to further evaluate the 3566 Generator Identifier <crash, target> causal relation A was crash by B as C targeted ... non-causal relation A was crash by B because C targeted ... Generator Identifier <order, attack> ... A ordered B to attack ... non-causal relation ... A order when B attack ... causal relation Dual reward feedback Dual reward feedback a) b) Figure 6: The modification of dual learning. diversity. As aforementioned, the task-relevance of new sentences on ECI is manifested in causality and well-formedness, while the diversity indicates the degree of generalization. As shown in Table 5, we note the sentences generated by LearnDA are equipped with the above three properties that are close to the labeled sentences. Specifically, the sentences produced by EDA has a certain degree of causality and diversity due to the lexical substitution assisted by external knowledge. However, they cannot well express the causality due to the grammatical irregularities. Correspondingly, new sentences generated via back translation are very similar to the original sentences, while the diversity is poor. 4.6 Case Study We conduct a case study to further investigate the effectiveness of our LearnDA. Figure 6 illustrates the modification process of dual learning. For example as a), given two causal events, the generator is expected to generate a causal sentence. However, the generator without dual learning produces a noncausal sentence. Fortunately, with dual learning, the identifier judges the generated sentence as a non-causal one and guides the generator to produce a causal sentence with the feedback. Similarly, as shown in b), given a causal sentence, the identifier is expected to output a causal relation, but no dual-trained one cannot do. Correspondingly, the generator constructs feedback of low confidence to guide the identifier to output a causal relation. 5 Conclusion This paper proposes a new learnable knowledgeguided data augmentation framework (LearnDA) to solve the data lacking problem on ECI. Our framework can leverage the duality between generation and identification via dual learning to generate task-related sentences for ECI. Moreover, our framework is knowledge guided and learnable. Our method achieves state-of-the-art performance on EventStoryLine and Causal-TimeBank datasets. Acknowledgments We thank anonymous reviewers for their insightful comments and suggestions. This work is supported by the National Key Research and Development Program of China (No.2018YFB1005100), the National Natural Science Foundation of China (No.U1936207, 61806201). This work is also supported by Beijing Academy of Artificial Intelligence (BAAI2019QN0301) and the joint project with Beijing Baidu Netcom Science Technology Co., Ltd. References Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich, Amir Kantor, George Kour, Segev Shlomov, Naama Tepper, and Naama Zwerdling. 2019. Not enough data? deep learning to the rescue! ArXiv, abs/1911.03118. Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998. The Berkeley FrameNet project. In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 1, pages 86–90, Montreal, Quebec, Canada. Association for Computational Linguistics. Brandon Beamer and Roxana Girju. 2009. Using a bigram event model to predict causal potential. In International Conference on Intelligent Text Processing and Computational Linguistics, pages 430–441. Springer. Antoine Bordes, Nicolas Usunier, Alberto GarciaDuran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In Advances in neural information processing systems, pages 2787–2795. Ruisheng Cao, Su Zhu, Chen Liu, Jieyu Li, and Kai Yu. 2019. Semantic parsing with dual learning. pages 51–64. Ruisheng Cao, Su Zhu, Chenyu Yang, Chen Liu, Rao Ma, Yanbin Zhao, Lu Chen, and Kai Yu. 2020. Unsupervised dual paraphrasing for two-stage semantic parsing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6806–6817, Online. Association for Computational Linguistics. Tommaso Caselli and Piek Vossen. 2017. The event StoryLine corpus: A new benchmark for causal and temporal relation extraction. In Proceedings of the 3567 Events and Stories in the News Workshop, pages 77– 86, Vancouver, Canada. Association for Computational Linguistics. Amit Chaudhary. 2020. A visual survey of data augmentation in nlp. Yubo Chen, Shulin Liu, Xiang Zhang, Kang Liu, and Jun Zhao. 2017. Automatically labeled data generation for large scale event extraction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 409–419, Vancouver, Canada. Association for Computational Linguistics. Fei Cheng and Yusuke Miyao. 2017. Classifying temporal relations by bidirectional LSTM over dependency paths. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1–6, Vancouver, Canada. Association for Computational Linguistics. Prafulla Kumar Choubey and Ruihong Huang. 2017. A sequential model for classifying temporal relations between intra-sentence events. pages 1796–1802. Claude Coulombe. 2018. Text data augmentation made simple by leveraging nlp cloud apis. ArXiv, abs/1812.04718. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Quang Do, Yee Seng Chan, and Dan Roth. 2011. Minimally supervised event causality identification. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 294–303, Edinburgh, Scotland, UK. Association for Computational Linguistics. Jesse Dunietz, Lori Levin, and Jaime Carbonell. 2015. Annotating causal language using corpus lexicography of constructions. In Proceedings of The 9th Linguistic Annotation Workshop, pages 188–196, Denver, Colorado, USA. Association for Computational Linguistics. Jesse Dunietz, Lori Levin, and Jaime Carbonell. 2017. The BECauSE corpus 2.0: Annotating causality and overlapping relations. In Proceedings of the 11th Linguistic Annotation Workshop, pages 95–104, Valencia, Spain. Association for Computational Linguistics. Lei Gao, Prafulla Kumar Choubey, and Ruihong Huang. 2019. Modeling document-level causal structures for event causal relation identification. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1808–1817, Minneapolis, Minnesota. Association for Computational Linguistics. Roxana Girju. 2003. Automatic detection of causal relations for question answering. In Proceedings of the ACL 2003 Workshop on Multilingual Summarization and Question Answering, pages 76–83, Sapporo, Japan. Association for Computational Linguistics. PDTB Research Group et al. 2008. The pdtb 2.0. Annotation Manual. Technical Report IRCS-08-01, Institute for Research in Cognitive Science, University of Pennsylvania. Hongyu Guo, Yongyi Mao, and Richong Zhang. 2019. Augmenting data with mixup for sentence classification: An empirical study. ArXiv, abs/1905.08941. Chikara Hashimoto, Kentaro Torisawa, Julien Kloetzer, Motoki Sano, Istv´an Varga, Jong-Hoon Oh, and Yutaka Kidawara. 2014. Toward future scenario generation: Extracting event causality exploiting semantic relation, context, and association features. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 987–997, Baltimore, Maryland. Association for Computational Linguistics. Christopher Hidey and Kathy McKeown. 2016. Identifying causal relations using parallel Wikipedia articles. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1424–1433, Berlin, Germany. Association for Computational Linguistics. Zhichao Hu, Elahe Rahimtoroghi, and Marilyn Walker. 2017. Inference of fine-grained event causality from blogs and films. pages 52–58. Zhichao Hu and Marilyn Walker. 2017. Inferring narrative causality between event pairs in films. pages 342–351. Jian Liu, Yubo Chen, and Jun Zhao. 2020. Knowledge enhanced event causality identification with mention masking generalizations. In IJCAI-20, pages 3608– 3614. International Joint Conferences on Artificial Intelligence Organization. Main track. George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39– 41. Paramita Mirza. 2014. Extracting temporal and causal relations between events. pages 10–17. Paramita Mirza, Rachele Sprugnoli, Sara Tonelli, and Manuela Speranza. 2014. Annotating causality in the TempEval-3 corpus. In Proceedings of the EACL 2014 Workshop on Computational Approaches to Causality in Language (CAtoCL), pages 10–19, Gothenburg, Sweden. Association for Computational Linguistics. 3568 Paramita Mirza and Sara Tonelli. 2014. An analysis of causality between events and its relation to temporal information. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2097– 2106, Dublin, Ireland. Dublin City University and Association for Computational Linguistics. Paramita Mirza and Sara Tonelli. 2016. CATENA: CAusal and TEmporal relation extraction from NAtural language texts. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 64–75, Osaka, Japan. The COLING 2016 Organizing Committee. Nasrin Mostafazadeh, Alyson Grealish, Nathanael Chambers, James Allen, and Lucy Vanderwende. 2016. CaTeRS: Causal and temporal relation scheme for semantic annotation of event structures. In Proceedings of the Fourth Workshop on Events, pages 51–61, San Diego, California. Association for Computational Linguistics. Jong-Hoon Oh, Kentaro Torisawa, Chikara Hashimoto, Motoki Sano, Stijn De Saeger, and Kiyonori Ohtake. 2013. Why-question answering using intra- and inter-sentential causal relations. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1733–1743, Sofia, Bulgaria. Association for Computational Linguistics. Jong-Hoon Oh, Kentaro Torisawa, Canasai Kruengkrai, Ryu Iida, and Julien Kloetzer. 2017. Multi-column convolutional neural networks with causality-attention for why-question answering. In Proceedings of the Tenth ACM International Conference on Web Search and Data Mining, pages 415– 424. ACM. Yannis Papanikolaou and A. Pierleoni. 2020. Dare: Data augmented relation extraction with gpt-2. ArXiv, abs/2004.13845. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Mehwish Riaz and Roxana Girju. 2010. Another look at causality: Discovering scenario-specific contingency relationships with no supervision. In 2010 IEEE Fourth International Conference on Semantic Computing, pages 361–368. IEEE. Mehwish Riaz and Roxana Girju. 2013. Toward a better understanding of causality between verbal events: Extraction and analysis of the causal power of verbverb associations. In Proceedings of the SIGDIAL 2013 Conference, pages 21–30, Metz, France. Association for Computational Linguistics. Mehwish Riaz and Roxana Girju. 2014a. In-depth exploitation of noun and verb semantics to identify causation in verb-noun pairs. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pages 161– 170, Philadelphia, PA, U.S.A. Association for Computational Linguistics. Mehwish Riaz and Roxana Girju. 2014b. Recognizing causality in verb-noun pairs via noun and verb semantics. In Proceedings of the EACL 2014 Workshop on Computational Approaches to Causality in Language (CAtoCL), pages 48–57, Gothenburg, Sweden. Association for Computational Linguistics. Dana Ruiter, Cristina Espa˜na-Bonet, and Josef van Genabith. 2019. Self-supervised neural machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1828–1834, Florence, Italy. Association for Computational Linguistics. Karin Kipper Schuler. 2005. Verbnet: A broadcoverage, comprehensive verb lexicon. Lei Shen and Yang Feng. 2020. CDL: Curriculum dual learning for emotion-controllable response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 556–566, Online. Association for Computational Linguistics. Shang-Yu Su, Chao-Wei Huang, and Yun-Nung Chen. 2019. Dual supervised learning for natural language understanding and generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5472–5477, Florence, Italy. Association for Computational Linguistics. Shang-Yu Su, Chao-Wei Huang, and Yun-Nung Chen. 2020. Towards unsupervised language understanding and generation by joint dual learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 671–680, Online. Association for Computational Linguistics. Mingming Sun, Xu Li, and Ping Li. 2018. Logician and orator: Learning from the duality between language and knowledge in open domain. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium. Association for Computational Linguistics. William Yang Wang and Diyi Yang. 2015. That’s so annoying!!!: A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using #petpeeve tweets. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2557–2563, Lisbon, Portugal. Association for Computational Linguistics. Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. In Proceedings of the 3569 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6382–6388, Hong Kong, China. Association for Computational Linguistics. Yingce Xia, Tao Qin, Wei Chen, Jiang Bian, Nenghai Yu, and Tie-Yan Liu. 2017. Dual supervised learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3789– 3798. JMLR. org. Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L. Yuille, and Quoc V. Le. 2019a. Adversarial examples improve image recognition. ArXiv, abs/1911.09665. Qizhe Xie, Zihang Dai, Eduard H. Hovy, Minh-Thang Luong, and Quoc V. Le. 2019b. Unsupervised data augmentation for consistency training. arXiv: Learning. Sen Yang, Dawei Feng, Linbo Qiao, Zhigang Kan, and Dongsheng Li. 2019. Exploring pre-trained language models for event extraction and generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5284–5294, Florence, Italy. Association for Computational Linguistics. Hai Ye, Wenjie Li, and Lu Wang. 2019. Jointly learning semantic parser and natural language generator via dual information maximization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2090–2101, Florence, Italy. Association for Computational Linguistics. Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In NIPS. Xinyu Zuo, Yubo Chen, Kang Liu, and Jun Zhao. 2020. KnowDis: Knowledge enhanced data augmentation for event causality detection via distant supervision. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1544–1550, Barcelona, Spain (Online). International Committee on Computational Linguistics. A Supplementary Experiment Results A.1 Statistics of Dual Augmented Data Annotated data Augmented data #causal ep. 1170 3588 #causal sent. 1770 10442 #Ave sent. 1.5 2.9 Table 6: Statistics of causal event pairs and causal sentences in labeled data (ESC) and dual augmented data. (#causal ep. denotes the number of causal event pairs after removing duplicates, #causal sent. denotes the number of causal sentences, #Ave sent. denotes the average number of causal sentences containing the same causal event pair.) As shown in Table 6, our dual augmented data is significantly more quantitative than the labeled data. Specifically, the causal event pairs are increased by 3.1 times, the causal sentences are increased by 5.9 times and the average number of causal sentences corresponding to each causal event pair is also increased. A.2 Effectiveness of Different Quantities of Augmented Training Data Ratio P R F1 1:1 37.3 64.7 47.3* 1:2 37.8 65.6 48.0* 1:3 37.0 64.8 47.1* 1:4 36.2 64.2 46.3* Table 7: Performance of identifier (BERT) trained with different ratios of labeled data and dual augmented data. * denotes a significant test at the level of 0.05. We change the quantity of dual augmented data for training to explore the influence of augmentation ratio on ECI. As shown in Table 7, when the ratio is 1:2, the effective knowledge brought by dual augmented data is maximized. And as the ratio increasing, the dual augmented data will bring noises, which obstructs the model to identify event causality and may change the data distribution from original data (Xie et al., 2019a). This suggests that too much augmented data is not better and that there is a trade-off between introducing knowledge and reducing noise. A.3 Effectiveness of Extracting Event Pairs with Different Filtering Ratios Table 8 tries to show the effectiveness of extracting event pairs with different filtering ratios on ECI. With the ratio of retained event pairs increasing, 3570 α P R F1 ∇ 30% 37.8 65.6 48.0* 40% 37.0 65.7 47.3* -0.7 50% 36.2 65.0 46.5* -1.5 Table 8: Performance of identifier (BERT) trained with different extracting event pairs filtered in different α. * denotes a significant test at the level of 0.05. the augmented data hurts ECI’s performance. This proves the effectiveness of filtering, which further improves the causality of the generated sentences. A.4 Effectiveness of Generated sentences with Different Filtering Ratios β P R F1 ∇ 50% 37.8 65.6 48.0* 60% 37.3 65.3 47.5* -0.5 70% 36.9 64.9 47.0* -1.0 80% 36.6 64.5 46.7* -1.3 Table 9: Performance of identifier (BERT) trained with new generated sentences filtered in different β. * denotes a significant test at the level of 0.05. Table 9 tries to show the effectiveness of generated sentences with different filtering ratios. With the ratio of retained generated sentences increasing, the contribution of filtered generated sentences for ECI decreases gradually. This proves the effectiveness of filtering, which can balance the overall quality of the sentences against diversity. B Supplementary Related Work B.1 Dual Learning For many Natural Language Processing (NLP) tasks, there exist many primal and dual tasks, such as open information narration (OIN) and open information extraction (OIE) (Sun et al., 2018), natural language understanding (NLU) and natural language generation (NLG) (Su et al., 2019, 2020), semantic parsing and natural language generation (Ye et al., 2019; Cao et al., 2019, 2020), link prediction and entailment graph induction (Cao et al., 2019), query-to-response and response-to-query generation (Shen and Feng, 2020) and so on. The duality between the primal task and the dual task is considered as a constraint that both problems must share the same joint probability mutually. Recently, inspired by Xia et al. (2017) who implemented the duality in a neural-based dual learning system, the above primal-dual tasks are implemented in two different ways: 1) providing additional labeled samples via bootstrapping, and 2) adding rewards at the training stage for each agent. We observe that the event causality identification and the sentence generation are dual to each other. Therefore, we apply a dual learning framework in the second way to optimize identification and generation interactively for generating ECI-related data. B.2 Data Augmentation for NLP The scarcity of annotated data is a thorny problem in machine learning. Unlike computer vision, the augmentation of text data in NLP is pretty rare. Existing text data augmentation methods for NLP tasks are almost task-independent frameworks and can be roughly summarized into the following categories (Chaudhary, 2020): (1) Lexical substitution tries to substitute words without changing the meaning (Zhang et al., 2015; Wei and Zou, 2019; Wang and Yang, 2015; Xie et al., 2019b); (2) Back translation tries to paraphrase a text while retraining the meaning (Xie et al., 2019b); (3) Text surface transformation tries to match transformations using regex (Coulombe, 2018); (4) Random noise injection tries to inject noise in the text to make the model more robust (Wei and Zou, 2019); (5) Generative method tries to generate additional training data while preserving the class label (Anaby-Tavor et al., 2019; Yang et al., 2019); (6) Distantly supervision and self-supervision try to introduce new training data from unlabeled text (Chen et al., 2017; Ruiter et al., 2019). As aforementioned, these frameworks cannot directly produce new suitable task-related examples for ECI. However, (1), (3), and (4) cannot guarantee the causality and wellformedness of new examples for ECI. Additionally, (2) and (5) are not easy to directly use external knowledge bases to generalize the event-related causal commonsense. Furthermore, (6) needs to design proprietary processing methods to generate ECI task-related training data. Zuo et al. (2020) solved the data lacking problem of ECI with the distantly supervised labeled training data. However, including the distant supervision, most of the existing text data augmentation methods for NLP tasks are task-independent frameworks. Therefore, we introduce a new learnable framework for augmenting task-related training data for ECI via dual learning enhanced with external knowledge. C Generation with ConceptNet To make a fair comparison, we introduce causalrelated events from ConceptNet based on causal3571 related concepts, and obtain the causal sentence via the method in KonwDis (Zuo et al., 2020) to further re-train MM (Liu et al., 2020). Specifically, firstly, we obtain triples based on cause-related semantic relations from ConceptNet, such as Causes, HasSubevent, HasFirstSubevent, HasLastSubevent, MotivatedByGoal, and CausesDesire relations. Secondly, we assemble any two events from obtained causal triples to generate causal event pairs set and filter them via the filter of KonwDis. Next, we employ filtered causal event pairs to collect preliminary noisy labeled sentences from external documents via the DistantAnnotator of KonwDis. Then, we use the CommonFilter of KnowDis assisted with causal commonsense knowledge to pick out labeled sentences that express causal semantics between events. Finally, the refined causal sentences are input into LearnDA to generated ECIrelated dual augmented training data and further train the MM to obtain MM+ConceptAug. D Main Experimental Environments and Other Parameters Settings D.1 Experimental Environments We deploy all models on a server with 250GB of memory and 4 TITAN Xp GPUs. Specifically, the configuration environment of the server is ubuntu 16.04, and our framework mainly depends on python 3.6.0 and PyTorch 1.0. D.2 Other Parameters Settings All the final hyper-parameters for evaluation are averaged after 3 independent tunings on the development set. Moreover, the whole dual learning framework which includes event causality identifier and knowledge guided sentence generator takes approximately 5 minutes per epoch when training. According to the early stop strategy, the training rounds for different folds are different, and it takes about 20-30 rounds.
|
2021
|
276
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3572–3581 August 1–6, 2021. ©2021 Association for Computational Linguistics 3572 Revisiting the Negative Data of Distantly Supervised Relation Extraction Chenhao Xie1,2, Jiaqing Liang1,2, Jingping Liu1, Chengsong Huang1, Wenhao Huang1, Yanghua Xiao1,3∗ 1Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University 2Shuyan Technology Inc., Shanghai, China 3Fudan-Aishu Cognitive Intelligence Joint Research Center, Shanghai, China {redreamality, l.j.q.light}@gmail.com {jpliu17, huangcs19, whhuang17, shawyh}@fudan.edu.cn Abstract Distantly supervision automatically generates plenty of training samples for relation extraction. However, it also incurs two major problems: noisy labels and imbalanced training data. Previous works focus more on reducing wrongly labeled relations (false positives) while few explore the missing relations that are caused by incompleteness of knowledge base (false negatives). Furthermore, the quantity of negative labels overwhelmingly surpasses the positive ones in previous problem formulations. In this paper, we first provide a thorough analysis of the above challenges caused by negative data. Next, we formulate the problem of relation extraction into as a positive unlabeled learning task to alleviate false negative problem. Thirdly, we propose a pipeline approach, dubbed RERE, that first performs sentence classification with relational labels and then extracts the subjects/objects. Experimental results show that the proposed method consistently outperforms existing approaches and remains excellent performance even learned with a large quantity of false positive samples. Source code is available online1. 1 Introduction Relational extraction is a crucial step towards knowledge graph construction. It aims at identifying relational triples from a given sentence in the form of ⟨subject, relation, object⟩, in short, ⟨s, r, o⟩. For example, given S1 in Figure 1, we hope to extract ⟨WILLIAM SHAKESPEARE, BIRTHPLACE, STRATFORD-UPON-AVON⟩. This task is usually modeled as a supervised learning problem and distant supervision (Mintz et al., 2009) is utilized to acquire large-scale training data. The core idea is to obtain training data ∗Corresponding author 1https://github.com/redreamality/ RERE-relation-extraction Figure 1: Illustration of distant supervision process. S2S5 are examples for four kinds of label noise. TP, FP, FN and PL mean true positive, false positive, false negative and partially labeled, respectively. “R-” or “E-” indicates whether the error occurs at relation-level or entitylevel. Bold tokens are ground-truth subjects/objects. Underlined tokens together with the relation in the third column are labeled by distant supervision. “NA” means no relation. is through automatically labeling a sentence with existing relational triples from a knowledge base (KB). For example, given a triple ⟨s, r, o⟩and a sentence, if the sentence contains both s and o, distant supervision methods regard ⟨s, r, o⟩as a valid sample for the sentence. If no relational triples are applicable, the sentence is labeled as “NA”. Despite the abundant training data obtained with distant supervision, nonnegligible errors also occur in the labels. There are two types of errors. In the first type, the labeled relation does not conform with the original meaning of sentence, and this type of error is referred to as false positive (FP). For example, in S2, the sentence “Shakespeare spent the last few years of his life in Stratford-upon-Avon.” does not express the relation BIRTHPLACE, thus being a FP. In the second type, large amounts of 3573 relations in sentences are missing due to the incompleteness of KB, which is referred to as false negative (FN). For instance, in S3, “Buffett was born in 1930 in Omaha, Nebraska.” is wrongly labeled as NA since there is no relation (e.g., BIRTHPLACE) between BUFFETT and OMAHA, NEBRASKA in the KB. Many efforts have been devoted to solving the FP problem, including pattern-based methods (Jia et al., 2019), multi-instance learning methods (Lin et al., 2016; Zeng et al., 2018a) and reinforcement learning methods (Feng et al., 2018). Significant improvements have been made. However, FN problem receives much less attention (Min et al., 2013; Xu et al., 2013; Roller et al., 2015). To the best of our knowledge, none existing work with deep neural networks to solve this problem. We argue that this problem is fatal in practice since there are massive FN cases in datasets. For example, there exist at least 33% and 35% FNs in NYT and SKE datasets, respectively. We will deeply analyze the problem in Section 2.1 Another huge problem in relation extraction is the overwhelming negative labels. As is widely acknowledged, information extraction tasks are highly imbalanced in class labels (Chowdhury and Lavelli, 2012; Lin et al., 2018; Li et al., 2020). In particular, the negative labels account for most of the labels in relation extraction under almost any problem formulation, which makes relation extraction a hard machine learning problem. We systematically analyze this in Section 2.2. In this paper, we address these challenges caused by negative data. Our main contribution can be summarized as follows. • We systematically compare the class distributions of different problem modeling and explain why first extract relation then entities, i.e., the third paradigm (P3) in Section 2.2, is superior to the others. • Based on the first point, we adopt P3 and propose a novel two-staged pipeline model dubbed RERE. It first detects relation at sentence level and then extracts entities for a specific relation. We model the false negatives in relation extraction as “unlabeled positives” and propose a multi-label collective loss function. • Our empirical evaluations show that the proposed method consistently outperforms existing approaches, and achieves excellent performance even learned with a large quantity of false positive samples. We also provide two carefully annotated test sets aiming at reducing the false negatives of previous annotation, namely, NYT21 and SKE21, with 370 and 1150 samples, respectively. 2 Problem Analysis and Pilot Experiments We use (ci, Ti) to denote a training instance, where ci is a sentence consisting of N tokens ci = [ci1, ..., ciN] labeled by a set of triples Ti = {⟨s, r, o⟩} from the training set D. For rigorous definition, [ci1, ..., ciN] can be viewed as an ordered set {(ci1, 1), ..., (ciN, N)} so that set operations can be applied. We assume r ∈R, where R is a finite set of all relations in D. Other model/taskspecific notations are defined after each problem formulation. We now clarify some terms used in the introduction and title without formal definition. A negative sample refers to a triple t /∈Ti. Negative label refers to the negative class label (e.g., usually “0” for binary classification), used for supervision with respect to task-specific models. Under different task formulation, the negative labels can be different. Negative data is a general term that includes both negative labels and negative samples. There are two kinds of false negatives. Relation-level false negative (S3 in Figure 1) refers to the situation where there exists t′ = ⟨s′, r′, o′⟩/∈Ti, but r′ is actually expressed by ci, and does not appear in other t ∈Ti. Similarly, Entity-level false negative (S4 and S5 in Figure 1) means r′ appears in other t ∈Ti. Imbalanced class distribution means that the quantity of negative labels is much larger than that of positive ones. 2.1 Addressing the False Negatives As shown in Table 1, the triples in NYT (SKE) datasets2 labeled by Freebase3 (BaiduBaike4) is 88,253 (409,767), while the ones labeled by Wikidata5 (CN-DBPedia6) are 58,135 (342,931). In other words, there exists massive FN matches if only labeled by one KB due to the incompleteness of KBs. Notably, we find that the FN rate is underestimated by previous researches (Min 2Detailed description of datasets is in Sec. 5.1 3(Bollacker et al., 2008) 4https://baike.baidu.com/ 5(Vrandecic and Kr¨otzsch, 2014) 6 (Xu et al., 2017) 3574 et al., 2013; Xu et al., 2013), based on the manual evaluation of which there are 15%-35% FN matches. This discrepancy may be caused by human error. In specific, a volunteer may accidentally miss some triples. For example, as pointed out by Wei et al. (2020, in Appendix C), the test set of NYT11 (Hoffmann et al., 2011) missed lots of triples, especially when multiple relations occur in a same sentence, though labeled by human. That also provides an evidence that FN’s are harder to discover than FP’s. NYT (English) SKE (Chinese) # Sentence 56,196 194,747 # Triples # Rels # Triples # Rels Original 88,253 23 409,767 49 Re-labeled 58,135 57 342,931 378 Intersection 13,848 18 121,326 46 Union 132,540 62 631,372 381 Original FNR ≥0.33 ≥0.35 Relabel FNR ≥0.56 ≥0.46 Table 1: Statistics of the quantity of distantly labeled relational triples by using different KB’s. The “original” refers to freebase for NYT and BaiduBaike for SKE. The “relabeled” means aligning using Wikidata and CNDBpedia to re-label NYT and SKE datasets. In specific, we consider triples with the same subject and object to be candidate triples and use a relation mapping table to determine whether the triples match. The intersection of SKE dataset has two values because the original relation has a one-to-many mapping with relations in CN-DBpedia. FNR stands for false negative rates, calculated by using the # Triples in Original (Re-labeled) divided by the union. 2.2 Addressing the Overwhelming Negative Labels We point out that some of the previous paradigms designed for relation extraction aggravate the imbalance and lead to inefficient supervision. The mainstream approaches for relation extraction mainly fall into three paradigms depending on what to extract first. P1 The first paradigm is a pipeline that begins with named entity recognition (NER) and then classifies each entity pair into different relations, i.e., [s, o then r]. It is adopted by many traditional approaches (Mintz et al., 2009; Chan and Roth, 2011; Zeng et al., 2014, 2015; Gormley et al., 2015; dos Santos et al., 2015; Lin et al., 2016). P2 The second paradigm first detects all possible subjects in a sentence then identifies objects with respect to each relation, i.e., [s then r, o]. Specific implementation includes modeling relation extraction as multi-turn question answering (Li et al., 2019), span tagging (Yu et al., 2020) and cascaded binary tagging (Wei et al., 2020). P3 The third paradigm first perform sentencelevel relation detection (cf. P1, which is at entity pair level.) then extract subjects and entities, i.e., [r then s, o]. This paradigm is largely unexplored. HRL (Takanobu et al., 2019) is hitherto the only work to apply this paradigm based on our literature review. We provide theoretical analysis of the output space and class prior with statistical support from three datasets (see Section 5.1 for description) of the three paradigms in Table 2. The second step of P1 can be compared with the first step of P3. Both of them find relation from a sentence (P1 with target entity pair given). Suppose a sentence contains m entities7, the classifier has to decide relation from O(m2) entity pairs, while in reality, relations are often sparse, i.e., O(m). In other words, most entity pairs in P1 do not form valid relation, thus resulting in a low class prior. The situation is even worse when the sentence contains more entities, such as in NYT11-HRL. For P2, we demonstrate with the problem formulation of CASREL (Wei et al., 2020). The difference of the first-step class prior between P2 and P3 depends on the result of comparison between # relations and average sentence length (i.e., |R| and ¯N), which varies in different scenarios/domains. However, π2 of P2 is extremely low, where a classifier has to decide from a space of |R| ∗¯N. In contrast, P3 only need to decide from 4 ∗¯N based on our task formulation (Section 3.1) Other task formulations include jointly extracting the relation and entities (Yu and Lam, 2010; Li and Ji, 2014; Miwa and Sasaki, 2014; Gupta et al., 2016; Katiyar and Cardie, 2017; Ren et al., 2017) and recently in the manner of sequence tagging (Zheng et al., 2017), sequence-to-sequence learning (Zeng et al., 2018b). In contrast to the aforementioned three paradigms, most of these methods actually provide an incomplete decision space that cannot handle all the situation of relation extrac7Below the same. 3575 Paradigm Theoretical NYT10-HRL NYT11-HRL SKE |R|=31, ¯ N= 39.08 |R|=11, ¯ N=39.46 |R|=51, ¯ N= 54.67 π1 π2 π1 π2 π1 π2 π1 π2 s, o then r – E[ P y |R| ] – 0.01421 – 0.00280 – 0.00494 s then r, o E[ P y ¯ N ] E[ P y ¯ N∗|R|] 0.0585 0.00093 0.0574 0.00257 0.0405 0.00067 r then s, o E[ P y |R| ] E[ P y 4∗¯ N ] 0.0390 0.00842 0.0826 0.00835 0.0344 0.00927 Table 2: Comparison of class prior under different relation extraction paradigms. |R| means the total number of relations and ¯N is the average sentence length. π1 (π2) refers to the class prior for the first (second) task in the pipeline. π1 for the first paradigm is omitted because it is often considered a preceding step. P y is the summation of 1’s in labels, of using which our intention is to represent the information a positive sample conveys. tion, for example, the overlapping one (Wei et al., 2020). 3 Solution Framework 3.1 Framework of RERE Given an instance (ci, Ti) from D, the goal of training is to maximize the likelihood defined in Eq. (1). It is decomposed into two components by applying the definition of conditional probability, formulated in Eq. (2). |D| Y i=1 Pr(Ti|ci; θ) (1) = |D| Y i=1 Y r∈Ti Pr(r|ci; θ) Y ⟨s,o⟩∈Ti|r Pr(s, o|r, ci; θ), (2) where we use r ∈Ti as a shorthand for r ∈{r | ⟨s, r, o⟩∈Ti}, which means that r occurs in the triple set w.r.t. ci; Similarly, s ∈Ti, ⟨s, o⟩∈Ti|r stands for s ∈{s | ⟨s, r, o⟩∈Ti|r} and ⟨s, o⟩∈ {⟨s, o⟩| ⟨s, r, o⟩∈Ti|r}, respectively. Ti|r represents a subset of Ti with a common relation r. 1[·] is an indicator function; 1[condition] = 1 when the condition happens. We denote by θ the model parameters. Under this decomposition, relational triple extraction task is formulated into two subtasks: relation classification and entity extraction. Relation Classification. As is discussed, building relation classifier at entity-pair level will introduce excessive negative samples and form a hard learning problem. Therefore, we alternatively model the relation classification at sentence level. Intuitively speaking, we hope that the model could capture what relation a sentence is expressing. We formalize it as a multi-label classification task. Pr(r|ci; θ) = |R| Y j=1 (ˆyj rc)1[yj rc=1](1 −ˆyj rc)1[yj rc=0], (3) where ˆyj rc is the probability that c is expressing rj, the j-th relation8. yj rc is the ground truth from the labeled data; yj rc = 1 is equivalent to rj ∈Ti while yj rc = 0 means the opposite. Entity Extraction. We then model entity extraction task. We observe that given the relation r and context ci, it naturally forms a machine reading comprehension (MRC) task (Chen, 2018), where (r, ci, s/o) naturally fits into the paradigm of (QUERY, CONTEXT, ANSWER). Particularly, the subjects and objects are continuous spans from ci, which falls into the category of span extraction. We adopt the boundary detection model with answer pointer (Wang and Jiang, 2017) as the output layer, which is widely used in MRC tasks. Formally, for a sentence of N tokens, Pr(s, o|r, ci; θ) = Y k∈K N Y n=1 (ˆyn,k ee )1[yn,k ee =1](1 −ˆyn,k ee )1[yn,k ee =0], (4) where K = {sstart, send, ostart, oend} represents the identifier of each pointer; ˆyn,k ee refers to the probability of n-th token being the start/end of the subject/object. yn,k ee is the ground truth from the training data; if ∃s ∈Ti|r occurs in ci at position from n to n + l, then yn,sstart ee = 1 and yn+l,send ee = 1, otherwise 0; the same applies for the objects. 8ˆyj rc is parameterized by θ, omitted in the equation for clarity, below the same. 3576 3.2 Advantages Our task formulation shows several advantages. By adopting P3 as paradigm, the first and foremost advantage of our solution is that it suffers less from the imbalanced classes (Section 2.2). Secondly, relation-level false negative is easy to recover. When modeled as a standard classification problem, many off-the-shelf methods on positive unlabeled learning can be leveraged. Thirdly, entity-level false negatives do not affect relation classification. Taking S5 in Figure 1 as an example, even though the BIRTHPLACE relation between WILLIAM SWARTZ and SCRANTON is missing, the relation classifier can still capture the signal from the other sample with a same relation, i.e., ⟨JOE BIDEN, BIRTHPLACE, SCRANTON ⟩. Fourthly, this kind of modeling is easy to update with new relations without the need of retraining a model from bottom up. Only relation classifier needs to be redesigned, while entity extractor can be updated in an online manner without modifying the model structure. Last but not the least, relation classifier can be regarded as a pruning step when applied to practical tasks. Many existing methods treat relation extraction as question answering (Li et al., 2019; Zhao et al., 2020). However, without first identifying the relation, they all need to iterate over all the possible relations and ask diverse questions. This results in extremely low efficiency where time consumed for predicting one sample may take up to |R| times larger than our method. 4 Our Model The relational triple extraction task decomposed in Eq. (2) inspires us to design a two-staged pipeline, in which we first detect relation at sentence level and then extract subjects/objects for each relation. The overall architecture of RERE is shown in Figure 2. 4.1 Sentence Classifier with Relational Label We first detect relation at sentence level. The input is a sequence of tokens c and we denote by ˆyrc = [ˆy1 rc, ˆy2 rc, ..., ˆy|R| rc ] the output vector of the model, which aims to estimate ˆyi rc in Eq. (3). We use BERT (Devlin et al., 2019) for English and RoBERTa (Liu et al., 2019) for Chinese, pretrained language models with multi-layer bidirectional Transformer structure (Vaswani et al., 2017), to encode the inputs9. Specifically, the input sequence xrc = [[CLS], ci, [SEP]], which is fed into BERT for generating a token representation matrix Hrc ∈RN×d, where d is the hidden dimension defined by pre-trained Transformers. We take h0 rc, which is the encoded vector of the first token [CLS], as the representation of the sentence. The final output of relation classification module ˆyrc is defined in Eq. (5). ˆyrc = σ(Wrch0 rc + brc), (5) where Wrc and brc are trainable model parameters, representing weights and bias, respectively; σ denotes the sigmoid activation function. 4.2 Relation-specific Entity Extractor After the relation detected at sentence-level, we extract subjects and objects for each candidate relation. We aim to estimate ˆyee = [0, 1]N×4, of which each element corresponds to ˆyn,k ee in Eq. (4), using a deep neural model. We take ˆyrc, the one-hot output vector of relation classifier, and generate query tokens q using each of the detected relations (i.e., the “1”s in ˆyrc). We are aware that many recent works (Li et al., 2019; Zhao et al., 2020) have studied how to generate diverse queries for the given relation, which have the potential of achieving better performance. Nevertheless, that is beyond the scope of this paper. To keep things simple, we use the surface text of a relation as the query. Next, the input sequence is constructed as xee = [[CLS], qi, [SEP], ci, [SEP]]. Like Section 4.1, we get the token representation matrix Hee ∈RN×d from BERT. The k-th output pointer of entity extractor is defined by ˆyk ee = σ(Wk eeHee + bk ee), (6) where k ∈{sstart, send, ostart, oend} is in accordance to Eq. (4); Wk ee and bk ee are the corresponding parameters. The final subject/object spans are generated by pairing the nearest sstart/ostart with send/oend. Next, all subjects are paired to the nearest object. If multiple objects occur before the next subject appears, all subsequent objects will be paired with it until next subject occurs. 9For convenience, we refer to the pre-trained Transformer as BERT hereinafter. 3577 The comic book character Aurakles was created by American artist Dick Dillin . Pre-trained Transformer Encoder [CLS] Pre-trained Transformer Encoder [CLS] [SEP] [SEP] 0 1 0 0 0 0 1 0 0 0 0 0 0 [SEP] Creator Nationality 1 1 1 1 1 1 1 1 sstart send ostart oend Query generation Entity Extractor Relation Classifier Figure 2: The overall architecture of RERE. In this example, there are two relations, NATIONALITYand CREATOR, can be found in the Relation Classifier, which will be sent to the Entity Extractor one by one along with the sentence. When The relation NATIONALITY is extracted, the Entity Extractor will find the position of the subject and object of Nationality. The word AMERICAN and DICK DILLIN will be found. The relation CREATOR will then be handled similarly. The values of grey blocks in ˆyee are zero. 4.3 Multi-label Collective Loss function In normal cases, the log-likelihood is taken as the learning objective. However, as is emphasized, there exist many false negative samples in the training data. Intuitively speaking, the negative labels cannot be simply considered as negative. Instead, a small portion of the negative labels should be considered as unlabeled positives and their influence towards the penalty should be eradicated. Therefore, we adopt cPU (Xie et al., 2020), a collective loss function that is designed for positive unlabeled learning (PU learning). To briefly review, cPU considers the learning objective to be the correctness under a surrogate function, ℓ(ˆy, y) = ln(c(ˆy, y)), (7) where they redefine the correctness function for PU learning as c(ˆy, y) = ( E[ˆy] if y = 1, 1 −|E[ˆy] −µ| otherwise, (8) where µ is the ratio of false negative data (i.e., the unlabeled positive in the original paper). We extend it to multi-label situation by embodying the original expectation at sample level. Due to the fact that class labels are highly imbalanced for our tasks, we introduce a class weight γ ∈(0, 1) to downweight the positive penalty. For relation classifier, ℓrc(ˆy, y) = −γrc ln( 1 |R| |R| X i=1 ˆyi rc]) if yi rc = 1 −ln(1 −| 1 |R| |R| X i=1 ˆyi rc −µrc|) otherwise. (9) For entity extractor, ℓee(ˆyk, yk) = −γee ln( N X n=1 ˆyn,k ee ]) if yn,k ee = 1 −ln(1 −| N X n=1 ˆyn,k ee −µee|) otherwise. (10) In practice, we set µ = π(τ + 1), where τ ≈ 1 −# labeled positive # all positive is the ratio of false negative and π is the class prior. Note that µ is not difficult to estimate for both relation classification and entity extraction task in practice. Besides various 3578 of methods in the PU learning (du Plessis et al., 2015; Bekker and Davis, 2018) for estimating it, an easy approximation is µ ≈π when π ≪τ, which happens to be the case for our tasks. 5 Experiments 5.1 Datasets Our experiments are conducted on these four datasets10. Some statistics of the datasets are provided in Table 1 and Table 2. In relation extraction, some datasets with the same names involve different preprocessing, which leads to unfair comparison. We briefly review all the datasets below and specify the operations to perform before applying each dataset. • NYT (Riedel et al., 2010). NYT is the very first version among all the NYT-related datasets. It is based on the articles in New York Times12. We use the sentences from it to conduct the pilot experiment in Table 1. However, 1) it contains duplicate samples, e.g., 1504 in the training set; 2) It only labels the last word of an entity, which will mislead the evaluation results. • NYT10-HRL. & NYT11-HRL. These two datasets are based on NYT. The difference is that they both contain complete entity mentions. NYT10 (Riedel et al., 2010) is the original one. and NYT11 (Hoffmann et al., 2011) is a small version of NYT10 with 53,395 training samples and a manually labeled test set of 368 samples. We refer to them as NYT10HRL and NYT11-HRL after preprocessed by HRL (Takanobu et al., 2019) where they removed 1) training relation not appearing in the testing and 2) “NA” sentences. These two steps are almost adopted by all the compared methods. To compare fairly, we use this version in evaluations. • NYT21. We provide relabel version of the test set of NYT11-HRL. The test set of NYT11HRL still have false negative problem. Most of the samples in the NYT11-HRL has only one relation. We manually added back the missing triples to the test set. 10We do not use WebNLG (Gardent et al., 2017) and ACE0411 because these datasets are not automatically labeled by distant supervision. WebNLG is constructed by natural language generation with triples. ACE04 is manually labeled. 12https://www.nytimes.com/ • SKE2019/SKE2113. SKE2019 is a dataset in Chinese published by Baidu. The reason we also adopt this dataset is that it is currently the largest dataset available for relation extraction. There are 194,747 sentences in the training set and 21,639 in the validation set. We manually labeled 1,150 sentences from the test set with 2,765 annotated triples, which we refer to as SKE21. No preprocessing for this dataset is needed. We provide this data for future research14. 5.2 Compared Methods and Metrics We evaluate our model by comparing with several models on the same datasets, which are SOTA graphical model MultiR (Hoffmann et al., 2011), joint models SPTree (Miwa and Bansal, 2016) and NovelTagging (Zheng et al., 2017), recent strong SOTA models CopyR (Zeng et al., 2018b), HRL (Takanobu et al., 2019), CasRel (Wei et al., 2020), TPLinker (Wang et al., 2020). We also provide the result of automatically aligning Wikidata/CN-KBpedia with the corpus, namely Match, as a baseline. To note, we only keep the intersected relations, otherwise it will result in low precision due to the false negative in the original dataset. We report standard micro Precision (Prec.), Recall (Rec.) and F1 score for all the experiments. Following the previous works (Takanobu et al., 2019; Wei et al., 2020), we adopt partial match on these data sets for fair comparison. We also provide the results of exact match results of the methods we implemented, and only exact match on SKE2019. 5.3 Overall Comparison We show the overall comparison result in Table 3. First, we observe that RERE consistently outperforms all the compared models. We find an interesting result that by purely aligning the database with the corpus, it already achieves surprisingly good overall result (surpassing MultiR) and relatively high precision (comparable to CoType in NYT11-HRL). However, the recall is quite low, which is consistent with our discussion in Section 2.1 that distant supervision leads to many false negatives. We also provide an ablation result where BERT is replaced with a bidirectional 13http://ai.baidu.com/broad/download? dataset=sked 14download url. 3579 NYT10-HRL NYT11-HRL NYT21 SKE21 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 KB Match 38.10 32.38 34.97 47.92 31.08 37.7 47.92 29.56 36.57 69.12 28.1 39.96 MultiR (Hoffmann et al., 2011) 32.8 30.6 31.7 SPTree (Miwa and Bansal, 2016) 49.2 55.7 52.2 52.2 54.1 53.1 NovelTagging (Zheng et al., 2017) 59.3 38.1 46.4 46.9 48.9 47.9 CoType (Ren et al., 2017) 48.6 38.6 43.0 CopyR (Zeng et al., 2018b) 56.9 45.2 50.4 34.7 53.4 42.1 HRL (Takanobu et al., 2019) 71.4 58.6 64.4 53.8 53.8 53.8 TPLinker (Wang et al., 2020)* 81.19 65.41 72.45 56.2 55.14 55.67 59.78 55.78 57.71 CasRel (Wei et al., 2020)* 77.7 68.8 73.0 50.1 58.4 53.9 58.64 56.62 57.61 RERE - LSTM 56.71 42.00 48.26 56.46 35.4 43.52 62.06 37.01 46.37 RERE 75.45 72.50 73.95 53.12 59.59 56.23 57.69 61.69 59.62 TPLinker (Wang et al., 2020)*(exact) 80.34 65.11 71.93 55.43 55.12 55.28 58.96 55.78 57.33 83.86 84.77 84.32 CasRel (Wei et al., 2020)*(exact) 75.12 65.72 70.11 47.88 55.13 51.25 55.06 54.49 54.78 86.94 85.96 86.45 RERE (exact) 74.90 71.97 73.4 52.40 58.91 55.47 56.97 60.93 58.88 90.44 84.20 87.21 Table 3: The main evaluation results of different models on NYT10-HRL, NYT11-HRL, and two hand labeled test sets NYT21 and SKE21 on by the compared method on the datasets. The results with only one decimal are quoted from (Wei et al., 2020). The methods with * are based on our re-implementation. Best partial (exact) match results are marked bold (underlined). 0.50 0.56 0.62 0.70 Recall 0.8 0.9 Precision NYT10-HRL CasRel-0.1 CasRel-0.3 CasRel-0.5 RERE-0.1 RERE-0.3 RERE-0.5 0.3 0.4 0.5 0.6 Recall 0.5 0.6 0.7 0.8 Precision NYT11-HRL CasRel-0.1 CasRel-0.3 CasRel-0.5 RERE-0.1 RERE-0.3 RERE-0.5 Figure 3: Precision-Recall Curve of RERE and CASREL under different false negative rates. Lines are better in the upper-right corner than the opposite. Note that the coordinates do not start from 0. LSTM encoder (Graves et al., 2013) with randomly initialized weights. From the results we discover that even without BERT, our framework achieves competitive results against the previous approaches such as CoType and CopyR. This further prove the effectiveness of our RERE framework. 5.4 How Robust is RERE against False Negatives? To further study how our model behaves when training data includes different quantity of false negatives, we conduct experiments on synthetic datasets. We construct five new training data by randomly removing triples with probability of 0.1, 0.3 and 0.5, simulating the situation of different FN rates. We show the precision-recall curves of our method in comparison with CASREL (Wei et al., 2020), the best performing competitor, in Figure 3. 1) The overall performance of RERE is superior to competitor models even when trained on a dataset with a 0.5 FN rate. 2) We show that the intervals of RERE between lines are smaller than CASREL, indicating that the performance decline under different FN rates of RERE is smaller. 3) The straight line before curves of our model means that there is no data point at the places where recall is very low. This means that our model is insensitive with the decision boundary and thus more robust. 6 Conclusion In this paper, we revisit the negative data in relation extraction task. We first show that the false negative rate is largely underestimated by previous researches. We then systematically compare three 3580 commonly adopted paradigms and prove that our paradigm suffers less from the overwhelming negative labels. Based on this advantage, we propose RERE, a pipelined framework that first detect relations at sentence level and then extract entities for each specific relation and provide a multi-label PU learning loss to recover false negatives. Empirical results show that RERE consistently outperforms the existing state-of-the-arts by a considerable gap, even when learned with large false negative rates. Acknowledgments This work is supported by National Key Research and Development Project (No. 2020AAA0109302), Shanghai Science and Technology Innovation Action Plan (No.19511120400) and Shanghai Municipal Science and Technology Major Project (No.2021SHZDZX0103). The authors would like to thank the anonymous reviewers for their constructive comments. References Jessa Bekker and Jesse Davis. 2018. Estimating the class prior in positive and unlabeled data through decision tree induction. In Proceedings of AAAI. Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of SIGMOD. Yee Seng Chan and Dan Roth. 2011. Exploiting syntactico-semantic structures for relation extraction. In Proceedings of ACL, pages 551–560. Danqi Chen. 2018. Neural Reading Comprehension and Beyond. Ph.D. thesis, Stanford University. Md. Faisal Mahbub Chowdhury and A. Lavelli. 2012. Impact of less skewed distributions on efficiency and effectiveness of biomedical relation extraction. In Proceedings of COLING. J. Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT. Jun Feng, Minlie Huang, Li Zhao, Yang Yang, and Xiaoyan Zhu. 2018. Reinforcement learning for relation classification from noisy data. In Proceedings of AAAI, volume 32. Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-Beltrachini. 2017. Creating training corpora for NLG micro-planners. In Proceedings of ACL. Matthew R Gormley, Mo Yu, and Mark Dredze. 2015. Improved relation extraction with feature-rich compositional embedding models. In Proceedings of ACL, pages 1774–1784. Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. 2013. Speech recognition with deep recurrent neural networks. In 2013 IEEE international conference on acoustics, speech and signal processing, pages 6645–6649. Pankaj Gupta, Hinrich Sch¨utze, and Bernt Andrassy. 2016. Table filling multi-task recurrent neural network for joint entity and relation extraction. In Proceedings of COLING, pages 2537–2547. R. Hoffmann, Congle Zhang, Xiao Ling, Luke Zettlemoyer, and Daniel S. Weld. 2011. Knowledge-based weak supervision for information extraction of overlapping relations. In Proceedings of ACL. Wei Jia, Dai Dai, Xinyan Xiao, and Hua Wu. 2019. ARNOR: Attention regularization based noise reduction for distant supervision relation classification. In Proceedings of ACL, pages 1399–1408. Arzoo Katiyar and Claire Cardie. 2017. Going out on a limb: Joint extraction of entity mentions and relations without dependency trees. In Proceedings of ACL, pages 917–928. Q. Li and Heng Ji. 2014. Incremental joint extraction of entity mentions and relations. In Proceedings of ACL. Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, F. Wu, and J. Li. 2020. Dice loss for data-imbalanced nlp tasks. In Proceedings of ACL. Xiaoya Li, Fan Yin, Zijun Sun, Xiayu Li, Arianna Yuan, Duo Chai, Mingxin Zhou, and J. Li. 2019. Entityrelation extraction as multi-turn question answering. In Proceedings of ACL. Hongyu Lin, Yaojie Lu, Xianpei Han, and Le Sun. 2018. Adaptive scaling for sparse detection in information extraction. In Proceedings of ACL, pages 1033–1043. Yankai Lin, Shiqi Shen, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2016. Neural relation extraction with selective attention over instances. In Proceedings of ACL, pages 2124–2133. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Bonan Min, Ralph Grishman, Li Wan, Chang Wang, and David Gondek. 2013. Distant supervision for relation extraction with an incomplete knowledge base. In Proceedings of HLT-NAACL. Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Proceedings of ACL. 3581 Makoto Miwa and Mohit Bansal. 2016. End-to-end relation extraction using lstms on sequences and tree structures. In Proceedings of ACL, pages 1105–1116. Makoto Miwa and Yutaka Sasaki. 2014. Modeling joint entity and relation extraction with table representation. In Proceedings of EMNLP, pages 1858–1869. Marthinus Christoffel du Plessis, Gang Niu, and Masashi Sugiyama. 2015. Class-prior estimation for learning from positive and unlabeled data. Machine Learning, 106:463–492. Xiang Ren, Zeqiu Wu, Wenqi He, Meng Qu, Clare R Voss, Heng Ji, Tarek F Abdelzaher, and Jiawei Han. 2017. Cotype: Joint extraction of typed entities and relations with knowledge bases. In Proceedings of WWW, pages 1015–1024. S. Riedel, Limin Yao, and A. McCallum. 2010. Modeling relations and their mentions without labeled text. In Proceedings of ECML/PKDD. Roland Roller, Eneko Agirre, Aitor Soroa, and Mark Stevenson. 2015. Improving distant supervision using inference learning. In Proceedings of ACL, pages 273–278. Cicero dos Santos, Bing Xiang, and Bowen Zhou. 2015. Classifying relations by ranking with convolutional neural networks. In Proceedings of ACL, pages 626– 634. Ryuichi Takanobu, Tianyang Zhang, Jiexi Liu, and Minlie Huang. 2019. A hierarchical framework for relation extraction with reinforcement learning. In Proceedings of AAAI, volume 33, pages 7072–7079. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of NeuroIPS, pages 6000– 6010. Denny Vrandecic and M. Kr¨otzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57:78–85. Shuohang Wang and Jing Jiang. 2017. Machine comprehension using match-lstm and answer pointer. In Proceedings of ICLR. Yucheng Wang, Bowen Yu, Yueyang Zhang, Tingwen Liu, Hongsong Zhu, and Limin Sun. 2020. TPLinker: Single-stage joint extraction of entities and relations through token pair linking. In Proceedings of COLING, pages 1572–1582. Zhepei Wei, Jianlin Su, Yue Wang, Y. Tian, and Yi Chang. 2020. A novel cascade binary tagging framework for relational triple extraction. In Proceedings of ACL. Chenhao Xie, Qiao Cheng, Jiaqing Liang, Lihan Chen, and Y. Xiao. 2020. Collective loss function for positive and unlabeled learning. ArXiv, abs/2005.03228. Bo Xu, Yong Xu, Jiaqing Liang, Chenhao Xie, Bin Liang, Wanyun Cui, and Y. Xiao. 2017. CNDBpedia: A never-ending chinese knowledge extraction system. In Proceedings of IEA/AIE. Wei Xu, Raphael Hoffmann, Le Zhao, and Ralph Grishman. 2013. Filling knowledge base gaps for distant supervision of relation extraction. In Proceedings of ACL, pages 665–670. Bowen Yu, Zhenyu Zhang, Xiaobo Shu, Tingwen Liu, Yubin Wang, Bin Wang, and Sujian Li. 2020. Joint extraction of entities and relations based on a novel decomposition strategy. In Proceedings of ECAI. Xiaofeng Yu and Wai Lam. 2010. Jointly identifying entities and extracting relations in encyclopedia text via a graphical model approach. In Proceedings of COLING, pages 1399–1407. Daojian Zeng, Kang Liu, Yubo Chen, and Jun Zhao. 2015. Distant supervision for relation extraction via piecewise convolutional neural networks. In Proceedings of EMNLP. Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, Jun Zhao, et al. 2014. Relation classification via convolutional deep neural network. In Proceedings of COLING, pages 2335–2344. Xiangrong Zeng, Shizhu He, Kang Liu, and Jun Zhao. 2018a. Large scaled relation extraction with reinforcement learning. In Proceedings of AAAI, volume 32. Xiangrong Zeng, Daojian Zeng, Shizhu He, Kang Liu, and Jun Zhao. 2018b. Extracting relational facts by an end-to-end neural model with copy mechanism. In Proceedings of ACL. Tianyang Zhao, Zhao Yan, Y. Cao, and Zhoujun Li. 2020. Asking effective and diverse questions: A machine reading comprehension based framework for joint entity-relation extraction. In Proceedings of IJCAI. Suncong Zheng, Feng Wang, Hongyun Bao, Yuexing Hao, Peng Zhou, and Bo Xu. 2017. Joint extraction of entities and relations based on a novel tagging scheme. In Proceedings of ACL, pages 1227–1236.
|
2021
|
277
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3582–3593 August 1–6, 2021. ©2021 Association for Computational Linguistics 3582 Knowing the No-match: Entity Alignment with Dangling Cases Zequn Sun1 Muhao Chen2,3 Wei Hu1 1State Key Laboratory for Novel Software Technology, Nanjing University, China 2Department of Computer Science, University of Southern California, USA 3Information Sciences Institute, University of Southern California, USA [email protected], [email protected], [email protected] Abstract This paper studies a new problem setting of entity alignment for knowledge graphs (KGs). Since KGs possess different sets of entities, there could be entities that cannot find alignment across them, leading to the problem of dangling entities. As the first attempt to this problem, we construct a new dataset and design a multi-task learning framework for both entity alignment and dangling entity detection. The framework can opt to abstain from predicting alignment for the detected dangling entities. We propose three techniques for dangling entity detection that are based on the distribution of nearest-neighbor distances, i.e., nearest neighbor classification, marginal ranking and background ranking. After detecting and removing dangling entities, an incorporated entity alignment model in our framework can provide more robust alignment for remaining entities. Comprehensive experiments and analyses demonstrate the effectiveness of our framework. We further discover that the dangling entity detection module can, in turn, improve alignment learning and the final performance. The contributed resource is publicly available to foster further research. 1 Introduction Knowledge graphs (KGs) have evolved to be the building blocks of many intelligent systems (Ji et al., 2020). Despite the importance, KGs are usually costly to construct (Paulheim, 2018) and naturally suffer from incompleteness (Gal´arraga et al., 2017). Hence, merging multiple KGs through entity alignment can lead to mutual enrichment of their knowledge (Chen et al., 2020), and provide downstream applications with more comprehensive knowledge representations (Trivedi et al., 2018; Chen et al., 2020). Entity alignment seeks to discover identical entities in different KGs, such as English entity Thailand and its French counterpart source KG entities target KG entities dangling entities in the target KG dangling entities in the source KG Figure 1: Illustration of entity alignment between two KGs with dangling cases. Paired red and black squares in the overlap region denote entity alignment while others are dangling entities without counterparts. Tha¨ılande. To tackle this important problem, literature has attempted with the embedding-based entity alignment methods (Chen et al., 2017; Wang et al., 2018; Cao et al., 2019; Fey et al., 2020; Wu et al., 2020a; Liu et al., 2020; Sun et al., 2020a). These methods jointly embed different KGs and put similar entities at close positions in a vector space, where the nearest neighbor search can retrieve entity alignment. Due to its effectiveness, embedding-based entity alignment has drawn extensive attention in recent years (Sun et al., 2020c). Nonetheless, to practically support the alignment of KGs as a real-world task, existing studies suffer one common problem of identifying entities without alignment across KGs (called dangling entities). Specifically, current methods are all built upon the assumption that any source entity has a counterpart in the target KG (Sun et al., 2020c), and are accordingly developed with learning resources that enforce the same assumption. Hence, given every entity in a source KG, a model always tends to predict a counterpart via the nearest neighbor search in the embedding space. However, since each KG may be independently created based on separate corpora (Lehmann et al., 2015) or contributed by different crowds (Speer et al., 2017; Carlson et al., 2010), it is natural for KGs to possess different sets 3583 of entities (Collarana et al., 2017), as illustrated in Fig. 1. Essentially, this problem overlooked in prior studies causes existing methods to fall short of distinguishing between matchable and dangling entities, hence hinders any of such methods to align KGs in a real-world scenario. Towards more practical solutions of entity alignment for KGs, we provide a redefinition of the task with the incorporation of dangling cases (§2.1), as the first contribution of this work. Given a source entity, our setting does not assume that it must have a counterpart in the target KG as what previous studies do. Instead, conducting entity alignment also involves identifying whether the counterpart of an entity actually exists in another KG. Hence, a system to tackle this realistic problem setting of entity alignment is also challenged by the requirement for justifying the validity of its prediction. To facilitate the research towards the new problem, the second contribution of this work is to construct a new dataset DBP2.0 for entity alignment with dangling cases (§2.2). As being discussed, existing benchmarks for entity alignment, including DBP15K (Sun et al., 2017), WK3L (Chen et al., 2017) and the more recent OpenEA (Sun et al., 2020c), are set with the constraint that any entity to be aligned should have a valid counterpart. We use the full DBpedia (Lehmann et al., 2015) to build a new dataset and the key challenge lies in that we need to guarantee the selected dangling entities actually do not have counterparts. We first extract two subgraphs with one-to-one entity alignment (i.e., all entities have counterparts). Then, we randomly remove some entities to make their left counterparts in the peer KG dangling. Although embedding-based entity alignment has been investigated for several years, handling with dangling entities has not been studied yet. As the third contribution, we present a multi-task learning framework for the proposed task (§3). It consists of two jointly optimized modules for entity alignment and dangling entity detection, respectively. While the entity alignment module can basically incorporate any existing techniques from prior studies (Sun et al., 2020c), in this paper, we experiment with two representative techniques, i.e., relational embedding based (Chen et al., 2017) and neighborhood aggregation based (Sun et al., 2020b) methods. For dangling entity detection, our framework incorporates an auxiliary learning objective, which seeks to learn a confidence metric for the inferred entity alignment. The principle to realize such metric learning is that the embeddings of dangling entities should be isolated and are distant from others. According to this principle, we exploit several techniques to distinguish between matchable and dangling entities based on their distance distribution with their neighbors (§3), including nearest neighbor classification, marginal ranking and background ranking (Dhamija et al., 2018). We conduct comprehensive experiments on the new DBP2.0 dataset, which demonstrate the proposed techniques to solve the dangling entity detection problem to different extents. Moreover, we observe that training the dangling detection model (marginal ranking) provides an effective indirect supervision that improves the detection of alignment for matchable entities. We hope our task, dataset and framework can foster further investigation of entity alignment techniques in the suggested real scenario, leading to more effective and practical solutions to this challenging but important problem. 2 Task and Dataset We hereby describe the problem setting of our task and introduce the new dataset. 2.1 Task Definition A KG is a set of relational triples T ⊆E × R × E, where E and R denote vocabularies of entities and relations, respectively. Without loss of generality, we consider entity alignment between two KGs, i.e., a source KG K1 =(T1, E1, R1) and a target KG K2 =(T2, E2, R2). Given a small set of seed entity alignment A12 = {(e1, e2) ∈E1 × E2∥e1 ≡e2} along with a small set of source entities D ⊂E1 known to have no counterparts as training data, the task seeks to find the remaining entity alignment. Different from the conventional entity alignment setting (Sun et al., 2017), a portion (with an anticipated quantity) of entities in E1 and E2 may have no counterparts. Our training and inference stages take such dangling entities into consideration. 2.2 Dataset Construction As discussed, previous testbeds for entity alignment do not contain dangling entities (Sun et al., 2017; Chen et al., 2018; Sun et al., 2020c). Therefore, we first create a new dataset to support the study of the proposed problem setting. Same as the widely used existing benchmark DBP15K (Sun 3584 Datasets # Entities # Rel. # Triples # Align. ZH-EN ZH 84,996 3,706 286,067 33,183 EN 118,996 3,402 586,868 JA-EN JA 100,860 3,243 347,204 39,770 EN 139,304 3,396 668,341 FR-EN FR 221,327 2,841 802,678 123,952 EN 278,411 4,598 1,287,231 Table 1: Statistics of the DBP2.0 dataset. et al., 2017), we choose DBpedia 2016-101 as the raw data source. Following DBP15K, we also use English (EN), French (FR), Japanese (JA) and Chinese (ZH) versions of DBpedia to build three entity alignment settings of ZH-EN, JA-EN and FR-EN. For each monolingual KG, the triples are extracted from the Infobox Data of DBpedia, where relations are not mapped to a unified ontology. The reference entity alignment data is from the inter-language links (ILLs) of DBpedia across these three bridges of languages. Such reference data is later used as alignment labels for training and testing, and also serves as references to recognize dangling entities. Construction. The key challenge of building our dataset lies in that we need to ensure the selected dangling entities are indeed without counterparts. Specifcally, we cannot simply regard entities without ILLs as dangling ones, since the ILLs are also incomplete (Chen et al., 2017). Under this circumstance, we use a two-step dataset extraction process, which first samples two subgraphs whose entities all have counterparts based on ILLs, and randomly removes a disjoint set of entities in the source and target graphs to make their counterparts dangling. For the first step, we iteratively delete unlinked entities and their triples from the source and target KGs until the left two subgraphs are one-to-one aligned. In the second step for entity removal, while the removed entities are disjoint in two KGs, the proportion of the removed entities also complies with the proportion of unaligned entities in each KG. Statistics and evaluation. Tab. 1 lists the statistics our dataset. The three entity alignment settings have different data scales and each is much larger than the same setting in DBP15K, thus can benefit better scalability analysis of models. For dangling entity detection, we split 30% of dangling entities for training, 20% for validation and others for test1Downloaded from https://wiki.dbpedia.org/ downloads-2016-10. The latest 2020 version has not provided updated data for some languages other than English when this study is conducted. ing. The splits of reference alignment follow the same partition ratio, which is also consistent with that of DBP15K to simulate the weak alignment nature of KGs (Chen et al., 2017; Sun et al., 2017). We also compare the degree distribution of matchable and dangling entities in our dataset against DBP15K in Fig. 7 of Appx. §A. We find the matchable and unlabeled entities in DBP15K have biased degree distribution, which has an adverse effect on dangling entity detection and leads to unreal evaluation. By contrast, in DBP2.0, matchable and dangling entities have similar degree distribution. 3 Entity Alignment with Dangling Cases We propose a multi-task learning framework for entity alignment with dangling cases, as illustrated in Fig. 2. It has two jointly optimized modules, i.e., entity alignment and dangling entity detection. The entity alignment module takes as input relational triples of two KGs (for KG embedding) and seed entity alignment (for alignment learning). As for the detection of dangling entities, the module uses a small number of labeled dangling entities to jump-start the learning of a confidence metric for distinguishing between matchable and dangling entities. In the inference stage for entity alignment, our framework is able to first identify and remove dangling entities, then predict alignment for those that are decided to be matchable. 3.1 Entity Alignment Our framework can incorporate any entity alignment technique. For the sake of generality, we consider two representative techniques in our framework. One technique is based on MTransE (Chen et al., 2017), which is among the earliest studies for embedding-based entity alignment. It employs the translational model TransE (Bordes et al., 2013) to embed KGs in separate spaces, meanwhile jointly learns a linear transformation between the embedding spaces to match entity counterparts. Specifically, given an entity pair (x1, x2) ∈A12, let x1 and x2 be their embeddings learned by the translational model. MTransE learns the linear transformation induced by a matrix M by minimizing ∥Mx1−x2∥, where ∥·∥denotes the L1 or L2 norm. The other technique is from AliNet (Sun et al., 2020b), which is one of the SOTA methods based on graph neural networks. AliNet encodes entities by performing a multi-hop neighborhood aggregation, seeking to cope with heteromorphism of 3585 alignment search source KG target KG seed entity alignment dangling source entities entity alignment dangling entity detection Input Learning Inference source entity verification remove dangling entities Output training data Figure 2: Framework of entity alignment w/ abstention. their neighborhood structures. For alignment learning, different from MTransE that only minimizes the transformed embedding distance, AliNet additionally optimizes a margin-based ranking loss for entity counterparts with negative samples. Specifically, let x be a matchable source entity in the seed entity alignment, and x′ is a randomly-sampled entity in the target KG, AliNet attempts to ensure ∥x −x′∥> λ1 > 0, where λ1 is a distance margin. 3.2 Dangling Entity Detection We propose three techniques to implement the dangling detection module based on the distribution of the nearest neighbor distance in embedding space. 3.2.1 NN Classification This technique is to train a binary classifier to distinguish between dangling entities (labeled 1, i.e., y = 1) and matchable ones (y = 0). Specifically, we experiment with a feed-forward network (FFN) classifier. Given a source entity x, its input feature representation is the difference vector between its embedding x and its transformed NN embedding xnn in the target KG embedding space2. The confidence of x being a dangling entity is given by p(y = 1|x) = sigmoid(FFN(Mx −xnn)). Let D be the training set of dangling source entities and A denotes the set of matchable entities in the training alignment data. For every x ∈D∪A, we minimize the cross-entropy loss: Lx = − yx log(p(y = 1|x)) + (1 −yx) log(1 −p(y = 1|x)) , (1) where yx denotes the truth label for entity x. In a real-world entity alignment scenario, the dangling entities and matchable ones usually differ greatly in quantity, leading to unbalanced label distribution. In that case, we apply label weights (Huang et al., 2016) to balance between the losses for both labels. 2We use transformed nearest neighbor (NN) to denote the the NN of a source KG entity after it is transformed to the target embedding space. 3.2.2 Marginal Ranking Considering that dangling entities are the noises for finding entity alignment based on embedding distance, we are motivated to let dangling entities have solitary representations in the embedding space, i.e., they should keep a distance away from their surrounding embeddings. Hence, we seek to put a distance margin between dangling entities and their sampled NNs. For every input dangling entity x ∈D, we minimize the following loss: Lx = max(0, λ −∥Mx −xnn∥), (2) where λ is a distance margin. This loss and the entity alignment loss (e.g., that of MTransE) conduct joint learning-to-rank, i.e., the distance between unaligned entities should be larger than that of aligned entities while dangling entities should have a lower ranking in the candidate list of any source entity. 3.2.3 Background Ranking In the two aforementioned techniques, searching for the NN of an entity is time-consuming. Furthermore, selecting an appropriate value for the distance margin of the second technique is not trivial. Based on empirical studies, we find that the margin has a significant influence on the final performance. Hence, we would like to find a more efficient and self-driven technique. Inspired by the open-set classification approach (Dhamija et al., 2018) that lets a classifier equally penalize the output logits for samples of classes that are unknown to training (i.e. background classes), we follow a similar principle and let the model equally enlarge the distance of a dangling entity from any sampled target-space entities. This method is to treat all dangling entities as the “background” of the embedding space, since they should be distant from matchable ones. We also decrease the scale of the dangling entity embeddings to further provide a separation between the embeddings of matchable and dangling entities. For the dangling entity x ∈D, let Xv x be the set of randomly-sampled target entities with size of v. The loss is defined as Lx = x′∈Xvx λx −∥Mx −x′∥ + α∥x∥, (3) where | · | denotes the absolute value and α is a weight hyper-parameter for balance. λx is the average distance, i.e., λx = 1 v x′∈Xvx ∥Mx −x′∥. This objective can push the relatively close entities away from the source entity without requiring a pre-defined distance margin. 3586 3.3 Learning and Inference The overall learning objective of the proposed framework is a combination of the entity alignment loss (e.g., MTransE’s loss) and one of the dangling entity detection loss as mentioned above. The two losses are optimized in alternate batches. More training details are presented in §4.1. Like the training phase, the inference phase is also separated into dangling entity detection and entity alignment. The way of inference for dangling entities differs with the employed technique. The NN classification uses the jointly trained FFN classifier to estimate whether the input entity is a dangling one. The marginal ranking takes the preset margin value in training as a confidence threshold, and decides whether an entity is a dangling one based on if its transformed NN distance is higher than the threshold. The inference of background ranking is similar to that of marginal ranking, with only the difference, by its design, to be that the confidence threshold is set as the average NN distance of entities in the target embedding space. After detecting dangling entities, the framework finds alignment in the remaining entities based on the transformed NN search among the matchable entities in the embedding space of the target KG. Accelerated NN search. The first and second techniques need to search NNs. We can use an efficient similarity search library Faiss (Johnson et al., 2017) for fast NN retrieval in large embedding space. We also maintain a cache to store the NNs of entities backstage and update it every ten training epochs. 4 Experiments In this section, we report our experimental results. We start with describing the experimental setups (§4.1). Next, we separately present the experimentation under two different evaluation settings (§4.2§4.3), followed by an analysis on the similarity score distribution of the obtained representations for matchable and dangling entities (§4.4). To faciliate the use of the contributed dataset and software, we have incorporated these resources into the OpenEA benchmark3 (Sun et al., 2020c). 4.1 Experimental Settings We consider two evaluation settings. One setting is for the proposed problem setting with dangling entities, for which we refer as the consolidated 3https://github.com/nju-websoft/OpenEA 41.8% 41.9% 41.3% 31.4% 31.6% 28.0% 20% 40% 60% ZH-EN JA-EN FR-EN DBP15K DBP2.0 Figure 3: Average neighbor overlap ratio of aligned entities in DBP15K and our DBP2.0. evaluation setting. We first detect and remove the dangling source entities and then search alignment for the left entities. For this evaluation setting, we also separately assess the performance of the dangling detection module. The other simplified setting follows that in previous studies (Sun et al., 2017, 2020c) where the source entities in test set all have counterparts in the target KG, so no dangling source entities are considered. In this relaxed evaluation setting, we seek to evaluate the effect of dangling entity detection on entity alignment and make our results comparable to previous work. Evaluation Protocol. For the relaxed evaluation setting, given each source entity, the candidate counterpart list is selected via NN search in the embedding space. The widely-used metrics on the ranking lists are Hits@k (k = 1, 10, H@k for short) and mean reciprocal rank (MRR). Higher H@k and MRR indicate better performance. For the consolidated setting, we report precision, recall and F1 for dangling entity detection. As for assessing the eventual performance of realistic entity alignment, since the dangling entity detection may not be perfect. it is inevitable for some dangling entities to be incorrectly sent to the entity alignment module for aligning, while some matchable ones may be wrongly excluded. In this case, H@k and MRR are not applicable for the consolidated entity alignment evaluation. Following a relevant evaluation setting for entity resolution in database (Mudgal et al., 2018; Ebraheem et al., 2018), we also use precision, recall and F1 as metrics. More specifically, if a source entity is dangling and is not identified by the detection module, the prediction is always regarded as incorrect. Similarly, if a matchable entity is falsely excluded by the dangling detection module, this test case is also regarded as incorrect since the alignment model has no chance to search for alignment. Otherwise, the alignment module searches for the NN of a source entity in the target embedding space and assesses if the predicated counterpart is correct. Model Configuration. As described in §3.2, our dangling detection module has three variants, i.e., 3587 Methods ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR H@1 H@10 MRR H@1 H@10 MRR H@1 H@10 MRR H@1 H@10 MRR H@1 H@10 MRR H@1 H@10 MRR MTransE .358 .675 .463 .353 .670 .461 .348 .661 .453 .342 .670 .452 .245 .524 .338 .247 .531 .342 w/ NNC .350 .668 .457 .356 .664 .460 .340 .657 .441 .336 .630 .445 .253 .539 .343 .251 .536 .343 w/ MR .378 .693 .487 .383 .699 .491 .373 .686 .476 .374 .707 .485 .259 .541 .348 .265 .553 .360 w/ BR .360 .678 .468 .357 .675 .465 .344 .660 .451 .346 .675 .456 .251 .525 .342 .249 .531 .343 AliNet .332 .594 .421 .359 .629 .451 .338 .596 .429 .363 .630 .455 .223 .473 .306 .246 .495 .329 w/ NNC .321 .598 .415 .335 .608 .428 .330 .602 .422 .344 .627 .439 .212 .467 .294 .230 .476 .312 w/ MR .343 .606 .433 .364 .637 .459 .349 .608 .438 .377 .646 .469 .230 .477 .312 .252 .502 .335 w/ BR .333 .599 .426 .357 .632 .451 .341 .608 .431 .369 .636 .461 .214 .468 .298 .238 .487 .321 Table 2: Entity alignment results (relaxed setting) of MTransE and AliNet on DBP2.0. Methods ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 MTransE NNC .604 .485 .538 .719 .511 .598 .622 .491 .549 .686 .506 .583 .459 .447 .453 .557 .543 .550 MR .781 .702 .740 .866 .675 .759 .799 .708 .751 .864 .653 .744 .482 .575 .524 .639 .613 .625 BR .811 .728 .767 .892 .700 .785 .816 .733 .772 .888 .731 .801 .539 .686 .604 .692 .735 .713 AliNet NNC .676 .419 .517 .738 .558 .634 .597 .482 .534 .761 .120 .207 .466 .365 .409 .545 .162 .250 MR .752 .538 .627 .828 .505 .627 .779 .580 .665 .854 .543 .664 .552 .570 .561 .686 .549 .609 BR .762 .556 .643 .829 .515 .635 .783 .591 .673 .846 .546 .663 .547 .556 .552 .674 .556 .609 Table 3: Dangling entity detection results on DBP2.0. NN classification (NNC), marginal ranking (MR), and background ranking (BR). We report the implementation details of the entity alignment module (w/ MTransE or AliNet) in Appendices B and C. We initialize KG embeddings and model parameters using the Xavier initializer (Glorot and Bengio, 2010), and use Adam (Kingma and Ba, 2015) to optimize the learning objectives with the learning rate 0.001 for MTransE and 0.0005 for AliNet. Note that we do not follow some methods to initialize with machine translated entity name embeddings (Wu et al., 2020a). As being pointed out by recent studies (Chen et al., 2021; Liu et al., 2021, 2020), this is necessary to prevent test data leakage. Entity similarity is measured by cross-domain similarity local scaling (Lample et al., 2018) for reduced hubness effects, as being consistent to recent studies (Sun et al., 2020b; Chen et al., 2021). We use a twolayer FFN in NNC. For MR, the margin is set as λ = 0.9 for MTransE and 0.2 for AliNet. BR randomly samples 20 target entities for each entity per epoch and α = 0.01. Training is terminated based on F1 results of entity alignment on validation data. 4.2 Relaxed Evaluation We first present the evaluation under the relaxed entity alignment setting based on Tab. 2. This setting only involves matchable source entities to test entity alignment, which is an ideal (but less realistic) scenario similar to prior studies (Sun et al., 2020c). We also examine if jointly learning to detect dangling entities can indirectly improve alignment. As observed, MTransE, even without dangling detection, can achieve promising performance on DBP2.0. The results are even better than those on DBP15K as reported by Sun et al. (2017). We attribute this phenomenon to the robustness of this simple embedding method and our improved implementation (e.g., more effective negative sampling). By contrast, although we have tried our best in tuning, the latest GNN-based AliNet falls behind MTransE. Unlike MTransE that learns entity embeddings from a first-order perspective (i.e., based on triple plausibility scores), AliNet represents an entity from a high-order perspective by aggregating its neighbor embeddings, and entities with similar neighborhood structures would have similar representations. However, the dangling entities in DBP2.0 inevitably become spread noises in entity neighborhoods. To further probe into this issue, we count the average neighbor overlap ratio of aligned entities in DBP15K and our DBP2.0. Given an entity alignment pair (x1, x2), let π(x1) and π(x2) be the sets of their neighboring entities respectively, where we also merge their aligned neighbors as one identity based on reference entity alignment. Then the neighbor overlap ratio of x1 and x2 is calculated as |π(x1)∩π(x2)|/|π(x1)∪π(x2)|. We average such a ratio for both DBP15K and DBP2.0 as given in Fig. 3. We can see that the three settings’ 3588 Methods ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 Prec. Rec. F1 MTransE NNC .164 .215 .186 .118 .207 .150 .180 .238 .205 .101 .167 .125 .185 .189 .187 .135 .140 .138 MR .302 .349 .324 .231 .362 .282 .313 .367 .338 .227 .366 .280 .260 .220 .238 .213 .224 .218 BR .312 .362 .335 .241 .376 .294 .314 .363 .336 .251 .358 .295 .265 .208 .233 .231 .213 .222 AliNet NNC .121 .193 .149 .085 .138 .105 .113 .146 .127 .067 .208 .101 .126 .148 .136 .086 .161 .112 MR .207 .299 .245 .159 .320 .213 .231 .321 .269 .178 .340 .234 .195 .190 .193 .160 .200 .178 BR .203 .286 .238 .155 .308 .207 .223 .306 .258 .170 .321 .222 .183 .181 .182 .164 .200 .180 Table 4: Entity alignment results on DBP2.0. 0.4 0.6 0.8 ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR NNC MR BR Figure 4: Accuracy of dangling entity detection. overlap ratios in DBP2.0 are all much lower than those in DBP15K. Thus, DBP2.0 poses additional challenges, as compared to DBP15K, specifically for those methods relying on neighborhood aggregation. Based on results and analysis, we argue that methods performing well on the previous synthetic entity alignment dataset may not robustly generalize to the more realistic dataset with dangling cases. The performance of both MTransE and AliNet is relatively worse on FR-EN, which has more entities (i.e., larger candidate search space) and a low neighborhood overlap ratio (therefore, more difficult to match entities based on neighborhood similarity). Meanwhile, we find that the dangling detection module can affect the performance of entity alignment. In details, MR consistently leads to improvement to both MTransE and AliNet. BR can also noticeably boost entity alignment on most settings. This shows that learning to isolate dangling entities from matchable ones naturally provides indirect help to discriminate the counterpart of a matchable entity from irrelevant ones. On the other hand, such indirect supervision signals may be consumed by the additional trainable parameters in NNC, causing its effect on entity alignment to be negligible. Overall, the observation here calls for more robust entity alignment methods and dangling detection techniques, and lead to further analysis (§4.3). 4.3 Consolidated Evaluation We now report the experiment on the more realistic consolidated evaluation setting. Tab. 3 gives the precision, recall and F1 results of dangling entity detection, and the final entity alignment performance is presented in Tab. 4. In addition, Fig. 4 0 20 40 60 ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR NNC MR BR Figure 5: Average training time (seconds) of one epoch for dangling entity detection (MTransE variants). shows the accuracy of dangling entity detection. We analyze the results from the following aspects. Dangling entity detection. Regardless of which alignment module is incorporated, NNC performs the worst (e.g., the low recall and accuracy around 0.5) among the dangling detection techniques, whereas BR generally performs the best. NNC determines whether an entity is dangling based on the difference vector of the entity embedding and its NN, instead of directly capturing the embedding distance which is observed to be more important based on the results by the other two techniques. By directly pushing dangling entities away from their NNs in the embedding space, both MR and BR offer much better performance. Besides, BR outperforms MR in most cases. By carefully checking their prediction results and the actual distance of NNs, we find that the induced distance margin in BR better discriminates dangling entities from matchable ones than the pre-defined margin. Efficiency. We compare the average epoch time of training the three dangling detection modules for MTransE in Fig. 5. We conduct the experiment using a workstation with an Intel Xeon E51620 3.50GHz CPU and a NVIDIA GeForce RTX 2080 Ti GPU. Since NNC and MR need to search for NNs of source entities, both techniques spend much more training time that is saved by random sampling in BR. Overall, BR is an effective and efficient technique for dangling entity detection. Entity alignment. Generally, for both MTransE and AliNet variants, MR and BR lead to better entity alignment results than NNC. MR and BR 3589 PDWFKDEOHHQWLWLHV GDQJOLQJHQWLWLHV Figure 6: Kernel density estimate plot of the test matchable and dangling entities’ similarity distribution with their nearest target neighbors in ZH-EN. obtain higher precision and recall performance on detecting dangling entities as listed in Tab. 3, resulting in less noise that enters the entity alignment stage. By contrast, NNC has a low accuracy and thus introduces many noises. As BR outperforms MR in dangling detection, it also achieves higher entity alignment results than MR on most settings. We also notice that MR in a few settings, MR offer comparible or slightly better performance than BR. This is because MR can enhance the learning of alignment modules (see §4.2 for detailed analysis), thus delivering improvement to the final performance. MTransE variants generally excels AliNet variants in both entity alignment (see Tab. 2) and dangling entity detection (see Tab. 3) than AliNet, similar to the observation in §4.2. Alignment direction. We find that the alignment direction makes a difference in both dangling entity detection and entity alignment. Using EN KG as the source is coupled with easier dangling detection than in other languages, as the most populated EN KG contributes more dangling entities and triples to training than other KGs. As for entity alignment, we find the observation to be quite the opposite, as using the EN KG as a source leads to noticeable drops in results. For example, the precision of MTransE-BR is 0.312 on ZH-EN, but only 0.241 on EN-ZH. This is because the EN KG has a larger portion of dangling entities. Although the dangling detection module performs well on the EN KG than on others, there are still much more dangling entities entering the alignment search stage, thus reducing the entity alignment precision. This observation suggests that choosing the alignment direction from a less populated KG to the more populated EN KG can be a more effective solution. 4.4 Similarity Score Distribution To illustrate how well the BR technique distinguishes between matchable and dangling entities, we plot in Fig. 6 the distribution of similarity scores of each test entity and its NN. The plot illustrates BR has the expected effect to isolate dangling entities from their NNs, whereas matchable entities are generally placed closer to their NNs. Yet, we can still see a modest overlap between the two NN similarity distributions of dangling and matchable entities, and a number of dangling entities still have a quite large NN similarity. This also reveals the fact that the proposed problem setting of entity alignment with dangling cases has many remaining challenges that await further investigation. 5 Related Work We discuss two topics of relevant work. 5.1 Entity Alignment Embedding-based entity alignment is first attempted in MTransE (Chen et al., 2017), which jointly learns a translational embedding model and a transform-based alignment model for two KGs. Later studies generally follow three lines of improvement. (i) The first line improves the embedding technique to better suit the alignment task, including contextual translation techniques (Sun et al., 2019), long-term dependency techniques (Guo et al., 2019) and neighborhood aggregation (or GNN-based) ones (Wang et al., 2018; Cao et al., 2019; Li et al., 2019; Sun et al., 2020b,a; Fey et al., 2020). (ii) The second line focuses on effective alignment learning with limited supervision. Some leverage semi-supervised learning techniques to resolve the training data insufficiency issue, including self-learning (Sun et al., 2018; Mao et al., 2020) and co-training (Chen et al., 2018). (iii) Another line of research seeks to retrieve auxiliary or indirect supervision signals from profile information or side features of entities, such as entity attributes (Sun et al., 2017; Trisedya et al., 2019; Zhang et al., 2019; Pei et al., 2019), literals (Wu et al., 2019, 2020b; Liu et al., 2020), free text (Chen et al., 2021), pre-trained language models (Yang et al., 2019; Tang et al., 2020) or visual modalities (Liu et al., 2021). Due to the large body of recent advances, we refer readers to a more comprehensive summarization in the survey (Sun et al., 2020c). 5.2 Learning with Abstention Learning with abstention is a fundamental machine learning, where the learner can opt to abstain from making a prediction if without enough decisive 3590 confidence (Cortes et al., 2016, 2018). Related techniques include thresholding softmax (Stefano et al., 2000), selective classification (Geifman and El-Yaniv, 2017), open-set classification with background classes (Dhamija et al., 2018) and out-ofdistribution detection (Liang et al., 2018; Vyas et al., 2018). The idea of learning with abstention also has applications in NLP, such as unanswerable QA, where correct answers of some questions are not stated in the given reference text (Rajpurkar et al., 2018; Zhu et al., 2019; Hu et al., 2019). To the best of our knowledge, our task, dataset, and the proposed dangling detection techniques are the first contribution to support learning with abstention for entity alignment and structured representation learning. 6 Conclusion and Future Work In this paper, we propose and study a new entity alignment task with dangling cases. We construct a dataset to support the study of the proposed problem setting, and design a multi-learning framework for both entity alignment and dangling entity detection. Three types of dangling detection techniques are studied, which are based on nearest neighbor classification, marginal ranking, and background ranking. Comprehensive experiments demonstrate the effectiveness of the method, and provide insights to foster further investigation on this new problem. We further find that dangling entity detection can, in turn, effectively provide auxiliary supervision signals to improve the performance of entity alignment. For future work, we plan to extend the benchmarking on DBP2.0 with results from more base models of entity alignment as well as more abstention inference techniques. Extending our framework to support more prediction tasks with abstention, such as entity type inference (Hao et al., 2019) and relation extraction (Alt et al., 2020), is another direction with potentially broad impact. Acknowledgments We thank the anonymous reviewers for their insightful comments. This work is supported by the National Natural Science Foundation of China (No. 61872172), and the Collaborative Innovation Center of Novel Software Technology & Industrialization. Muhao Chen’s work is supported by the National Science Foundation of United States Grant IIS-2105329. References Christoph Alt, Aleksandra Gabryszak, and Leonhard Hennig. 2020. Tacred revisited: A thorough evaluation of the tacred relation extraction task. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 1558–1569. Antoine Bordes, Nicolas Usunier, Alberto Garc´ıaDur´an, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In Proceedings of the 27th Annual Conference on Neural Information Processing Systems (NeurIPS), pages 2787–2795. Yixin Cao, Zhiyuan Liu, Chengjiang Li, Zhiyuan Liu, Juanzi Li, and Tat-Seng Chua. 2019. Multi-channel graph neural network for entity alignment. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 1452–1461. Andrew Carlson, Justin Betteridge, Bryan Kisiel, Burr Settles, Estevam Hruschka, and Tom Mitchell. 2010. Toward an architecture for never-ending language learning. In Proceedings of the 24th AAAI Conference on Artificial Intelligence (AAAI). Muhao Chen, Weijia Shi, Ben Zhou, and Dan Roth. 2021. Cross-lingual Entity Alignment with Incidental Supervision. In Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics (EACL). Muhao Chen, Yingtao Tian, Kai-Wei Chang, Steven Skiena, and Carlo Zaniolo. 2018. Co-training embeddings of knowledge graphs and entity descriptions for cross-lingual entity alignment. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pages 3998–4004. Muhao Chen, Yingtao Tian, Mohan Yang, and Carlo Zaniolo. 2017. Multilingual knowledge graph embeddings for cross-lingual knowledge alignment. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (IJCAI), pages 1511– 1517. Xuelu Chen, Muhao Chen, Changjun Fan, Ankith Uppunda, Yizhou Sun, and Carlo Zaniolo. 2020. Multilingual knowledge graph completion via ensemble knowledge transfer. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3227–3238. Diego Collarana, Mikhail Galkin, Ignacio Traverso Rib´on, Christoph Lange, Maria-Esther Vidal, and S¨oren Auer. 2017. Semantic data integration for knowledge graph construction at query time. In Proceedings of the 11th IEEE International Conference on Semantic Computing (ICSC), pages 109–116. Corinna Cortes, Giulia DeSalvo, Claudio Gentile, Mehryar Mohri, and Scott Yang. 2018. Online learning with abstention. In Proceedings of the 35th Inter3591 national Conference on Machine Learning (ICML), pages 1067–1075. Corinna Cortes, Giulia DeSalvo, and Mehryar Mohri. 2016. Boosting with abstention. In Proceedings of the 30th Annual Conference on Neural Information Processing Systems (NeurIPS), pages 1660–1668. Akshay Raj Dhamija, Manuel G¨unther, and Terrance E. Boult. 2018. Reducing network agnostophobia. In Proceedings of the 32nd Annual Conference on Neural Information Processing Systems (NeurIPS), pages 9175–9186. Muhammad Ebraheem, Saravanan Thirumuruganathan, Shafiq R. Joty, Mourad Ouzzani, and Nan Tang. 2018. Distributed representations of tuples for entity resolution. Proceedings of the VLDB Endowment, 11(11):1454–1467. Matthias Fey, Jan Eric Lenssen, Christopher Morris, Jonathan Masci, and Nils M. Kriege. 2020. Deep graph matching consensus. In Proceedings of the 8th International Conference on Learning Representations (ICLR). Luis Gal´arraga, Simon Razniewski, Antoine Amarilli, and Fabian M. Suchanek. 2017. Predicting completeness in knowledge bases. In Proceedings of the 10th ACM International Conference on Web Search and Data Mining (WSDM), pages 375–383. Yonatan Geifman and Ran El-Yaniv. 2017. Selective classification for deep neural networks. In Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NeurIPS), pages 4878– 4887. Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the 13th International Conference on Artificial Intelligence and Statistics (AISTATS), pages 249–256. Lingbing Guo, Zequn Sun, and Wei Hu. 2019. Learning to exploit long-term relational dependencies in knowledge graphs. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 2505–2514. Junheng Hao, Muhao Chen, Wenchao Yu, Yizhou Sun, and Wei Wang. 2019. Universal representation learning of knowledge bases by jointly embedding instances and ontological concepts. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 1709–1719. Minghao Hu, Furu Wei, Yuxing Peng, Zhen Huang, Nan Yang, and Dongsheng Li. 2019. Read + Verify: Machine reading comprehension with unanswerable questions. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI), pages 6529– 6537. Chen Huang, Yining Li, Chen Change Loy, and Xiaoou Tang. 2016. Learning deep representation for imbalanced classification. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 5375–5384. Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Marttinen, and Philip S. Yu. 2020. A survey on knowledge graphs: Representation, acquisition and applications. CoRR, abs/2002.00388. Jeff Johnson, Matthijs Douze, and Herv´e J´egou. 2017. Billion-scale similarity search with gpus. CoRR, abs/1702.08734. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR). Guillaume Lample, Alexis Conneau, Marc’Aurelio Ranzato, Ludovic Denoyer, and Herv´e J´egou. 2018. Word translation without parallel data. In Proceedings of the 6th International Conference on Learning Representations (ICLR). Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N. Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick van Kleef, S¨oren Auer, and Christian Bizer. 2015. DBpedia - A large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6(2):167–195. Chengjiang Li, Yixin Cao, Lei Hou, Jiaxin Shi, Juanzi Li, and Tat-Seng Chua. 2019. Semi-supervised entity alignment via joint knowledge embedding model and cross-graph model. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2723–2732. Shiyu Liang, Yixuan Li, and R Srikant. 2018. Enhancing the reliability of out-of-distribution image detection in neural networks. In Proceedings of the 6th International Conference on Learning Representations (ICLR). Fangyu Liu, Muhao Chen, Dan Roth, and Nigel Collier. 2021. Visual Pivoting for (Unsupervised) Entity Alignment. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI). Zhiyuan Liu, Yixin Cao, Liangming Pan, Juanzi Li, and Tat-Seng Chua. 2020. Exploring and evaluating attributes, values, and structures for entity alignment. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6355–6364. Xin Mao, Wenting Wang, Huimin Xu, Man Lan, and Yuanbin Wu. 2020. Mraea: an efficient and robust entity alignment approach for cross-lingual knowledge graph. In Proceedings of the 13th International Conference on Web Search and Data Mining (WSDM), pages 420–428. 3592 Sidharth Mudgal, Han Li, Theodoros Rekatsinas, AnHai Doan, Youngchoon Park, Ganesh Krishnan, Rohit Deep, Esteban Arcaute, and Vijay Raghavendra. 2018. Deep learning for entity matching: A design space exploration. In Proceedings of the 2018 International Conference on Management of Data (SIGMOD), pages 19–34. Heiko Paulheim. 2018. How much is a triple? estimating the cost of knowledge graph creation. In Proceedings of the 17th International Semantic Web Conference (ISWC). Shichao Pei, Lu Yu, Robert Hoehndorf, and Xiangliang Zhang. 2019. Semi-supervised entity alignment via knowledge graph embedding with awareness of degree difference. In Proceedings of the World Wide Web Conference (WWW), pages 3130–3136. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL), pages 784–789. Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI), volume 31. Claudio De Stefano, Carlo Sansone, and Mario Vento. 2000. To reject or not to reject: that is the questionan answer in case of neural classifiers. IEEE Transactions on Systems, Man, and Cybernetics - Part C: Applications and Reviews, 30(1):84–94. Zequn Sun, Muhao Chen, Wei Hu, Chengming Wang, Jian Dai, and Wei Zhang. 2020a. Knowledge association with hyperbolic knowledge graph embeddings. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5704–5716. Zequn Sun, Wei Hu, and Chengkai Li. 2017. Cross-lingual entity alignment via joint attributepreserving embedding. In Proceedings of the 16th International Semantic Web Conference (ISWC), pages 628–644. Zequn Sun, Wei Hu, Qingheng Zhang, and Yuzhong Qu. 2018. Bootstrapping entity alignment with knowledge graph embedding. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pages 4396–4402. Zequn Sun, JiaCheng Huang, Wei Hu, Muhao Chen, Lingbing Guo, and Yuzhong Qu. 2019. Transedge: Translating relation-contextualized embeddings for knowledge graphs. In Proceedings of the 18th International Semantic Web Conference (ISWC), pages 612–629. Zequn Sun, Chengming Wang, Wei Hu, Muhao Chen, Jian Dai, Wei Zhang, and Yuzhong Qu. 2020b. Knowledge graph alignment network with gated multi-hop neighborhood aggregation. In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI), pages 222–229. Zequn Sun, Qingheng Zhang, Wei Hu, Chengming Wang, Muhao Chen, Farahnaz Akrami, and Chengkai Li. 2020c. A benchmarking study of embedding-based entity alignment for knowledge graphs. Proceedings of the VLDB Endowment, 13(11):2326–2340. Xiaobin Tang, Jing Zhang, Bo Chen, Yang Yang, Hong Chen, and Cuiping Li. 2020. BERT-INT: A bertbased interaction model for knowledge graph alignment. In Proceedings of the 29th International Joint Conference on Artificial Intelligence (IJCAI), pages 3174–3180. Bayu Distiawan Trisedya, Jianzhong Qi, and Rui Zhang. 2019. Entity alignment between knowledge graphs using attribute embeddings. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI), pages 297–304. Rakshit Trivedi, Bunyamin Sisman, Xin Luna Dong, Christos Faloutsos, Jun Ma, and Hongyuan Zha. 2018. LinkNBed: Multi-graph representation learning with entity linkage. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL), pages 252–262. Apoorv Vyas, Nataraj Jammalamadaka, Xia Zhu, Dipankar Das, Bharat Kaul, and Theodore L Willke. 2018. Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. In Proceedings of the European Conference on Computer Vision (ECCV), pages 550–564. Zhichun Wang, Qingsong Lv, Xiaohan Lan, and Yu Zhang. 2018. Cross-lingual knowledge graph alignment via graph convolutional networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 349–357. Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, and Dongyan Zhao. 2019. Jointly learning entity and relation representations for entity alignment. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 240–249. Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, and Dongyan Zhao. 2020a. Neighborhood matching network for entity alignment. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 6477–6487. Yuting Wu, Xiao Liu, Yansong Feng, Zheng Wang, and Dongyan Zhao. 2020b. Neighborhood matching network for entity alignment. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 6477–6487. 3593 Hsiu-Wei Yang, Yanyan Zou, Peng Shi, Wei Lu, Jimmy Lin, and Xu Sun. 2019. Aligning cross-lingual entities with multi-aspect information. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4431–4441. Qingheng Zhang, Zequn Sun, Wei Hu, Muhao Chen, Lingbing Guo, and Yuzhong Qu. 2019. Multi-view knowledge graph embedding for entity alignment. In Proceedings of the 28th International Joint Conference (IJCAI), pages 5429–5435. Haichao Zhu, Li Dong, Furu Wei, Wenhui Wang, Bing Qin, and Ting Liu. 2019. Learning to ask unanswerable questions for machine reading comprehension. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 4238–4248. Appendices A Degree Distribution Fig. 7 shows the degree distribution of the matchable and dangling entities in our dataset against DBP15K. Although DBP15K contains some entities that are not labeled to have counterparts, by checking the ILLs in the recent update of DBpedia, we find many of these entities to have counterparts in the target KG. Hence, these entities in DBP15k cannot act as dangling entities that are key to the more realistic evaluation protocol being proposed in this work. From the comparison, we can see that these unlabeled entities in DBP15K have much fewer triples than matchable entities. This biased degree distribution will have an adverse effect on dangling entity detection and lead to unreal evaluation. By contrast, in our dataset, matchable and dangling entities have similar degree distribution. Figure 7: Degree distribution of matchable and dangling entities in DBP15K FR-EN and our FR-EN. B Configuration of MTransE and AliNet For entity alignment, we experiment with MTransE (Chen et al., 2017) and the SOTA method AliNet (Sun et al., 2020b). The implementation of our 0.2 0.4 0.6 0.8 ZH-EN EN-ZH JA-EN EN-JA FR-EN EN-FR NNC MR BR Figure 8: Recall@10 results of entity alignment. framework is extended based on OpenEA (Sun et al., 2020c). We adopt the truncated negative sampling method by BootEA (Sun et al., 2018) to generate negative triples for MTransE and negative alignment links for AliNet, which leads to improved performance. The embedding size is 128 for MTransE and 256 for AliNet. The batch size of MTransE is 20, 480 on ZH-EN and JA-EN, and 102, 400 on FR-EN. The batch size of AliNet is 8, 192 on ZH-EN and JA-EN, and 20, 480 on FREN. λ1 = 1.4 in AliNet. C Hyper-parameter Settings We select each hyper-parameter setting within a wide range of values as follows: • Learning rate: {0.0001, 0.0002, 0.0005, 0.001} • Embedding dimension: {64, 128, 256, 512} • Batch size: {4096, 8192, 10240, 20480, 102400} • # FNN layers: {1, 2, 3, 4} • # Random targets: {1, 10, 20, 30, 40, 50} • λ: {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0} D Recall@10 of Entity Alignment Fig. 8 gives the recall@10 results of the MTransE variants with dangling entity detection in the consolidated evaluation setting. We can see that the recall@10 results on FR-EN are lower than that on ZH-EN and JA-EN, which is similar to the observation in entity alignment §4.3. From the results, we think existing embedding-based entity alignment methods are still far from being usable in practice.
|
2021
|
278
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3594–3608 August 1–6, 2021. ©2021 Association for Computational Linguistics 3594 Superbizarre Is Not Superb: Derivational Morphology Improves BERT’s Interpretation of Complex Words Valentin Hofmann*‡, Janet B. Pierrehumbert†*, Hinrich Schütze‡ *Faculty of Linguistics, University of Oxford †Department of Engineering Science, University of Oxford ‡Center for Information and Language Processing, LMU Munich [email protected] Abstract How does the input segmentation of pretrained language models (PLMs) affect their interpretations of complex words? We present the first study investigating this question, taking BERT as the example PLM and focusing on its semantic representations of English derivatives. We show that PLMs can be interpreted as serial dual-route models, i.e., the meanings of complex words are either stored or else need to be computed from the subwords, which implies that maximally meaningful input tokens should allow for the best generalization on new words. This hypothesis is confirmed by a series of semantic probing tasks on which DelBERT (Derivation leveraging BERT), a model with derivational input segmentation, substantially outperforms BERT with WordPiece segmentation. Our results suggest that the generalization capabilities of PLMs could be further improved if a morphologically-informed vocabulary of input tokens were used. 1 Introduction Pretrained language models (PLMs) such as BERT (Devlin et al., 2019), GPT-2 (Radford et al., 2019), XLNet (Yang et al., 2019), ELECTRA (Clark et al., 2020), and T5 (Raffel et al., 2020) have yielded substantial improvements on a range of NLP tasks. What linguistic properties do they have? Various studies have tried to illuminate this question, with a focus on syntax (Hewitt and Manning, 2019; Jawahar et al., 2019) and semantics (Ethayarajh, 2019; Ettinger, 2020; Vuli´c et al., 2020). One common characteristic of PLMs is their input segmentation: PLMs are based on fixed-size vocabularies of words and subwords that are generated by compression algorithms such as bytepair encoding (Gage, 1994; Sennrich et al., 2016) and WordPiece (Schuster and Nakajima, 2012; Wu et al., 2016). The segmentations produced by these sw x y superbizarre neg applausive pos ##iza superb ##rre BERT p(y|sw(x)) = .149 (a) BERT (sw) sd x y superbizarre neg applausive pos super bizarre BERT p(y|sd(x)) = .931 (b) DelBERT (sd) Figure 1: Basic experimental setup. BERT with WordPiece segmentation (sw) mixes part of the stem bizarre with the prefix super, creating an association with superb (left panel). DelBERT with derivational segmentation (sd), on the other hand, separates prefix and stem by a hyphen (right panel). The two likelihoods are averaged across 20 models trained with different random seeds. The average likelihood of the true class is considerably higher with DelBERT than with BERT. While superbizarre has negative sentiment, applausive is an example of a complex word with positive sentiment. algorithms are linguistically questionable at times (Church, 2020), which has been shown to worsen performance on certain downstream tasks (Bostrom and Durrett, 2020; Hofmann et al., 2020a). However, the wider implications of these findings, particularly with regard to the generalization capabilities of PLMs, are still poorly understood. Here, we address a central aspect of this issue, namely how the input segmentation affects the semantic representations of PLMs, taking BERT as the example PLM. We focus on derivationally complex words such as superbizarre since they exhibit systematic patterns on the lexical level, providing an ideal testbed for linguistic generalization. At the same time, the fact that low-frequency and out-of-vocabulary words are often derivationally complex (Baayen and Lieber, 1991) makes our work relevant in practical settings, especially when many one-word expressions are involved, e.g., in query processing (Kacprzak et al., 2017). 3595 The topic of this paper is related to the more fundamental question of how PLMs represent the meaning of complex words in the first place. So far, most studies have focused on methods of representation extraction, using ad-hoc heuristics such as averaging the subword embeddings (Pinter et al., 2020; Sia et al., 2020; Vuli´c et al., 2020) or taking the first subword embedding (Devlin et al., 2019; Heinzerling and Strube, 2019; Martin et al., 2020). While not resolving the issue, we lay the theoretical groundwork for more systematic analyses by showing that PLMs can be regarded as serial dual-route models (Caramazza et al., 1988), i.e., the meanings of complex words are either stored or else need to be computed from the subwords. Contributions. We present the first study examining how the input segmentation of PLMs, specifically BERT, affects their interpretations of derivationally complex English words. We show that PLMs can be interpreted as serial dualroute models, which implies that maximally meaningful input tokens should allow for the best generalization on new words. This hypothesis is confirmed by a series of semantic probing tasks on which derivational segmentation substantially outperforms BERT’s WordPiece segmentation. This suggests that the generalization capabilities of PLMs could be further improved if a morphologically-informed vocabulary of input tokens were used. We also publish three large datasets of derivationally complex words with corresponding semantic properties.1 2 How Are Complex Words Processed? 2.1 Complex Words in Psycholinguistics The question of how complex words are processed has been at the center of psycholinguistic research over the last decades (see Leminen et al. (2019) for a recent review). Two basic processing mechanisms have been proposed: storage, where the meaning of complex words is listed in the mental lexicon (Manelis and Tharp, 1977; Butterworth, 1983; Feldman and Fowler, 1987; Bybee, 1988; Stemberger, 1994; Bybee, 1995; Bertram et al., 2000a), and computation, where the meaning of complex words is inferred based on the meaning of stem and affixes (Taft and Forster, 1975; Taft, 1979, 1981, 1988, 1991, 1994; Rastle et al., 2004; Taft, 2004; Rastle and Davis, 2008). 1We make our code and data available at https:// github.com/valentinhofmann/superbizarre. In contrasting with single-route frameworks, dual-route models allow for a combination of storage and computation. Dual-route models are further classified by whether they regard the processes of retrieving meaning from the mental lexicon and computing meaning based on stem and affixes as parallel, i.e., both mechanisms are always activated (Frauenfelder and Schreuder, 1992; Schreuder and Baayen, 1995; Baayen et al., 1997, 2000; Bertram et al., 2000b; New et al., 2004; Kuperman et al., 2008, 2009), or serial, i.e., the computation-based mechanism is only activated when the storagebased one fails (Laudanna and Burani, 1985; Burani and Caramazza, 1987; Caramazza et al., 1988; Burani and Laudanna, 1992; Laudanna and Burani, 1995; Alegre and Gordon, 1999). Outside the taxonomy presented so far are recent models that assume multiple levels of representation as well as various forms of interaction between them (Rácz et al., 2015; Needle and Pierrehumbert, 2018). In these models, sufficiently frequent complex words are stored together with representations that include their internal structure. Complex-word processing is driven by analogical processes over the mental lexicon (Rácz et al., 2020). 2.2 Complex Words in NLP and PLMs Most models of word meaning proposed in NLP can be roughly assigned to either the single-route or dual-route approach. Word embeddings that represent complex words as whole-word vectors (Deerwester et al., 1990; Mikolov et al., 2013a,b; Pennington et al., 2014) can be seen as single-route storage models. Word embeddings that represent complex words as a function of subword or morpheme vectors (Schütze, 1992; Luong et al., 2013) can be seen as single-route computation models. Finally, word embeddings that represent complex words as a function of subword or morpheme vectors as well as whole-word vectors (Botha and Blunsom, 2014; Qiu et al., 2014; Bhatia et al., 2016; Bojanowski et al., 2017; Athiwaratkun et al., 2018; Salle and Villavicencio, 2018) are most closely related to parallel dual-route approaches. Where are PLMs to be located in this taxonomy? PLMs represent many complex words as wholeword vectors (which are fully stored). Similarly to how character-based models represent word meaning (Kim et al., 2016; Adel et al., 2017), they can also store the meaning of frequent complex words that are segmented into subwords, i.e., frequent sub3596 word collocations, in their model weights. When the complex-word meaning is neither stored as a whole-word vector nor in the model weights, PLMs compute the meaning as a compositional function of the subwords. Conceptually, PLMs can thus be interpreted as serial dual-route models. While the parallelism has not been observed before, it follows logically from the structure of PLMs. The key goal of this paper is to show that the implications of this observation are borne out empirically. As a concrete example, consider the complex words stabilize, realize, finalize, mobilize, tribalize, and templatize, which are all formed by adding the verbal suffix ize to a nominal or adjectival stem. Taking BERT, specifically BERTBASE (uncased) (Devlin et al., 2019), as the example PLM, the words stabilize and realize have individual tokens in the input vocabulary and are hence associated with whole-word vectors storing their meanings, including highly lexicalized meanings as in the case of realize. By contrast, the words finalize and mobilize are segmented into final, ##ize and mob, ##ili, ##ze, which entails that their meanings are not stored as whole-word vectors. However, both words have relatively high absolute frequencies of 2,540 (finalize) and 6,904 (mobilize) in the English Wikipedia, the main dataset used to pretrain BERT (Devlin et al., 2019), which means that BERT can store their meanings in its model weights during pretraining.2 Notice this is even possible in the case of highly lexicalized meanings as for mobilize. Finally, the words tribalize and templatize are segmented into tribal, ##ize and te, ##mp, ##lat, ##ize, but as opposed to finalize and mobilize they do not occur in the English Wikipedia. As a result, BERT cannot store their meanings in its model weights during pretraining and needs to compute them from the meanings of the subwords. Seeing PLMs as serial dual-route models allows for a more nuanced view on the central research question of this paper: in order to investigate semantic generalization we need to investigate the representations of those complex words that activate the computation-based route. The words that do so are the ones whose meaning is neither stored as a whole-word vector nor in the model weights 2Previous research suggests that such lexical knowledge is stored in the lower layers of BERT (Vuli´c et al., 2020). and hence needs to be computed compositionally as a function of the subwords (tribalize and templatize in the discussed examples). We hypothesize that the morphological validity of the segmentation affects the representational quality in these cases, and that the best generalization is achieved by maximally meaningful tokens. It is crucial to note this does not imply that the tokens have to be morphemes, but the segmentation boundaries need to coincide with morphological boundaries, i.e., groups of morphemes (e.g., tribal in the segmentation of tribalize) are also possible.3 For tribalize and templatize, we therefore expect the segmentation tribal, ##ize (morphologically valid since all segmentation boundaries are morpheme boundaries) to result in a representation of higher quality than the segmentation te, ##mp, ##lat, ##ize (morphologically invalid since the boundaries between te, ##mp, and ##lat are not morpheme boundaries). On the other hand, complex words whose meanings are stored in the model weights (finalize and mobilize in the discussed examples) are expected to be affected by the segmentation to a much lesser extent: if the meaning of a complex word is stored in the model weights, it should matter less whether the specific segmentation activating that meaning is morphologically valid (final, ##ize) or not (mob, ##ili, ##ze).4 3 Experiments 3.1 Setup Analyzing the impact of different segmentations on BERT’s semantic generalization capabilities is not straightforward since it is not clear a priori how to measure the quality of representations. Here, we devise a novel lexical-semantic probing task: we use BERT’s representations for complex words to predict semantic dimensions, specifically sentiment and topicality (see Figure 1). For sentiment, given the example complex word superbizarre, the task is to predict that its sentiment is negative. For topicality, given the example complex word isotopize, the task is to predict that it is used in physics. We confine ourselves to binary predic3This is in line with substantial evidence from linguistics showing that frequent groups of morphemes can be treated as semantic wholes (Stump, 2017, 2019). 4We expect the distinction between storage and computation of complex-word meaning for PLMs to be a continuum. While the findings presented here are consistent with this view, we defer a more in-depth analysis to future work. 3597 Class 1 Class 2 Dataset Dimension |D| Class Examples Class Example Amazon Sentiment 239,727 neg overpriced, crappy pos megafavorite, applausive ArXiv Topicality 97,410 phys semithermal, ozoneless cs autoencoded, rankable Reddit Topicality 85,362 ent supervampires, spoilerful dis antirussian, immigrationism Table 1: Dataset characteristics. The table provides information about the datasets such as the relevant semantic dimensions with their classes and example complex words. |D|: number of complex words; neg: negative; pos: positive; phys: physics; cs: computer science; ent: entertainment; dis: discussion. tion, i.e., the probed semantic dimensions always consist of two classes (e.g., positive and negative). The extent to which a segmentation supports a solution of this task is taken as an indicator of its representational quality. More formally, let D be a dataset consisting of complex words x and corresponding classes y that instantiate a certain semantic dimension (e.g., sentiment). We denote with s(x) = (t1, . . . , tk) the segmentation of x into a sequence of k subwords. We ask how s impacts the capability of BERT to predict y, i.e., how p(y|(s(x)), the likelihood of the true semantic class y given a certain segmentation of x, depends on different choices for s. The two segmentation methods we compare in this study are BERT’s standard WordPiece segmentation (Schuster and Nakajima, 2012; Wu et al., 2016), sw, and a derivational segmentation that segments complex words into stems and affixes, sd. 3.2 Data Since existing datasets do not allow us to conduct experiments following the described setup, we create new datasets in a weakly-supervised fashion that is conceptually similar to the method proposed by Mintz et al. (2009): we employ large datasets annotated for sentiment or topicality, extract derivationally complex words, and use the dataset labels to establish their semantic classes. For determining and segmenting derivationally complex words, we use the algorithm introduced by Hofmann et al. (2020b), which takes as input a set of prefixes, suffixes, and stems and checks for each word in the data whether it can be derived from a stem using a combination of prefixes and suffixes.5 The algorithm is sensitive to morpho-orthographic rules of English (Plag, 2003), e.g., when the suf5The distinction between inflectionally and derivationally complex words is notoriously fuzzy (Haspelmath and Sims, 2010; ten Hacken, 2014). We try to exclude inflection as far as possible (e.g., by removing problematic affixes such as ing) but are aware that a clear separation does not exist. fix ize is removed from isotopize, the result is isotope, not isotop. We follow Hofmann et al. (2020a) in using the prefixes, suffixes, and stems in BERT’s WordPiece vocabulary as input to the algorithm. This means that all tokens used by the derivational segmentation are in principle also available to the WordPiece segmentation, i.e., the difference between sw and sd does not lie in the vocabulary per se but rather in the way the vocabulary is used. See Appendix A.1 for details about the derivational segmentation. To get the semantic classes, we compute for each complex word which fraction of texts containing the word belongs to one of two predefined sets of dataset labels (e.g., reviews with four and five stars for positive sentiment) and rank all words accordingly. We then take the first and third tertiles of complex words as representing the two classes. We randomly split the words into 60% training, 20% development, and 20% test. In the following, we describe the characteristics of the three datasets in greater depth. Table 1 provides summary statistics. See Appendix A.2 for details about data preprocessing. Amazon. Amazon is an online e-commerce platform. A large dataset of Amazon reviews has been made publicly available (Ni et al., 2019).6 We extract derivationally complex words from reviews with one or two (neg) as well as four or five stars (pos), discarding three-star reviews for a clearer separation (Yang and Eisenstein, 2017). ArXiv. ArXiv is an open-access distribution service for scientific articles. Recently, a dataset of all papers published on ArXiv with associated metadata has been released.7 For this study, we extract all articles from physics (phys) and computer science (cs), which we identify using ArXiv’s subject classification. We choose physics and computer 6https://nijianmo.github.io/amazon/ index.html 7https://www.kaggle.com/ Cornell-University/arxiv 3598 Amazon ArXiv Reddit Model Dev Test Dev Test Dev Test DelBERT .635 ± .001 .639 ± .002 .731 ± .001 .723 ± .001 .696 ± .001 .701 ± .001 BERT .619 ± .001 .624 ± .001 .704 ± .001 .700 ± .002 .664 ± .001 .664 ± .003 Stem .572 ± .003 .573 ± .003 .705 ± .001 .697 ± .001 .679 ± .001 .684 ± .002 Affixes .536 ± .008 .539 ± .008 .605 ± .001 .603 ± .002 .596 ± .001 .596 ± .001 Table 2: Results. The table shows the average performance as well as standard deviation (F1) of 20 models trained with different random seeds. Best result per column highlighted in gray, second-best in light gray. Figure 2: Convergence analysis. The upper panels show the distributions of the number of epochs after which the models reach their maximum validation performance. The lower panels show the trajectories of the average validation performance (F1) across epochs. The plots are based on 20 models trained with different random seeds. The convergence statistics for DelBERT and BERT are directly comparable because the optimal learning rate is the same (see Appendix A.3). DelBERT models reach their performance peak faster than BERT models. science since we expect large topical distances for these classes (compared to alternatives such as mathematics and computer science). Reddit. Reddit is a social media platform hosting discussions about various topics. It is divided into smaller communities, so-called subreddits, which have been shown to be a rich source of derivationally complex words (Hofmann et al., 2020c). Hofmann et al. (2020a) have published a dataset of derivatives found on Reddit annotated with the subreddits in which they occur.8 Inspired by a content-based subreddit categorization scheme,9 we define two groups of subreddits, an entertainment set (ent) consisting of the subreddits anime, DestinyTheGame, funny, Games, gaming, leagueoflegends, movies, Music, pics, and videos, as well as a discussion set (dis) consisting of the subred8https://github.com/valentinhofmann/ dagobert 9https://www.reddit.com/r/ TheoryOfReddit/comments/1f7hqc/the_200_ most_active_subreddits_categorized_by dits askscience, atheism, conspiracy, news, Libertarian, politics, science, technology, TwoXChromosomes, and worldnews, and extract all derivationally complex words occurring in them. We again expect large topical distances for these classes. Given that the automatic creation of the datasets necessarily introduces noise, we measure human performance on 100 randomly sampled words per dataset, which ranges between 71% (Amazon) and 78% (ArXiv). These values can thus be seen as an upper bound on performance. 3.3 Models We train two main models on each binary classification task: BERT with the standard WordPiece segmentation (sw) and BERT using the derivational segmentation (sd), a model that we refer to as DelBERT (Derivation leveraging BERT). BERT and DelBERT are identical except for the way in which they use the vocabulary of input tokens (but the vocabulary itself is also identical for both models). 3599 Figure 3: Frequency analysis. The plots show the average performance (accuracy) of 20 BERT and DelBERT models trained with different random seeds for complex words of low (f ≤5), mid (5 < f ≤500), and high (f > 500) frequency. On all three datasets, BERT performs similarly or better than DelBERT for complex words of high frequency but worse for complex words of low and mid frequency. The specific BERT variant we use is BERTBASE (uncased) (Devlin et al., 2019). For the derivational segmentation, we follow previous work by Hofmann et al. (2020a) in separating stem and prefixes by a hyphen. We further follow Casanueva et al. (2020) and Vuli´c et al. (2020) in mean-pooling the output representations for all subwords, excluding BERT’s special tokens. The mean-pooled representation is then fed into a two-layer feed-forward network for classification. To examine the relative importance of different types of morphological units, we train two additional models in which we ablate information about stems and affixes, i.e., we represent stems and affixes by the same randomly chosen input embedding.10 We finetune BERT, DelBERT, and the two ablated models on the three datasets using 20 different random seeds. We choose F1 as the evaluation measure. See Appendix A.3 for details about implementation and hyperparameters. 3.4 Results DelBERT (sd) outperforms BERT (sw) by a large margin on all three datasets (Table 2). It is interesting to notice that the performance difference is larger for ArXiv and Reddit than for Amazon, indicating that the gains in representational quality are particularly large for topicality. What is it that leads to DelBERT’s increased performance? The ablation study shows that models using only stem information already achieve relatively high performance and are on par or even better than the BERT models on ArXiv and Reddit. However, the DelBERT models still perform substantially better than the stem models on all three datasets. The gap is particularly pronounced 10For affix ablation, we use two different input embeddings for prefixes and suffixes. for Amazon, which indicates that the interaction between the meaning of stem and affixes is more complex for sentiment than for topicality. This makes sense from a linguistic point of view: while stems tend to be good cues for the topical associations of a complex word, sentiment often depends on semantic interactions between stems and affixes. For example, while the prefix un turns the sentiment of amusing negative, it turns the sentiment of biased positive. Such effects involving negation and antonymy are known to be challenging for PLMs (Ettinger, 2020; Kassner and Schütze, 2020) and might be one of the reasons for the generally lower performance on Amazon.11 The performance of models using only affixes is much lower. 3.5 Quantitative Analysis To further examine how BERT (sw) and DelBERT (sd) differ in the way they infer the meaning of complex words, we perform a convergence analysis. We find that the DelBERT models reach their peak in performance faster than the BERT models (Figure 2). This is in line with our interpretation of PLMs as serial dual-route models (see Section 2.2): while DelBERT operates on morphological units and can combine the subword meanings to infer the meanings of complex words, BERT’s subwords do not necessarily carry lexical meanings, and hence the derivational patterns need to be stored by adapting the model weights. This is an additional burden, leading to longer convergence times and substantially worse overall performance. Our hypothesis that PLMs can use two routes 11Another reason for the lower performance on sentiment is that the datasets were created automatically (see Section 3.2), and hence many complex words do not directly carry information about sentiment or topicality. The density of such words is higher for sentiment than topicality since the topic of discussion affects the likelihoods of most content words. 3600 (a) Topicality prediction (b) Sentiment prediction Figure 4: Accuracy increase of DelBERT compared to BERT for prefixes. The plots show the accuracy increase as a function of the proportion of morphologically incorrect WordPiece segmentations (topicality prediction) and as ordered boxplot pairs centered on the median accuracy of BERT (sentiment prediction). Negative values mean that the DelBERT models have a lower accuracy than the BERT models for a certain prefix. to process complex words (storage in weights and compositional computation based on input embeddings), and that the second route is blocked when the input segmentation is not morphological, suggests the existence of frequency effects: BERT might have seen frequent complex words multiple times during pretraining and stored their meaning in the model weights. This is less likely for infrequent complex words, making the capability to compositionally infer the meaning (i.e., the computation route) more important. We therefore expect the difference in performance between DelBERT (which should have an advantage on the computation route) and BERT to be larger for infrequent words. To test this hypothesis, we split the complex words of each dataset into three bins of low (f ≤5), mid (5 < f ≤500), and high (f > 500) absolute frequencies, and analyze how the performance of BERT and DelBERT differs on the three bins. For this and all subsequent analyses, we merge development and test sets and use accuracy instead of F1 since it makes comparisons across small sets of data points more interpretable. The results are in line with our hypothesis (Figure 3): BERT performs worse than DelBERT on complex words of low and mid frequencies but achieves very similar (ArXiv, Reddit) or even better (Amazon) accuracies on high-frequency complex words. These results strongly suggest that two different mechanisms are involved, and that BERT has a disadvantage for complex words that do not have a high frequency. At the same time, the slight advantage of BERT on high-frequency complex words indicates that it has high-quality representations of these words in its weights, which DelBERT cannot exploit since it uses a different segmentation. We are further interested to see whether the affix type has an impact on the relative performance of BERT and DelBERT. To examine this question, we measure the accuracy increase of DelBERT as compared to BERT for individual affixes, averaged across datasets and random seeds. We find that the increase is almost twice as large for prefixes (µ = .023, σ = .017) than for suffixes (µ = .013, σ = .016), a difference that is shown to be significant by a two-tailed Welch’s t-test (d = .642, t(82.97) = 2.94, p < .01).12 Why is having access to the correct morphological segmentation more advantageous for prefixed than suffixed complex words? We argue that there are two key factors at play. First, the WordPiece tokenization sometimes generates the morphologically correct segmenta12We use a Welch’s instead of Student’s t-test since it does not assume that the distributions have equal variance. 3601 Dataset x y sd(x) µp sw(x) µp Amazon applausive pos applause, ##ive .847 app, ##laus, ##ive .029 superannoying neg super, -, annoying .967 super, ##ann, ##oy, ##ing .278 overseasoned neg over, -, seasoned .956 overseas, ##oned .219 ArXiv isotopize phy isotope, ##ize .985 iso, ##top, ##ize .039 antimicrosoft cs anti, -, microsoft .936 anti, ##mic, ##ros, ##oft .013 inkinetic phy in, -, kinetic .983 ink, ##ine, ##tic .035 Reddit prematuration dis premature, ##ation .848 prem, ##at, ##uration .089 nonmultiplayer ent non, -, multiplayer .950 non, ##mu, ##lt, ##ip, ##layer .216 promosque dis pro, -, mosque .961 promo, ##sque .066 Table 3: Error analysis. The table gives example complex words that are consistently classified correctly by DelBERT and incorrectly by BERT. x: complex word; y: semantic class; sd(x): derivational segmentation; µp: average likelihood of true semantic class across 20 models trained with different random seeds; sw(x): WordPiece segmentation. For the complex words shown, µp is considerably higher with DelBERT than with BERT. tion, but it does so with different frequencies for prefixes and suffixes. To detect morphologically incorrect segmentations, we check whether the WordPiece segmentation keeps the stem intact, which is in line with our definition of morphological validity (Section 2.2) and provides a conservative estimate of the error rate. For prefixes, the WordPiece tokenization is seldom correct (average error rate: µ = .903, σ = .042), whereas for suffixes it is correct about half the time (µ = .503, σ = .213). Hence, DelBERT gains a greater advantage for prefixed words. Second, prefixes and suffixes have different linguistic properties that affect the prediction task in unequal ways. Specifically, whereas suffixes have both syntactic and semantic functions, prefixes have an exclusively semantic function and always add lexical-semantic meaning to the stem (Giraudo and Grainger, 2003; Beyersmann et al., 2015). As a result, cases such as unamusing where the affix boundary is a decisive factor for the prediction task are more likely to occur with prefixes than suffixes, thus increasing the importance of a morphologically correct segmentation.13 Given the differences between sentiment and topicality prediction, we expect variations in the relative importance of the two identified factors: (i) in the case of sentiment the advantage of sd should be maximal for affixes directly affecting sentiment; (ii) in the case of topicality its advantage should be the larger the higher the proportion of incorrect segmentations for a particular affix, and hence the more frequent the cases where DelBERT has access to the stem while BERT does not. To test this hypothesis, we focus on pre13Notice that there are suffixes with similar semantic effects (e.g., less), but they are less numerous. dictions for prefixed complex words. For each dataset, we measure for individual prefixes the accuracy increase of the DelBERT models as compared to the BERT models, averaged across random seeds, as well as the proportion of morphologically incorrect segmentations produced by WordPiece. We then calculate linear regressions to predict the accuracy increases based on the proportions of incorrect segmentations. This analysis shows a significant positive correlation for ArXiv (R2 = .304, F(1, 41) = 17.92, p < 0.001) and Reddit (R2 = .270, F(1, 40) = 14.80, p < 0.001) but not for Amazon (R2 = .019, F(1, 41) = .80, p = .375), which is in line with our expectations (Figure 4a). Furthermore, ranking the prefixes by accuracy increase for Amazon confirms that the most pronounced differences are found for prefixes that can change the sentiment such as non, anti, mal, and pseudo (Figure 4b). 3.6 Qualitative Analysis Besides quantitative factors, we are interested in identifying qualitative contexts in which DelBERT has a particular advantage compared to BERT. To do so, we filter the datasets for complex words that are consistently classified correctly by DelBERT and incorrectly by BERT. Specifically, we compute for each word the average likelihood of the true semantic class across DelBERT and BERT models, respectively, and rank words according to the likelihood difference between both model types. Examining the words with the most extreme differences, we observe three classes (Table 3). First, the addition of a suffix is often connected with morpho-orthographic changes (e.g., the deletion of a stem-final e), which leads to a segmentation of the stem into several subwords 3602 since the truncated stem is not in the WordPiece vocabulary (applausive, isotopize, prematuration). The model does not seem to be able to recover the meaning of the stem from the subwords. Second, the addition of a prefix has the effect that the word-internal (as opposed to word-initial) form of the stem would have to be available for proper segmentation. Since this form rarely exists in the WordPiece vocabulary, the stem is segmented into several subwords (superannoying, antimicrosoft, nonmultiplayer). Again, it does not seem to be possible for the model to recover the meaning of the stem. Third, the segmentation of prefixed complex words often fuses the prefix with the first characters of the stem (overseasoned, inkinetic, promosque). This case is particularly detrimental since it not only makes it difficult to recover the meaning of the stem but also creates associations with unrelated meanings, sometimes even opposite meanings as in the case of superbizarre. The three classes thus underscore the difficulty of inferring the meaning of complex words from the subwords when the wholeword meaning is not stored in the model weights and the subwords are not morphological. 4 Related Work Several recent studies have examined how the performance of PLMs is affected by their input segmentation. Tan et al. (2020) show that tokenizing inflected words into stems and inflection symbols allows BERT to generalize better on non-standard inflections. Bostrom and Durrett (2020) pretrain RoBERTa with different tokenization methods and find tokenizations that align more closely with morphology to perform better on a number of tasks. Ma et al. (2020) show that providing BERT with character-level information also leads to enhanced performance. Relatedly, studies from automatic speech recognition have demonstrated that morphological decomposition improves the perplexity of language models (Fang et al., 2015; Jain et al., 2020). Whereas these studies change the vocabulary of input tokens (e.g., by adding special tokens), we show that even when keeping the pretrained vocabulary fixed, employing it in a morphologically correct way leads to better performance.14 14There are also studies that analyze morphological aspects of PLMs without a focus on questions surrounding segmentation (Edmiston, 2020; Klemen et al., 2020). Most NLP studies on derivational morphology have been devoted to the question of how semantic representations of derivationally complex words can be enhanced by including morphological information (Luong et al., 2013; Botha and Blunsom, 2014; Qiu et al., 2014; Bhatia et al., 2016; Cotterell and Schütze, 2018), and how affix embeddings can be computed (Lazaridou et al., 2013; Kisselew et al., 2015; Padó et al., 2016). Cotterell et al. (2017), Vylomova et al. (2017), and Deutsch et al. (2018) propose sequence-to-sequence models for the generation of derivationally complex words. Hofmann et al. (2020a) address the same task using BERT. In contrast, we analyze how different input segmentations affect the semantic representations of derivationally complex words in PLMs, a question that has not been addressed before. 5 Conclusion We have examined how the input segmentation of PLMs, specifically BERT, affects their interpretations of derivationally complex words. Drawing upon insights from psycholinguistics, we have deduced a conceptual interpretation of PLMs as serial dual-route models, which implies that maximally meaningful input tokens should allow for the best generalization on new words. This hypothesis was confirmed by a series of semantic probing tasks on which DelBERT, a model using derivational segmentation, consistently outperformed BERT using WordPiece segmentation. Quantitative and qualitative analyses further showed that BERT’s inferior performance was caused by its inability to infer the complex-word meaning as a function of the subwords when the complex-word meaning was not stored in the weights. Overall, our findings suggest that the generalization capabilities of PLMs could be further improved if a morphologically-informed vocabulary of input tokens were used. Acknowledgements This work was funded by the European Research Council (#740516) and the Engineering and Physical Sciences Research Council (EP/T023333/1). The first author was also supported by the German Academic Scholarship Foundation and the Arts and Humanities Research Council. We thank the reviewers for their helpful comments. 3603 References Heike Adel, Ehsaneddin Asgari, and Hinrich Schütze. 2017. Overview of character-based models for natural language processing. In International Conference on Computational Linguistics and Intelligent Text Processing (CICLing) 18. Maria Alegre and Peter Gordon. 1999. Frequency effects and the representational status of regular inflections. Journal of Memory and Language, 40:41–61. Ben Athiwaratkun, Andrew Wilson, and Anima Anandkumar. 2018. Probabilistic fasttext for multi-sense word embeddings. In Annual Meeting of the Association for Computational Linguistics (ACL) 56. R. Harald Baayen, Ton Dijkstra, and Robert Schreuder. 1997. Singulars and plurals in Dutch: Evidence for a parallel dual-route model. Journal of Memory and Language, 37:94–117. R. Harald Baayen and Rochelle Lieber. 1991. Productivity and English derivation: A corpus-based study. Linguistics, 29(5). R. Harald Baayen, Robert Schreuder, and Richard Sproat. 2000. Morphology in the mental lexicon: A computational model for visual word recognition. In Frank van Eynde and Dafydd Gibbon, editors, Lexicon development for speech and language processing, pages 267–293. Springer, Dordrecht. Raymond Bertram, Matti Laine, R. Harald Baayen, Robert Schreuder, and Jukka Hyönä. 2000a. Affixal homonymy triggers full-form storage, even with inflected words, even in a morphologically rich language. Cognition, 74:B13–B25. Raymond Bertram, Robert Schreuder, and R. Harald Baayen. 2000b. The balance of storage and computation in morphological processing: The role of word formation type, affixal homonymy, and productivity. Journal of Experimental Psychology: Learning, Memory, and Cognition, 26(2):489–511. Elisabeth Beyersmann, Johannes C. Ziegler, and Jonathan Grainger. 2015. Differences in the processing of prefixes and suffixes revealed by a letter-search task. Scientific Studies of Reading, 19(5):360–373. Parminder Bhatia, Robert Guthrie, and Jacob Eisenstein. 2016. Morphological priors for probabilistic neural word embeddings. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2016. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135–146. Kaj Bostrom and Greg Durrett. 2020. Byte pair encoding is suboptimal for language model pretraining. In Findings of Empirical Methods in Natural Language Processing (EMNLP) 2020. Jan A. Botha and Phil Blunsom. 2014. Compositional morphology for word representations and language modelling. In International Conference on Machine Learning (ICML) 31. Cristina Burani and Alfonso Caramazza. 1987. Representation and processing of derived words. Language and Cognitive Processes, 2(3-4):217–227. Cristina Burani and Alessandro Laudanna. 1992. Units of representation for derived words in the lexicon. In Ram Frost and Leonard Katz, editors, Orthography, phonology, morphology, and meaning, pages 361– 376. North-Holland, Amsterdam. Brian Butterworth. 1983. Lexical representation. In Brian Butterworth, editor, Language production: Development, writing and other language processes, pages 257–294. Academic Press, London. Joan Bybee. 1988. Morphology as lexical organization. In Michael Hammond and Michael Noonan, editors, Theoretical approaches to morphology: Approaches in modern linguistics, pages 119–141. Academic Press, San Diego, CA. Joan Bybee. 1995. Regular morphology and the lexicon. Language and Cognitive Processes, 10(425455). Alfonso Caramazza, Alessandro Laudanna, and Cristina Romani. 1988. Lexical access and inflectional morphology. Cognition, 28(297-332). Iñigo Casanueva, Tadas Temˇcinas, Daniela Gerz, Matthew Henderson, and Ivan Vuli´c. 2020. Efficient intent detection with dual sentence encoders. In Workshop on Natural Language Processing for Conversational AI 2. Kenneth Church. 2020. Emerging trends: Subwords, seriously? Natural Language Engineering, 26(3):375–382. Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In International Conference on Learning Representations (ICLR) 8. Ryan Cotterell and Hinrich Schütze. 2018. Joint semantic synthesis and morphological analysis of the derived word. Transactions of the Association for Computational Linguistics, 6:33–48. Ryan Cotterell, Ekaterina Vylomova, Huda Khayrallah, Christo Kirov, and David Yarowsky. 2017. Paradigm completion for derivational morphology. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2017. Scott Deerwester, Susan T. Dumais, George Furnas, Thomas Landauer, and Richard Harshman. 1990. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6):391–407. 3604 Daniel Deutsch, John Hewitt, and Dan Roth. 2018. A distributional and orthographic aggregation model for English derivational morphology. In Annual Meeting of the Association for Computational Linguistics (ACL) 56. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL HTL) 2019. Jesse Dodge, Suchin Gururangan, Dallas Card, Roy Schwartz, and Noah A. Smith. 2019. Show your work: Improved reporting of experimental results. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2019. Daniel Edmiston. 2020. A systematic analysis of morphological content in BERT models for multiple languages. In arXiv 2004.03032. Kawin Ethayarajh. 2019. How contextual are contextualized word representations? comparing the geometry of BERT, ELMo, and GPT-2 embeddings. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2019. Allyson Ettinger. 2020. What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models. Transactions of the Association for Computational Linguistics, 8:34–48. Hao Fang, Mari Ostendorf, Peter Baumann, and Janet B. Pierrehumbert. 2015. Exponential language modeling using morphological features and multitask learning. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(12):2410– 2421. Laurie B. Feldman and Carol A. Fowler. 1987. The inflected noun system in Serbo-Croatian: Lexical representation of morphological structure. Memory and Cognition, 15(1):1–12. Uli H. Frauenfelder and Robert Schreuder. 1992. Constraining psycholinguistic models of morphological processing and representation: The role of productivity. In Geert Booij and Jaap van Marle, editors, Yearbook of morphology 1991, volume 26, pages 165– 183. Kluwer, Dordrecht. Philip Gage. 1994. A new algorithm for data compression. The C Users Journal, 12(2):23–38. Hélène Giraudo and Jonathan Grainger. 2003. On the role of derivational affixes in recognizing complex words: Evidence from masked priming. In R. Harald Baayen and Robert Schreuder, editors, Morphological structure in language processing, pages 209– 232. De Gruyter, Berlin. Pius ten Hacken. 2014. Delineating derivation and inflection. In Rochelle Lieber and Pavol Štekauer, editors, The Oxford handbook of derivational morphology, pages 10–25. Oxford University Press, Oxford. Bo Han and Timothy Baldwin. 2011. Lexical normalisation of short text messages: Makn sens a #twitter. In Annual Meeting of the Association for Computational Linguistics (ACL) 49. Martin Haspelmath and Andrea D. Sims. 2010. Understanding morphology. Routledge, New York, NY. Benjamin Heinzerling and Michael Strube. 2019. Sequence tagging with contextual and non-contextual subword representations: A multilingual evaluation. In Annual Meeting of the Association for Computational Linguistics (ACL) 57. John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL HTL) 2019. Valentin Hofmann, Janet B. Pierrehumbert, and Hinrich Schütze. 2020a. DagoBERT: Generating derivational morphology with a pretrained language model. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2020. Valentin Hofmann, Janet B. Pierrehumbert, and Hinrich Schütze. 2020b. Predicting the growth of morphological families from social and linguistic factors. In Annual Meeting of the Association for Computational Linguistics (ACL) 58. Valentin Hofmann, Hinrich Schütze, and Janet B. Pierrehumbert. 2020c. A graph auto-encoder model of derivational morphology. In Annual Meeting of the Association for Computational Linguistics (ACL) 58. Abhilash Jain, Aku Rouhe, Stig-Arne Grönroos, and Mikko Kurimo. 2020. Finnish ASR with deep transformer models. In Conference of the International Speech Communication Association (INTERSPEECH) 21. Ganesh Jawahar, Benoit Sagot, and Djamé Seddah. 2019. What does BERT learn about the structure of language? In Annual Meeting of the Association for Computational Linguistics (ACL) 57. Emilia Kacprzak, Laura M. Koesten, Luis-Daniel Ibáñez, Elena Simperl, and Jeni Tennison. 2017. A query log analysis of dataset search. In International Conference on Web Engineering (ICWE) 17. Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Annual Meeting of the Association for Computational Linguistics (ACL) 58. 3605 Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. 2016. Character-aware neural language models. In Conference on Artificial Intelligence (AAAI) 30. Diederik P. Kingma and Jimmy L. Ba. 2015. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR) 3. Max Kisselew, Sebastian Padó, Alexis Palmer, and Jan Šnajder. 2015. Obtaining a better understanding of distributional models of german derivational morphology. In International Conference on Computational Semantics (IWCS) 11. Matej Klemen, Luka Krsnik, and Marko RobnikŠikonja. 2020. Enhancing deep neural networks with morphological information. In arXiv 2011.12432. Victor Kuperman, Raymond Bertram, and R. Harald Baayen. 2008. Morphological dynamics in compound processing. Language and Cognitive Processes, 23(7-8):1089–1132. Victor Kuperman, Robert Schreuder, Raymond Bertram, and R. Harald Baayen. 2009. Reading of polymorphemic Dutch compounds: Towards a multiple route model of lexical processing. Journal of Experimental Psychology: Human Perception and Performance, 35(3):876–895. Alessandro Laudanna and Cristina Burani. 1985. Address mechanisms to decomposed lexical entries. Linguistics, 23(5). Alessandro Laudanna and Cristina Burani. 1995. Distributional properties of derivational affixes: Implications for processing. In Laurie B. Feldman, editor, Morphological aspects of language processing, pages 345–364. Lawrence Erlbaum, Hillsdale, NJ. Angeliki Lazaridou, Marco Marelli, Roberto Zamparelli, and Marco Baroni. 2013. Compositional-ly derived representations of morphologically complex words in distributional semantics. In Annual Meeting of the Association for Computational Linguistics (ACL) 51. Alina Leminen, Eva Smolka, Jon Duñabeitia, and Christos Pliatsikas. 2019. Morphological processing in the brain: The good (inflection), the bad (derivation) and the ugly (compounding). Cortex, 116:4–44. Minh-Thang Luong, Richard Socher, and Christopher D. Manning. 2013. Better word representations with recursive neural networks for morphology. In Conference on Computational Natural Language Learning (CoNLL) 17. Wentao Ma, Yiming Cui, Chenglei Si, Ting Liu, Shijin Wang, and Guoping Hu. 2020. CharBERT: Character-aware pre-trained language model. In International Conference on Computational Linguistics (COLING) 28. Leon Manelis and David A. Tharp. 1977. The processing of affixed words. Memory and Cognition, 5(6):690–695. Louis Martin, Benjamin Muller, Pedro J. Suárez, Yoann Dupont, Laurent Romary, de la Clergerie, Éric V., Djamé Seddah, and Benoit Sagot. 2020. CamemBERT: A tasty French language model. In Annual Meeting of the Association for Computational Linguistics (ACL) 58. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. In arXiv 1301.3781. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013b. Distributed representations of words and phrases and their compositionality. In Advances in Neural Information Processing Systems (NIPS) 26. Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In Annual Meeting of the Association for Computational Linguistics (ACL) 47. Jeremy M. Needle and Janet B. Pierrehumbert. 2018. Gendered associations of english morphology. Journal of the Association for Laboratory Phonology, 9(1):119. Boris New, Marc Brysbaert, Juan Segui, Ludovic Ferrand, and Kathleen Rastle. 2004. The processing of singular and plural nouns in french and english. Journal of Memory and Language, 51(4):568–585. Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fined-grained aspects. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2019. Sebastian Padó, Aurélie Herbelot, Max Kisselew, and Jan Šnajder. 2016. Predictability of distributional semantics in derivational word formation. In International Conference on Computational Linguistics (COLING) 26. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global vectors for word representation. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2014. Yuval Pinter, Cassandra L. Jacobs, and Jacob Eisenstein. 2020. Will it unblend? In Findings of Empirical Methods in Natural Language Processing (EMNLP) 2020. Ingo Plag. 2003. Word-formation in English. Cambridge University Press, Cambridge, UK. Siyu Qiu, Qing Cui, Jiang Bian, Bin Gao, and Tie-Yan Liu. 2014. Co-learning of word representations and morpheme representations. In International Conference on Computational Linguistics (COLING) 25. 3606 Péter Rácz, Clay Beckner, Jennifer Hay, and Janet B. Pierrehumbert. 2020. Morphological convergence as on-line lexical analogy. Language, 96(4):735– 770. Péter Rácz, Janet B. Pierrehumbert, Jennifer Hay, and Viktória Papp. 2015. Morphological emergence. In Brian MacWhinney and William O’Grady, editors, The handbook of language emergence, pages 123– 146. Wiley, Hoboken, NJ. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67. Kathleen Rastle and Matthew H. Davis. 2008. Morphological decomposition based on the analysis of orthography. Language and Cognitive Processes, 23(7-8):942–971. Kathleen Rastle, Matthew H. Davis, and Boris New. 2004. The broth in my brother’s brothel: Morpho-orthographic segmentation in visual word recognition. Psychonomic Bulletin and Review, 11(6):1090–1098. Alexandre Salle and Aline Villavicencio. 2018. Incorporating subword information into matrix factorization word embeddings. In Workshop on Subword/Character LEvel Models 2. Robert Schreuder and R. Harald Baayen. 1995. Modeling morphological processing. In Laurie B. Feldman, editor, Morphological aspects of language processing, pages 131–154. Lawrence Erlbaum, Hillsdale, NJ. Mike Schuster and Kaisuke Nakajima. 2012. Japanese and Korean voice search. In International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 37. Hinrich Schütze. 1992. Word space. In Advances in Neural Information Processing Systems (NIPS) 5, pages 895–902. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Annual Meeting of the Association for Computational Linguistics (ACL) 54. Suzanna Sia, Ayush Dalmia, and Sabrina J. Mielke. 2020. Tired of topic models? Clusters of pretrained word embeddings make for fast and good topics too! In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2020. Joseph P. Stemberger. 1994. Rule-less morphology at the phonology-lexicon interface. In Susan D. Lima, Roberta Corrigan, and Gregory Iverson, editors, The reality of linguistic rules, pages 147–169. John Benjamins, Amsterdam. Gregory Stump. 2017. Rule conflation in an inferentialrealizational theory of morphotactics. Acta Linguistica Academica, 64(1):79–124. Gregory Stump. 2019. Some sources of apparent gaps in derivational paradigms. Morphology, 29(2):271– 292. Marcus Taft. 1979. Recognition of affixed words and the word frequency effect. Memory and Cognition, 7(4):263–272. Marcus Taft. 1981. Prefix stripping revisited. Journal of Verbal Learning and Verbal Behavior, 20:289– 297. Marcus Taft. 1988. A morphological-decomposition model of lexical representation. Linguistics, 26:657– 667. Marcus Taft. 1991. Reading and the mental lexicon. Lawrence Erlbaum, Hove, UK. Marcus Taft. 1994. Interactive-activation as a framework for understanding morphological processing. Language and Cognitive Processes, 9(3):271–294. Marcus Taft. 2004. Morphological decomposition and the reverse base frequency effect. The Quarterly Journal of Experimental Psychology, 57(4):745– 765. Marcus Taft and Kenneth I. Forster. 1975. Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behavior, 14:638–647. Samson Tan, Shafiq Joty, Lav R. Varshney, and MinYen Kan. 2020. Mind your inflections! Improving NLP for non-standard Englishes with base-inflection encoding. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2020. Ivan Vuli´c, Edoardo M. Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. 2020. Probing pretrained language models for lexical semantics. In Conference on Empirical Methods in Natural Language Processing (EMNLP) 2020. Ekaterina Vylomova, Ryan Cotterell, Timothy Baldwin, and Trevor Cohn. 2017. Context-aware prediction of derivational word-forms. In Conference of the European Chapter of the Association for Computational Linguistics (EACL) 15. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc Le V, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Łukasz Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith 3607 Stevens, George Kurian, Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. In arXiv 1609.08144. Yi Yang and Jacob Eisenstein. 2017. Overcoming language variation in sentiment analysis with social attention. Transactions of the Association for Computational Linguistics, 5:295–307. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems (NeurIPS) 33. A Appendices A.1 Derivational Segmentation Let A be a set of derivational affixes and S a set of stems. To determine the derivational segmentation of a word w, we employ an iterative algorithm. Define the set BA 1 of w as the words that remain when one derivational affix from A is removed from w. For example, unlockable can be segmented into un, lockable and unlock, able so BA 1 (unlockable) = {lockable, unlock} (we assume that un and able are in A). We then iteratively create BA i+1(w) = S b∈BA i (w) BA 1 (b), i.e., we iteratively remove affixes from w. We stop as soon as BA i+1(w) ∩S ̸= ∅. The element in this intersection, together with the used affixes from A, forms the derivational segmentation of w.15 If there is no i such that BA i+1(w)∩S ̸= ∅, w does not have a derivational segmentation. The algorithm is sensitive to most morpho-orthographic rules of English (Plag, 2003), e.g., when the suffix ize is removed from isotopize, the resulting word is isotope, not isotop. In this paper, we follow Hofmann et al. (2020a) in using BERT’s prefixes, suffixes, and stems as input to the algorithm. Specifically, we assign 46 productive prefixes and 44 productive suffixes in BERT’s vocabulary to A and all fully alphabetic words with more than 3 characters in BERT’s vocabulary (excluding stopwords and affixes) to S, resulting in a total of 20,259 stems. This means that we only consider derivational segmentations that are possible given BERT’s vocabulary. 15If |BA i+1(w) ∩S| > 1 (rarely the case in practice), the element with the lowest number of suffixes is chosen. A.2 Data Preprocessing We exclude texts written in a language other than English and remove strings containing numbers as well as hyperlinks. We follow Han and Baldwin (2011) in reducing repetitions of more than three letters (niiiiice) to three letters. A.3 Hyperparameters The feed-forward network has a ReLU activation after the first layer and a sigmoid activation after the second layer. The first layer has 100 dimensions. We apply dropout of 0.2 after the first layer. All other hyperparameters are as for BERTBASE (uncased) (Devlin et al., 2019). The number of trainable parameters is 109,559,241. We use a batch size of 64 and perform grid search for the number of epochs n ∈ {1, . . . , 20} and the learning rate l ∈ {1 × 10−6, 3 × 10−6, 1 × 10−5, 3 × 10−5} (selection criterion: F1 score). We tune l on Reddit (80 hyperparameter search trials per model type) and use the best configuration (which is identical for all model types) for 20 training runs with different random seeds on all three datasets (20 hyperparameter search trials per model type, dataset, and random seed). Models are trained with binary cross-entropy as the loss function and Adam (Kingma and Ba, 2015) as the optimizer. Experiments are performed on a GeForce GTX 1080 Ti GPU (11GB). Table 4 lists statistics of the validation performance over hyperparameter search trials and provides information about best hyperparameter configurations as well as runtimes.16 See also Section 3.5 and particularly Figure 2 in the main text, where we present a detailed analysis of the convergence behavior of the two main model types examined in this study (DelBERT and BERT). 16Since expected validation performance (Dodge et al., 2019) may not be correct for grid search, we report mean and standard deviation of the performance instead. 3608 Amazon ArXiv Reddit Model µ σ n l τ µ σ n l τ µ σ n l τ DelBERT .627 .007 6.75 3e-06 67.73 .725 .006 11.45 3e-06 28.69 .687 .006 5.45 3e-06 25.56 BERT .612 .006 7.30 3e-06 66.18 .693 .015 17.05 3e-06 28.04 .657 .007 9.25 3e-06 25.06 Stem .556 .016 9.85 3e-06 67.43 .699 .005 8.15 3e-06 28.56 .670 .006 6.00 3e-06 25.39 Affixes .519 .008 5.55 3e-06 67.70 .599 .004 7.50 3e-06 28.43 .593 .003 9.35 3e-06 25.49 Table 4: Validation performance statistics and hyperparameter search details. The table shows the mean (µ) and standard deviation (σ) of the validation performance (F1) on all hyperparameter search trials, the number of epochs (n) and learning rate (l) with the best validation performance, and the runtime (τ) in minutes for one full hyperparameter search (20 trials). The numbers are averaged across 20 training runs with different random seeds.
|
2021
|
279
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 328–339 August 1–6, 2021. ©2021 Association for Computational Linguistics 328 Multimodal Sentiment Detection Based on Multi-channel Graph Neural Networks Xiaocui Yang, Shi Feng, Yifei Zhang, Daling Wang School of Computer Science and Engineering, Northeastern University, China [email protected], {fengshi, wangdaling, zhangyifei}@cse.neu.edu.cn Abstract With the popularity of smartphones, we have witnessed the rapid proliferation of multimodal posts on various social media platforms. We observe that the multimodal sentiment expression has specific global characteristics, such as the interdependencies of objects or scenes within the image. However, most previous studies only considered the representation of a single image-text post and failed to capture the global co-occurrence characteristics of the dataset. In this paper, we propose Multi-channel Graph Neural Networks with Sentiment-awareness (MGNNS) for imagetext sentiment detection. Specifically, we first encode different modalities to capture hidden representations. Then, we introduce multichannel graph neural networks to learn multimodal representations based on the global characteristics of the dataset. Finally, we implement multimodal in-depth fusion with the multi-head attention mechanism to predict the sentiment of image-text pairs. Extensive experiments conducted on three publicly available datasets demonstrate the effectiveness of our approach for multimodal sentiment detection. 1 Introduction The tasks of extracting and analyzing sentiments embedded in data have attracted substantial attention from both academic and industrial communities (Zhang et al., 2018; Yue et al., 2018). With the increased use of smartphones and the bloom of social media such as Twitter, Tumblr and Weibo, users can post multimodal tweets (e.g., text, image, and video) about diverse events and topics to convey their feelings and emotions. Therefore, multimodal sentiment analysis has become a popular research topic in recent years (Kaur and Kautish, 2019; Soleymani et al., 2017). As shown in Fig. 1, sentiment is no longer expressed by a pure modality in the multimodal scenario but rather by the com(a) We have a fun day on the beach! (Positive) (b) We have a nice day on a deserted beach. (Positive) Figure 1: Multimodal posts with global characteristics. Two posts express the user’s positive sentiment from multimodal data that has global characteristics, including the “have a fun/nice day” phrase, the ocean scene, and the beach scene. bined expressions of multiple modalities (e.g., text, image, etc.). In contrast to unimodal data, multimodal data consist of more information and make the user’s expression more vivid and interesting. We focus on multimodal sentiment detection for image-text pairs in social media posts. The problem of image-text mismatch and flaws in social media data, such as informality, typos, and a lack of punctuation, pose a fundamental challenge for the effective representation of multimodal data for the sentiment detection task. To tackle this challenge, Xu et al. (2017; 2017) constructed different networks for multimodal sentiment analysis, such as a Hierarchical Semantic Attentional Network (HSAN) and a Multimodal Deep Semantic Network (MDSN). Xu et al. (2018) and Yang et al. (2020) proposed a Co-Memory network (Co-Mem) and a Multi-view Attentional Network (MVAN) models, respectively, introducing memory networks to realize the interaction between modalities. The above methods treat each image-text post in the dataset as a single instance, and feature dependencies across instances are neglected or modeled implicitly. In fact, social media posts have specific global co-occurring characteristics, i.e., co329 occurring words, objects, or scenes, which tend to share similar sentiment orientations and emotions. For example, the co-occurrences of the words “have a fun/nice day” and of the bright scenes “ocean/beach” in the two images in Fig. 1 imply a strong relationship between these features and positive sentiment. How to more effectively make use of the feature co-occurrences across instances and capture the global characteristics of the data remain a great challenge. We propose a Multi-channel Graph Neural Networks model with Sentiment-awareness (MGNNS) for multimodal sentiment analysis that consists of three stages. (i) Feature extraction. For text modality, we encode the text and obtain a text memory bank; for image modality, we first extract objects and scenes and then capture the image’ semantic features from a multiview perspective. (ii) Feature representation. We employ a Graph Neural Network (GNN) for text modality based on the global shared matrices, i.e., one text graph based on word co-occurrence is built based on the whole dataset. Specifically, we first connect word nodes within an appropriate small window in the text. After that, we update the node representation by itself as well as neighbor nodes. For image modality, it is believed that different views of an image, such as the beach (Scene view) and person (Object view) in Fig. 1(a), can reflect a user’s emotions (Xu and Mao, 2017). The existing literature usually models the relationship between the scenes and objects within an image, failing to capture the rich co-occurrence information from the perspective of the whole dataset. In contrast, we explicitly build two graphs for scenes and objects according to the co-occurrences in the datasets and propose Graph Convolutional Network (GCN) models over the two graphs to represent the images. In general, to tackle the isolated feature problem, we build multiple graphs for different modalities, with each GNN acting as a channel, and propose a Multi-channel Graph Neural Networks (MultiGNN) module to capture the in-depth global characteristics of the data. This multi-channel based method can provide complementary representation from different sources (George and Marcel, 2021; George et al., 2019; Islam et al., 2019). (iii) Feature fusion. Previous studies usually directly connect multimodal representations, without considering multimodal interactions (Wang et al., 2020a; Xu, 2017; Xu and Mao, 2017). In this stage, we realize the pairwise interaction of text and image modalities from different channels through the use of the Multimodal Multi-head Attention Interaction (MMAI) module and obtain the fusion representation. Our main contributions are summarized as follows: • We propose a novel MGNNS framework that models the global characteristics of the dataset to handle the multimodal sentiment detection task. To the best of our knowledge, we are the first to apply GNN to the image-text multimodal sentiment detection task. • We construct the MMAI module from different channels to realize in-depth multimodal interaction. • We conduct extensive experiments on three publicly available datasets, and the results show that our model outperforms the stateof-the-art methods. 2 Related Work 2.1 Multimodal Sentiment Analysis For convenience, multimodal polarity analysis and emotion analysis are unified to form multimodal sentiment analysis. Traditional machine learning methods are adopted to address the multimodal sentiment analysis task (P´erez-Rosas et al., 2013; You et al., 2016). Recently, deep learning models have also achieved promising results for this task. For the video dataset, Wang et al. (2020b) proposed a novel method, TransModality, to fuse multimodal features with end-to-end translation models; Zhang et al. (2020) leveraged semi-supervised variational autoencoders to mine more information from unlabeled data; and Hazarika et al. (2020) constructed a novel framework, MISA, which projects each modality to two distinct subspaces: modalityinvariant and modality-specific subspaces. There is a massive amount image-text data on social platforms, and thus, image-text multimodal sentiment analysis has attracted the attention of many researchers. Xu et al. constructed different networks for multimodal sentiment analysis—HSAN (2017), MDSN (2017) and Co-Mem (2018). Yang et al. (2020) built an image-text emotion dataset, named TumEmo, and further proposed MVAN for multimodal emotion analysis. 330 2.2 Graph Neural Network The Graph Neural Network has achieved promising results for text classification, multi-label recognition, and multimodal tasks. For text classification, a novel neural network called Graph Neural Network (GNN), and its variants have been rapidly developed, and their performance is better than that of traditional methods, such as Text GCN (Yao et al., 2019), TensorGCN (Liu et al., 2020), and TextLevelGNN (Huang et al., 2019). The GCN is also introduced in the multi-label image recognition task to model the label dependencies (Chen et al., 2019). Recently, Graph Convolutional Network has been applied in different multimodal tasks, such as Visual Dialog (Guo et al., 2020; Khademi, 2020), multimodal fake news detection (Wang et al., 2020a), and Visual Question Answering (VQA) (Hudson and Manning, 2019; Khademi, 2020). Jiang et al. (2020) applied a novel KnowledgeBridge Graph Network (KBGN) in modeling the relations among the visual dialogue cross-modal information in fine granularity. Wang et al. (2020a) proposed a novel Knowledge-driven Multimodal Graph Convolutional Network (KMGCN) to model semantic representations for fake news detection. However, the KMGCN extracted visual words as visual information and did not make full use of the global information of the image. Khademi (2020) introduced a new neural network architecture, a Multimodal Neural Graph Memory Network (MNGMN), for VQA, which model constructed a visual graph network based on the bounding-boxes, which produced overlapping parts that might provide redundant information. For the image-text dataset, we found that certain words often appear in a text post simultaneously, and different objects or scenes within an image have specific co-occurrences that indicate certain sentiments. We explicitly model these global characteristics of the dataset through the use of a multichannel GNN. 3 Proposed Model Fig. 2 illustrates the overall architecture of our proposed MGNNS model for multimodal sentiment detection that consists of three modules: the encoding module, the Multi-GNN module, and the multimodal interaction module. We first encode text and image input into hidden representations. Then, we introduce GNN from different channels to learn multiple modal representations. In this paper, the channels are the Text-GNN (TG) module, the Image-GCN-Scene (IGS) module, and the Image-GCN-Object (IGO) module. Finally, we realize the in-depth interactions between different modalities by multimodal multi-head attention. 3.1 Problem Formalization The goal of our model is to identify which sentiment is expressed by an image-text post. Given a set of multimodal posts from social media, P = {(T1, V1), ..., (TN, VN)}, where Ti is the text modality and Vi is the corresponding visual information, N represents the number of posts. We need to learn the model f : P →L to classify each post (Ti, Vi) into the predefined categories Li. For polarity classification, Li ∈ {Positive, Neutral, Negative}; for emotion classification, Li ∈ {Angry, Bored, Calm, Fear, Happy, Love, Sad }. 3.2 Encoding For text modality, we first encode words by GloVe (Pennington et al., 2014) to obtain the embedding vector and then obtain the text memory bank, Mt, by BiGRU (Cho et al., 2014): Mt = fBiGRU(Embedding(T)), Mt ∈RLt×2dt, (1) where T is a text sequence, Lt is the maximum length of a padded text sequence, and dt is the dimension of hidden units in the BiGRU. For image modality, we extract image features from both the object and scene views to capture sufficient information. We believe that there are interdependencies between different objects or scenes in an image. To explicitly model this co-occurrence, we first extract objects O = {o1, ..., olo} by YOLOv3 (Farhadi and Redmon, 2018), and extract scenes S = {s1, ..., sls} by VGG-Place (Zhou et al., 2017). Finally, we obtain the object and scene memory banks with the pretrained ResNet (He et al., 2016). Thus, if an input image V has a 448×448 resolution and is split into 14×14 = 196 visual blocks of the same size, then each block is represented by a 2,048-dimensional vector. Mx = fx ResNet(V ), Mx ∈RLx×dx, (2) where x ∈{Object, Scene}, Lx = 196, and dx = 2, 048. 331 We have a fun day on the beach! Glove Embedding we have fun day beach the on a ocean happy there an life Text_GCN Bi-GRU Text Memory Bank Scene_ResNet Extract Scenes bridge beach ocean coast castle ImageScene Memory Bank … 𝐶!"#$# = 365 𝐷!"#$# × Sentiment-awareness scene feature K-V Emotion-aware Multihead-Attention Q Sentiment Embedding Matrix Extract Objects Object_ResNet person building boat bus castle ImageObject Memory Bank … 𝐶%&'#"( = 80 𝐷%&'#"( × Emotion-aware Multihead-Attention K-V Q Sentiment Embedding Matrix Multimodal Multi-head Attention Interaction Emotion Label Predicting ×𝑳𝒐 Text feature ×𝑳𝒔 Image-GCN- Scene (IGS) Text-GNN (TG) Sentiment-awareness scene feature Image-GCN- Object (IGO) … … … … … … Figure 2: The framework of the proposed Multi-channel Graph Neural Networks with Sentiment-awareness (MGNNS) for multimodal sentiment detection. The channels are Text-GNN (TG) for text modality, Image-GCNScene (IGS) for image scene modality, and Image-GCN-Object (IGO) for image object modality. Note that we delete the stopwords during data preprocessing so that the words “a” and “the” do not have connections. 3.3 Multi-channel Graph Neural Networks In this subsection, we present our proposed MultiGNN module. As Fig. 2 shows, this module consists of the TG channel (middle), the IGO channel (right), and the IGS channel (left). Text GNN: As shown in the middle of Fig. 2, motivated by (Huang et al., 2019), we learn text representation through the Text Level GNN. For text with lt words T = {w1, ..., wk, ..., wlt}, where the kth word, wk, is initialized by glove embedding rt k ∈Rd, d = 300. We build the graph of the textbased vocabulary of the training dataset, which is defined as follows: Nt = {wk|k ∈[1, lt]}. (3) We build edges between wk and wj when the number of co-occurrences of two words is not less than 2. Et = {et k,j|wk ∈[w1, wlt]; wj ∈[wk−ws, wk+ws]}, (4) where Nt and Et are the set of nodes and edges of the text graph, respectively. The word representations in Nt and the edge weights in Et are taken from global shared matrices built based on vocabulary and the edge set of the dataset, respectively. That is, the representations of the same nodes and weights of the edges are shared globally. et k,j is initialized by point-wise mutual information (PMI) (Wang et al., 2020a) and is learned in the training process. ws is the hyperparameter sliding window size, which indicates how many adjacent nodes are connected to each word in the text graph. Then, we update the node representation based on its original representations and neighboring nodes by the message passing mechanism (MPM) (Gilmer et al., 2017), which is defined as follows: At k = max j∈Nws k et kjrt k, (5) rt k ′ = αrt k + (1 −α)At k, (6) where At k ∈Rd is the aggregated information from neighboring nodes from node k−ws to k+ws, and max is the reduction function. α is the trainable variable that indicates how much original information of the node should be kept, and rt k ′ ∈Rd is the updated representation of node k. Finally, we can calculate the new representation of text T as follows: T ′ = lt X k=1 rt k ′ (7) Image GCN: In this module, we explicitly model interdependence within lx scenes or objects by IGX, as shown on the left and right sides of Fig. 332 2, respectively. The graph of the image is defined as follows: Nx = {xp|p ∈[1, lx]}, (8) where Nx ∈RCx is the set of nodes of IGX; x or X ∈{Object, Scene}, Cx = 80 when x = Object, and Cx = 365 when x = Scene. To build the edges of IGX, we first build the global shared co-occurrence matrix-based dataset: Ex = {ex p,q|p ∈[1, lx] , q ∈[1, lx]}, (9) where Ex ∈RCx×Cx is the co-occurrence matrix; edge weight ex p,q indicates the co-occurrence times of xp and xq in the dataset. Then, we calculate the conditional probability for node p as follows: P x p,q = ex p,q/N x p , when q ̸= p (10) where Nx p denotes the occurrence times of xp in the dataset. Note that P x p,q ̸= P x q,p. As mentioned by (Chen et al., 2019), the simple correlation above may suffer several drawbacks. We further build the binary co-occurrence matrix: Bx p,q = ( 1, if P x p,q ≥β 0, if P x p,q ≤β , (11) where β is the hyperparameter used to filter noisy edges. It is obvious that the role of the central node is different from that of neighboring nodes, so we need to further calculate the weight of the edge: Rx p,q = ( 1 −γ, if p = q γ/PCx q=1Bx p,q, if p ̸= q , (12) where Rx ∈ RCx×Cx is the weighted cooccurrence matrix, and hyperparameter γ indicates the importance of neighboring nodes. Finally, we input node Nx and edge Rx of the image into the graph convolutional network. Like in (Kipf and Welling, 2016), every layer can be calculated as follows: Hx L+1 = h(c RxHx LW x L), (13) where Hx L ∈RCx×dx, Hx L+1 ∈RCx×dx′ , W x L ∈ Rdx×dx′ , and c Rx ∈RCx×Cx is the normalized representation of Rx; h(·) is a non-linear operation. When L = 1, Hx 1 is the word-embedding vector of Nx. K-V Text-Guided Image Scene Attention Q Text feature Image-X memory bank Add & Norm Feed Forward Add & Norm Text Memory Bank K-V Image Scene -Guided Text Attention Add & Norm Feed Forward Add & Norm Q Sentiment-awareness X feature Fused feature 𝑁!"# 𝑁#"! Figure 3: The MMAI module illustrates the process of multimodal interaction from four channels, X ∈ {Object, Scene}. We take the interaction process between text and image scene channels as an example to demonstrate this for convenience. The dotted arrows are the outputs of the other two channels after the interactions. By stacking multiple GCN layers, we can explicitly learn and model the complex interdependence of the nodes. Then, we obtain the image representation with objects or scenes dependencies: Ix = MaxPooling(Mx)(Hx L+1)T, Ix ∈RCx. (14) But, we cannot capture the relationship between nodes and sentiments. Therefore, we learn the sentiment-awareness image representation through multi-head attention (Vaswani et al., 2017). Att = softmax( QKT √dk )V, (15) EIx = MH(Q, K, V ) = Concat(head1, ..., headH)W O where headh = Att(QW Q h , KW K h , V W V h ), (16) where MH(·) is multi-head attention; W Q h ∈ Rd×dk, W K h ∈Rdmodel×dk, W V h ∈Rdmodel×dv, and W O ∈RHdv×d; and H = 5, dmodel = 300, dk = dv = 60. Q ∈Rls×d is a sentiment embedding matrix built based on the label set ls = 3 for polarity classification and ls = 7 for emotion classification; K = V = IxW I, W I ∈ RCx×dmodel, K, V ∈Rdmodel. 3.4 Multimodal Interaction Motivated by the Transformer (Vaswani et al., 2017) prototype, we design a Multimodal Multihead Attention Interaction (MMAI) module that can effectively learn the interaction between text 333 modality and image modality by multiple channels, as shown in Fig. 3. We employ the MMAI to obtain the Text guided Image-X representations and Image-X guided Text representations, X ∈{Object, Scene}. For the Text-guided Image-X attention, OTgX N+1 = LN(MH(Q = HTgX N , K = V = Mx) + HTgX N ), (17) HTgX N+1 = LN(FFN(OTgX N+1) + OTgX N+1), (18) where LN(·) is layer normalization, and FFN(·) is the feed-forward network. When N = 1, HTgX 1 = T ′, as in Eq. 7. For the Image-X-guided Text attention, OXgT N+1 = LN(MH(Q = HXgT N , K = V = Mt) + HXgT N ), (19) HXgT N+1 = LN(FFN(OXgT N+1) + OXgT N+1), (20) when N = 1, HXgT 1 = EIx, as in Eq. 16. For MH, H = 4, dmodel = 512, dk = dv = 128. The fused multimodal representation is as follows: Rm = [HTgO N ⊕HTgS N ⊕HOgT N ⊕HSgT N ], where ⊕is a concatenation operation. 3.5 Sentiment Detection Finally, we feed the above fused representation, Rm, into the top fully connected layer and employ the softmax function for sentiment detection. Lm = softmax(wsRm + bs), Lm ∈Rls, (21) where ws and bs are the parameters of the fully connected layer. 4 Experiments We conduct experiments on three multimodal sentiment datasets from social media platforms, MVSASingle, MVSA-Multiple (Niu et al., 2016), and TumEmo (Yang et al., 2020), and compare our MGNNS model with a number of unimodal and multimodal approaches. 4.1 Datasets MVSA-Single and MVSA-Multiple are two different scale image-text sentiment datasets crawled from Twitter1. TumEmo is a multimodal weaksupervision emotion dataset containing a large 1https://twitter.com Dataset Train Val Test All MVSA-S 3,608 451 452 4,511 MVSA-M 13,618 1,703 1,703 17,024 TumEmo 156,204 19,525 19,536 195,265 Table 1: Statistics of the different datasets. amount of image-text data crawled from Tumblr2. The statistics of these datasets are given in Appendix A; and for a fair comparison, we adopt the same data preprocessing method as that of Yang (Yang et al., 2020). The corresponding details are shown in Appendix B. 4.2 Experimental Setup Parameter MVSA-∗ TumEmo Learning rate 4e −5 5e −5 ws 4 5 Object-β 0.4 0.4 Scene-β 0.3 0.5 γ 0.2 0.2 Lx 2 2 NTgX 1 1 NXgT 1 1 Table 2: Parameter settings of the different datasets. We adopt the cross-entropy loss function and Adam optimizer. In the process of extracting objects and scenes, we reserve the objects with the probability greater than 0.5 and the top-5 scenes, respectively. The other parameters are listed in Table 2, ∗∈{Single, Multiple}. We use Accuracy (Acc) and F1-score (F1) as evaluation metrics. All models are implemented with PyTorch. 4.3 Baselines We compare our model with multimodal sentiment models with the same modalities and the unimodal baseline models. Unimodal Baselines: For text modality, CNN (Kim, 2014) and Bi-LSTM (Zhou et al., 2016) are well-known models for text classification tasks, and BiACNN (Lai et al., 2015) incorporates the CNN and BiLSTM models with an attention mechanism for text sentiment analysis. TGNN (Huang et al., 2019) is a text-level graph neural network for text classification. For image modality, OSDA (Yang 2http://tumblr.com 334 Modality Model MVSA-Single MVSA-Multiple TumEmo Acc F1 Acc F1 Acc F1 Text CNN 0.6819 0.5590 0.6564 0.5766 0.6154 0.4774 BiLSTM 0.7012 0.6506 0.6790 0.6790 0.6188 0.5126 BiACNN 0.7036 0.6916 0.6847 0.6319 0.6212 0.5016 TGNN 0.7034 0.6594 0.6967 0.6180 0.6379 0.6362 Image OSDA 0.6675 0.6651 0.6662 0.6623 0.4770 0.3438 SGN 0.6620 0.6248 0.6765 0.5864 0.4353 0.4232 OGN 0.6659 0.6191 0.6743 0.6010 0.4564 0.4446 DuIG 0.6822 0.6538 0.6819 0.6081 0.4636 0.4561 ImageText HSAN 0.6988 0.6690 0.6796 0.6776 0.6309 0.5398 MDSN 0.6984 0.6963 0.6886 0.6811 0.6418 0.5692 Co-Mem 0.7051 0.7001 0.6992 0.6983 0.6426 0.5909 MVAN‡ 0.7298‡ 0.7139‡ 0.7183‡ 0.7038‡ 0.6553‡ 0.6543‡ MGNNS 0.7377 0.7270 0.7249 0.6934 0.6672 0.6669 Table 3: Experiment results of Acc and F1 on three datasets. ‡ represents the reproductive operation. et al., 2020) is an image sentiment analysis model based on multiple views. Note that the SGN, OGN, and DuIG are variants of our model and rely only on image modality. SGN and OGN are the image graph convolutional neural networks based on scenes and objects for image sentiment analysis, respectively. DuIG is the image graph convolutional neural network with dual views, e.g., Object and Scene. Muiltimodal Baselines: HSAN (Xu, 2017) is a hierarchical semantic attentional network based on image captions for multimodal sentiment analysis. MDSN (Xu and Mao, 2017) is a deep semantic network with attention for multimodal sentiment analysis. Co-Mem (Xu et al., 2018) is a co-memory network for iteratively modeling the interactions between multiple modalities. MVAN (Yang et al., 2020) is a multi-view attentional network that utilizes a memory network for multimodal emotion analysis. This model achieves state-of-the-art performance on image-text multimodal sentiment classification tasks. 4.4 Experimental Results and Analysis The experimental results of the baseline methods and our model are shown in Table 3, where MGNNS denotes that our model is based on multichannel graph neural networks3. We can make the following observations. First, 3The source codes are available for use at https:// github.com/YangXiaocui1215/MGNNS. our model (MGNNS) is competitive with the other strong baseline models on the three datasets. Note that the data distribution of MVSA-∗is extremely unbalanced. Thus, we reproduce the MVAN model with ACC and Weighted-F1 metrics instead of the Micro-F1 metric used in the original paper, which is more realistic. Second, the multimodal sentiment analysis models perform better than most of the unimodal sentiment analysis models on all three datasets. Moreover, the segmental indictors are difficult to capture for images owing to the low information density, and the sentiment analysis on the image modality achieves the worst results. Finally, the TGNN unimodal model outperforms the HSAN multimodal model, indicating that the GNN has excellent performance in sentiment analysis. 4.5 Ablation Experiments We conduct ablation experiments on the MGNNS model to demonstrate the effectiveness of different modules. Table 4 shows that the whole MGNNS model achieves the best performance among all models. To show the performance of the MultiGNN module, we replace the Text-GNN with the CNN, as well as the Image-GCN with the pretrained ResNet. The removal of the MMAI module (w/o MMAI) and Multi-GNN module (w/o MGNN) adversely affect the model results, which indicates that these modules are useful for multimodal sentiment analysis. By replacing the MMAI module with the CoAtt (Lu et al., 2016) module 335 Datasets Model Acc F1 MVSA-Single w/o MGNN 0.7010 0.6847 w/o MMAI 0.7108 0.6879 +CoAtt 0.7255 0.6986 w/o Scene 0.7304 0.6988 w/o Object 0.7034 0.6900 MGNNS 0.7377 0.7270 MVSA-Multiple w/o MGNN 0.7019 0.6752 w/o MMAI 0.7128 0.6792 +CoAtt 0.7210 0.6849 w/o Scene 0.7170 0.6797 w/o Object 0.7110 0.6848 MGNNS 0.7249 0.6934 TumEmo w/o MGNN 0.6553 0.6547 w/o MMAI 0.6370 0.6347 +CoAtt 0.6624 0.6606 w/o Scene 0.6618 0.6593 w/o Object 0.6592 0.6584 MGNNS 0.6672 0.6669 Table 4: Ablation experiment results. (+CoAtt), the model performance is found to be slightly worse than that of the MGNNS module. This further illustrates the importance of multimodal interactions and the superiority of the MMAI module. When one of the object views (w/o Object) or scene views (w/o Scene) is removed, the performance of the model declines, which indicates that both views of the image are effective for multimodal sentiment analysis. 4.6 Transferability Experiment In the Multi-GNN module, we build multiple graphs for different modalities based on the dataset. For different datasets, the graphs built by the unimodal model are different. However, can graph capture from one dataset (e.g., MVSA-Single) have positive effects on other datasets (e.g., TumEmo)? In this subsection, we will verify the transferability of the model through experiments. As Table 5 shows, the following conclusions can be drawn: (i) Regardless of the modality, such as text or image, compared to introducing the graph constructed based on own dataset, the experimental results calculated based on graphs transferred from other datasets are worse. This is mainly because each dataset has unique global characteristics, the experimental results based on transferred graphs are slightly worse. (ii) However, due to the commonality of datasets when expressing the same emotions, the results of the transferred models are not completely worse. For example, the same scenes and objects can appear in different images in different datasets simultaneously for image modalities. Therefore, graphs from different datasets have transferability and can be used for other datasets. (iii) For different datasets, the experimental results of “X2Y-Text” are worse than those of “X2Y-Image”. That is, the text graph has worse transferability. The reason for this may be that text graphs with various nodes are created based on the vocabulary of different datasets. Two situations in the transferred text graph will seriously affect the results: fewer nodes will lose information, and more nodes will provide redundant information. (iv) When the dataset gap is relatively wide, the transferability of text graphs is worse. For example, from the larger datasets transfer to the smallest dataset, including T2S-Text and M2S-Text, experimental results show a drop of 2.45% and 2.69%, respectively; from the smaller datasets transfer to the most largest dataset, including S2T-Text and M2T-Text, experimental results show a significant drop of 4.81% and 4.09%, respectively. 4.7 Hyperparameter Settings Hyperparameter ws: To obtain adequate information from neighboring nodes in the TGNN, we conduct experiments under different settings for hyperparameter ws in Eq. 4, the related results of which are shown in Fig. 4. The best ws selection varies among different datasets since the average text length of TumEmo is longer compared to other data. The TGNN cannot obtain sufficient information from neighboring nodes with ws values that are too small, while larger values may degrade the performance due to the redundant information provided by neighboring nodes. (a) Comparisons on MVSA-∗ (b) Comparisons on TumEmo Figure 4: Acc comparisons with different values of ws. MS is MVSA-Single, MM is MVSA-Multiple, and T is TumEmo. 336 Model MVSA-Single Model MVSA-Multiple Model TumEmo Acc F1 Acc F1 Acc F1 M2S-Text 0.7132 0.6985 S2M-Text 0.7146 0.6912 S2T-Text 0.6191 0.6202 T2S-Text 0.7108 0.6939 T2M-Text 0.7110 0.6752 M2T-Text 0.6263 0.6239 M2S-Image 0.7206 0.6901 S2M-Image 0.7177 0.6795 S2T-Image 0.6635 0.6611 T2S-Image 0.7255 0.7027 T2M-Image 0.7183 0.6848 M2T-Image 0.6625 0.6615 MGNNS 0.7377 0.7270 MGNNS 0.7249 0.6934 MGNNS 0.6672 0.6669 Table 5: Transferability experiment results of Acc and F1 on different datasets. S, M and T denote MVSA-Single, MVSA-Multiple, and TumEmo, respectively. For “Z” modality, “X2Y-Z” represents that the graph that is built based on the “X” dataset is transfered to the “Y” dataset, where Z ∈{Text, Image}, X ∈{MVSA-Single, MVSAMultiple, TumEmo}, and Y ∈{MVSA-Single, MVSA-Multiple, TumEmo}. For example, “M2S-Text” represents that the text graph that is built based on the MVSA-Multiple dataset is transferred to the MVSA-Single dataset. Hyperparameter β: We vary the values of hyperparameter β in Eq. 11 for the binary cooccurrence matrix from different views, the results of which are shown in Fig. 5. We find that the best β value is different for different views in different datasets. For MVSA-∗, the smaller β value can reserve more edges to capture more information since the scene co-occurrence matrix is sparser than that in the object view. For TumEmo with a large amount of data, preserving the top-5 scenes produces many noise edges, so the value of scene-β is greater than that of MVSA-∗. β (a) Comparisons of cbject view on MVSA-∗ β (b) Comparisons of scene view on MVSA-∗ β (c) Comparisons of object view on TumEmo β (d) Comparisons of scene view on TumEmo Figure 5: Acc comparisons with different β values. Hyperparameter γ: As Fig. 6 shows, the model receives the best performance for the three datasets when γ is 0.2. When γ is smaller, the neighboring nodes do not receive enough attention; in contrast, their own information is not fully utilized. γ (a) Comparisons on MVSA-∗ γ (b) Comparisons on TumEmo Figure 6: Acc comparisons with different γ values. 5 Conclusions This paper proposes a novel model, MGNNS, that is built based on the global characteristics of the dataset for multimodal sentiment detection tasks. As far as we know, this is the first application of graph neural networks in image-text multimodal sentiment analysis. The experimental results on publicly available datasets demonstrated that our proposed model is competitive with strong baseline models. In future work, we plan to construct a model that adopts the advantages of the GNN and pretrained models such as BERT, VisualBERT, and etc. We want to design a reasonable algorithm to characterize the quality of the objects and scenes selected from the image and further improve the representation ability of the model. Acknowledgments The project is supported by the National Key R&D Program of China (2018YFB1004700) and by the National Natural Science Foundation of China (61772122, 61872074, U1811261). 337 References Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. 2019. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5177–5186. Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724– 1734. Ali Farhadi and Joseph Redmon. 2018. Yolov3: An incremental improvement. Computer Vision and Pattern Recognition, cite as. Anjith George and Sebastien Marcel. 2021. Learning one class representations for face presentation attack detection using multi-channel convolutional neural networks. IEEE Transactions on Information Forensics and Security, 16:361–375. Anjith George, Zohreh Mostaani, David Geissenbuhler, Olegs Nikisins, Andr´e Anjos, and S´ebastien Marcel. 2019. Biometric face presentation attack detection with multi-channel convolutional neural network. IEEE Transactions on Information Forensics and Security, 15:42–55. Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. 2017. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, pages 1263–1272. Dan Guo, Hui Wang, Hanwang Zhang, Zheng-Jun Zha, and Meng Wang. 2020. Iterative contextaware graph inference for visual dialog. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10055–10064. Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. 2020. Misa: Modality-invariant and -specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122–1131. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778. Lianzhe Huang, Dehong Ma, Sujian Li, Xiaodong Zhang, and Houfeng Wang. 2019. Text level graph neural network for text classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3442–3448. Drew A. Hudson and Christopher D. Manning. 2019. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6700–6709. Jumayel Islam, Robert E Mercer, and Lu Xiao. 2019. Multi-channel convolutional neural network for twitter emotion and sentiment recognition. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1355–1365. Xiaoze Jiang, Siyi Du, Zengchang Qin, Yajing Sun, and Jing Yu. 2020. Kbgn: Knowledge-bridge graph network for adaptive vision-text reasoning in visual dialogue. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1265–1273. Ramandeep Kaur and Sandeep Kautish. 2019. Multimodal sentiment analysis: A survey and comparison. International Journal of Service Science, Management, Engineering, and Technology (IJSSMET), 10(2):38–58. Mahmoud Khademi. 2020. Multimodal neural graph memory networks for visual question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7177– 7188. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751. Thomas N. Kipf and Max Welling. 2016. Semisupervised classification with graph convolutional networks. In ICLR (Poster). Siwei Lai, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Recurrent convolutional neural networks for text classification. In Twenty-ninth AAAI conference on artificial intelligence. Xien Liu, Xinxin You, Xiao Zhang, Ji Wu, and Ping Lv. 2020. Tensor graph convolutional networks for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8409–8416. Jiasen Lu, Jianwei Yang, Dhruv Batra, and Devi Parikh. 2016. Hierarchical question-image co-attention for visual question answering. In Advances in Neural Information Processing Systems, volume 29, pages 289–297. Teng Niu, Shiai Zhu, Lei Pang, and Abdulmotaleb El Saddik. 2016. Sentiment analysis on multi-view social data. In International Conference on Multimedia Modeling, pages 15–27. Springer. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference 338 on empirical methods in natural language processing (EMNLP), pages 1532–1543. Ver´onica P´erez-Rosas, Rada Mihalcea, and LouisPhilippe Morency. 2013. Utterance-level multimodal sentiment analysis. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 973–982. Mohammad Soleymani, David Garcia, Brendan Jou, Bj¨orn Schuller, Shih-Fu Chang, and Maja Pantic. 2017. A survey of multimodal sentiment analysis. Image and Vision Computing, 65:3–14. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, volume 30, pages 5998–6008. Youze Wang, Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. 2020a. Fake news detection via knowledge-driven multimodal graph convolutional networks. In Proceedings of the 2020 International Conference on Multimedia Retrieval, pages 540–547. Zilong Wang, Zhaohong Wan, and Xiaojun Wan. 2020b. Transmodality: An end2end fusion method with transformer for multimodal sentiment analysis. In Proceedings of The Web Conference 2020, pages 2514–2520. Nan Xu. 2017. Analyzing multimodal public sentiment based on hierarchical semantic attentional network. In 2017 IEEE International Conference on Intelligence and Security Informatics (ISI), pages 152–154. IEEE. Nan Xu and Wenji Mao. 2017. Multisentinet: A deep semantic network for multimodal sentiment analysis. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2399–2402. ACM. Nan Xu, Wenji Mao, and Guandan Chen. 2018. A comemory network for multimodal sentiment analysis. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 929–932. Xiaocui Yang, Shi Feng, Daling Wang, and Yifei Zhang. 2020. Image-text multimodal emotion classification via multi-view attentional network. IEEE Transactions on Multimedia. Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Graph convolutional networks for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7370–7377. Quanzeng You, Jiebo Luo, Hailin Jin, and Jianchao Yang. 2016. Cross-modality consistent regression for joint visual-textual sentiment analysis of social multimedia. In Proceedings of the Ninth ACM international conference on Web search and data mining, pages 13–22. ACM. Lin Yue, Weitong Chen, Xue Li, Wanli Zuo, and Minghao Yin. 2018. A survey of sentiment analysis in social media. Knowledge and Information Systems, pages 1–47. Dong Zhang, Shoushan Li, Qiaoming Zhu, and Guodong Zhou. 2020. Multi-modal sentiment classification with independent and interactive knowledge via semi-supervised learning. IEEE Access, 8:22945–22954. Lei Zhang, Shuai Wang, and Bing Liu. 2018. Deep learning for sentiment analysis: A survey. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 8(4):e1253. Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2017. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452–1464. Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. 2016. Attention-based bidirectional long short-term memory networks for relation classification. In Proceedings of the 54th annual meeting of the association for computational linguistics (volume 2: Short papers), pages 207– 212. A Dataset A.1 MVSA-Single and MVSA-Multiple The statistics for the MVSA-Simple and MVSAMultiple datasets are listed in Table 1, showing that the various categories are highly unbalanced. MVSA-Single and MVSA-Multiple have different data distributions. Dataset Sentiment Train Val Test All MVSASimple Positive 2,146 268 269 2,683 Neutral 376 47 47 470 Negative 1,086 136 136 1,358 All 3,608 451 452 4,511 MVSAMultiple Positive 9,054 1,132 1,132 11,318 Neutral 3,526 441 441 4,408 Negative 1,038 130 130 1,298 All 13,618 1,703 1,703 17,024 Table 6: Number of Instances for Each Sentiment on the MVSA-∗Dataset. 339 Emotion Train Val Test All Angry 11,635 1,454 1,455 14,544 Bored 25,826 3,228 3,229 32,283 Calm 14,487 1,811 1,811 18,109 Fearful 16,211 2,026 2,027 20,264 Happy 40,214 5,027 5,026 50,267 Loving 27,609 3,451 3,451 34,511 Sad 20,222 2,528 2,527 25,277 All 156,204 19,525 19,536 195,265 Table 7: Number of Instances of Each Emotion on the TumEmo Dataset. A.2 TumEmo The statistics for the TumEmo dataset are listed in Table 2, containing a large number of image-text posts labeled by emotion. B Preprocessing Data The text data contain many useless characters for sentiment analysis, such as URLs, stopwords, and punctuation. We need to preprocess text data to enhance the effectiveness of multimodal emotion detection. We perform data preprocessing as follows: • remove the “URL”, as in“http://...”; • remove the stopwords, such as “a, an, the, and etc. ”; • remove the useless punctuation, including periods, commas, semicolons, etc; • remove the hashtag and its content (#content); In particular, the TumEmo dataset uses #emotion as a weakly supervised label. • remove the posts for which the text length is less than 3.
|
2021
|
28
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3609–3624 August 1–6, 2021. ©2021 Association for Computational Linguistics 3609 BERT is to NLP what AlexNet is to CV: Can Pre-Trained Language Models Identify Analogies? Asahi Ushio, Luis Espinosa-Anke, Steven Schockaert, Jose Camacho-Collados Cardiff NLP, School of Computer Science and Informatics Cardiff University, United Kingdom {UshioA,Espinosa-AnkeL,SchockaertS1,CamachoColladosJ}@cardiff.ac.uk Abstract Analogies play a central role in human commonsense reasoning. The ability to recognize analogies such as “eye is to seeing what ear is to hearing”, sometimes referred to as analogical proportions, shape how we structure knowledge and understand language. Surprisingly, however, the task of identifying such analogies has not yet received much attention in the language model era. In this paper, we analyze the capabilities of transformer-based language models on this unsupervised task, using benchmarks obtained from educational settings, as well as more commonly used datasets. We find that off-the-shelf language models can identify analogies to a certain extent, but struggle with abstract and complex relations, and results are highly sensitive to model architecture and hyperparameters. Overall the best results were obtained with GPT-2 and RoBERTa, while configurations using BERT were not able to outperform word embedding models. Our results raise important questions for future work about how, and to what extent, pre-trained language models capture knowledge about abstract semantic relations.1 1 Introduction One of the most widely discussed properties of word embeddings has been their surprising ability to model certain types of relational similarities in terms of word vector differences (Mikolov While the title is probably self-explanatory, this is a small note explaining it. BERT is to NLP what AlexNet is to CV is making an analogy on what the BERT and AlexNet models represented for Natural Language Processing (NLP) and Computer Vision (CV), respectively. They both brought a paradigm shift in how research was undertaken in their corresponding disciplines and this is what the analogy refers to. 1Source code and data to reproduce our experimental results are available in the following repository: https://github.com/asahi417/ analogy-language-model Query: word:language Candidates: (1) paint:portrait (2) poetry:rhythm (3) note:music (4) tale:story (5) week:year Table 1: An example analogy task from the SAT dataset. The third candidate is the answer to the query. et al., 2013a; Vylomova et al., 2016; Allen and Hospedales, 2019; Ethayarajh et al., 2019). The underlying assumption is that when “a is to b what c is to d” the word vector differences b −a and d −c are expected to be similar, where we write x for the embedding of a word x. While this assumption holds for some types of syntactic relations, for semantic relations this holds to a much more limited degree than was suggested in early work (Linzen, 2016; Schluter, 2018). Moreover, the most commonly used benchmarks have focused on specific and well-defined semantic relations such as “capital of”, rather than the more abstract notion of relational similarity that is often needed for solving the kind of psychometric analogy problems that can be found in IQ tests and educational settings. An example of such a problem is shown in Table 1. Given the central role of analogy in human cognition, it is nonetheless important to understand the extent to which NLP models are able to solve these more abstract analogy problems. Besides its value as an intrinsic benchmark for lexical semantics, the ability to recognize analogies is indeed important in the contexts of human creativity (Holyoak et al., 1996), innovation (Hope et al., 2017), computational creativity (Goel, 2019) and education (Pardos and Nam, 2020). Analogies are also a prerequisite to build AI systems for the legal domain (Ashley, 1988; Walton, 2010) and are used in machine learning (Miclet et al., 2008; Hug et al., 3610 2016; H¨ullermeier, 2020) and for ontology alignment (Raad and Evermann, 2015), among others. Within NLP, however, the task of recognizing analogies has received relatively little attention. To solve such problems, Turney (2005) proposed Latent Relational Analysis (LRA), which was essentially designed as a relational counterpart to Latent Semantic Analysis (Landauer and Dumais, 1997). Somewhat surprisingly, perhaps, despite the substantial progress that word embeddings and language models (LMs) have enabled in NLP, LRA still represents the current state-of-the-art in solving abstract word analogy problems. When going beyond a purely unsupervised setting, however, GPT-3 was recently found to obtain slightly better results (Brown et al., 2020). The aim of this paper is to analyze the ability of pre-trained LMs to recognize analogies. Our focus is on the zero-shot setting, where LMs are used without fine-tuning. To predict whether two word pairs (a, b) and (c, d) are likely to be analogical, we need a prompt, i.e. a template that is used to construct the input to the LM, and a scoring function. We extensively analyze the impact of both of these choices, as well as the differences between different LMs. When the prompt and scoring function are carefully calibrated, we find that GPT-2 can outperform LRA, standard word embeddings as well as the published results for GPT-3 in the zero-shot setting. However, we also find that these results are highly sensitive to the choice of the prompt, as well as two hyperparameters in our scoring function, with the optimal choices not being consistent across different datasets. Moreover, using BERT leads to considerably weaker results, underperforming even standard word embeddings in all of the considered configurations. These findings suggest that while transformer-based LMs learn relational knowledge to a meaningful extent, more work is needed to understand how such knowledge is encoded, and how it can be exploited. 2 Related work 2.1 Understanding Pre-trained LMs Since their recent dominance in standard NLP benchmarks (Peters et al., 2018a; Devlin et al., 2019; Liu et al., 2019), pre-trained language models have been extensively studied. This has mainly been done through probing tasks, which are aimed at understanding the knowledge that is implicitly captured by their parameters. After the initial focus on understanding pre-trained LSTM-based LMs (Peters et al., 2018b), attention has now shifted toward transformer-based models. The main aspects that have been studied in recent years are syntax (Goldberg, 2019; Saphra and Lopez, 2019; Hewitt and Manning, 2019; van Schijndel et al., 2019; Jawahar et al., 2019; Tenney et al., 2019b) and semantics (Ettinger, 2019; Tenney et al., 2019a). For a more complete overview on analyses of the different properties of transformer-based LMs, we refer to Rogers et al. (2021). Despite the rise in probing analyses for LMs and the importance of analogical reasoning in human cognition, understanding the analogical capabilities of LMs remains understudied. The most similar works have focused on capturing relational knowledge from LMs (in particular the type of information available in knowledge graphs). For instance, Petroni et al. (2019) analyzed to what extent LMs could fill manually-defined templates such as “Dante was born in [MASK]”. Follow-up works extended this initial approach by automatically generating templates and fine-tuning LMs on them (Bouraoui et al., 2020; Jiang et al., 2020), showing an improved performance. In this paper, we focus on the analogical knowledge that is encoded in pre-trained LMs, without the extra step of fine-tuning on additional data. 2.2 Word Analogy Probing Word analogies have been used as a standard intrinsic evaluation task for measuring the quality of word embeddings. Mikolov et al. (2013b) showed that word embeddings, in particular Word2vec embeddings, were able to solve analogy problems by simple vector operations (e.g. king - man + woman = queen). The motivation for this task dates back to the connectionism theory (Feldman and Ballard, 1982) in cognitive science. In particular, neural networks were thought to be able to model emergent concepts (Hopfield, 1982; Hinton, 1986) by learning distributed representations across an embedding space (Hinton et al., 1986), similar to the properties that word embeddings displayed in the analogy task. More recent works have proposed new mathematical theories and experiments to understand the analogical capabilities of word embeddings, attempting to understand their linear algebraic structure (Arora et al., 2016; Gittens et al., 2017; Allen and Hospedales, 2019) or by explicitly studying their compositional nature (Levy and 3611 Goldberg, 2014; Paperno and Baroni, 2016; Ethayarajh et al., 2019; Chiang et al., 2020). However, recent works have questioned the impressive results displayed by word embeddings in this task. In many cases simple baselines excluding the input pair (or query) were competitive (Linzen, 2016). Simultaneously, some researchers have found that many relationships may not be retrieved in the embedding space by simple linear transformations (Drozd et al., 2016; Bouraoui et al., 2018) and others argued that the standard evaluation procedure has limitations (Schluter, 2018). New datasets and measures have also been introduced to address some of these issues (Gladkova et al., 2016; Fournier et al., 2020). Finally, in the context of bias detection, for which analogies have been used as a proxy (Bolukbasi et al., 2016), it has also been found that word analogies may misguide or hide the real relationships existing in the vector space (Gonen and Goldberg, 2019; Nissim et al., 2020). As far as language models are concerned, word analogies have not been explored to the same extent as for word embeddings. Recently, Brown et al. (2020) evaluated the unsupervised capabilities of GPT-3 by evaluating it on the SAT analogies dataset (Turney et al., 2003), which we also include in our evaluation (see Section 3.2). However, the evaluation is limited to a single dataset (i.e., SAT) and model (i.e., GPT-3), and the general capabilities of language models were not investigated. Despite their limitations, analogy tests remain appealing for evaluating the ability of embeddings and language models to identify abstract relationships. To mitigate the aforementioned methodological issues, in this work we rely on analogy tests from educational resources, where the task is to complete analogical proportions, given only the first word pair. In contrast, word embedding models have mostly been evaluated using a predictive task, in which three of the four words are given. Moreover, the considered datasets are focused on abstract analogies, whereas the most commonly used datasets only include well-defined semantic relations such as “capital of”. For completeness, however, we also show results on these standard datasets. We furthermore experiment with several simple baselines to understand possible artifacts present in the different datasets. 3 Word Analogies In this section, we describe the word analogy formulation that is used for our experiments (Section 3.1). Subsequently, we provide an overview of the datasets used in our experiments (Section 3.2). 3.1 Task Description We frame the analogy task in terms of analogical proportions (Prade and Richard, 2017). Given a query word pair (hq, tq) and a list of candidate answer pairs {(hi, ti)}n i=1, the goal is to find the candidate answer pair that has the most similar relation to the query pair. Table 1 shows a sample query and candidate answers drawn from one of the datasets used in our evaluation (see Section 3.2). 3.2 Analogy Datasets We split analogy datasets in two types, based on how the analogy problems were constructed. 3.2.1 Psychometric Analogy Tests Word analogy tests are commonly used in assessments of linguistic and cognitive ability. For instance, in the past, such tests were included in the SAT exams, which are a US college admission test. Turney et al. (2003) collected a benchmark of 374 word analogy problems, consisting primarily of problems from these SAT tests. Aimed at college applicants, these problems are designed to be challenging for humans. A key challenge for NLP systems is that solving these problems often requires identifying fine-grained semantic differences between word pairs that belong to the same coarse-grained relation. For instance, in the case of Table 1, we could say that “a year consists of weeks” like “language consists of words”, but the week-year pair is nonetheless less similar to wordlanguage than note-music. Another analogy benchmark was constructed by Boteanu and Chernova (2015), who used word analogy problems from an educational resource2. They used in particular UNIT 2 of the analogy problems from the educational site. These problems have the same form as those from the SAT benchmark, but rather than college applicants, they are aimed at children in grades 4 to 12 from the US school system (i.e. from age 9 onwards). In this paper, we will also include this UNIT 2 benchmark. Moreover, we have collected another benchmark from 2https://www.englishforeveryone.org/ Topics/Analogies.html 3612 Dataset Data size No. No. (val / test) candidates groups SAT 37 / 337 5 2 UNIT 2 24 / 228 5,4,3 9 UNIT 4 48 / 432 5,4,3 5 Google 50 / 500 4 2 BATS 199 / 1799 4 3 Table 2: High-level statistics of the analogy datasets after unification: data size, number of candidates and number of group partitions. the UNIT 4 problems on the same website. These UNIT 4 problems are organised in 5 difficulty levels: high-beginning, low-intermediate, highintermediate, low-advanced and high-advanced. The low-advanced level is stated to be at the level of the SAT tests, whereas the high-advanced level is stated to be at the level of the GRE test (which is used for admission into graduate schools). 3.2.2 Lexical Semantics Benchmarks Since the introduction of Word2vec (Mikolov et al., 2013a), the problem of modelling analogies has been commonly used as an intrinsic benchmark for word embedding models. However, the datasets that have been used in that context are focused on well-defined and relatively coarse-grained relations. The Google analogy dataset (Mikolov et al., 2013b) has been one of the most commonly used benchmarks for intrinsic evaluation of word embeddings. This dataset contains a mix of semantic and morphological relations such as capital-of and singular-plural, respectively. However, its coverage has been shown to be limiting, and BATS (Gladkova et al., 2016) was developed in an attempt to address its main shortcomings. BATS includes a larger number of concepts and relations, which are split into four categories: lexicographic, encyclopedic, and derivational and inflectional morphology. As pointed out above, these datasets were tailored to the evaluation of word embeddings in a predictive setting. To provide an evaluation setting which is comparable to the benchmarks obtained from human analogy tests, we constructed word analogy problems from the Google and BATS datasets, by choosing for each correct analogy pair a number of negative examples. The resulting benchmark thus follows the same format as described in Section 3.1. To obtain sufficiently challenging negative examples, for each query pair (e.g. Paris-France) we extracted three negative inFigure 1: Solving a word analogy problem by selecting one with the highest LM score among the candidates. stances: (1) two random words from the head of the input relation type (e.g. Rome-Oslo); (2) two random words from the tail of the input relation type (e.g. Germany-Canada); (3) a random word pair from a relation type of the same high-level category as the input relation type (e.g. Argentina-peso).3 3.2.3 Unification and Statistics Table 2 provides an overview of our datasets. The instances from each dataset are organised into groups. In the case of Google and BATS, these groups refer to the relation types (e.g. semantic or morphological in the case of Google). In the case of UNIT 2 and UNIT 4, the groups refer to the difficulty level. For the SAT dataset, we consider two groups, capturing whether the instances come from an actual SAT test or not. Finally, we randomly sample 10% of each group in each dataset to construct a validation set, and regard the remaining data as the test set. 4 Methodology In this section, we explain our strategy for using pretrained LMs to solve analogy problems without fine-tuning. First, in Section 4.1 we explain how each relation pair is converted into a natural sentence to be fed into the LM. In Section 4.2, we then discuss a number of scoring functions that can be used to select the most plausible answer candidate. Finally, we take advantage of the fact that analogical proportion is invariant to particular permutations, which allows for a natural extension of the proposed scoring functions (Section 4.3). Figure 1 shows a high-level overview of our methodology. 4.1 Relation Pair Prompting We define a prompting function Tt(w1, w2, w3, w4) that takes four placeholders and a template type t, 3In order to avoid adding various correct answers to the query, we avoided adding negative pairs from all country-of type relations, and from similar lexicographic relations in the BATS dataset with more than one relation type, namely antonyms, synonyms, meronyms and hyponyms. 3613 and returns a sentence in which the placeholders were replaced by the words w1, w2, w3, and w4. For instance, given a query “word:language” and a candidate “note:music”, the prompting function produces Tto-as(“word”, “language”, “note”, “music”) = “word is to language as note is to music” where we use the template type to-as here. Using manually specified template types can result in a sub-optimal textual representation. For this reason, recent studies have proposed autoprompting strategies, which optimize the template type on a training set (Shin et al., 2020), paraphrasing (Jiang et al., 2020), additional prompt generation model (Gao et al., 2020), and corpus-driven template mining (Bouraoui et al., 2020). However, none of these approaches can be applied to unsupervised settings. Thus, we do not explore auto-prompting methods in this work. Instead, we will consider a number of different template types in the experiments, and assess the sensitivity of the results to the choice of template type. 4.2 Scoring Function Perplexity. We first define perplexity, which is widely used as a sentence re-ranking metric (Chan et al., 2016; Gulcehre et al., 2015). Given a sentence x, for autoregressive LMs such as LSTM based models (Zaremba et al., 2014) and GPTs (Radford et al., 2018, 2019; Brown et al., 2020), perplexity can be computed as f(x) = exp − m X j=1 log Pauto(xj|xj−1) (1) where x is tokenized as [x1...xm] and Pauto(x|x) is the likelihood from an autoregressive LM’s next token prediction. For masked LMs such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), we instead use pseudoperplexity, which is defined as in (1) but with Pmask(xj|x\j) instead of Pauto(xj|xj−1), where x\j = [x1 . . . xj1〈mask〉xj+1 . . . xm] and Pmask(xj|x\j) is the pseudo-likelihood (Wang and Cho, 2019) that the masked token is xj. PMI. Although perplexity is well-suited to capture the fluency of a sentence, it may not be the best choice to test the plausibility of a given analogical proportion candidate. As an alternative, we propose a scoring function that focuses specifically Figure 2: Positive and negative permutations for a relation pair (a:b)-(c:d). on words from the two given pairs. To this end, we propose to use an approximation of point-wise mutual information (PMI), based on perplexity. PMI is defined as the difference between a conditional and marginal log-likelihood. In our case, we consider the conditional likelihood of ti given hi and the query pair (recall from Section 3.1 that h and t represent the head and tail of a given word pair, respectively), i.e. P(ti|hq, tq, hi), and the marginal likelihood over hi, i.e. P(ti|hq, tq). Subsequently, the PMI-inspired scoring function is defined as r(ti|hi, hq, tq) = log P(ti|hi, hq, tq) −α · log P(ti|hq, tq) (2) where α is a hyperparameter to control the effect of the marginal likelihood. The PMI score corresponds to the specific case where α = 1. However, Davison et al. (2019) found that using a hyperparameter to balance the impact of the conditional and marginal probabilities can significantly improve the results. The probabilities in (2) are estimated by assuming that the answer candidates are the only possible word pairs that need to be considered. By relying on this closed-world assumption, we can estimate marginal probabilities based on perplexity, which we found to give better results than the masking based strategy from Davison et al. (2019). In particular, we estimate these probabilities as P(ti|hq, tq, hi) = − f (Tt(hq, tq, hi, ti)) nP k=1 f (Tt(hq, tq, hi, tk)) P(ti|hq, tq) = − nP k=1 f (Tt(hq, tq, hk, ti)) nP k=1 nP l=1 f (Tt(hq, tq, hk, tl)) 3614 where n is the number of answer candidates for the given query. Equivalently, since PMI is symmetric, we can consider the difference between the logs of P(hi|hq, tq, ti) and P(hi|hq, tq). While this leads to the same PMI value in theory, due to the way in which we approximate the probabilities, this symmetric approach will lead to a different score. We thus combine both scores with an aggregation function Ag. This aggregation function takes a list of scores and outputs an aggregated value. As an example, given a list [1, 2, 3, 4], we write Amean([1, 2, 3, 4]) = 2.5 for the mean and Aval1([1, 2, 3, 4]) = 1 for the first element. Given such an aggregation function, we define the following PMI-based score sPMI(ti, hi|hq, tq) = Ag (r) (3) where we consider basic aggregation operations over the list r = [r(ti|hi, hq, tq), r(hi|ti, hq, tq)], such as the mean, max, and min value. The choice of using only one of the scores r(ti|hi, hq, tq), r(hi|ti, hq, tq) is viewed as a special case, in which the aggregation function g simply returns the first or the second item. mPPL. We also experiment with a third scoring function, which borrows ideas from both perplexity and PMI. In particular, we propose the marginal likelihood biased perplexity (mPPL) defined as smPPL(ti, hi|hq, tq) = log sPPL(ti, hi|hq, tq) −αt · log P(ti|hq, tq) −αh · log P(hi|hq, tq) where αt and αh are hyperparameters, and sPPL is a normalized perplexity defined as sPPL(ti, hi|hq, tq) = − f (Tt(hq, tq, hi, ti)) nP k=1 f (Tt(hq, tq, hk, tk)) . The mPPL score extends perplexity with two bias terms. It is motivated from the insight that treating α as a hyperparameter in (2) can lead to better results than fixing α = 1. By tuning αt and αh, we can essentially influence to what extent answer candidates involving semantically similar words to the query pair should be favored. 4.3 Permutation Invariance The formalization of analogical proportions dates back to Aristotle (Barbot et al., 2019). According to the standard axiomatic characterization, whenever we have an analogical proportion a : b :: c : d (meaning “a is to b what c is to d”), it also holds that c : d :: a : b and a : c :: b : d are analogical proportions. It follows from this that for any given analogical proportion a : b :: c : d there are eight permutations of the four elements a, b, c, d that form analogical proportions. These eight permutations, along with the 16 “negative permutations”, are shown in Figure 2. To take advantage of the different permutations of analogical proportions, we propose the following Analogical Proportion (AP) score: AP(hq, tq, hi, ti) = Agpos(p) −β · Agneg(n) (4) p = [s(a, b|c, d)](a:b,c:d)∈P n = [s(a, b|c, d)](a:b,c:d)∈N where P and N correspond to the list of positive and negative permutations of the candidate analogical proportion hq : tq :: hi : ti in the order shown in Figure 2, β is a hyperparameter to control the impact of the negative permutations, and s(a, b|c, d) is a scoring function as described in Section 4.2. Here Agpos and Agneg refer to the aggregation functions that are used to combine the scores for the positive and negative permutations respectively, where these aggregation functions are defined as in Section 4.2. To solve an analogy problem, we simply choose the answer candidate that results in the highest value of AP(ti, hi, hq, tq). 5 Evaluation In this section, we evaluate language models on the five analogy datasets presented in Section 3. 5.1 Experimental Setting We consider three transformer-based LMs of a different nature: two masked LMs, namely BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), and GPT-2, as a prominent example of an autoregressive language model. Each pretrained model was fetched from the Huggingface transformers library (Wolf et al., 2019), from which we use bert-large-cased, roberta-large, and gpt2-xl respectively. For parameter selection, we run grid search on β, α, αh, αt, t, g, gpos, and gneg for each model and select the configuration which achieves the best accuracy on each validation set. We experiment with the three scoring functions presented in Section 4.2, i.e., sPPL (perplexity), 3615 Model Score Tuned SAT U2 U4 Google BATS Avg LM BERT sPPL 32.9 32.9 34.0 80.8 61.5 48.4 ✓ 39.8 41.7 41.0 86.8 67.9 55.4 sPMI 27.0 32.0 31.2 74.0 59.1 44.7 ✓ 40.4 42.5 27.8 87.0 68.1 53.2 smPPL ✓ 41.8 44.7 41.2 88.8 67.9 56.9 GPT-2 sPPL 35.9 41.2 44.9 80.4 63.5 53.2 ✓ 50.4 48.7 51.2 93.2 75.9 63.9 sPMI 34.4 44.7 43.3 62.8 62.8 49.6 ✓ 51.0 37.7 50.5 91.0 79.8 62.0 smPPL ✓ 56.7 50.9 49.5 95.2 81.2 66.7 RoBERTa sPPL 42.4 49.1 49.1 90.8 69.7 60.2 ✓ 53.7 57.0 55.8 93.6 80.5 68.1 sPMI 35.9 42.5 44.0 60.8 60.8 48.8 ✓ 51.3 49.1 38.7 92.4 77.2 61.7 smPPL ✓ 53.4 58.3 57.4 93.6 78.4 68.2 WE FastText 47.8 43.0 40.7 96.6 72.0 60.0 GloVe 47.8 46.5 39.8 96.0 68.7 59.8 Word2vec 41.8 40.4 39.6 93.2 63.8 55.8 Base PMI 23.3 32.9 39.1 57.4 42.7 39.1 Random 20.0 23.6 24.2 25.0 25.0 23.6 Table 3: Accuracy results on each analogy dataset, categorized into language models (LM), word embeddings (WE), and baselines (Base). All LMs use the analogical proportion (AP) function described in Section 4.3. The default configuration for AP includes α = αh = αt = β = 0, gpos = g = val1, and t = to-as. Note that sPPL = smPPL with the default configuration. Average accuracy (Avg) across datasets is included in the last column. sPMI and smPPL. Possible values for each hyperparameter (including the selection of six prompts and an ablation test on the scoring function) and the best configurations that were found by grid search are provided in the appendix. As baseline methods, we also consider three pre-trained word embedding models, which have been shown to provide competitive results in analogy tasks, as explained in Section 2.2: Word2vec (Mikolov et al., 2013a), GloVe (Pennington et al., 2014), and FastText (Bojanowski et al., 2017). For the word embedding models, we simply represent word pairs by taking the difference between their embeddings4. We then choose the answer candidate with the highest cosine similarity to the query in terms of this vector difference. To put the results into context, we also include two simple statistical baselines. First, we report the expected random performance. Second, we use a method based on each word pair’s PMI in a given corpus. We then select the answer candidate with the highest 4Vector differences have been found to be the most robust encoding method in the context of word analogies (Hakami and Bollegala, 2017). PMI as the prediction. Note that the query word pair is completely ignored in this case. This PMI score is the well-known word-pair association metric introduced by Church and Hanks (1990) for lexicographic purposes (specifically, collocation extraction), which compares the probability of observing two words together with the probabilities of observing them independently (chance). The PMI scores in our experiments were computed using the English Wikipedia with a fixed window size 10. 5.2 Results Table 3 shows our main results. As far as the comparison among LMs is concerned, RoBERTa and GPT-2 consistently outperform BERT. Among the AP variants, smPPL achieves substantially better results than sPMI or sPPL in most cases. We also observe that word embeddings perform surprisingly well, with FastText and GloVe outperforming BERT on most datasets, as well as GPT-2 and RoBERTa with default hyperparameters. FastText achieves the best overall accuracy on the Google dataset, confirming that this dataset is particularly well-suited to word embeddings (see Section 2.2). 3616 Model Score Tuned Accuracy LM BERT sPPL 32.6 ✓ 40.4* sPMI 26.8 ✓ 41.2* smPPL ✓ 42.8* GPT-2 sPPL 41.4 ✓ 56.2* sPMI 34.7 ✓ 56.8* sPPL ✓ 57.8* RoBERTa sPPL 49.6 ✓ 55.8* sPMI 42.5 ✓ 54.0* smPPL ✓ 55.8* GPT-3 Zero-shot 53.7 Few-shot ✓ 65.2* LRA 56.4 WE FastText 49.7 GloVe 48.9 Word2vec 42.8 Base PMI 23.3 Random 20.0 Table 4: Accuracy results for the full SAT dataset. Results marked with * are not directly comparable as they were tuned on full data (for our models) or use training data (for GPT-3 few-shot). These results are included to provide an upper bound only. Results in italics were taken from the original papers. In order to compare with published results from prior work, we carried out an additional experiment on the full SAT dataset (i.e., without splitting it into validation and test). Table 4 shows the results. GPT3 (Brown et al., 2020) and LRA (Turney, 2005) are added for comparison. Given the variability of the results depending on the tuning procedure, we have also reported results of configurations that were tuned on the entire set, to provide an upper bound on what is possible within the proposed unsupervised setting. This result shows that even with optimal hyperparameter values, LMs barely outperform the performance of the simpler LRA model. GPT-3 similarly fails to outperform LRA in the zero-shot setting. 6 Analysis We now take a closer look into our results to investigate parameter sensitivity, the correlation between model performance and human difficulty levels, and possible dataset artifacts. The following analysis focuses on smPPL as it achieved the best results among the LM based scoring functions. Figure 3: Box plot of the relative improvement on test accuracy in each dataset over all configurations of smPPL grouped by gpos. Here valk corresponds to kth positive permutation shown in Figure 2. Parameter Sensitivity We found that optimal values of the parameters α and β are highly dependent on the dataset, while other parameters such as the template type t vary across LMs. On the other hand, as shown in Figure 3, the optimal permutations of the templates are relatively consistent, with the original ordering a : b :: c : d typically achieving the best results. The results degrade most for permutations that mix the two word pairs (e.g. a : c :: b : d). In the appendix we include an ablation study for the sensitivity and relevance of other parameters and design choices. Difficulty Levels To increase our understanding of what makes an analogy problem difficult for LMs, we compare the results for each difficulty level.5 Recall from Section 3.2 that the U2 and U4 datasets come from educational resources and are split by difficulty level. Figure 4 shows the results of all LMs (tuned setting), FastText and the PMI baseline according to these difficulty levels. Broadly speaking, we can see that instances that are harder for humans are also harder for the considered models. The analogies in the most difficult levels are generally more abstract (e.g. witness : testimony :: generator : electricity), or contain obscure or infrequent words (e.g. grouch : cantakerous :: palace : ornate).6 5For SAT, Google and BATS, there are no difficulty levels available, but we show the results split by high-level categories in the appendix. We also note that the number of candidates in U2 and U4 vary from three to five, so results per difficulty level are not fully comparable. However, they do reflect the actual difficulty of the educational tests. 6In the appendix we include more examples with errors made by RoBERTa in easy instances. 3617 Figure 4: Test accuracy in U2 and U4 per difficulty level. LMs use smPPL with the best configuration tuned in the corresponding validation sets. Hypothesis Only Recently, several researchers have found that standard NLP benchmarks, such as SNLI (Bowman et al., 2015) for language inference, contain several annotation artifacts that makes the task simpler for automatic models (Poliak et al., 2018; Gururangan et al., 2018). One of their most relevant findings is that models which do not even consider the premise can reach high accuracy. More generally, these issues have been found to be problematic in NLP models (Linzen, 2020) and neural networks more generally (Geirhos et al., 2020). According to the results shown in Table 3, we already found that the PMI baseline achieved a non-trivial performance, even outperforming BERT in a few settings and datasets. This suggests that several implausible negative examples are included in the analogy datasets. As a further exploration of such artifacts, here we analyse the analogue of a hypothesis-only baseline. In particular, for this analysis, we masked the head or tail of the candidate answer in all evaluation instances. Then, we test the masked language models with the same AP conMask SAT U2 U4 Google BATS BERT full 41.8 44.7 41.2 88.8 67.9 head 31.8 28.1 34.3 72.0 62.4 tail 33.5 31.6 38.2 64.2 63.1 RoBERTa full 53.4 58.3 57.4 93.6 78.4 head 38.6 37.7 41.0 60.6 54.5 tail 35.6 37.3 40.5 55.8 64.2 Table 5: Accuracy results by masking head or tail of the candidate answers. Results in the top row correspond to the full model without masking. figuration and tuning on these artificially-modified datasets.As can be seen in Table 5, a non-trivial performance is achieved for all datasets, which suggests that the words from the answer pair tend to be more similar to the words from the query than the words from negative examples. 7 Conclusion In this paper, we have presented an extensive analysis of the ability of language models to identify analogies. To this end, we first compiled datasets with psychometric analogy problems from educational resources, covering a wide range of difficulty levels and topics. We also recast two standard benchmarks, the Google and BATS analogy datasets, into the same style of problems. Then, we proposed standard techniques to apply language models to the unsupervised task of solving these analogy problems. Our empirical results shed light on the strengths and limitations of various models. To directly answer the question posed in the title, our conclusion is that language models can identify analogies to a certain extent, but not all language models are able to achieve a meaningful improvement over word embeddings (whose limitations in analogy tasks are well documented). On the other hand, when carefully tuned, some language models are able to achieve state-of-the-art results. We emphasize that results are highly sensitive to the chosen hyperparameters (which define the scoring function and the prompt among others). Further research could focus on the selection of these optimal hyperparameters, including automatizing the search or generation of prompts, along the lines of Bouraoui et al. (2020) and Shin et al. (2020), respectively. Finally, clearly LMs might still be able to learn to solve analogy tasks when given appropriate training data, which is an aspect that we leave for future work. 3618 References Carl Allen and Timothy Hospedales. 2019. Analogies explained: Towards understanding word embeddings. In International Conference on Machine Learning, pages 223–231. Sanjeev Arora, Yuanzhi Li, Yingyu Liang, Tengyu Ma, and Andrej Risteski. 2016. A latent variable model approach to pmi-based word embeddings. Transactions of the Association for Computational Linguistics, 4:385–399. Kevin D Ashley. 1988. Arguing by analogy in law: A case-based model. In Analogical reasoning, pages 205–224. Springer. Nelly Barbot, Laurent Miclet, and Henri Prade. 2019. Analogy between concepts. Artificial Intelligence, 275:487–539. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association of Computational Linguistics, 5(1):135–146. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, pages 4349–4357. Adrian Boteanu and Sonia Chernova. 2015. Solving and explaining analogy questions using semantic networks. In Proceedings of the AAAI Conference on Artificial Intelligence. Zied Bouraoui, Jose Camacho-Collados, and Steven Schockaert. 2020. Inducing relational knowledge from bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7456– 7463. Zied Bouraoui, Shoaib Jameel, and Steven Schockaert. 2018. Relation induction in word embeddings revisited. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1627– 1637, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Annual Conference on Neural Information Processing Systems. William Chan, Navdeep Jaitly, Quoc Le, and Oriol Vinyals. 2016. Listen, attend and spell: A neural network for large vocabulary conversational speech recognition. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4960–4964. IEEE. Hsiao-Yu Chiang, Jose Camacho-Collados, and Zachary Pardos. 2020. Understanding the source of semantic regularities in word embeddings. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 119–131, Online. Association for Computational Linguistics. Kenneth Church and Patrick Hanks. 1990. Word association norms, mutual information, and lexicography. Computational linguistics, 16(1):22–29. Joe Davison, Joshua Feldman, and Alexander M Rush. 2019. Commonsense knowledge mining from pretrained models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1173– 1178. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Aleksandr Drozd, Anna Gladkova, and Satoshi Matsuoka. 2016. Word embeddings, analogies, and machine learning: Beyond king-man+ woman= queen. In Proceedings of coling 2016, the 26th international conference on computational linguistics: Technical papers, pages 3519–3530. Kawin Ethayarajh, David Duvenaud, and Graeme Hirst. 2019. Towards understanding linear word analogies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3253–3262. Allyson Ettinger. 2019. What bert is not: Lessons from a new suite of psycholinguistic diagnostics for language models. Transactions of the Association for Computational Linguistics, 8:34–48. Jerome A. Feldman and Dana H. Ballard. 1982. Connectionist models and their properties. Cognitive Science, 6(3):205–254. 3619 Louis Fournier, Emmanuel Dupoux, and Ewan Dunbar. 2020. Analogies minus analogy test: measuring regularities in word embeddings. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 365–375, Online. Association for Computational Linguistics. Tianyu Gao, Adam Fisch, and Danqi Chen. 2020. Making pre-trained language models better few-shot learners. arXiv preprint arXiv:2012.15723. Robert Geirhos, J¨orn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. 2020. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665–673. Alex Gittens, Dimitris Achlioptas, and Michael W Mahoney. 2017. Skip-gram- zipf+ uniform= vector additivity. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 69–76. Anna Gladkova, Aleksandr Drozd, and Satoshi Matsuoka. 2016. Analogy-based detection of morphological and semantic relations with word embeddings: what works and what doesn’t. In Proceedings of the Student Research Workshop at NAACL, pages 8–15. Ashok Goel. 2019. Computational design, analogy, and creativity. In Computational Creativity, pages 141–158. Springer. Yoav Goldberg. 2019. Assessing bert’s syntactic abilities. arXiv preprint arXiv:1901.05287. Hila Gonen and Yoav Goldberg. 2019. Lipstick on a pig: Debiasing methods cover up systematic gender biases in word embeddings but do not remove them. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 609–614. Caglar Gulcehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Loic Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2015. On using monolingual corpora in neural machine translation. arXiv preprint arXiv:1503.03535. Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. Annotation artifacts in natural language inference data. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 107–112, New Orleans, Louisiana. Association for Computational Linguistics. Huda Hakami and Danushka Bollegala. 2017. Compositional approaches for representing relations between words: A comparative study. KnowledgeBased Systems, 136:172–182. John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138, Minneapolis, Minnesota. Association for Computational Linguistics. Geoffrey E. Hinton. 1986. Learning distributed representations of concepts. In Proceedings of the eighth annual conference of the cognitive science society, volume 1, page 12. Amherst, MA. Geoffrey E. Hinton, James L. McClelland, and David E. Rumelhart. 1986. Distributed representations. Parallel distributed processing: explorations in the microstructure of cognition, vol. 1, pages 77–109. Keith J Holyoak, Keith James Holyoak, and Paul Thagard. 1996. Mental leaps: Analogy in creative thought. MIT press. Tom Hope, Joel Chan, Aniket Kittur, and Dafna Shahaf. 2017. Accelerating innovation through analogy mining. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 235–243. John J. Hopfield. 1982. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8):2554–2558. Nicolas Hug, Henri Prade, Gilles Richard, and Mathieu Serrurier. 2016. Analogical classifiers: a theoretical perspective. In Proceedings of the Twentysecond European Conference on Artificial Intelligence, pages 689–697. Eyke H¨ullermeier. 2020. Towards analogy-based explanations in machine learning. In International Conference on Modeling Decisions for Artificial Intelligence, pages 205–217. Ganesh Jawahar, Benoˆıt Sagot, and Djam´e Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651–3657, Florence, Italy. Association for Computational Linguistics. Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423–438. Thomas K. Landauer and Susan T. Dumais. 1997. A solution to Plato’s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge. Psychological Review, 104(2):211. Omer Levy and Yoav Goldberg. 2014. Linguistic regularities in sparse and explicit word representations. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning, 3620 pages 171–180, Ann Arbor, Michigan. Association for Computational Linguistics. Tal Linzen. 2016. Issues in evaluating semantic spaces using word analogies. In Proceedings of the 1st Workshop on Evaluating Vector-Space Representations for NLP, pages 13–18. Tal Linzen. 2020. How can we accelerate progress towards human-like linguistic generalization? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5210– 5217, Online. Association for Computational Linguistics. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. Laurent Miclet, Sabri Bayoudh, and Arnaud Delhay. 2008. Analogical dissimilarity: definition, algorithms and two experiments in machine learning. Journal of Artificial Intelligence Research, 32:793– 824. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013a. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119. Tomas Mikolov, Wen-tau Yih, and Geoffrey Zweig. 2013b. Linguistic regularities in continuous space word representations. In Proceedings of HLTNAACL, pages 746–751. Malvina Nissim, Rik van Noord, and Rob van der Goot. 2020. Fair is better than sensational: Man is to doctor as woman is to doctor. Computational Linguistics, 46(2):487–497. Denis Paperno and Marco Baroni. 2016. When the whole is less than the sum of its parts: How composition affects pmi values in distributional semantic vectors. Computational Linguistics, 42(2):345–350. Zachary A. Pardos and Andrew J. H. Nam. 2020. A university map of course knowledge. PLoS ONE, 15(9). Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of EMNLP, pages 1532–1543. Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018a. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227–2237, New Orleans, Louisiana. Association for Computational Linguistics. Matthew Peters, Mark Neumann, Luke Zettlemoyer, and Wen-tau Yih. 2018b. Dissecting contextual word embeddings: Architecture and representation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1499–1509, Brussels, Belgium. Association for Computational Linguistics. Fabio Petroni, Tim Rockt¨aschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 2463–2473. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180–191, New Orleans, Louisiana. Association for Computational Linguistics. Henri Prade and Gilles Richard. 2017. Analogical proportions and analogical reasoning-an introduction. In International Conference on Case-Based Reasoning, pages 16–32. Springer. Elie Raad and Joerg Evermann. 2015. The role of analogy in ontology alignment: A study on lisa. Cognitive Systems Research, 33:1–16. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2021. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842–866. Naomi Saphra and Adam Lopez. 2019. Understanding learning dynamics of language models with SVCCA. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3257–3267, Minneapolis, Minnesota. Association for Computational Linguistics. Marten van Schijndel, Aaron Mueller, and Tal Linzen. 2019. Quantity doesn’t buy quality syntax with neural language models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5831–5837, Hong Kong, China. Association for Computational Linguistics. 3621 Natalie Schluter. 2018. The word analogy testing caveat. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 242–246. Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222–4235, Online. Association for Computational Linguistics. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019a. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593– 4601, Florence, Italy. Association for Computational Linguistics. Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R. Thomas McCoy, Najoung Kim, Benjamin Van Durme, Samuel R. Bowman, Dipanjan Das, and Ellie Pavlick. 2019b. What do you learn from context? probing for sentence structure in contextualized word representations. In Proceeding of the 7th International Conference on Learning Representations (ICLR). Peter D. Turney. 2005. Measuring semantic similarity by latent relational analysis. In Proc. of IJCAI, pages 1136–1141. Peter D. Turney, Michael L. Littman, Jeffrey Bigham, and Victor Shnayder. 2003. Combining independent modules in lexical multiple-choice problems. In Recent Advances in Natural Language Processing III, pages 101–110. Ekaterina Vylomova, Laura Rimell, Trevor Cohn, and Timothy Baldwin. 2016. Take and took, gaggle and goose, book and read: Evaluating the utility of vector differences for lexical relation learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1671– 1682. Douglas Walton. 2010. Similarity, precedent and argument from analogy. Artificial Intelligence and Law, 18(3):217–246. Alex Wang and Kyunghyun Cho. 2019. BERT has a mouth, and it must speak: BERT as a Markov random field language model. In Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, pages 30–36, Minneapolis, Minnesota. Association for Computational Linguistics. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2019. Huggingface’s transformers: State-of-the-art natural language processing. ArXiv, abs/1910.03771. Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. 2014. Recurrent neural network regularization. arXiv preprint arXiv:1409.2329. A Experimental Details In our grid search to find the optimal configuration for each dataset and language model, each parameter was selected within the values shown in Table 6. As the coefficient of marginal likelihood α, αh, αt, we considered negative values as well as we hypothesized that the marginal likelihood could be beneficial for LMs as a way to leverage lexical knowledge of the head and tail words. Additionally, Table 7 shows the set of custom templates (or prompts) used in our experiments. Finally, Tables 8, 9, and 10 include the best configuration based on each validation set in for sPMI, smPPL and the hypothesis-only baseline, respectively. Parameter Value α -0.4, -0.2, 0, 0.2, 0.4 αh -0.4, -0.2, 0, 0.2, 0.4 αt -0.4, -0.2, 0, 0.2, 0.4 β 0, 0.2, 0.4, 0.6, 0.8, 1.0 g max,mean,min,val1,val2 gpos max,mean,min,val1,...,val8 gneg max,mean,min,val1,...,val16 Table 6: Hyperparameters with each search space. Type Template to-as [w1] is to [w2] as [w3] is to [w4] to-what [w1] is to [w2] What [w3] is to [w4] rel-same The relation between [w1] and [w2] is the same as the relation between [w3] and [w4]. what-to what [w1] is to [w2], [w3] is to [w4] she-as She explained to him that [w1] is to [w2] as [w3] is to [w4] as-what As I explained earlier, what [w1] is to [w2] is essentially the same as what [w3] is to [w4]. Table 7: Custom templates used in our experiments. Each has four placeholders [w1, ..., w4] and they are fulfilled by words from a relation pair. 3622 Data g α gpos gneg β t BERT SAT val2 -0.4 val5 val12 0.4 what-to U2 val2 -0.4 mean mean 0.6 what-to U4 val1 0.4 max val7 1.0 rel-same Google val1 -0.4 val1 val11 0.4 she-as BATS val1 -0.4 val11 val1 0.4 she-as GPT-2 SAT val2 -0.4 val3 val1 0.6 rel-same U2 val2 0.0 val4 val4 0.6 rel-same U4 val2 -0.4 mean mean 0.6 rel-same Google val1 0.0 mean val11 0.4 as-what BATS val1 -0.4 val1 val6 0.4 rel-same RoBERTa SAT min -0.4 min val7 0.2 as-what U2 min 0.4 mean val4 0.6 what-to U4 val2 0.0 mean val4 0.8 to-as Google val1 -0.4 val1 val6 0.4 what-to BATS max -0.4 mean val11 0.6 what-to Table 8: The best configuration of sPMI score. Data αh αt gpos gneg β t BERT SAT -0.2 -0.4 val5 val5 0.2 what-to U2 0.0 -0.2 mean mean 0.8 she-as U4 -0.2 0.4 val7 min 0.4 to-as Google 0.4 -0.2 val5 val12 0.6 she-as BATS 0.0 0.0 val8 min 0.4 what-to GPT-2 SAT -0.4 0.2 val3 val1 0.8 rel-same U2 -0.2 0.2 mean mean 0.8 as-what U4 -0.2 0.2 mean mean 0.8 rel-same Google -0.2 -0.4 mean mean 0.8 rel-same BATS 0.4 -0.4 val1 val5 0.8 rel-same RoBERTa SAT 0.2 0.2 val5 val11 0.2 as-what U2 0.4 0.4 val1 val4 0.4 what-to U4 0.2 0.2 val1 val1 0.4 as-what Google 0.2 0.2 val1 val6 0.2 what-to BATS 0.2 -0.2 val5 val11 0.4 what-to Table 9: The best configuration of smPPL score. B Additional Ablation Results We show a few more complementary results to our main experiments. B.1 Alternative Scoring Functions As alternative scoring functions for LM, we have tried two other scores: PMI score based on masked token prediction (Davison et al., 2019) (Mask PMI) and cosine similarity between the embedding difference of a relation pair similar to what used in word-embedding models. For embedding method, we give a prompted sentence to LM to get the last layer’s hidden state for each word in the given pair and we take the difference between them, which we regard as the embedding vector for the pair. Finally we pick up the most similar candidate in terms of the cosine similarity with the query embedding. TaMask Data gpos t BERT head SAT val5 to-what U2 val5 to-as U4 mean to-as Google val5 she-as BATS val5 to-as tail SAT val3 what-to U2 val7 to-what U4 val4 rel-same Google val7 as-what BATS val7 to-as RoBERTa head SAT val5 as-what U2 val5 rel-same U4 val7 she-as Google val5 what-to BATS val5 she-as tail SAT mean what-to U2 val7 rel-same U4 mean what-to Google val7 as-what BATS val7 what-to Table 10: The best configurations for hypothesis-only scores. ble 11 shows the test accuracy on each dataset. As one can see, AP scores outperform other methods with a great margin. Score SAT U2 U4 Google BATS BERT embedding 24.0 22.4 26.6 28.2 28.3 Mask PMI 25.2 23.3 31.5 61.2 46.2 sPMI 40.4 42.5 27.8 87.0 68.1 smPPL 41.8 44.7 41.2 88.8 67.9 RoBERTa embedding 40.4 42.5 27.8 87.0 68.1 Mask PMI 43.0 36.8 39.4 69.2 58.3 sPMI 51.3 49.1 38.7 92.4 77.2 smPPL 53.4 58.3 57.4 93.6 78.4 Table 11: Test accuracy tuned on each validation set. B.2 Parameter Sensitivity: template type t Figure 5 shows the box plot of relative improvement across all datasets grouped by t and the results indicate that there is a mild trend that certain templates tend to perform well, but not significant universal selectivity can be found across datasets. B.3 Parameter Sensitivity: aggregation method gneg Figure 6 shows the box plot of relative improvement across all datasets grouped by gneg. Unlike gpos we show in Figure 3, they do not give a strong signals over datasets. 3623 Figure 5: Box plot of the relative improvement on test accuracy in each dataset over all configurations of smPPL grouped by template type. Figure 6: Box plot of the relative improvement on test accuracy in each dataset over all configurations of smPPL grouped by gneg. Here valk corresponds to kth positive permutation shown in Figure 2. B.4 Relation Types in BATS/Google Figure 7 shows the results of different language models with the smPPL scoring function on the different categories of the BATS and Google datasets. C Error Analysis Table 12 shows all examples from the U2 dataset of the easiest difficuly (i.e. grade 4), which were misclassified by RoBERTa, with smPPL tuned on the validation set. We can see a few typical issues with word embeddings and language models. For instance, in the first example, the model confuses the antonym pair right:wrong with synonymy. In the second example, we have that someone who is poor lacks money, while someone who is hungry lacks food. However, the selected candidate pair is hungy:water rather than hungry:food, which is Figure 7: BATS (top) and Google (bottom) results split by high-level categories. presumably chosen because water is assumed to be a near-synonym of food. In the third example (wrench:tool), the hypnernymy relation is confused with a meronymy relation in the selected candidate tree:forest. In the last three examples, the model has selected answers which seem reasonable. In the fourth example, beautiful:pretty, terrible:bad and brave:valiant can all be considered to be synonym pairs. In the fifth example, vehicle:transport is clearly the correct answer, but the pair song:sing is nonetheless relationally similar to shield:protect. In the last example, we can think of being sad as an emotional state, like being sick is a health state, which provides some justification for the predicted answer. On the other hand, the gold answer is based on the argument that someone who is sick lacks health like someone who is scared lacks courage. 3624 Query Candidates hilarious:funny right:wrong, hard:boring, nice:crazy, great:good poor:money tired:energy, angry:emotion, hot:ice, hungry:water wrench:tool cow:milk, radio:sound, tree:forest, carrot:vegetable beautiful:pretty terrible:bad, brave:valiant, new:old, tall:skinny shield:protect computer:talk, vehicle:transport, pencil:make, song:sing sick:health sad:emotion, tall:intelligence, scared:courage, smart:energy Table 12: Model prediction examples from RoBERTa with smPPL tuned on the validation set. Gold answers are shown in bold, while the model predictions are underlined.
|
2021
|
280
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3625–3640 August 1–6, 2021. ©2021 Association for Computational Linguistics 3625 Exploring the Representation of Word Meanings in Context: A Case Study on Homonymy and Synonymy Marcos Garcia CiTIUS – Research Center in Intelligent Technologies Universidade de Santiago de Compostela, Galiza [email protected] Abstract This paper presents a multilingual study of word meaning representations in context. We assess the ability of both static and contextualized models to adequately represent different lexical-semantic relations, such as homonymy and synonymy. To do so, we created a new multilingual dataset that allows us to perform a controlled evaluation of several factors such as the impact of the surrounding context or the overlap between words, conveying the same or different senses. A systematic assessment on four scenarios shows that the best monolingual models based on Transformers can adequately disambiguate homonyms in context. However, as they rely heavily on context, these models fail at representing words with different senses when occurring in similar sentences. Experiments are performed in Galician, Portuguese, English, and Spanish, and both the dataset (with more than 3,000 evaluation items) and new models are freely released with this study. 1 Introduction Contrary to static vector models, which represent the different senses of a word in a single vector (Erk, 2012; Mikolov et al., 2013), contextualized models generate representations at token-level (Peters et al., 2018; Devlin et al., 2019), thus being an interesting approach to model word meaning in context. In this regard, several studies have shown that clusters produced by some contextualized word embeddings (CWEs) are related to different senses of the same word (Reif et al., 2019; Wiedemann et al., 2019), or that similar senses can be aligned in cross-lingual experiments (Schuster et al., 2019). However, more systematic evaluations of polysemy (i.e., word forms that have different related meanings depending on the context (Apresjan, 1974)), have shown that even though CWEs present some correlations with human judgments (Nair et al., 2020), they fail to predict the similarity of the various senses of a polysemous word (Haber and Poesio, 2020). As classical datasets to evaluate the capabilities of vector representations consist of single words without context (Finkelstein et al., 2001) or heavily constrained expressions (Kintsch, 2001; Mitchell and Lapata, 2008), new resources with annotations of words in free contexts have been created, including both graded similarities (Huang et al., 2012; Armendariz et al., 2020) or binary classification of word senses (Pilehvar and Camacho-Collados, 2019; Raganato et al., 2020). However, as these datasets largely include instances of polysemy, they are difficult to solve even for humans (in fact, the highest reported human upper bound is about 80%) as the nuances between different senses depend on non-linguistic factors such as the annotator procedure or the target task (Tuggy, 1993; Kilgarriff, 1997; Hanks, 2000; Erk, 2010). In this paper, we rely on a more objective and simple task to assess how contextualized approaches (both neural network models and contextualized methods of distributional semantics) represent word meanings in context. In particular, we observe whether vector models can identify unrelated meanings represented by the same word form (homonymy) and the same sense conveyed by different words (synonymy). In contrast to polysemy, there is a strong consensus concerning the representation of homonymous senses in the lexicon, and it has been shown that homonyms are cognitively processed differently than polysemous words (Klepousniotou et al., 2012; MacGregor et al., 2015). In this regard, exploratory experiments in English suggest that some CWEs correctly model homonymy, approximating the contextualized vectors of a homonym to those of its paraphrases (Lake and Murphy, 2020), and showing stronger correlation with human judgments to those 3626 of polysemous words (Nair et al., 2020). However, as homonyms convey unrelated meanings depending on the context, it is not clear whether the good performance of CWEs actually derives from the contextualization process or simply from the use of explicit lexical cues present in the sentences. Taking the above into account, we have created a new multilingual dataset (in Galician, Portuguese, English, and Spanish) with more than 3,000 evaluation items. It allows for carrying out more than 10 experiments and controlling factors such as the surrounding context, the word overlap, and the sense conveyed by different word forms. We use this resource to perform a systematic evaluation of contextualized word meaning representations. We compare different strategies using both static embeddings and current models based on deep artificial neural networks. The results suggest that the best monolingual models based on Transformers (Vaswani et al., 2017) can identify homonyms having different meanings adequately. However, as they strongly rely on the surrounding context, words with different meanings are represented very closely when they occur in similar sentences. Apart from the empirical conclusions and the dataset, this paper also contributes with new BERT and fastText models for Galician.1 Section 2 presents previous studies about word meaning representation. Then, Section 3 introduces the new dataset used in this paper. In Section 4 we describe the models and methods to obtain the vector representations. Finally, the experiments and results are discussed in Section 5, while Section 6 draws some conclusions of our study. 2 Related Work A variety of approaches has been implemented to compute word meaning in context by means of standard methods of distributional semantics (Schütze, 1998; Kintsch, 2001; McDonald and Brew, 2004; Erk and Padó, 2008). As compositional distributional models construct sentence representations from their constituents vectors, they take into account contextualization effects on meaning (Mitchell and Lapata, 2008; Baroni and Zamparelli, 2010; Baroni, 2013). However, these approaches often have scalability problems as their representations grow exponentially with the size of the sentences. Therefore, the datasets used to 1Dataset, models, and code are available at https:// github.com/marcospln/homonymy_acl21/. evaluate them are composed of highly restricted phrases (Grefenstette and Sadrzadeh, 2011). The rise of artificial neural networks on natural language processing popularized the use of vector representations, and the remarkable performance of neural language models (Melamud et al., 2016; Peters et al., 2018) led to a productive line of research exploring to what extent these models represent linguistic knowledge (Rogers et al., 2020). However, few of these works have focused on lexical semantics, and most of the relevant results in this field come from evaluations in downstream tasks. In this regard, Wiedemann et al. (2019) found that clusters of BERT embeddings (Devlin et al., 2019) seem to be related to word senses, while Schuster et al. (2019) observed that clusters of polysemous words correspond to different senses in a cross-lingual alignment of vector representations. Probing LSTMs on lexical substitution tasks, Aina et al. (2019) showed that these architectures rely on the lexical information from the input embeddings, and that the hidden states are biased towards contextual information. On an exploration of the geometric representations of BERT, Reif et al. (2019) found that different senses of a word tend to appear separated in the vector space, while several clusters seem to correspond to similar senses. Recently, Vuli´c et al. (2020) evaluated the performance of BERT models on several lexical-semantic tasks in various languages, including semantic similarity or word analogy. The results show that using special tokens ([CLS] or [SEP]) hurts the quality of the representations, and that these tend to improve across layers until saturation. As this study uses datasets of single words (without context), typelevel representations are obtained by averaging the contextualized vectors over various sentences. There are several resources to evaluate word meaning in free contexts, such as the Stanford Contextual Word Similarity (Huang et al., 2012) and CoSimLex (Armendariz et al., 2020), both representing word similarity on a graded scale, or the Word-in-Context datasets (WiC), focused on binary classifications (i.e., each evaluation item contains two sentences with the same word form, having the same or different senses) (Pilehvar and CamachoCollados, 2019; Raganato et al., 2020). These datasets include not only instances of homonymy but mostly of polysemous words. In this regard, studies on polysemy using Transformers have obtained diverse results: Haber and Poesio (2020) 3627 found that BERT embeddings correlate better with human ratings of co-predication than with similarity between word senses, thus suggesting that these representations encode more contextual information than word sense knowledge. Nevertheless, the results of Nair et al. (2020) indicate that BERT representations are correlated with human scores of polysemy. An exploratory experiment of the latter study also shows that BERT discriminates between polysemy and homonymy, which is also suggested by other pilot evaluations reported by Lake and Murphy (2020) and Yu and Ettinger (2020). Our study follows this research line pursuing objective and unambiguous lexical criteria such as the representation of homonyms and synonyms. In this context, there is a broad consensus in the psycholinguistics literature regarding the representation of homonyms as different entries in the lexicon (in contrast to polysemy, for which there is a long discussion on whether senses of polysemous words are stored as a single core representation or as independent entries (Hogeweg and Vicente, 2020)). In fact, several studies have shown that homonyms are cognitively processed differently from polysemous words (Klepousniotou et al., 2012; Rabagliati and Snedeker, 2013). In contrast to the different senses of polysemous words, which are simultaneously activated, the meanings of homonyms are in conflict during processing, with the not relevant ones being deactivated by the context (MacGregor et al., 2015). To analyze how vector models represent homonymy and synonymy in context, we have built a new multilingual resource with a strong inter-annotator agreement, presented below. 3 A New Multilingual Resource of Homonymy and Synonymy in Context This section briefly describes some aspects of lexical semantics relevant to our study, and then presents the new dataset used in the paper. Homonymy and homography: Homonymy is a well-known type of lexical ambiguity that can be described as the relation between distinct and unrelated meanings represented by the same word form, such as match, meaning for instance ‘sports game’ or ‘stick for lighting fire’. In contrast to polysemy (where one lexeme conveys different related senses depending on the context, e.g., newspaper as an organization or as a set of printed pages), it is often assumed that homonyms are different lexemes that have the same lexical form (Cruse, 1986), and therefore they are stored as independent entries in the lexicon (Pustejovsky, 1998). There are two main criteria for homonymy identification: Diachronically, homonyms are lexical items that have different etymologies but are accidentally represented by the same word form, while a synchronic perspective strengthens unrelatedness in meaning. Even if both approaches tend to identify similar sets of homonyms, there may be ambiguous cases that are diachronically but not synchronically related (e.g., two meanings of banco –‘bench’ and ‘financial institution’– in Portuguese or Spanish could be considered polysemous as they derive from the same origin,2 but as this is a purely historical association, most speakers are not aware of the common origin of both senses). In this study, we follow the synchronic perspective, and consider homonymous meanings those that are clearly unrelated (e.g., they unambiguously refer to completely different concepts) regardless of their origin. It is worth mentioning that as we are dealing with written text we are actually analyzing homographs (different lexemes with the same spelling) instead of homonyms. Thus, we discard instances of phonologically identical words which are written differently, such as the Spanish hola ‘hello’ and ola ‘wave’, both representing the phonological form /ola/. Similarly, we include words with the same spelling representing different phonological forms, e.g., the Galician-Portuguese sede, which corresponds to both /sede/ ‘thirst’, and /sEde/ ‘headquarters’. In this paper, homonymous senses are those unrelated meanings conveyed by the same (homonym) word form. For instance, coach may have two homonymous senses (‘bus’ and ‘trainer’), which can be conveyed by other words (synonyms) in different contexts (e.g., by bus or trainer). Structure of the dataset: We have created a new resource to investigate how vector models represent word meanings in context. In particular, we want to observe whether they capture (i) different senses conveyed by the same word form (homonymy), and (ii) equivalent senses expressed by different words (synonymy). The resource contains controlled sentences so that it allows us to observe how the context and word overlap affect word representations. To allow for different comparisons with the same 2In fact, several dictionaries organize them in a single entry: https://dicionario.priberam.org/banco, https://dle.rae.es/banco. 3628 Sense Sentences 1-3 Sentence 4 Sentence 5 (1) We’re going to the airport by coach. [. . . ] the coach was badly delayed by roadworks. They had to travel everywhere by bus. We’re going to the airport by bus. We’re going to the airport by bicycle. (2) That man was appointed as the new coach. She has recently joined the amateur team as coach. They need a new trainer for the young athletes. That man was appointed as the new trainer. That man was appointed as the new president. Table 1: Example sentences for two senses of coach in English (‘bus’ and ‘trainer’). Sentences 1 to 3 include, in the same context, the target word, a synonym, and a word with a different sense (in italic), respectively. Sentences 4 and 5 contain the target word and a synonym in different contexts, respectively. and different contexts, we have included five sentences for each meaning (see Table 1 for examples): three sentences containing the target word, a synonym, and a word with a different sense, all of them in the same context (sentences 1 to 3), and two additional sentences with the target word and a synonym, representing the same sense (sentences 4 and 5, respectively). Thus, for each sense we have four sentences (1, 2, 4, 5) with a word conveying the same sense (both in the same and in different contexts) and another sentence (3) with a different word in the same context as sentences 1 and 2. From this structure, we can create datasets of sentence triples, where the target words of two of them convey the same sense, and the third one has a different meaning. Thus, we can generate up to 48 triples for each pair of senses (24 in each direction: sense 1 vs. sense 2, and vice-versa). These datasets allow us to evaluate several semantic relations at the lexical level, including homonymy, synonymy, and various combinations of homonymous senses. Interestingly, we can control for the impact of the context (e.g., are contextualized models able to distinguish between different senses occurring in the same context, or do they incorporate excessive contextual information into the word vectors?), the word overlap (e.g., can a model identify different senses of the same word form depending on the context, or it strongly depends on lexical cues?), or the POS-tag (e.g., are homonyms with different POS-tags easily disambiguated?). Construction of the dataset: We compiled data for four languages: Galician, Portuguese, Spanish, and English.3 We tried to select sentences compatible with the different varieties of the same language 3Galician is generally considered a variety of a single (Galician-)Portuguese language. However, they are divided in this resource, as Galician has recently been standardized using a Spanish-based orthography that formally separates it from Portuguese (Samartim, 2012). (e.g., with the same meaning in UK and US English, or in Castilian and Mexican Spanish). However, we gave priority to the European varieties when necessary (e.g., regarding spelling variants). The dataset was built using the following procedure: First, language experts (one per language) compiled lists of homonyms using dedicated resources for language learning, together with WordNet and other lexicographic data (Miller, 1995; Montraveta and Vázquez, 2010; Guinovart, 2011; Rademaker et al., 2014). Only clear and unambiguous homonyms were retained (i.e., those in the extreme of the homonymy-polysemy-vagueness scale (Tuggy, 1993)). These homonyms were then enriched with frequency data from large corpora: Wikipedia and SLI GalWeb (Agerri et al., 2018) for Galician, and a combination of Wikipedia and Europarl for English, Spanish and Portuguese (Koehn, 2005). From these lists, each linguist selected the most frequent homonyms, annotating them as ambiguous at type or token level (absolute homonymy and partial homonymy in Lyons’ terms (Lyons, 1995)). As a substantial part were nounverb pairs, only a few of these were included. For each homonym, the language experts selected from corpora two sentences (1 and 4) in which the target words were not ambiguous.4 They then selected a synonym that could be used in sentence 1 without compromising grammaticality (thus generating sentence 2), and compiled an additional sentence for it (5), trying to avoid further lexical ambiguities in this process.5 For each homonym, the linguists selected a word with a different meaning (for sen4Sentences were selected, adapted, and simplified using GDEX-inspired constraints (Kilgarriff et al., 2008) (i.e., avoiding high punctuation ratios, unnecessary subordinate clauses, etc.), which resulted in the creation of new sentences. 5In most cases, this synonym is the same as that of sentence 2, but this is not always the case. Besides, in some cases we could not find words conveying the same sense, for which we do not have sentences 2 and 5. 3629 Language Hom. Senses Sent. Triples Pairs WiC κ Galician 22 47 (4) 227 1365 823 197 0.94 English 14 30 (5) 138 709 463 129 0.96 Portuguese 11 22 (1) 94 358 273 81 0.96 Spanish 10 23 (3) 105 645 391 101 0.95 Total 57 122 564 3077 1950 508 0.94 Table 2: Characteristics of the dataset. First three columns display the number of homonyms (Hom), senses, and sentences (Sent), respectively. Senses in parentheses are the number of homonymous pairs with different POStags). Center columns show the size of the evaluation data in three formats: triples, pairs, and WiC-like pairs, followed by the Cohen’s κ agreements and their micro-average. The total number of homonyms and senses is the sum of the language-specific ones, regardless of the fact that some senses occur in more than one language. tence 3), trying to maximize the following criteria: (i) to refer unambiguously to a different concept, and to preserve (ii) semantic felicity and (iii) grammaticality. The size of the final datasets varies depending on the initial lists and on the ease of finding synonyms in context. Results: Apart from the sentence triples explained above, the dataset structure allows us to create evaluation sets with different formats, such as sentence pairs to perform binary classifications as in the WiC datasets. Table 2 shows the number of homonyms, senses, and sentences of the multilingual resource, together with the size of the evaluation datasets in different formats. As the original resource was created by one annotator per language, we ensured its quality as follows: We randomly extracted sets of 50 sentence pairs and gave them to other annotators (5 for Galician, and 1 for each of the other three varieties, all of them native speakers of the target language). We then computed the Cohen’s κ inter-annotator agreement (Cohen, 1960) between the original resource and the outcome of this second annotation (see the right column of Table 2). We obtained a microaverage κ = 0.94 across languages, a result which supports the task’s objectivity. Nevertheless, it is worth noting that few sentences have been carefully modified after this analysis, as it has shown that several misclassifications were due to the use of an ambiguous synonym. Thus, it is likely that the final resource has higher agreement values. 4 Models and Methods This section introduces the models and procedures to obtain vector representations followed by the evaluation method. 4.1 Models We have used static embeddings and CWEs based on Transformers, comparing different ways of obtaining the vector representations in both cases: Static embeddings: We have used skip-gram fastText models of 300 dimensions (Bojanowski et al., 2017).6 For English and Spanish, we have used the official vectors trained on Wikipedia. For Portuguese, we have used the model provided by Hartmann et al. (2017), and for Galician we have trained a new model (see Appendix C for details).7 Contextualized embeddings: We have evaluated multilingual and monolingual models:8 Multilingual models: We have used the official multilingual BERT (mBERT cased, 12 layers) (Devlin et al., 2019), XLM-RoBERTa (Base, 12 layers) (Conneau et al., 2020), and DistilBERT (DistilmBERT, 6 layers) (Sanh et al., 2019). Monolingual models: For English, we have used the official BERT-Base model (uncased). For Portuguese and Spanish, BERTimbau (Souza et al., 2020) and BETO (Cañete et al., 2020) (both cased). For Galician, we trained two BERT models (with 6 and 12 layers; see Appendix C). 4.2 Obtaining the vectors Static models: These are the methods used to obtain the representations from the static models: Word vector (WV): Embedding of the target word (homonymous senses with the same word form will have the same representation). 6In preliminary experiments we also used word2vec and GloVe models, obtaining slightly lower results than fastText. 7These Portuguese and Galician models obtained better results (0.06 on average) than the official ones. 8To make a fair comparison we prioritized base models (12 layers), but we also report results for large (24 layers) and 6 layers models when available. 3630 Language Exp1 Exp2 Exp3 Exp4 Total Full Galician 122 105 183 149 278 229 135 135 718 618 1365 1157 English 77 52 89 58 144 91 68 68 378 269 709 494 Portuguese 45 41 37 37 80 74 41 41 203 193 358 342 Spanish 65 49 87 71 146 110 59 59 357 289 645 517 Table 3: Number of instances of each experiment and language. Numbers on the right of each column are those triples where the three target words belong to the same morphosyntactic category (left values are the total number of triples). Total are the sums of the four experiments, while Full refers to all the instances of the dataset. Sentence vector (Sent): Average embedding of the whole sentence. Syntax (Syn): Up to four different representations obtained by adding the vector of the target word to those of their syntactic heads and dependents. This method is based on the assumption that the syntactic context of a word characterizes its meaning, providing relevant information for its contextualized representation (e.g., in ‘He swims to the bank’, bank may be disambiguated by combining its vector with the one of swim).9 Appendix D describes how heads and dependents are selected. Contextualized models: For these models, we have evaluated the following approaches: Sentence vector (Sent): Vector of the sentence built by averaging all words (except for the special tokens [CLS] and [SEP]), each of them represented by the standard approach of concatenating the last 4 layers (Devlin et al., 2019). Word vector (WV): Embedding of the target word, combining the vectors of the last 4 layers. We have evaluated two operations: vector concatenation (Cat), and addition (Sum). Word vector across layers (Lay): Vector of the target word on each layer. This method allows us to explore the contextualization effects on each layer. Vectors of words split into several sub-words are obtained by averaging the embeddings of their components. Similarly, MWEs vectors are the average of the individual vectors of their components, both for static and for contextualized embeddings. 4.3 Measuring sense similarities Given a sentence triple where two of the target words (a and b) have the same sense and the third (c) a different one, we evaluate a model as follows (in a similar way as other studies (Kintsch, 2001; Lake and Murphy, 2020)): First, we obtain 9We have also evaluated a contextualization method using selectional preferences inspired by Erk and Padó (2008), but the results were almost identical to those of the WV approach. three cosine similarities between the vector representations: sim1 = cos(a, b); sim2 = cos(a, c); sim3 = cos(b, c). Then, an instance is labeled as correct if those words conveying the same sense (a and b) are closer together than the third one (c). In other words, sim1 > sim2 and sim1 > sim3: Otherwise, the instance is considered as incorrect. 5 Evaluation This section presents the experiments performed using the new dataset and discusses their results. 5.1 Experiments Among all the potential analyses of our data, we have selected four evaluations to assess the behavior of a model by controlling factors such as the context and the word overlap: Homonymy (Exp1): The same word form in three different contexts, two of them with the same sense (e.g., coach in sentences [1:1, 1:4, 2:1]10 in Table 1). This test evaluates if a model correctly captures the sense of a unique word form in context. Hypothesis: Static embeddings will fail as they produce the same vector in the three cases, while models that adequately incorporate contextual cues should correctly identify the outlier sense. Synonyms of homonymous senses (Exp2): A word is compared with its synonym and with the synonym of its homonym, all three in different contexts (e.g., coach=bus̸=trainer in [1:1, 1:5, 2:2]). This test assesses if there is a bias towards one of the homonymous senses, e.g., the most frequent one (MacGregor et al., 2015). Hypothesis: Models with this type of bias may fail, so as in Exp1, they should also appropriately incorporate contextual information to represent these examples. Synonymy vs homonymy (Exp3): We compare a word to its synonym and to a homonym, all in 10First and second digits refer to the sense and sentence ids. 3631 different contexts (e.g., coach=bus̸=coach in [1:1, 1:5, 2:1]). Here we evaluate whether a model adequately represents both (i) synonymy in context –two word forms with the same sense in different contexts– and (ii) homonymy –one of the former word forms having a different meaning. Hypothesis: Models relying primarily on lexical knowledge are likely to represent homonyms closer than synonyms (giving rise to an incorrect output), but those integrating contextual information will be able to model the three representations correctly. Synonymy (Exp4): Two synonyms vs. a different word (and sense), all of them in the same context (e.g., [2:1, 2:2, 2:3]). It assesses to what extent the context affects word representations of different word forms. Hypothesis: Static embeddings may pass this test as they tend to represent typelevel synonyms closely in the vector space. Highly contextualized models might be puzzled as different meanings (from different words) occur in the same context, so that the models should have an adequate trade-off between lexical and contextual knowledge. Table 3 displays the number of sentence triples for each experiment as well as the total number of triples of the dataset. To focus on the semantic knowledge encoded in the vectors –rather than on the morphosyntactic information–, we have evaluated only those triples in which the target words of the three sentences have the same POS-tag (numbers on the right).11 Besides, we have also carried out an evaluation on the full dataset. 5.2 Results and discussion Table 4 contains a summary of the results of each experiment in the four languages. For reasons of clarity, we include only fastText embeddings and the best contextualized model (BERT). Results for all models and languages can be seen in Appendix A. BERT models have the best performance overall, both on the full dataset and on the selected experiments, except for Exp4 (in which the three sentences share the context) where the static models outperform the contextualized representations. In Exp1 and Exp2, where the context plays a crucial role, fastText models correctly labeled between 50%/60% of the examples (depending on the language and vector type, with better results 11On average, BERT-base models achieved 0.24 higher results (Add) when tested on all the instances (including different POS-tags) of the four experiments. for Sent and Syn). For BERT, the best accuracy surpasses 0.98 (Exp1 in English), with an average across languages of 0.78, and where word vectors outperform sentence representations. These high results and the fact that WVs work better in general than Sent may be indicators that Transformers are properly incorporating contextual knowledge. Solving Exp3 requires both dealing with contextual effects and homonymy (as two words have the same form but different meaning) so that static embeddings hardly achieve 0.5 accuracy (Sent, with lower results for both WV and Syn). BERT’s performance is also lower than in Exp1 and Exp2, with an average of 0.67 and Sent beating WVs in most cases, indicating that the word vectors are not adequately representing the target senses. Finally, fastText obtains better results than BERT on Exp4 (where the three instances have the same context), reaching 0.81 in Spanish with an average across languages of 0.64 (always with WVs). BERT’s best performance is 0.41 (in two languages) with an average of 0.42, suggesting that very similar contexts may confound the model. To shed light on the contextualization process of Transformers, we have analyzed their performance across layers. Figure 1 shows the accuracy curves (vs. the macro-average Sent and WV vectors of the contextualized and static embeddings) for five Transformers models on Galician, the language with the largest dataset (see Appendix A for equivalent figures for the other languages). In Exp1 to Exp3 the best accuracies are obtained at upper layers, showing that word vectors appropriately incorporate contextual information. This is true especially for the monolingual BERT versions, as the multilingual models’ representations show higher variations. Except for Galician, Exp1 has better results than Exp2, as the former primarily deals with context while the latter combines contextualization with lexical effects. In Exp3 the curves take longer to rise as initial layers rely more on lexical than on contextual information. Furthermore, except for English (which reaches 0.8), the performance is low even in the best hidden layers (≈0.4). In Exp4 (with the same context in the three sentences), contextualized models cannot correctly represent the word senses, being surpassed in most cases by the static embeddings. Finally, we have observed how Transformers representations vary across the vector space. Figure 2 shows the UMAP visualizations (McInnes et al., 3632 Model Vec. Exp1 Exp2 Exp3 Exp4 Macro Micro Full Galician BERT-base Sent 0.695 0.758 0.751 0.178 0.596 0.618 0.727 Cat 0.705 0.799 0.293 0.422 0.555 0.513 0.699 fastText Sent 0.562 0.685 0.476 0.141 0.466 0.468 0.618 WV 0.21 0.564 0 0.526 0.325 0.286 0.461 Syn (3) 0.533 0.658 0.197 0.185 0.393 0.362 0.567 English BERT-base Sent 0.788 0.655 0.736 0.221 0.6 0.599 0.7 Add 0.981 0.81 0.758 0.441 0.748 0.732 0.839 fastText Sent 0.596 0.5 0.505 0.147 0.437 0.431 0.543 WV 0.308 0.552 0.033 0.574 0.366 0.335 0.48 Syn (3) 0.442 0.69 0.231 0.176 0.385 0.357 0.546 Portuguese BERT-base Sent 0.683 0.432 0.635 0.22 0.493 0.518 0.564 Add 0.854 0.541 0.378 0.366 0.535 0.508 0.67 fastText Sent 0.61 0.622 0.527 0.171 0.482 0.487 0.55 WV 0.024 0.541 0 0.634 0.3 0.244 0.453 Syn (3) 0.659 0.459 0.176 0.195 0.372 0.337 0.508 Spanish BERT-base Sent 0.755 0.592 0.536 0.186 0.517 0.516 0.595 Add 0.857 0.704 0.409 0.441 0.603 0.564 0.74 fastText Sent 0.449 0.338 0.445 0.085 0.329 0.346 0.429 WV 0.122 0.62 0.018 0.814 0.393 0.346 0.479 Syn (3) 0.367 0.577 0.173 0.237 0.339 0.318 0.553 Table 4: Summary of the BERT and fastText results. Macro and Micro refer to the macro-average and microaverage results across the four experiments, respectively. Full are the micro-average values on the whole dataset. Figure 1: Results across layers and models for Galician. Sent and WV (dashed) are macro-average values. MacroAvg|Syn is the macro-average per layer (Transformers) and the macro-average of the Syn strategy (fastText). 2018) of the contextualization processes of Exp1 and Exp3 examples in English. In 2a, the similar vectors of match in layer 1 are being contextualized across layers, producing a suitable representation since layer 7. However, 2b shows how the model is not able to adequately represent match close to its 3633 (a) Exp1: Sentence 2: “Chelsea have a match with United next week.”. Sentence 3: “You should always strike a match away from you.” (b) Exp3: Sentence 2: “A game consists of two halves lasting 45 minutes, meaning it is 90 minutes long.”. Sentence 3: “He was watching a football stadium.” Figure 2: UMAP visualizations of word contextualization across layers (1 to 12) in Exp1 and Exp3 in English (BERT-base). In both cases, sentence 1 is “He was watching a football match.”, and the target word in sentence 3 is the outlier. synonym game, as the vectors seem to incorporate excessive information (or at least limited lexical knowledge) from the context. Additional visualizations in Galician can be found in Appendix B. In sum, the experiments performed in this study allow us to observe how different models generate contextual representations. In general, our results confirm previous findings which state that Transformers models increasingly incorporate contextual information across layers. However, we have also found that this process may deteriorate the representation of the individual words, as it may be incorporating excessive contextual information, as suggested by Haber and Poesio (2020). 6 Conclusions and Further Work This paper has presented a systematic study of word meaning representation in context. Besides static word embeddings, we have assessed the ability of state-of-the-art monolingual and multilingual models based on the Transformers architecture to identify unambiguous cases of homonymy and synonymy. To do so, we have presented a new dataset in four linguistic varieties that allows for controlled evaluations of vector representations. The results of our study show that, in most cases, the best contextualized models adequately identify homonyms conveying different senses in various contexts. However, as they strongly rely on the surrounding contexts, they misrepresent words having different senses in similar sentences. In further work, we plan to extend our dataset with multiword expressions of different degrees of idiomaticity and to include less transparent –but still unambiguous– contexts of homonymy. Finally, we also plan to systematically explore how multilingual models represent homonymy and synonymy in cross-lingual scenarios. Acknowledgments We would like to thank the anonymous reviewers for their valuable comments, and NVIDIA Corporation for the donation of a Titan Xp GPU. This research is funded by a Ramón y Cajal grant (RYC2019-028473-I) and by the Galician Government (ERDF 2014-2020: Call ED431G 2019/04). References Rodrigo Agerri, Xavier Gómez Guinovart, German Rigau, and Miguel Anxo Solla Portela. 2018. Developing new linguistic resources and tools for the Galician language. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). Laura Aina, Kristina Gulordava, and Gemma Boleda. 2019. Putting words in context: LSTM language models and lexical ambiguity. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3342–3348, Florence, Italy. Association for Computational Linguistics. Ju D Apresjan. 1974. Regular polysemy. Linguistics, 12(142):5–32. 3634 Carlos Santos Armendariz, Matthew Purver, Matej Ulˇcar, Senja Pollak, Nikola Ljubeši´c, and Mark Granroth-Wilding. 2020. CoSimLex: A resource for evaluating graded word similarity in context. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 5878–5886, Marseille, France. European Language Resources Association. Marco Baroni. 2013. Composition in distributional semantics. Language and Linguistics Compass, 7(10):511–522. Marco Baroni and Roberto Zamparelli. 2010. Nouns are vectors, adjectives are matrices: Representing adjective-noun constructions in semantic space. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1183–1193, Cambridge, MA. Association for Computational Linguistics. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135–146. José Cañete, Gabriel Chaperon, Rodrigo Fuentes, JouHui Ho, Hojin Kang, and Jorge Pérez. 2020. Spanish Pre-Trained BERT Model and Evaluation Data. In PML4DC at ICLR 2020. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440– 8451, Online. Association for Computational Linguistics. David Alan Cruse. 1986. Lexical semantics. Cambridge University Press. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Katrin Erk. 2010. What is word meaning, really? (and how can distributional models help us describe it?). In Proceedings of the 2010 Workshop on GEometrical Models of Natural Language Semantics, pages 17–26, Uppsala, Sweden. Association for Computational Linguistics. Katrin Erk. 2012. Vector space models of word meaning and phrase meaning: A survey. Language and Linguistics Compass, 6(10):635–653. Katrin Erk and Sebastian Padó. 2008. A structured vector space model for word meaning in context. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 897–906, Honolulu, Hawaii. Association for Computational Linguistics. Lev Finkelstein, Evgeniy Gabrilovich, Yossi Matias, Ehud Rivlin, Zach Solan, Gadi Wolfman, and Eytan Ruppin. 2001. Placing search in context: The concept revisited. In Proceedings of the 10th international conference on World Wide Web, pages 406– 414. Marcos Garcia and Pablo Gamallo. 2010. Análise Morfossintáctica para Português Europeu e Galego: Problemas, Soluções e Avaliação. Linguamática, 2(2):59–67. Edward Grefenstette and Mehrnoosh Sadrzadeh. 2011. Experimental support for a categorical compositional distributional model of meaning. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1394–1404, Edinburgh, Scotland, UK. Association for Computational Linguistics. Xavier Gómez Guinovart. 2011. Galnet: WordNet 3.0 do galego. Linguamática, 3(1):61–67. Janosch Haber and Massimo Poesio. 2020. Assessing polyseme sense similarity through co-predication acceptability and contextualised embedding distance. In Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics, pages 114–124, Barcelona, Spain (Online). Association for Computational Linguistics. Patrick Hanks. 2000. Do Word Meanings Exist? Computers and the Humanities, 34:205–215. Nathan Hartmann, Erick Fonseca, Christopher Shulby, Marcos Treviso, Jéssica Silva, and Sandra Aluísio. 2017. Portuguese word embeddings: Evaluating on word analogies and natural language tasks. In Proceedings of the 11th Brazilian Symposium in Information and Human Language Technology, pages 122–131, Uberlândia, Brazil. Sociedade Brasileira de Computação. Lotte Hogeweg and Agustin Vicente. 2020. On the nature of the lexicon: The status of rich lexical meanings. Journal of Linguistics, 56(4):865–891. Eric Huang, Richard Socher, Christopher Manning, and Andrew Ng. 2012. Improving word representations via global context and multiple word prototypes. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 873–882, Jeju Island, Korea. Association for Computational Linguistics. 3635 Adam Kilgarriff. 1997. I don’t believe in word senses. Computers and the Humanities, 31(2):91–113. Adam Kilgarriff, Milos Husák, Katy McAdam, Michael Rundell, and Pavel Rychl`y. 2008. GDEX: Automatically finding good dictionary examples in a corpus. In Proceedings of the XIII EURALEX international congress, pages 425–432. Documenta Universitaria Barcelona, Spain. Walter Kintsch. 2001. Predication. Cognitive science, 25(2):173–202. Ekaterini Klepousniotou, G Bruce Pike, Karsten Steinhauer, and Vincent Gracco. 2012. Not all ambiguous words are created equal: An EEG investigation of homonymy and polysemy. Brain and language, 123(1):11–21. Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In Conference Proceedings: the tenth Machine Translation Summit, volume 5, pages 79–86. AAMT. Yuri Kuratov and Mikhail Arkhipov. 2019. Adaptation of Deep Bidirectional Multilingual Transformers for Russian Language. Computational Linguistics and Intellectual Technologies, 18:333–339. Brenden M. Lake and Gregory L. Murphy. 2020. Word meaning in minds and machines. ArXiv preprint: 2008.01766. John Lyons. 1995. Linguistic semantics: An introduction. Cambridge University Press. Lucy J MacGregor, Jennifer Bouwsema, and Ekaterini Klepousniotou. 2015. Sustained meaning activation for polysemous but not homonymous words: Evidence from EEG. Neuropsychologia, 68:126–138. Scott McDonald and Chris Brew. 2004. A distributional model of semantic context effects in lexical processing. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04), pages 17–24, Barcelona, Spain. Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. UMAP: Uniform Manifold Approximation and Projection. Journal of Open Source Software, 3(29):861. Oren Melamud, Jacob Goldberger, and Ido Dagan. 2016. context2vec: Learning generic context embedding with bidirectional LSTM. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 51–61, Berlin, Germany. Association for Computational Linguistics. Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. In Workshop Proceedings of the International Conference on Learning Representations (ICLR) 2013. ArXiv preprint arXiv:1301.3781. George A Miller. 1995. WordNet: a lexical database for English. Communications of the ACM, 38(11):39–41. Jeff Mitchell and Mirella Lapata. 2008. Vector-based models of semantic composition. In Proceedings of ACL-08: HLT, pages 236–244, Columbus, Ohio. Association for Computational Linguistics. Ana María Fernández Montraveta and Gloria Vázquez. 2010. La construcción del WordNet 3.0 en español. La lexicografía en su dimensión teórica, pages 201– 220. Sathvik Nair, Mahesh Srinivasan, and Stephan Meylan. 2020. Contextualized word embeddings encode aspects of human-like word sense knowledge. In Proceedings of the Workshop on the Cognitive Aspects of the Lexicon, pages 129–141, Online. Association for Computational Linguistics. Lluís Padró and Evgeny Stanilovsky. 2012. FreeLing 3.0: Towards wider multilinguality. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2473–2479, Istanbul, Turkey. European Language Resources Association (ELRA). Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227–2237, New Orleans, Louisiana. Association for Computational Linguistics. Mohammad Taher Pilehvar and Jose CamachoCollados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267–1273, Minneapolis, Minnesota. Association for Computational Linguistics. James Pustejovsky. 1998. The Generative Lexicon. The MIT Press. Hugh Rabagliati and Jesse Snedeker. 2013. The truth about chickens and bats: Ambiguity avoidance distinguishes types of polysemy. Psychological science, 24(7):1354–1360. Alexandre Rademaker, Valeria de Paiva, Gerard de Melo, Livy Real, and Maira Gatti. 2014. OpenWordNet-PT: A project report. In Proceedings of the Seventh Global Wordnet Conference, pages 383–390, Tartu, Estonia. University of Tartu Press. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text 3636 transformer. Journal of Machine Learning Research, 21(140):1–67. Alessandro Raganato, Tommaso Pasini, Jose CamachoCollados, and Mohammad Taher Pilehvar. 2020. XL-WiC: A multilingual benchmark for evaluating semantic contextualization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7193–7206, Online. Association for Computational Linguistics. Emily Reif, Ann Yuan, Martin Wattenberg, Fernanda B Viegas, Andy Coenen, Adam Pearce, and Been Kim. 2019. Visualizing and Measuring the Geometry of BERT. In Advances in Neural Information Processing Systems, volume 32, pages 8594–8603. Curran Associates, Inc. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the Association for Computational Linguistics, 8:842–866. Roberto Samartim. 2012. Língua somos: A construção da ideia de língua e da identidade coletiva na galiza (pré-) constitucional. In Novas achegas ao estudo da cultura galega II: enfoques socio-históricos e lingüístico-literarios, pages 27–36. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. In Proceedings of the 5th Workshop on Energy Efficient Machine Learning and Cognitive Computing at NeurIPS 2019, Vancouver, Canada. Tal Schuster, Ori Ram, Regina Barzilay, and Amir Globerson. 2019. Cross-lingual alignment of contextual word embeddings, with applications to zeroshot dependency parsing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1599–1613, Minneapolis, Minnesota. Association for Computational Linguistics. Hinrich Schütze. 1998. Automatic word sense discrimination. Computational Linguistics, 24(1):97–123. Fábio Souza, Rodrigo Nogueira, and Roberto Lotufo. 2020. BERTimbau: pretrained BERT models for Brazilian Portuguese. In 9th Brazilian Conference on Intelligent Systems, BRACIS, Rio Grande do Sul, Brazil, October 20-23 (to appear). Milan Straka, Jana Straková, and Jan Hajic. 2019. UDPipe at SIGMORPHON 2019: Contextualized embeddings, regularization with morphological categories, corpora merging. In Proceedings of the 16th Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 95–103, Florence, Italy. Association for Computational Linguistics. David Tuggy. 1993. Ambiguity, polysemy, and vagueness. Cognitive linguistics, 4(3):273–290. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. ArXiv preprint arXiv:1706.03762. Ivan Vuli´c, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. 2020. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222–7240, Online. Association for Computational Linguistics. Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzmán, Armand Joulin, and Edouard Grave. 2020. CCNet: Extracting high quality monolingual datasets from web crawl data. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4003–4012, Marseille, France. European Language Resources Association. Gregor Wiedemann, Steffen Remus, Avi Chawla, and Chris Biemann. 2019. Does BERT Make Any Sense? Interpretable Word Sense Disambiguation with Contextualized Embeddings. In Proceedings of the 15th Conference on Natural Language Processing (KONVENS 2019): Long Papers, pages 161– 170, Erlangen, Germany. German Society for Computational Linguistics & Language Technology. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Lang Yu and Allyson Ettinger. 2020. Assessing phrasal representation and composition in transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4896–4907, Online. Association for Computational Linguistics. 3637 Appendices A Complete results Figure 3 and Table 5 include the results for all languages and models. We also include large variants (BERT and XLM-RoBERTa) when available. For static embeddings, we report results for the best Syn setting, which combines up to three syntactically related words with the target word (see Appendix D). Figure 3: Results across layers and models for English (top), Portuguese (middle), and Spanish (bottom). Sent and WV (dashed) are macro-average values. MacroAvg|Syn is the macro-average per layer (Transformers) and the macro-average of the Syn strategy (fastText). 3638 Model Vec. E1 E2 E3 E4 Ma Mi F E1 E2 E3 E4 Ma Mi F E1 E2 E3 E4 Ma Mi F E1 E2 E3 E4 Ma Mi F Galician Portuguese Spanish English BERT Sent 0.7 0.76 0.75 0.18 0.6 0.62 0.73 0.68 0.43 0.64 0.22 0.49 0.52 0.56 0.76 0.59 0.54 0.19 0.52 0.52 0.6 0.79 0.66 0.74 0.22 0.6 0.6 0.7 Cat 0.71 0.8 0.29 0.42 0.56 0.51 0.7 0.85 0.51 0.38 0.37 0.53 0.5 0.66 0.86 0.7 0.41 0.44 0.6 0.56 0.74 0.96 0.83 0.76 0.43 0.74 0.73 0.84 Add 0.7 0.8 0.28 0.42 0.55 0.51 0.7 0.85 0.54 0.38 0.37 0.54 0.51 0.67 0.86 0.7 0.41 0.44 0.6 0.56 0.74 0.98 0.81 0.76 0.44 0.75 0.73 0.84 BERT2 Sent 0.61 0.6 0.59 0.16 0.49 0.5 0.64 0.68 0.49 0.69 0.22 0.52 0.55 0.6 – – – – – – – 0.89 0.59 0.8 0.27 0.63 0.64 0.7 Cat 0.62 0.71 0.3 0.2 0.46 0.43 0.65 0.95 0.51 0.38 0.46 0.58 0.54 0.68 – – – – – – – 1 0.81 0.78 0.57 0.79 0.78 0.87 Add 0.61 0.71 0.29 0.2 0.45 0.43 0.65 0.95 0.49 0.37 0.46 0.57 0.53 0.68 – – – – – – – 1 0.81 0.8 0.57 0.8 0.78 0.87 mBERT Sent 0.48 0.4 0.49 0.16 0.38 0.39 0.53 0.63 0.43 0.57 0.17 0.45 0.47 0.54 0.51 0.41 0.41 0.19 0.38 0.38 0.5 0.65 0.57 0.77 0.16 0.54 0.55 0.61 Cat 0.57 0.61 0.23 0.22 0.41 0.38 0.62 0.73 0.46 0.16 0.15 0.38 0.34 0.54 0.61 0.45 0.23 0.24 0.38 0.35 0.63 0.83 0.62 0.53 0.27 0.56 0.54 0.73 Add 0.57 0.62 0.21 0.22 0.4 0.37 0.61 0.73 0.49 0.14 0.17 0.38 0.34 0.55 0.63 0.44 0.22 0.25 0.39 0.35 0.63 0.83 0.62 0.54 0.27 0.56 0.54 0.73 XLM-b Sent 0.52 0.51 0.49 0.16 0.42 0.43 0.54 0.51 0.3 0.41 0.2 0.35 0.36 0.45 0.51 0.44 0.46 0.19 0.4 0.41 0.51 0.6 0.62 0.69 0.22 0.53 0.54 0.63 Cat 0.56 0.54 0.22 0.38 0.42 0.39 0.56 0.63 0.46 0.24 0.34 0.42 0.39 0.61 0.67 0.62 0.23 0.54 0.52 0.46 0.69 0.83 0.59 0.23 0.27 0.48 0.43 0.69 Add 0.55 0.54 0.2 0.39 0.42 0.38 0.56 0.63 0.51 0.22 0.34 0.43 0.39 0.61 0.67 0.62 0.23 0.56 0.52 0.47 0.69 0.81 0.55 0.23 0.29 0.47 0.43 0.68 XLM-l Sent 0.42 0.34 0.42 0.16 0.33 0.34 0.44 0.49 0.43 0.35 0.15 0.35 0.35 0.44 0.49 0.48 0.39 0.2 0.39 0.39 0.47 0.54 0.5 0.55 0.22 0.45 0.45 0.58 Cat 0.48 0.5 0.22 0.42 0.4 0.37 0.49 0.73 0.49 0.39 0.32 0.48 0.47 0.58 0.84 0.63 0.46 0.71 0.66 0.62 0.76 0.71 0.6 0.49 0.41 0.55 0.54 0.62 Add 0.46 0.51 0.2 0.43 0.4 0.37 0.5 0.73 0.51 0.38 0.32 0.49 0.47 0.58 0.84 0.66 0.46 0.71 0.67 0.62 0.77 0.73 0.6 0.51 0.46 0.57 0.56 0.64 DmBERT Sent 0.51 0.49 0.5 0.16 0.42 0.43 0.57 0.51 0.43 0.47 0.1 0.38 0.39 0.5 0.51 0.44 0.45 0.12 0.38 0.39 0.51 0.67 0.55 0.79 0.24 0.56 0.58 0.63 Cat 0.52 0.52 0.07 0.24 0.34 0.29 0.51 0.68 0.32 0 0.22 0.31 0.25 0.47 0.61 0.49 0.01 0.34 0.36 0.3 0.53 0.69 0.52 0.24 0.28 0.43 0.4 0.63 Add 0.54 0.56 0.07 0.26 0.36 0.31 0.51 0.71 0.35 0 0.22 0.32 0.26 0.47 0.61 0.52 0.01 0.37 0.38 0.31 0.54 0.69 0.53 0.21 0.37 0.45 0.41 0.63 fastT Sent 0.56 0.69 0.48 0.14 0.47 0.47 0.62 0.61 0.62 0.53 0.17 0.48 0.49 0.55 0.45 0.34 0.45 0.09 0.33 0.35 0.43 0.6 0.5 0.51 0.15 0.44 0.43 0.54 WV 0.21 0.56 0 0.53 0.33 0.29 0.46 0.02 0.54 0 0.63 0.3 0.24 0.45 0.12 0.62 0.02 0.81 0.39 0.35 0.48 0.31 0.55 0.03 0.57 0.37 0.34 0.48 Syn (3) 0.53 0.66 0.2 0.19 0.39 0.36 0.57 0.66 0.46 0.18 0.2 0.37 0.34 0.51 0.37 0.58 0.17 0.24 0.34 0.32 0.55 0.44 0.69 0.23 0.18 0.39 0.36 0.55 Table 5: Complete results for the four languages. BERT are BERT-Base models, and BERT2 refers to a second BERT model for each language (small for Galician, and large for Portuguese and English). XLM-b and XLM-l are XLM-RoBERTa base and large models, respectively. DmBERT is the multilingual version of DistilBERT, and fastT the fastText embeddings. Ma and Mi refer to the macro-average and micro-average results across the four experiments, respectively. F are the micro-average values on the whole dataset. 3639 B Contextualization process (a) Sent. 1: “Ten que haber algún erro nos cálculos porque o resultado non é correcto.” Sent. 2: “Segundo os meus cálculos acabaremos en tres días.” Sent. 3: “Tivo varios cálculos biliares.” (b) Sent. 1: “De sobremesa tomou queixo con marmelo.” Sentence 2: “Fomos a unhas xornadas gastronómicas do queixo.” Sentence 3: “Achegouse a ela e pasoulle a man polo queixo.” (c) Sentence 1: “Eran tantos que parecían un banco de xurelos.” Sent.2: “Desde a rocha víanse pequenos cardumes de robaliza.” Sentence 3: “Este asento de pedra é algo incómodo.” (d) Sent.1: “Apuntou todos os números de teléfono na axenda.” Sentence 2: “Anotou todos os números de teléfono na axenda.” Sentence 3: “Riscou todos os números de teléfono na axenda.”. (e) Sent. 1: “Vai ter lugar a elección da próxima sede dos Xogos Olímpicos.” Sent. 2: “A localización do evento será decidida esta semana.” Sent. 3: “Vou á fonte por auga, que teño sede.” (f) Sentence 1: “Encántalle comer o bolo de pan antes da sopa.” Sentence 2: “O molete tiña a codia un pouco dura.” Sentence 3: “Para atraeren as robalizas iscaban bolo vivo.” Figure 4: Examples in Galician using BERT-base (English translations of the sentences in Appendix E). First row shows examples of Ex1. In Figure 4a cálculos is correctly contextualized since layer 3. In Figure 4b, the outlier sense of queixo is not correctly contextualized in any layer. Second row shows examples of Exp2 (4c) and Exp4 (4d). In Figure 4c, the synonymys banco and cardume are closer to the outlier asento in layer 1 (and from 4 to 7), but the contextualization process is not able to correctly represent the senses in the vector space. In Figure 4d, the result is correct from layer 7 to 11, but in general the representations of words in similar sentences point towards a similar region. Third row incudes examples of Exp3. In Figure 4e, the occurrences of the homonym sede are correctly contextualized as the one in the first sentence approaches its synonym localización in upper layers. The equivalent example of Figure 4f is not adequately solved by the model, as both senses of bolo are notoriously distanct from molete, synonym of the first homonymous sense. 3640 C Galician models Training corpus: We combined the SLI GalWeb (Agerri et al., 2018), CC-100 (Wenzek et al., 2020), the Galician Wikipedia (April 2020 dump), and other news corpora crawled from the web. Following Raffel et al. (2020), sentences with a high ratio of punctuation and symbols, and duplicates were removed. The final corpus has 555M words (633M tokens tokenized with FreeLing (Padró and Stanilovsky, 2012; Garcia and Gamallo, 2010)). The corpus was divided into 90%/10% splits for train and development. fastText model: We trained a fastText skip-gram model for 15 iterations with 300 dimensions, window size of 5, negative sampling of 25, and a minimum word frequency of 5. We used the same 90% split used to train the BERT models, but with automatic tokenization (≈600M tokens). BERT models: We used the 90% train split of the corpus (with the original tokenization) to train two BERT models, with 6 and 12 layers: BERT-small (6 layers): This model has been trained from scratch using a vocabulary of 52,000 (sub-)words and a batch size of 208. It has been training during 1M steps (≈20 epochs) in 14 days. BERT-base (12 layers): Following Kuratov and Arkhipov (2019), we initialized the model from the official pre-trained mBERT, therefore having the same vocabulary size (119,547). We trained it on the Galician corpus during 600k steps (≈13 epochs in 28 days) with a batch size of 198. Both models were trained with the Transformers library (Wolf et al., 2020) on a single NVIDIA Titan XP GPU (12GB), a block size of 128, a learning rate of 0.0001, a masked language modeling (MLM) probability of 0.15, and a weight decay of 0.01. They have been trained only with the MLM objective. D Syntax (Syn method) To get the heads and dependents of each target word we have used the following hierarchies: For nouns: HeadV erb (the head verb, if any)> DepV erb (dependents of the head verb with one of the following relations: obj, nmod, obl)> DepAdj (a dependent adjective)> DepNoun (a dependent noun). For verbs: Head (only if it is a verb or a noun)> Obj (its direct object, if any)> Arg (a dependent with one of these relations: nsubj, nmod, obl). Using these hierarchies we have evaluated representations built by adding from 1 to 4 vectors to the one of each target word. As shown in Table 5, combining 3 syntactically related words to the target one obtains the best results. For the experiments, we have parsed the datasets using the 2.5 Universal Dependencies models provided by UDPipe (Straka et al., 2019). E English translations (Figure 4) Figure 4a, sentence 1: “There must be some error in the calculations because the result is incorrect”. Sentence 2: “According to my calculations we will finish in three days”. Sentence 3: “[He/she] had several gallstones”. Figure 4b, sentence 1: “For dessert [he/she] ate cheese with quince”. Sentence 2: “We went to a cheese gastronomy days”. Sentence 3: “[He/She] approached her and ran his hand over her chin”. Figure 4c, sentence 1: “They were so many that they looked like a school of mackerel”. Sentence 2: “From the rock small shoals of sea bass could be seen”. Sentence 3: “This stone seat is somewhat uncomfortable”. Figure 4d, sentences 1 and 2: “[He/She] wrote down all the phone numbers on the phone book.” Sentence 3: “[He/She] crossed out all the phone numbers on the phone book”. Figure 4e, sentence 1: “The choice of the next venue for the Olympics will take place”. Sentence 2: “The location of the event will be decided this week”. Sentence 3: “I’ll get water from the spring, I am thirsty”. Figure 4f, sentence 1: “[He/She] loves to eat the bread cake before soup”. Sentence 2: “The bread had a slightly hard crust”. Sentence 3: “They used live sand lance to attrack sea bass”.
|
2021
|
281
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3641–3651 August 1–6, 2021. ©2021 Association for Computational Linguistics 3641 Measuring Fine-Grained Domain Relevance of Terms: A Hierarchical Core-Fringe Approach Jie Huang1,3 Kevin Chen-Chuan Chang1,3 Jinjun Xiong2,3 Wen-mei Hwu1,3 1University of Illinois at Urbana-Champaign, USA 2IBM Thomas J. Watson Research Center, USA 3IBM-Illinois Center for Cognitive Computing Systems Research (C3SR), USA {jeffhj, kcchang, w-hwu}@illinois.edu [email protected] Abstract We propose to measure fine-grained domain relevance– the degree that a term is relevant to a broad (e.g., computer science) or narrow (e.g., deep learning) domain. Such measurement is crucial for many downstream tasks in natural language processing. To handle longtail terms, we build a core-anchored semantic graph, which uses core terms with rich description information to bridge the vast remaining fringe terms semantically. To support a finegrained domain without relying on a matching corpus for supervision, we develop hierarchical core-fringe learning, which learns core and fringe terms jointly in a semi-supervised manner contextualized in the hierarchy of the domain. To reduce expensive human efforts, we employ automatic annotation and hierarchical positive-unlabeled learning. Our approach applies to big or small domains, covers head or tail terms, and requires little human effort. Extensive experiments demonstrate that our methods outperform strong baselines and even surpass professional human performance.1 1 Introduction With countless terms in human languages, no one can know all terms, especially those belonging to a technical domain. Even for domain experts, it is quite challenging to identify all terms in the domains they are specialized in. However, recognizing and understanding domain-relevant terms is the basis to master domain knowledge. And having a sense of domains that terms are relevant to is an initial and crucial step for term understanding. In this paper, as our problem, we propose to measure fine-grained domain relevance, which is defined as the degree that a term is relevant to a 1The code and data, along with several term lists with domain relevance scores produced by our methods are available at https://github.com/jeffhj/ domain-relevance. given domain, and the given domain can be broad or narrow– an important property of terms that has not been carefully studied before. E.g., deep learning is a term relevant to the domains of computer science and, more specifically, machine learning, but not so much to others like database or compiler. Thus, it has a high domain relevance for the former domains but a low one for the latter. From another perspective, we propose to decouple extraction and evaluation in automatic term extraction that aims to extract domain-specific terms from texts (Amjadian et al., 2018; H¨atty et al., 2020). This decoupling setting is novel and useful because it is not limited to broad domains where a domain-specific corpus is available, and also does not require terms must appear in the corpus. A good command of domain relevance of terms will facilitate many downstream applications. E.g., to build a domain taxonomy or ontology, a crucial step is to acquire relevant terms (Al-Aswadi et al., 2019; Shang et al., 2020). Also, it can provide or filter necessary candidate terms for domain-focused natural language tasks (Huang et al., 2020). In addition, for text classification and recommendation, the domain relevance of a document can be measured by that of its terms. We aim to measure fine-grained domain relevance as a semantic property of any term in human languages. Therefore, to be practical, the proposed model for domain relevance measuring must meet the following requirements: 1) covering almost all terms in human languages; 2) applying to a wide range of broad and narrow domains; and 3) relying on little or no human annotation. However, among countless terms, only some of them are popular ones organized and associated with rich information on the Web, e.g., Wikipedia pages, which we can leverage to characterize the domain relevance of such “head terms.” In contrast, there are numerous “long-tail terms”– those not as 3642 frequently used– which lack descriptive information. As Challenge 1, how to measure the domain relevance for such long-tail terms? On the other hand, among possible domains of interest, only those broad ones (e.g., physics, computer science) naturally have domain-specific corpora. Many existing works (Velardi et al., 2001; Amjadian et al., 2018; H¨atty et al., 2020) have relied on such domain-specific corpora to identify domain-specific terms by contrasting their distributions to general ones. In contrast, those fine-grained domains (e.g., quantum mechanics, deep learning)– which can be any topics of interest– do not usually have a matching corpus. As Challenge 2, how to achieve good performance for a fine-grained domain without assuming a domain-specific corpus? Finally, automatic learning usually requires large amounts of training data. Since there are countless terms and plentiful domains, human annotation is very time-consuming and laborious. As Challenge 3, how to reduce expensive human efforts when applying machine learning methods to our problem? As our solutions, we propose a hierarchical corefringe domain relevance learning approach that addresses these challenges. First, to deal with longtail terms, we design the core-anchored semantic graph, which includes core terms which have rich description and fringe terms without that information. Based on this graph, we can bridge the domain relevance through term relevance and include any term in evaluation. Second, to leverage the graph and support fine-grained domains without relying on domain-specific corpora, we propose hierarchical core-fringe learning, which learns the domain relevance of core and fringe terms jointly in a semi-supervised manner contextualized in the hierarchy of the domain. Third, to reduce human effort, we employ automatic annotation and hierarchical positive-unlabeled learning, which allow to train our model with little even no human effort. Overall, our framework consists of two processes: 1) the offline construction process, where a domain relevance measuring model is trained by taking a large set of seed terms and their features as input; 2) the online query process, where the trained model can return the domain relevance of query terms by including them in the core-anchored semantic graph. Our approach applies to a wide range of domains and can handle any query, while nearly no human effort is required. To validate the effectiveness of our proposed methods, we conduct extensive experiments on various domains with different settings. Results show our methods significantly outperform well-designed baselines and even surpass human performance by professionals. 2 Related Work The problem of domain relevance of terms is related to automatic term extraction, which aims to extract domain-specific terms from texts automatically. Compared to our task, automatic term extraction, where extraction and evaluation are combined, possesses a limited application and has a relatively large dependence on corpora and human annotation, so it is limited to several broad domains and may only cover a small number of terms. Existing approaches for automatic term extraction can be roughly divided into three categories: linguistic, statistical, and machine learning methods. Linguistic methods apply human-designed rules to identify technical/legal terms in a target corpus (Handler et al., 2016; Ha and Hyland, 2017). Statistical methods use statistical information, e.g., frequency of terms, to identify terms from a corpus (Frantzi et al., 2000; Nakagawa and Mori, 2002; Velardi et al., 2001; Drouin, 2003; Meijer et al., 2014). Machine learning methods learn a classifier, e.g., logistic regression classifier, with manually labeled data (Conrado et al., 2013; Fedorenko et al., 2014; H¨atty et al., 2017). There also exists some work on automatic term extraction with Wikipedia (Vivaldi et al., 2012; Wu et al., 2012). However, terms studied there are restricted to terms associated with a Wikipedia page. Recently, inspired by distributed representations of words (Mikolov et al., 2013a), methods based on deep learning are proposed and achieve state-ofthe-art performance. Amjadian et al. (2016, 2018) design supervised learning methods by taking the concatenation of domain-specific and general word embeddings as input. H¨atty et al. (2020) propose a multi-channel neural network model that leverages domain-specific and general word embeddings. The techniques behind our hierarchical corefringe learning methods are related to research on graph neural networks (GNNs) (Kipf and Welling, 2017; Hamilton et al., 2017); hierarchical text classification (Vens et al., 2008; Wehrmann et al., 2018; Zhou et al., 2020); and positive-unlabeled learning (Liu et al., 2003; Elkan and Noto, 2008; Bekker and Davis, 2020). 3643 few-shot learning quantum chemistry ⋯ 0.877 0.001 ⋯ core fringe query terms domain relevance CFL HiCFL machine learning deep learning few-shot learning quantum mechanics ⋯ seed terms model training graph construction offline online Figure 1: The overview of the framework. In this figure, machine learning is a core term associated with a Wikipedia page, few-shot learning is a fringe term included in the offline core-anchored semantic graph, and quantum chemistry is a fringe term included in the online process. Best viewed in color. 3 Methodology We study the Fine-Grained Domain Relevance of terms, which is defined as follows: Definition 1 (Fine-Grained Domain Relevance) The fine-grained domain relevance of a term is the degree that the term is relevant to a given domain, and the given domain can be broad or narrow. The domain relevance of terms depends on many factors. In general, a term with higher semantic relevance, broader meaning scope, and better usage possesses a higher domain relevance regarding the target domain. To measure the fine-grained domain relevance of terms, we propose a hierarchical corefringe approach, which includes an offline training process and can handle any query term in evaluation. The overview of the framework is illustrated in Figure 1. 3.1 Core-Anchored Semantic Graph There exist countless terms in human languages; thus it is impractical to include all terms in a system initially. To build the offline system, we need to provide seed terms, which can come from knowledge bases or be extracted from broad, large corpora by existing term/phrase extraction methods (Handler et al., 2016; Shang et al., 2018). In addition to providing seed terms, we should also give some knowledge to machines so that they can differentiate whether a term is domain-relevant or not. To this end, we can leverage the description information of terms. For instance, Wikipedia contains a large number of terms (the surface form of page titles), where each term is associated with a Wikipedia article page. With this page information, humans can easily judge whether a term is domain-relevant or not. In Section 3.3, we will show the labeling can even be done completely automatically. However, considering the countless terms, the number of terms that are well-organized and associated with rich description is small. How to measure the fine-grained domain relevance of terms without rich information is quite challenging for both machines and humans. Fortunately, terms are not isolated, while complex relations exist between them. If a term is relevant to a domain, it must also be relevant to some domain-relevant terms and vice versa. This is to say, we can bridge the domain relevance of terms through term relevance. Summarizing the observations, we divide terms into two categories: core terms, which are terms associated with rich description information, e.g., Wikipedia article pages, and fringe terms, which are terms without that information. We assume, for each term, there exist some relevant core terms that share similar domains. If we can find the most relevant core terms for a given term, its domain relevance can be evaluated with the help of those terms. To this end, we can utilize the rich information of core terms for ranking. Taking Wikipedia as an example, each core term is associated with an article page, so they can 3644 be returned as the ranking results (result term) for a given term (query term). Considering the data resources, we use the built-in Elasticsearch based Wikipedia search engine2 (Gormley and Tong, 2015). More specifically, we set the maximum number of links as k (5 as default). For a query term v, i.e., any seed term, we first achieve the top 2k Wikipedia pages with exact match. For each result term u in the core, we create a link from u to v. If the number of links is smaller than k, we do this process again without exact match and build additional links. Finally, we construct a term graph, named Core-Anchored Semantic Graph, where nodes are terms and edges are links between terms. In addition, for terms that are not provided initially, we can also handle them as fringe terms and connect them to core terms in evaluation. In this way, we can include any term in the graph. 3.2 Hierarchical Core-Fringe Learning In this section, we aim to design learning methods to learn the fine-grained domain relevance of core and fringe terms jointly. In addition to using the term graph, we can achieve features of both core and fringe terms based on their linguistic and statistical properties (Terryn et al., 2019; Conrado et al., 2013) or distributed representations (Mikolov et al., 2013b; Yu and Dredze, 2015). We assume the labels, i.e., domain-relevant or not, of core terms are available, which can be achieved by an automatic annotation mechanism introduced in Section 3.3. As stated above, if a term is highly relevant to a given domain, it must also be highly relevant to some other terms with a high domain relevance and vice versa. Therefore, to measure the domain relevance of a term, in addition to using its own features, we aggregate its neighbors’ features. Specifically, we propagate the features of terms via the term graph and use the label information of core terms for supervision. In this way, core and fringe terms help each other, and the domain relevance is learned jointly. The propagation process can be achieved by graph convolutions (Hammond et al., 2011). We first apply the vanilla graph convolutional networks (GCNs) (Kipf and Welling, 2017) in our framework. The graph convolution operation (GCNConv) at the l-th layer is formulated as the 2https://en.wikipedia.org/w/index.php? search following aggregation and update process: h(l+1) i = φ X j∈Ni∪{i} 1 cij W (l) c h(l) j + b(l) c , (1) where Ni is the neighbor set of node i. cij is the normalization constant. h(l) j ∈Rd(l)×1 is the hidden state of node j at the l-th layer, with d(l) being the number of units; h(0) j = xj, which is the feature vector of node j. W (l) c ∈Rd(l+1)×d(l) is the trainable weight matrix at the l-th layer, and b(l) c is the bias vector. φ(·) is the nonlinearity activation function, e.g., ReLU(·) = max(0, ·). Since core terms are labeled as domain-relevant or not, we can use the labels to calculate the loss: L = − X i∈Vcore (yi log zi + (1 −yi) log(1 −zi)), (2) where yi is the label of node i regarding the target domain, and zi = σ(ho i ), with ho i being the output of the last GCNConv layer for node i and σ(·) being the sigmoid function. The weights of the model are trained by minimizing the loss. The relative domain relevance is obtained as s = z. Combining with the overall framework, we get the first domain relevance measuring model, CFL, i.e., Core-Fringe Domain Relevance Learning. CFL is useful to measure the domain relevance for broad domains such as computer science. For domains with relatively narrow scopes, e.g., machine learning, we can also leverage the label information of domains at the higher level of the hierarchy, e.g., CS →AI →ML, which is based on the idea that a domain-relevant term regarding the target domain should also be relevant to the parent domain. Inspired by related work on hierarchical multi-label classification (Vens et al., 2008; Wehrmann et al., 2018), we introduce a hierarchical learning method considering both global and local information. We first apply lc GCNConv layers according to Eq. (1) and get the output of the last GCNConv layer, which is h(lc) i . In order not to confuse, we omit the subscript that identifies the node number. For each domain in the hierarchy, we introduce a hierarchical global activation ap. The activation at the (l + 1)-th level of the hierarchy is given as a(l+1) p = φ(W (l) p [a(l) p ; h(lc)] + b(l) p ), (3) where [·; ·] indicates the concatenation of two vectors; a(1) p = φ(W (0) p h(lc) + b(0) p ). The global in3645 formation is produced after a fully connected layer: zp = σ(W (lp) p a(lp) p + b(lp) p ), (4) where lp is the total number of hierarchical levels. To achieve the local information for each level of the hierarchy, the model first generates the local hidden state a(l) q by a fully connected layer: a(l) q = φ(W (l) t a(l) p + b(l) t ). (5) The local information at the l-th level of the hierarchy is then produced as z(l) q = σ(W (l) q a(l) q + b(l) q ). (6) In our core-fringe framework, all the core terms are labeled at each level of the hierarchy. Therefore, the loss of hierarchical learning is computed as Lh = ϵ(zp, y(lp)) + lp X l=1 ϵ(z(l) q , y(l)), (7) where y(l) denotes the labels regarding the domain at the l-th level of the hierarchy and ϵ(z, y) is the binary cross-entropy loss described in Eq. (2). In testing, The relative domain relevance s is calculated as s = α · zp + (1 −α) · (z(1) q ◦z(2) q , ..., z(lp) q ), (8) where ◦denotes element-wise multiplication. α is a hyperparameter to balance the global and local information (0.5 as default). Combining with our general framework, we refer to this model as HiCFL, i.e., Hierarchical CFL. Online Query Process. If seed terms are provided by extracting from broad, large corpora relevant to the target domain, most terms of interest will be already included in the offline process. In evaluation, for terms that are not provided initially, our model treats them as fringe terms. Specifically, when receiving such a term, the model connects it to core terms by the method described in Section 3.1. With its features (e.g., compositional term embeddings) or only its neighbors’ features (when features cannot be generated directly), the trained model can return the domain relevance of any query. 3.3 Automatic Annotation and Hierarchical Positive-Unlabeled Learning Automatic Annotation. For the fine-grained domain relevance problem, human annotation is very time-consuming and laborious because the number of core terms is very large regarding a wide range of domains. Fortunately, in addition to building the term graph, we can also leverage the rich information of core terms for automatic annotation. In the core-anchored semantic graph constructed with Wikipedia, each core term is associated with a Wikipedia page, and each page is assigned one or more categories. All the categories form a hierarchy, furthermore providing a category tree. For a given domain, we can first traverse from a root category and collect some gold subcategories. For instance, for computer science, we treat category: subfields of computer science3 as the root category and take categories at the first three levels of it as gold subcategories. Then we collect categories for each core term and examine whether the term itself or one of the categories is a gold subcategory. If so, we label the term as positive. Otherwise, we label it as negative. We can also combine gold subcategories from some existing domain taxonomies and extract the categories of core terms from the text description, which usually contains useful text patterns like “x is a subfield of y”. Hierarchical Positive-Unlabeled Learning. According to the above methods, we can learn the finegrained domain relevance of terms for any domain as long as we can collect enough gold subcategories for that domain. However, for domains at the low level of the hierarchy, e.g., deep learning, a category tree might not be available in Wikipedia. To deal with this issue, we apply our learning methods in a positive-unlabeled (PU) setting (Bekker and Davis, 2020), where only a small number of terms, e.g., 10, are labeled as positive, and all the other terms are unlabeled. We use this setting based on the following consideration: if a user is interested in a specific domain, it is quite easy for her to give some important terms relevant to that domain. Benefiting from our hierarchical core-fringe learning approach, we can still obtain labels for domains at the high level of the hierarchy with the automatic annotation mechanism. Therefore, all the negative examples of the last labeled hierarchy can be used as reliable negatives for the target domain. For instance, if the target domain is deep learning, which is in the CS →AI →ML →DL hierarchy, we consider all the non-ML terms as the reliable negatives for DL. Taking the positively 3https://en.wikipedia.org/wiki/ Category:Subfields_of_computer_science 3646 labeled examples and the reliable negatives for supervision, we can learn the domain relevance of terms by our proposed HiCFL model contextualized in the hierarchy of the domain. 4 Experiments In this section, we evaluate our model from different perspectives. 1) We compare with baselines by treating some labeled terms as queries. 2) We compare with human professionals by letting humans and machines judge which term in a query pair is more relevant to a target domain. 3) We conduct intuitive case studies by ranking terms according to their domain relevance. 4.1 Experimental Setup Datasets and Preprocessing. To build the system, for offline processing, we extract seed terms from the arXiv dataset (version 6)4. As an example, for computer science or its sub-domains, we collect the abstracts in computer science according to the arXiv Category Taxonomy5, and apply phrasemachine to extract terms (Handler et al., 2016) with lemmatization and several filtering rules: frequency > 10; length ≤6; only contain letters, numbers, and hyphen; not a stopword or a single letter. We select three broad domains, including computer science (CS), physics (Phy), and mathematics (Math); and three narrow sub-domains of them, including machine learning (ML), quantum mechanics (QM), and abstract algebra (AA), with the hierarchies CS →AI →ML, Phy →mechanics → QM, and Math →algebra →AA. Each broad domain and its sub-domains share seed terms because they share a corpus. To achieve gold subcategories for automatic annotation (Section 3.3), we collect subcategories at the first three levels of a root category (e.g., category: subfields of physics) for broad domains (e.g., physics); or the first two levels for narrow domains, e.g., category: machine learning for machine learning. Table 1 reports the total sizes and the ratios that are core terms. Baselines. Since our task on fine-grained domain relevance is new, there is no existing baseline for model comparison. We adapt the following models on relevant tasks in our setting with additional inputs (e.g., domain-specific corpora): 4https://www.kaggle.com/ Cornell-University/arxiv 5https://arxiv.org/category_taxonomy domain #terms core ratio CS ML 113,038 27.7% Phy QM 416,431 12.1% Math AA 103,984 26.4% Table 1: The statistics of the data. • Relative Domain Frequency (RDF): Since domain-relevant terms usually occur more in a domain-specific corpus, we apply a statistical method using freqs(w)/freqg(w) to measure the domain relevance of term w, where freqs(·) and freqg(·) denote the frequency of occurrence in the domain-specific/general corpora respectively. • Logistic Regression (LR): Logistic regression is a standard supervised learning method. We use core terms with labels (domain-relevant or not) as training data, where features are term embeddings trained by a general corpus. • Multilayer Perceptron (MLP): MLP is a standard neural neural-based model. We train MLP using embeddings trained with a domain-specific corpus or a general corpus as term features, respectively. We also concatenate the two embeddings as features (Amjadian et al., 2016, 2018). • Multi-Channel (MC): Multi-Channel (H¨atty et al., 2020) is the state-of-the-art model for automatic term extraction, which is based on a multi-channel neural network that takes domainspecific and general corpora as input. Training. For all supervised learning methods, we apply automatic annotation in Section 3.3, i.e., we automatically label all the core terms for model training. In the PU setting, we remove labels on target domains. Only 20 (10 in the case studies) domain-relevant core terms are randomly selected as the positives, with the remaining terms unlabeled. In training, all the negative examples at the previous level of the hierarchy are used as reliable negatives. Implementation Details. Though our proposed methods are independent of corpora, some baselines (e.g., MC) require term embeddings trained from general/domain-specific corpora. For easy and fair comparison, we adopt the following approach to generate term features. We consider each term as a single token, and apply word2vec CBOW (Mikolov et al., 2013a) with negative sampling, where dimensionality is 100, window size is 5, and number of negative samples is 5. The training cor3647 Computer Science Physics Mathematics ROC-AUC PR-AUC ROC-AUC PR-AUC ROC-AUC PR-AUC RDF SG 0.714 0.417 0.736 0.496 0.694 0.579 LR G 0.802±0.000 0.535±0.000 0.822±0.000 0.670±0.000 0.854±0.000 0.769±0.000 MLP S 0.819±0.003 0.594±0.003 0.853±0.001 0.739±0.004 0.868±0.000 0.803±0.001 MLP G 0.863±0.001 0.674±0.002 0.874±0.001 0.761±0.003 0.904±0.001 0.846±0.002 MLP SG 0.867±0.001 0.667±0.002 0.875±0.001 0.765±0.002 0.904±0.001 0.843±0.003 MC SG 0.868±0.002 0.664±0.006 0.877±0.003 0.768±0.004 0.903±0.001 0.843±0.002 CFL G 0.885±0.001 0.712±0.002 0.905±0.000 0.812±0.002 0.918±0.001 0.870±0.002 CFL C 0.883±0.001 0.708±0.002 0.901±0.000 0.800±0.001 0.919±0.001 0.879±0.002 S and G indicate the corpus used. S: domain-specific corpus, G: general corpus, SG: both. C means the pre-trained compositional GloVe embeddings are used. Table 2: Results for broad domains. pus can be a general one (the entire arXiv corpus, denoted as G), or a domain-specific one (the subcorpus in the branch of the corresponding domain, denoted as S). We also apply compositional GloVe embeddings (Pennington et al., 2014) (elementwise addition of the pre-trained 100d word embeddings, denoted as C) as non-corpus-specific features of terms for reference. For all the neural network-based models, we use Adam (Kingma and Ba, 2015) with learning rate of 0.01 for optimization, and adopt a fixed hidden dimensionality of 256 and a fixed dropout ratio of 0.5. For the learning part of CFL and HiCFL, we apply two GCNConv layers and use the symmetric graph for training. To avoid overfitting, we adopt batch normalization (Ioffe and Szegedy, 2015) right after each layer (except for the output layer) and before activation and apply dropout (Hinton et al., 2012) after the activation. We also try to add regularizations for MLP and MC with full-batch or mini-batch training, and select the best architecture. To construct the core-anchored semantic graph, we set k as 5. All experiments are run on an NVIDIA Quadro RTX 5000 with 16GB of memory under the PyTorch framework. The training of CFL for the CS domain can finish in 1 minute. We report the mean and standard deviation of the test results corresponding to the best validation results with 5 different random seeds. 4.2 Comparison to Baselines To compare with baselines, we separate a portion of core terms as queries for evaluation. Specifically, for each domain, we use 80% labeled terms for training, 10% for validation, and 10% for testing (with automatic annotation). Terms in the validation and testing sets are treated as fringe terms. By doing this, the evaluation can represent the general performance for all fringe terms to some extent. And the model comparison is fair since the rich information of terms for evaluation is not used in training. We also create a test set with careful human annotation on machine learning to support our overall evaluation, which contains 2000 terms, with half for evaluation and half for testing. As evaluation metrics, we calculate both ROCAUC and PR-AUC with automatic or manually created labels. ROC-AUC is the area under the receiver operating characteristic curve, and PR-AUC is the area under the precision-recall curve. If a model achieves higher values, most of the domainrelevant terms are ranked higher, which means the model has a better measurement on the domain relevance of terms. Table 2 and Table 3 show the results for three broad/narrow domains respectively. We observe our proposed CFL and HiCFL outperform all the baselines, and the standard deviations are low. Compared to MLP, CFL achieves much better performance benefiting from the core-anchored semantic graph and feature aggregation, which demonstrates the domain relevance can be bridged via term relevance. Compared to CFL, HiCFL works better owing to hierarchical learning. In the PU setting– the situation when automatic annotation is not applied to the target domain, although only 20 positives are given, HiCFL still achieves satisfactory performance and significantly outperforms all the baselines (Table 4). The PR-AUC scores on the manually created test 3648 Machine Learning Quantum Mechanics Abstract Algebra ROC-AUC PR-AUC ROC-AUC PR-AUC ROC-AUC PR-AUC LR G 0.917±0.000 0.346±0.000 0.879±0.000 0.421±0.000 0.872±0.000 0.525±0.000 MLP S 0.902±0.001 0.453±0.009 0.903±0.001 0.545±0.004 0.910±0.000 0.641±0.007 MLP G 0.932±0.001 0.562±0.010 0.922±0.001 0.587±0.014 0.923±0.000 0.658±0.006 MLP SG 0.928±0.001 0.574±0.011 0.923±0.000 0.574±0.007 0.925±0.001 0.673±0.004 MC SG 0.928±0.002 0.554±0.007 0.924±0.001 0.590±0.003 0.924±0.001 0.685±0.005 CFL G 0.950±0.002 0.627±0.013 0.950±0.000 0.678±0.003 0.938±0.001 0.751±0.009 HiCFL G 0.965±0.003 0.645±0.014 0.957±0.001 0.691±0.003 0.942±0.002 0.769±0.006 S and G indicate the corpus used. S: domain-specific corpus, G: general corpus, SG: both. Table 3: Results for narrow domains. Machine Learning Quantum Mechanics Abstract Algebra ROC-AUC PR-AUC ROC-AUC PR-AUC ROC-AUC PR-AUC LR G 0.860±0.000 0.206±0.000 0.788±0.000 0.280±0.000 0.833±0.000 0.429±0.000 MLP S 0.804±0.003 0.144±0.003 0.767±0.009 0.260±0.005 0.804±0.006 0.421±0.010 MLP G 0.836±0.005 0.234±0.016 0.813±0.006 0.295±0.011 0.842±0.003 0.467±0.011 MLP SG 0.844±0.003 0.230±0.015 0.796±0.008 0.291±0.011 0.839±0.006 0.463±0.013 MC SG 0.852±0.006 0.251±0.019 0.795±0.014 0.303±0.017 0.861±0.004 0.547±0.006 CFL G 0.918±0.001 0.441±0.009 0.897±0.002 0.408±0.004 0.887±0.002 0.563±0.018 HiCFL G 0.940±0.008 0.508±0.026 0.897±0.004 0.421±0.014 0.915±0.002 0.648±0.009 Table 4: Results for narrow domains (PU learning). PR-AUC PR-AUC (PU) LR G 0.509±0.000 0.449±0.000 MLP S 0.550±0.017 0.113±0.010 MLP G 0.586±0.016 0.299±0.027 MLP SG 0.590±0.005 0.217±0.013 MC SG 0.603±0.016 0.281±0.012 CFL G 0.703±0.017 0.525±0.013 HiCFL G 0.755±0.011 0.581±0.036 Table 5: Results (PR-AUC) for machine learning with manual labeling. set without and with the PU setting are reported in Table 5. We observe that the results are generally consistent with results reported in Table 3 and Table 4, which indicates the evaluation with core terms can work just as well. 4.3 Comparison to Human Performance In this section, we aim to compare our model with human professionals in measuring the fine-grained domain relevance of terms. Because it is difficult for humans to assign a score representing doML-AI ML-CS AI-CS Human 0.698±0.087 0.846±0.074 0.716±0.115 HiCFL 0.854±0.017 0.932±0.007 0.768±0.023 Table 6: Accuracies of domain relevance comparison. main relevance directly, we generate term pairs as queries and let humans judge which one in a pair is more relevant to machine learning. Specifically, we create 100 ML-AI, ML-CS, and AI-CS pairs respectively. Taking ML-AI as an example, each query pair consists of an ML term and an AI term, and the judgment is considered right if the ML term is selected. The human annotation is conducted by five senior students majoring in computer science and doing research related to terminology. Because there is no clear boundary between ML, AI, and CS, it is possible that a CS term is more relevant to machine learning than an AI term. However, the overall trend is that the higher the accuracy, the better the performance. From Table 6, we observe that HiCFL far outperforms human performance. 3649 The depth of the background color indicates the domain relevance. The darker the color, the higher the domain relevance (annotated by the authors); * indicates the term is a core term, otherwise it is a fringe term. 1-10 101-110 1001-1010 10001-10010 100001-100010 supervised learning* adversarial machine learning* regularization strategy method for detection tumor region convolutional neural network* temporal-difference learning* weakly-supervised approach gait parameter mutual trust machine learning* restricted boltzmann machine learned embedding stochastic method inherent problem deep learning* backpropagation through time* node classification problem recommendation diversity healthcare system* semi-supervised learning* svms non-convex learning numerical experiment two-phase* q-learning* word2vec* sample-efficient learning second-order method posetrack reinforcement learning* rbms cnn-rnn model landmark dataset half* unsupervised learning* hierarchical clustering* deep bayesian general object detection mfcs recurrent neural network* stochastic gradient descent* classification score cold-start recommendation borda count* generative adversarial network* svm* classification algorithm* similarity of image diverse way Table 7: Ranking results for machine learning with HiCFL. Given positives (10): deep learning, neural network, deep neural network, deep reinforcement learning, multilayer perceptron, convolutional neural network, recurrent neural network, long short-term memory, backpropagation, activation function. 1-10 101-110 1001-1010 10001-10010 100001-100010 convolutional neural network* discriminative loss multi-task deep learning low light image law enforcement agency* recurrent neural network* dropout regularization self-supervision face dataset case of channel artificial neural network* semantic segmentation* state-of-the-art deep learning algorithm estimation network release* feedforward neural network* mask-rcnn generative probabilistic model method on benchmark datasets ahonen* deep learning* probabilistic neural network* translation model distributed constraint electoral control neural network* pretrained network probabilistic segmentation gradient information runge* generative adversarial network* discriminator model handwritten digit classification model on a variety many study multilayer perceptron* sequence-to-sequence learning deep learning classification model constraint mean value* long short-term memory* autoencoders multi-task reinforcement learning automatic detection efficient beam neural architecture search* conditional variational autoencoder skip-gram* feature redundancy pvt* Table 8: Ranking results for deep learning with HiCFL (PU learning). Although we have reduced the difficulty, the task is still very challenging for human professionals. 4.4 Case Studies We interpret our results by ranking terms according to their domain relevance regarding machine learning or deep learning, with hierarchy CS → AI →ML →DL. For CS-ML, we label terms with automatic annotation. For DL, we create 10 DL terms manually as the positives for PU learning. Table 7 and Table 8 show the ranking results (1-10 represents terms ranked 1st to 10th). We observe the performance is satisfactory. For ML, important concepts such as supervised learning, unsupervised learning, and deep learning are ranked very high. Also, terms ranked before 1010th are all good domain-relevant terms. For DL, although only 10 positives are provided, the ranking results are quite impressive. E.g., unlabeled positive terms like artificial neural network, generative adversarial network, and neural architecture search are ranked very high. Besides, terms ranked 101st to 110th are all highly relevant to DL, and terms ranked 1001st to 1010th are related to ML. 5 Conclusion We introduce and study the fine-grained domain relevance of terms– an important property of terms that has not been carefully studied before. We propose a hierarchical core-fringe domain relevance learning approach, which can cover almost all terms in human languages and various domains, while requires little or even no human annotation. We believe this work will inspire an automated solution for knowledge management and help a wide range of downstream applications in natural language processing. It is also interesting to integrate our methods to more challenging tasks, for example, to characterize more complex properties of terms even understand terms. Acknowledgments We thank the anonymous reviewers for their valuable comments and suggestions. This material is based upon work supported by the National Science Foundation IIS 16-19302 and IIS 16-33755, Zhejiang University ZJU Research 083650, IBMIllinois Center for Cognitive Computing Systems Research (C3SR) - a research collaboration as part of the IBM Cognitive Horizon Network, grants from eBay and Microsoft Azure, UIUC OVCR CCIL Planning Grant 434S34, UIUC CSBS Small Grant 434C8U, and UIUC New Frontiers Initiative. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the funding agencies. 3650 References Fatima N Al-Aswadi, Huah Yong Chan, and Keng Hoon Gan. 2019. Automatic ontology construction from text: a review from shallow to deep learning trend. Artificial Intelligence Review, pages 1–28. Ehsan Amjadian, Diana Inkpen, T Sima Paribakht, and Farahnaz Faez. 2018. Distributed specificity for automatic terminology extraction. Terminology. International Journal of Theoretical and Applied Issues in Specialized Communication, 24(1):23–40. Ehsan Amjadian, Diana Inkpen, Tahereh Paribakht, and Farahnaz Faez. 2016. Local-global vectors to improve unigram terminology extraction. In Proceedings of the 5th International Workshop on Computational Terminology, pages 2–11. Jessa Bekker and Jesse Davis. 2020. Learning from positive and unlabeled data: a survey. Mach. Learn., 109(4):719–760. Merley Conrado, Thiago Pardo, and Solange Oliveira Rezende. 2013. A machine learning approach to automatic term extraction using a rich feature set. In Proceedings of the 2013 NAACL HLT Student Research Workshop, pages 16–23. Patrick Drouin. 2003. Term extraction using nontechnical corpora as a point of leverage. Terminology, 9(1):99–115. Charles Elkan and Keith Noto. 2008. Learning classifiers from only positive and unlabeled data. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 213–220. Denis Fedorenko, N Astrakhantsev, and D Turdakov. 2014. Automatic recognition of domain-specific terms: an experimental evaluation. Proceedings of the Institute for System Programming, 26(4):55–72. Katerina Frantzi, Sophia Ananiadou, and Hideki Mima. 2000. Automatic recognition of multi-word terms:. the c-value/nc-value method. International journal on digital libraries, 3(2):115–130. Clinton Gormley and Zachary Tong. 2015. Elasticsearch: the definitive guide: a distributed real-time search and analytics engine. ” O’Reilly Media, Inc.”. Althea Ying Ho Ha and Ken Hyland. 2017. What is technicality? a technicality analysis model for eap vocabulary. Journal of English for Academic Purposes, 28:35–49. William L Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 1025–1035. David K Hammond, Pierre Vandergheynst, and R´emi Gribonval. 2011. Wavelets on graphs via spectral graph theory. Applied and Computational Harmonic Analysis, 30(2):129–150. Abram Handler, Matthew Denny, Hanna Wallach, and Brendan O’Connor. 2016. Bag of what? simple noun phrase extraction for text analysis. In Proceedings of the First Workshop on NLP and Computational Social Science, pages 114–124. Anna H¨atty, Michael Dorna, and Sabine Schulte im Walde. 2017. Evaluating the reliability and interaction of recursively used feature classes for terminology extraction. In Proceedings of the student research workshop at the 15th conference of the European chapter of the association for computational linguistics, pages 113–121. Anna H¨atty, Dominik Schlechtweg, Michael Dorna, and Sabine Schulte im Walde. 2020. Predicting degrees of technicality in automatic terminology extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2883–2889. Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. 2012. Improving neural networks by preventing coadaptation of feature detectors. arXiv preprint arXiv:1207.0580. Jie Huang, Zilong Wang, Kevin Chen-Chuan Chang, Wen-mei Hwu, and Jinjun Xiong. 2020. Exploring semantic capacity of terms. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Sergey Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pages 448–456. Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. Proceedings of the 3rd International Conference on Learning Representations. Thomas N. Kipf and Max Welling. 2017. Semisupervised classification with graph convolutional networks. In Proceedings of International Conference on Learning Representations. Bing Liu, Yang Dai, Xiaoli Li, Wee Sun Lee, and Philip S Yu. 2003. Building text classifiers using positive and unlabeled examples. In Third IEEE International Conference on Data Mining, pages 179– 186. IEEE. Kevin Meijer, Flavius Frasincar, and Frederik Hogenboom. 2014. A semantic approach for extracting domain taxonomies from text. Decision Support Systems, 62:78–93. 3651 Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119. Hiroshi Nakagawa and Tatsunori Mori. 2002. A simple but powerful automatic term extraction method. In COLING-02: COMPUTERM 2002: Second International Workshop on Computational Terminology. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Chao Shang, Sarthak Dash, Md Faisal Mahbub Chowdhury, Nandana Mihindukulasooriya, and Alfio Gliozzo. 2020. Taxonomy construction of unseen domains via graph-based cross-domain knowledge transfer. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2198–2208. Jingbo Shang, Jialu Liu, Meng Jiang, Xiang Ren, Clare R Voss, and Jiawei Han. 2018. Automated phrase mining from massive text corpora. IEEE Transactions on Knowledge and Data Engineering, 30(10):1825–1837. Ayla Rigouts Terryn, V´eronique Hoste, and Els Lefever. 2019. In no uncertain terms: a dataset for monolingual and multilingual automatic term extraction from comparable corpora. Language Resources and Evaluation, pages 1–34. Paola Velardi, Michele Missikoff, and Roberto Basili. 2001. Identification of relevant terms to support the construction of domain ontologies. In Proceedings of the ACL 2001 Workshop on Human Language Technology and Knowledge Management. Celine Vens, Jan Struyf, Leander Schietgat, Saˇso Dˇzeroski, and Hendrik Blockeel. 2008. Decision trees for hierarchical multi-label classification. Machine learning, 73(2):185. Jorge Vivaldi, Luis Adri´an Cabrera-Diego, Gerardo Sierra, and Mar´ıa Pozzi. 2012. Using wikipedia to validate the terminology found in a corpus of basic textbooks. In LREC, pages 3820–3827. Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. 2018. Hierarchical multi-label classification networks. In International Conference on Machine Learning, pages 5075–5084. Wenjuan Wu, Tao Liu, He Hu, and Xiaoyong Du. 2012. Extracting domain-relevant term using wikipedia based on random walk model. In 2012 Seventh ChinaGrid Annual Conference, pages 68–75. IEEE. Mo Yu and Mark Dredze. 2015. Learning composition models for phrase embeddings. Transactions of the Association for Computational Linguistics, 3:227– 242. Jie Zhou, Chunping Ma, Dingkun Long, Guangwei Xu, Ning Ding, Haoyu Zhang, Pengjun Xie, and Gongshen Liu. 2020. Hierarchy-aware global model for hierarchical text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1106–1117.
|
2021
|
282
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3652–3665 August 1–6, 2021. ©2021 Association for Computational Linguistics 3652 HERALD: An Annotation Efficient Method to Detect User Disengagement in Social Conversations Weixin Liang1 Stanford University [email protected] Kai-Hui Liang1 Columbia University [email protected] Zhou Yu Columbia University [email protected] Abstract Open-domain dialog systems have a usercentric goal: to provide humans with an engaging conversation experience. User engagement is one of the most important metrics for evaluating open-domain dialog systems, and could also be used as real-time feedback to benefit dialog policy learning. Existing work on detecting user disengagement typically requires hand-labeling many dialog samples. We propose HERALD, an efficient annotation framework that reframes the training data annotation process as a denoising problem. Specifically, instead of manually labeling training samples, we first use a set of labeling heuristics to label training samples automatically. We then denoise the weakly labeled data using the Shapley algorithm. Finally, we use the denoised data to train a user engagement detector. Our experiments show that HERALD improves annotation efficiency significantly and achieves 86% user disengagement detection accuracy in two dialog corpora. Our implementation is available at https:// github.com/Weixin-Liang/HERALD/. 1 Introduction Evaluation metrics heavily influence a field’s research direction. The ultimate goal of open-domain dialog systems is to provide an enjoyable experience to users. Previous research mainly focuses on optimizing automatic dialog evaluation metrics such as BLEU, which models the distance between the system responses and a limited number of references available. However, it has been shown that these metrics correlate poorly with human judgments (Liu et al., 2016). Open-domain dialog system evaluation has long been one of the most difficult challenges in the dialog community for several reasons: (1) The goal of 1Equal Contribution. dialog evaluation should be to evaluate users’ conversational experience. Existing automatic evaluation metrics such as BLEU are mostly constrained to a static corpus, and do not capture the user experience in a realistic interactive setting. (2) Currently, self-reported user ratings are widely used to evaluate open-domain dialogs. However, self-reported ratings suffer from bias and variance among different users (Liang et al., 2020e). Although we could tell which dialog system is better by running statistical tests on a large number of noisy ratings, it is challenging to locate dialogs with bad performance reliably. Only by identifying these bad dialogs effectively can we correct errors in these samples to improve dialog system quality. User engagement has been recognized as one of the essential metrics for open-domain dialog evaluation (Ram et al., 2018). Previous research also confirms that incorporating user engagement as real-time feedback benefits dialog policy learning (Yu et al., 2016). One of the most costly bottlenecks of learning to detect user disengagement is to annotate many turn-level user engagement labels (Ghazarian et al., 2020). In addition, the data annotation process becomes more expensive and challenging for privacy-sensitive dialog corpora, due to the privacy concerns in crowdsourcing (Xia and McKernan, 2020). To improve annotation efficiency, we reframe the training data annotation process as a denoising problem. Specifically, instead of manually labeling each training datum, we automatically label the training samples with a set of labeling heuristics. The heuristic functions primarily consist of regular expressions (Regexes) and incorporate open-sourced natural language understanding (NLU) services. Since the automatically generated labels might contain noise, we then denoise the labeled data using the Shapley algorithm (Jia et al., 2019a,b). We use the Shapley algorithm to 3653 quantify the contribution of each training datum, so that we can identify the noisy data points with negative contribution and then correct their labels. Our experiments show that HERALD achieves 86% accuracy in user disengagement detection in two dialog corpora. Our proposed framework HERALD is conceptually simple and suitable for a wide range of application scenarios: First, since our model could detect user engagement in real-time (i.e., after each user utterance), our model could be plugged into existing dialog systems as a real-time user experience monitor module. In this way, dialog systems could detect and react to user’s disengagement in both open-domain dialogs (Yu et al., 2016) and taskoriented dialogs (Yu et al., 2017). During training, our model could also be used as real-time feedback to benefit dialog policy learning (Yi et al., 2019). Second, HERALD could quantify user engagement and be used as an automatic dialog evaluation metric. It could locate dialogs with poor user experience reliably to improve dialog system quality (Ghazarian et al., 2020; Choi et al., 2019). Third, user engagement is an essential objective of dialog systems, but few dialog datasets with user engagement ratings are available. Our heuristic functions, combined with the proposed workflow, can be readily deployed to annotate new dialog datasets. 2 Related Work 2.1 Open-Domain Dialog System Evaluation Open-domain dialog system evaluation is a longlasting challenge. It has been shown that existing automatic dialog evaluation metrics correlate poorly with human judgments (Liu et al., 2016; Lowe et al., 2017; Novikova et al., 2017). A wellknown reason is that these automatic dialog evaluation metrics rely on modeling the distance between the generated response and a limited number of references available. The fundamental gap between the open-ended nature of the conversations and the limited references (Gupta et al., 2019) is not addressed in methods that are lexical-level based (Papineni et al., 2002; Lin, 2004; Banerjee and Lavie, 2005), embedding based (Rus and Lintean, 2012; Forgues et al., 2014), perplexity based (Adiwardana et al., 2020), or learning based (Tao et al., 2018; Lowe et al., 2017). Mehri and Eskénazi (2020) simulate user response using DialogGPT and evaluate the probability of user complaint. Given the limitations above, self-reported user ratings are widely used to evaluate open-domain dialogs. However, self-reported ratings suffer from bias and variance among different users (Venkatesh et al., 2018). Denoising human ratings is still an open research problem (Liang et al., 2020e; Li et al., 2019). 2.2 User Engagement in Dialogs User engagement is commonly defined as the user’s willingness to continue conversing with the dialog system (Yu et al., 2016, 2017). Existing work on measuring user engagement primarily resorts to human rating (Yi et al., 2019; Hancock et al., 2019), or proxy metrics. Example proxy metrics include conversation length like number of dialog turns (Venkatesh et al., 2018; Ram et al., 2018), and conversational breadth like topical diversity (Guo et al., 2018). Sporadic attempts have been made to detecting user disengagement in dialogs (Yu et al., 2004; Ghazarian et al., 2020; Choi et al., 2019). A major bottleneck of these methods is that they require hand-labeling many dialog samples for individual datasets. Although Liang et al. (2020e) denoise user self-reported ratings with the Shapley algorithm for dialog system evaluation, their method cannot be directly applied to dialogs without user ratings as in our setting. Our work is focusing on the problem that it is expensive and difficult to obtain user ratings. The core insight of our work is to reframe the training data annotation process as a process of denoising labels created by heuristic functions pre-defined. To the best of our knowledge, we are the first to combine automatic data labeling with the Shapley algorithm to perform dialog evaluation. Our method could potentially generalize to other classification tasks if different weak labelers are provided. 2.3 Learning from Weak Supervision Learning from weak supervision reduces annotation costs by utilizing noisy but cost-efficient labels (Ratner et al., 2020, 2016; Liang et al., 2020e). One of the most popular forms of weak supervision is distant supervision, in which the records of an external knowledge base are heuristically aligned with data points to produce noisy labels for relationship extraction tasks (Bunescu and Mooney, 2007; Mintz et al., 2009; Hancock et al., 2018). Other applications of weak supervision to scene graph prediction (Krishna et al., 2019), intent classification (Mallinar et al., 2019), and medical imag3654 Figure 1: Schematic of the HERALD two-stage workflow. Stage 1: Auto-label training data with Heuristic Functions. We first design heuristics rules for detecting user disengagement by investigating multiple dialog corpora. The heuristics rules are implemented as heuristic functions based on regular expressions and dialog acts. Then, we use the heuristic function to label the training set automatically. Stage 2: Denoise weakly-labeled training data with Shapley Algorithm. We calculate the Shapley value for each data point and correct the noisy data points with negative Shapely values by flipping their labels. Finally, we fine-tune the model on the denoised training data. ing (Varma et al., 2017) have observed similar benefits in annotation efficiency. Unlike the existing work, we leverage weak supervision to improve annotation efficiency for detecting user disengagement in social conversations. 3 Problem Formulation We defined engagement as the degree to which users are willing to continue conversing with the dialog system Yu et al. (2016, 2017). We focus on identifying the dialog turns with “disengaged” user response, since they usually indicate poor conversation experience. We formulate the user engagement prediction as a binary classification problem: Our goal is to learn a parameterized user engagement predictor Mθ that, given a dialog turn (along with its dialog context) x ∈X, predicts the turn-level user engagement label y ∈Y = {0, 1}, where label y = 1 means “disengaged” and y = 0 means “engaged”. We start from an unlabeled train set Dtrain = {xi}Ntrain 1 without any label yi. The test set Dtest = {(xi, yi)}Ntest 1 contains the ground-truth label yi. The development set Ddev has a similar structure as the test set Dtest but the development set can be much smaller than a train set (i.e., Ndev ≪Ntrain), making it economical to obtain. Following the general architecture of neural classifiers, we formulate our model Mθ = M(φ, f) = f(φ(x)): Here BERT (Devlin et al., 2019)-based φ is a text encoder that maps each dialog turn x to a feature space φ(x) ∈Rd. f is the final linear layer with softmax activation. 4 Data To ensure our framework is generalized to various corpora, we investigate multiple open-domain dialog datasets ranging from ASR-based (Gunrock (Liang et al., 2020a)) to text-based (ConvAI2 (Dinan et al., 2019), Blender (Roller et al., 2020), and Meena (Adiwardana et al., 2020)) dialog systems. Gunrock Movie Dataset Gunrock Movie dataset consists of dialog data collected from Gunrock, an ASR-based open-domain social chatbot originally designed for Amazon Alexa Prize (Liang et al., 2020a). The Gunrock dataset comes from a user study where in-lab users were recruited to carry on conversations. We have consent to use the data and we also removed any sensitive information in the conversation. Two dialog experts (co-authors of this paper) randomly annotated 134 dialogs and split them evenly into the test set and development set. In total, the experts labeled 519 turn-level disengaging user responses and 2,312 engaging user responses. They reached a high inter-annotator agreement score (Cohen, 1968) with kappa κ = 0.78. The training set contains 276 unlabeled dialogs, with 5644 dialog turns. In addition, we ensure that the data annotation is independent of the labeling heuristics collection, so there is no data leakage problem. A full example dialog can be found in Appendix A.4. ConvAI2 Dataset ConvAI2 dataset contains text-based dialog collected from the second Conver3655 Labeling Heuristics Coverage (%) Example Disengaged User Responses Heuristics Group Disengaged intents Gunrock ConvAI2 (1) Complain system responses Complain system repetition 1.93 1.95 { You already asked me that. | I already told you. Remember? } Complain system ignoring them { You’re not listening. | You didn’t answer my question. } Complain system misunderstanding { I never said I don’t eat my favorite seafood. } Not understanding system { What are you talking about? } Curse system { You’re dumb. } Express frustration { Sigh. } (2) Dislike current topic Express negative opinion 1.90 3.45 { I don’t like music. | It’s boring. } Show low interests { I don’t care. } (3) Request to end topic or conversation Request topic change 5.20 2.92 { Let’s talk about something else. } Request termination { Stop. | Bye. } (4) End with non-positive responses End with negative answer 20.13 4.86 { No. | I have not. } End with unsure answer { I don’t know. | I don’t remember. | Well, maybe. } End with back-channeling { Yeah. | Okay. } End with hesitation { Hmm... | That’s a hard one, let me think. } Table 1: Our labeling heuristics designed to capture user disengagement in dialogs. A dialog turn is considered disengaged if any of the heuristic rules apply to the user responses. sational Intelligence (ConvAI) Challenge (Dinan et al., 2019). We select dialogs from the main eight participated chatbots (Bot 1, 2, 3, 4, 6, 9, 11) and exclude dialogs that are one-sided or shorter than three turns. The dialog experts annotated 207 dialogs in total. The dialogs are evenly distributed over all the eight bots to ensure system diversity, and are randomly sampled within each bot. The annotated data consist of 209 disengaging turns and 1684 non-disengaging turns. They reached a high inter-annotator agreement score (Cohen, 1968) with kappa κ = 0.76. We split the annotated dialogs evenly into the test set and develop set. The training set contains 2,226 dialogs, with 18,306 dialog turns. Google Meena Dataset Meena (Adiwardana et al., 2020) is the largest end-to-end neural chatbot so far, trained on 867M public domain social media conversations. We study the 93 example Human-Menna conversations released by Google. Facebook Blender Dataset The Blender bot (Roller et al., 2020) is an open-domain chatbot with several conversational skills: providing engaging talking points and listening to their partners, displaying knowledge, empathy, and personality appropriately while maintaining a consistent persona. We study the 108 example Human-Blender conversations released by Facebook. 5 Method Our goal is to train a user engagement detector with minimum data annotation efforts. Traditional supervised learning paradigms require annotating many training samples. In addition, it requires additional data annotation to extend the model to a new dialog corpus. To reduce annotation work, we propose HERALD, a two-stage pipeline that annotates large-scale training data efficiently and accurately (Figure 1). Instead of hand-labeling training data points, we use heuristic functions to label each training datum automatically. The heuristic functions are built upon a set of user disengagement heuristics rules. Since the training data are automatically labeled, their labels would be noisy. We then clean the noisy training data with Shapley algorithm (Ghorbani and Zou, 2019) to improve the labeling accuracy. The Shapley algorithm denoises training data by identifying data with wrong labels and flip their labels. Finally, as we received clean training data, we use them to fine-tune a BERTbased model and obtain the final user disengagement detection model. 5.1 Stage 1: Auto-label Training Data with Heuristic Functions Since labeling large-scale training data is timeconsuming, we propose heuristic labeling functions to label training data automatically. The heuristic functions focus on detecting disengagement from user responses, as it directly indicates poor user experience. To build the heuristics functions, we first summarize the heuristic rules shared among users. We investigate the disengaged dialog turns from the four datasets mentioned above and identify four groups of user disengagement patterns: “complain system responses”, “dislike current topics”, “terminate or change topics”, and “end with non-positive responses” (Table 1). We then discuss the implementation of heuristics functions. 3656 5.1.1 Disengagement Heuristic Rules Group 1: Complain system responses. Complaints are an evident sign of user disengagement. We identify six related disengaged intents. The first three intents (“complain system repetition”, “complain system ignoring them” and “complain system misunderstanding”) usually appear when the bot makes errors like repeating the same content, ignoring, forgetting, and misunderstanding the user’s response. In these cases, users express their disengagement by indicating the bot’s error (e.g. “You already told me that”, “You’re not listening”). Another intent “not understanding system” happens when users cannot understand the system’s response (e.g. “I don’t know what you’re talking about.”). In the last two intents, users reveal negative emotions by cursing the system (e.g. “you’re dumb”) or express frustration (e.g. “sigh”) about the conversation. Group 2: Dislike current topics. When discussing a given topic, users might show their disengagement by expressing negative opinions or low interest. For example, given the bot’s response, “I write romantic novels under a pen name. ”, for users who are not interested in reading, users might say “reading is boring”, “I don’t like to read”, or “I’m not interested in this”. We also make sure to handle the corner cases where the user utterance should be labeled as engaged but contains negative opinions. For instance, to respond to the bot’s question, “do you want to not work?”, a user might say, “Yes. my job is boring. I have to work with mail”. Though the user mentions a negative feeling (“boring”), the user agrees with the bot and shares further information. Group 3: Terminate or change topics Group 3 considers the cases where users express disengagement to the current topic in a more straightforward fashion. For example, if users are not interested in the current topic, instead of just expressing their dislike to it, they may request to switch topics with “Let’s talk about something else”. In some cases, users might show strong disengagement by requesting to end the conversation if the user is no longer interested in continuing the conversation. Group 4: End with non-positive responses A more subtle but common clue of disengagement is when users end the response with non-positive content. For example, non-positive responses like “I don’t know”, “No”, “Yeah”, “uh”, “Probably”, imply that users do not have much to talk about the current topic. To keep the precision of our heuristics high, we carefully consider the counterexamples. One case is that the user follows up with more responses such as questions (e.g., Bot: “Have you seen any movies lately? ”, User: “No. Have you?”), and opinion (e.g. Bot: “What’s your favorite animation movie?”, User: “I don’t know, but it might actually be frozen two. My sister loves it.”) in the same dialog turn. These turns should not be labeled as disengaged since the user is still interested in sharing more content or asking followup questions. Therefore, we take a conservative approach: we label the dialog turn as disengaged only if no more responses follow the non-positive response. 5.1.2 Heuristic Functions Implementation Next, we discuss how to use heuristic functions to auto-label disengaged user utterances. First, we split user responses into segments since user responses may consist of multiple units with different semantic meanings. We use NLTK Sentence Tokenizer for text-based system, and a segmentation model (Chen et al., 2018) for ASR (Automatic Speech Recognition)-based system as the segmentation tool. We then apply the heuristic functions on each segment to detect disengaged intents. For heuristic groups 1 to 3, if any segment contains a disengaged intent, the user response is auto-labeled as disengaged. For heuristic group 4 (“End with non-positive responses”), we assign disengaged labels only if the disengaged intents are detected in the last segment. We detect disengaged intents with Regexes. The benefit of using Regexes is that they have minimum dependencies and are easy to modify. We design Regexes for each intent. Following common Regexes complexity metrics (Luo et al., 2018), our Regexes for each intent contains 43.9 Regexes groups and 87.7 or clauses on average. Our framework also supports incorporating additional resources to improve the intent detection accuracy for automatic training data labeling. For example, we can enhance the recall of Regexes intent detection by incorporating existing deep learning-based NLU (Natural Language Understanding) models. Specifically, we re-purpose an open-sourced dialog act classification model (Yu and Yu, 2021) to enhance disengagement intent detection: we select 6 out of the 23 supported dialog act labels that are associated with disen3657 gaged intents, and map each selected dialog act label to the heuristic groups. The dialog act “complaint” is mapped to the heuristic group “complain system repetition”;“closing” is mapped to the disengaged intent “request termination”; “hold” to “hesitation”;“other_answers” to “unsure answer”; “back-channeling” to “back-channeling”, and “neg_answer“ to ‘negative answer‘”. If a user utterance is detected with disengaged intent by either Regexes or the deep learning model, then the utterance is auto-labeled as disengaged. 5.2 Stage 2: Denoise with Shapley Algorithm & Fine-tune Overview Next, we denoise the labeled data using Shapley algorithm (Ghorbani and Zou, 2019). Shapley algorithm has been studied in the cooperative game theory (Dubey, 1975) and economics (Gul, 1989) as a fair distribution method. Shapley algorithm computes a Shapley value for each training datum, which quantifies the contribution of each training datum to the prediction and performance of a deep network. Low Shapley value data capture outliers and corruptions. Therefore, we can identify and denoise the incorrectly labeled data by computing their Shapley values and finetune the model on the cleaned training set. Shapley Algorithm Shapley algorithm comes originally from cooperative game theory (Dubey, 1975). Consider a cooperative game with n players D = {1, ..., n} and a utility function v : 2[n] →R which assigns a reward to each of 2n subsets of players: v(S ) is the reward if the players in subset S ⊆D cooperate. Shapley value defines a unique scheme to distribute the total gains generated by the coalition of all players v(D) with a set of appealing mathematical properties. In our setting, we can consider Dtrain = {(xi, yi)}Ntrain 1 as Ntrain players. We define the utility function v(S ) as the performance on the development set Ddev. The Shapley value for player i is defined as the average marginal contribution of {(xi, yi)} to all possible subsets that are formed by other players (Jia et al., 2019a,b): si = 1 N X S ⊆Dtrain\{xi} 1 N−1 |S | [v(S ∪{xi}) −v(S )] As suggested by the definition of Shapley value, computing Shapley value requires an exponentially large number of computations to enumerate O(2Ntrain) possible subsets and train the model Mθ on each subset, which is intractable. Inspired by (Jia et al., 2019a,b), HERALD tackles this issue by reducing the deep model Mθ to a Knearest neighbors (KNN) model and then apply the closed-form solution of Shapley value on KNN: We reduce our BERT-based classification model Mθ = M(φ, f) = f(φ(x)) to a KNN by first finetuning Mθ on the auto-labeled training samples. We then use the feature extractor φ to map each training datum to the feature space {φ(xi)}Ntrain 1 . We construct a KNN classifier in the feature space to compute the closed-form Shapley value. Next, we discuss the closed-form solution of Shapley value. We first consider a special case where the development set Ddev only contains one datum Ddev = {(xdev, ydev)}. Given any nonempty subset S ⊆Dtrain, we use the KNN classifier to classify xdev. To do this, we sort the data points in the training set {xi}Ntrain 1 based on their euclidean distance in the feature space φ(x) to the datum in the development set xdev, yielding (xα1, xα2, ..., xα|S |) with xα1, ..., xαK as the top-K most similar data points to xdev. The KNN classifier outputs the probability of xdev taking the label ydev as P[xdev → ydev] = 1 K PK k=1 1[yαk = ydev], where αk is the index of the kth nearest neighbor. We define the utility function as the likelihood of the correct label: ν(S ) = 1 K min{K,|S |} X k=1 1[yαk(S ) = ydev] (1) Jia et al. (2019a,b) proves that the Shapley value of each training point sαi can be calculated recursively in O(N log N) time as follows: sαN = 1[yαN = ydev] N sαi = sαi+1 + min{K, i} i × K 1[yαi=ydev]−1[yαi+1=ydev] The above result for a single point in Ddev could be readily extended to the multiple-point case, in which the utility function is defined by ν(S ) = 1 Ndev Ndev X j=1 1 K min{K,|S |} X k=1 1[yα(j) k (S ) = ydev,j] where α(j) k (S ) is the index of the kth nearest neighbor in S to xdev,j. Jia et al. (2019a,b) also prove that the Shapley value in this case is the average of the Shapley value for every single dev point. Denoising Procedure Our denoising procedure works as follows: (1) We first fine-tune our BERTbased classification model Mθ = M(φ, f) = f(φ(x)) 3658 No. Method Gunrock Movie ConvAI2 bACC F2Score bACC F2Score (1) Heuristics 78.32 65.09 76.58 58.16 (2) Heuristics (regex only) 62.81 35.46 72.04 49.90 (3) Heuristics (NLU only) 72.68 56.32 63.62 32.86 (4) Heuristics w/o Group 1 78.21 64.88 71.20 48.44 (5) Heuristics w/o Group 2 77.96 64.49 75.45 56.22 (6) Heuristics w/o Group 3 71.52 55.36 71.96 49.80 (7) Heuristics w/o Group 4 58.34 23.97 68.32 42.68 (8) BERT(dev) 73.98 60.74 74.97 55.40 (9) BERT(Auto) 80.55 71.77 78.76 63.13 (10) BERT(Auto+dev) 80.73 72.16 80.46 64.54 (11) HERALD 86.17* 80.01* 86.22* 70.49* Table 2: Evaluation results comparison among variants of HERALD. * indicates that the model is statistically significantly better than baseline models. All numbers in the table are in percentage. on the auto-labeled training samples. This step injects the knowledge in the labeling heuristic into the model Mθ. (2) We then map each auto-labeled training datum to the feature space {φ(xi)}Ntrain 1 , since we want to apply the closed-form KNN formula of Shapley value in the feature space. (3) Next, for a binary classification problem, we duplicate each training datum 2 times with labels [0, 1]. This generates a large training set Dlarge with 2 × Ntrain data points, and we note that the origin training set Dtrain is a subset of Dlarge, since Dlarge enumerates all C possible labels for each each training datum. (4) We then calculate Shapley value for the 2 × Ntrain data points in Dlarge using the closed-form KNN formula. (5) We remove the data with negative Shapley value in Dlarge, and get a cleaned training set Dclean. The duplicate-and-remove procedure “flips” the labels of the noisy data points with low Shapley value. (6) Finally, we fine-tune the classification model Mθ on Dclean to get the final user disengagement detection model. To sum up, the Shapley value quantifies the contribution of each training datum. Low Shapley value data capture outliers and corruptions that are not consistent with the distribution of other data points. We identify and correct these outliers and corruptions to provide a clean training set. 6 Experiments Model Setup We use K = 10 for the KNN Classifier. We use BERT (Devlin et al., 2019) as the text encoder φ of our classification model Mθ = M(φ, f) = f(φ(x)). Additional implementation details are included in Appendix. Model Comparisons and Ablations We compare HERALD to its several ablations (Table 2) and evaluate the performance on the test set. We report balanced accuracy (bACC) and Fβ Score with β = 2 (Baeza-Yates et al., 1999). (1) Heuristics uses the labeling heuristic function with both Regex and dialog act to predict the test set. (2) Heuristics (Regex only) uses the labeling heuristic function only with Regex to predict on the test set. (3) Heuristics (NLU only) uses the labeling heuristic function only with NLU. (4-7) show the ablation of the heuristics function prediction baseline by excluding each heuristic group. (8) BERT(dev) finetunes BERT on the expert-annotated development set. (9) BERT(Auto) fine-tunes BERT on the autolabeled training samples. (10) BERT(Auto+dev) fine-tunes BERT on both the auto-labeled training samples and the development set. (11) HERALD reports the performance of the final model trained on Dclean. Results Our first takeaway is that our labeling heuristics produce decent predictions and generalize to different datasets. As shown in Table 2, Heuristics prediction (Heuristic, 78.32%, 76.58%) is better than the BERT-based model with limited training samples (BERT(dev), 73.98%, 74.94%) on both datasets. It also shows that our labeling heuristics are generalizable to different corpora. Our second takeaway is that learning from a large number of noisy labels works better than learning from a limited number of clean labels. As shown in Table 2, BERT fine-tuned on the autolabeled training set (BERT(Auto), 80.55, 78.76) outperforms BERT fine-tuned on clean but small development set (BERT(dev), 73.98, 74.94) by a large margin. In addition, we also observe that the BERT model fine-tuned on the auto labeled training data (BERT(Auto), 80.55%, 78.76%) generalizes beyond the labeling heuristics (Heuristics, 78.32%, 76.58%). Our third takeaway is that using the expertannotated development set for denoising is more efficient than using the development set as additional training data. After fine-tuning BERT on the weakly labeled training data (BERT(Auto), 80.55%, 78.76%), having an additional fine-tuning step using the development set slightly improves the model’s performance (BERT(Auto+dev), 80.73%, 80.46%). In contrast, 3659 using the development set for the Shapley denoising algorithm gives a significant performance gain (HERALD, 86.17%, 86.22%). Figure 2: Removing data with low Shapley values (Shapley with Ktest = 1, 5, 10, 25, 50) improves the performance of the KNN in Gunrock Movie Dataset while removing data with high Shapley values and retain data with low Shapley values (“RetainHurtful”) leads to worse performance. Annotation Cost The cost of annotating the DEV set is small for the Shapley algorithm. For Gunrock Movie Dataset, we used 67 annotated dialogs as the DEV set. For ConvAI2, we used 52 annotated dialogs as the DEV set. The annotation takes less than 1 hour in both cases, which is negligible compared to the cost of annotating all training data. Heuristics Group Analysis We perform ablation studies to analyze the importance of each of the four heuristics groups in Table 1. As shown in Table 2, excluding heuristics group 4 leads to the most significant performance drop in both datasets (Heuristics w/o Group 4, 58.34%, 68.32%), indicating that “end with non-positive response” is the most prevalent form of user disengagement. In addition, each heuristics group has different importance in different datasets. For example, dropping heuristics group 1 (“complain system responses”) only leads to a marginal performance drop on the Gunrock Movie dataset but incurs a significant performance drop on the ConvAI2 dataset. We also notice that heuristic group 4 (“End with non-positive responses”) plays a more critical role in the Gunrock Movie dataset than in the ConvAI2 dataset. This might be mainly due to the difference between ASR-based (Gunrock Movie) and text-based (ConvAI2) systems. When asked an open-ended question in ASR-based systems, since users have less time to think, they are more likely to reply with responses such as “I’m not sure”, “let me think”. While in text-based systems (ConvAI2), users have more time to think and formulate their responses. Hence, heuristics group 4 covering these responses happen more in Gunrock Movie than ConvAI2. Generalizability of Heuristic Functions The results show that our heuristic functions are generalized to both ASR-based and text-based systems. As indicated in Table 2, our Regexes reach a decent accuracy of 62.81% and 72.04% on the expert annotated test set respectively on Gunrock Movie and ConvAI2 dataset, and thus can serve as a relatively reliable source for auto-labeling. In addition, although the dialog act model (MIDAS) is initially designed for ASR-based systems and thus has a better performance on the Gunrock Movie data, it should be generalizable to other ASR-based systems, as the six selected dialog acts are general and independent of topics. Therefore, the combination of dialog acts and Regexes should be sufficient to be applied to various corpora. Figure 3: An example dialog turn from the Gunrock Movie dataset with an incorrect auto label “nondisengaged” identified by data Shapley. In this case, the user actually says “I don’t wanna talk about movies anymore,” but an ASR error happens, and thus the labeling heuristics fail to capture this dialog turn. Figure 4: An example dialog turn from Gunrock Movie dataset that is incorrectly auto-labeled as “disengaged” because the labeling heuristics see the negative word “disagree”. This data point is also identified and corrected by data Shapley. Shapley Value Analysis We also present an analysis to show how Shapley denoising works, as shown in Figure 2. We examine the Shapley value for each training datum in Stage 2. We first show two example dialog turns from the Gunrock Movie dataset with a negative Shapley value in Figure 3 and Figure 4. In Figure 3, the dialog turn is incorrectly auto-labeled as “non-disengaged”. This is because an ASR error happens, and the user utterance “I don’t wanna talk about movies anymore” 3660 is transcribed as “I wanna talk about movies anymore”. In Figure 4, the user says, “Oh I disagree. I think the movie was fantastic!”. The labeling heuristics see the negative word “disagree” and auto-label this turn as “disengaged”. Both data points are with negative Shapley values and are corrected in Stage 3. Next, we present a quantitative analysis of Shapley value. According to the Shapley value, we remove data points one by one, starting from the least valuable (low Shapley values) to the most valuable (high Shapley values). Each time, after removing the data point, we create new KNN classifier models on the remaining dialog turns and labels and evaluate them on the test set with expert annotations. As shown in Figure 2, removing training data with low Shapley values increases the performance to a certain point before convergence for K of all choices. We observe a similar trend when re-training a model on the remaining data. In contrast, removing data randomly or removing data starting from high Shapley values decreases the performance on the test set (“Random” and “RetainHurtful” in Figure 2). This shows that low Shapley value data effectively capture outliers and corruptions, which further justifies our design choice of denoising with Shapley value. Alternative Data Valuation Methods We also explored alternative methods to data Shapley like influence function (Koh and Liang, 2017) and TracIn (Pruthi et al., 2020): on Gunrock Movie, Influence Functions and TracIn achieve 82.96% and 83.15% accuracy, respectively. Both methods outperform BERT(Auto+dev) (80.73%) significantly but perform slightly worse than HERALD (86.17%). Overall, results show that our data annotation workflow also works well with other data valuation methods. Figure 5: An error case where the low engagement dialog turn that is not captured by HERALD. Error Analysis Figure 5 shows an error example of HERALD, where both the labeling heuristics and the Shapley algorithm fail to identify this turn as low engagement. In this example, the chatbot system asks whether the user is interested in movies, but the user does not directly answer the question. Instead, the user says “I have a question for you social bot”, indicating that the user does not like the current topic and wants to talk about something else. HERALD fails to identify this dialog turn as low engagement, partly because the Regexes in the “request topic change” heuristic rule does not cover this example. One way to fix this error is to upgrade the Regexes. A more general solution is to consider the chatbot system’s expectations on user responses conditioned on the chatbot’s question. If the chatbot receives an “unexpected” user response, then the user is probably not interested in discussing the current topic. 7 Conclusion The ultimate chatbot evaluation metric should be user-centric, as chatbots are there to provide humans with enjoyable experiences. Previously detecting user disengagement typically requires annotating many dialog samples for each individual dataset. We propose a two-stage pipeline HERALD to automatically label and denoise training data and, at the same time, build a user disengagement detector. Our experiment shows that HERALD significantly reduces the annotation cost of a new corpus. HERALD’s disengagement detection results highly correlate with expert judgments on user disengagement in both datasets (86.17% bACC in Gunrock Movie, 86.22% in ConvAI2). Acknowledgments We thank ACL 2021 chairs and reviewers for their review efforts and constructive feedback. We would also like to thank Yu Li and Minh Nguyen for revising the Regexes. References Daniel Adiwardana, Minh-Thang Luong, David R. So, Jamie Hall, Noah Fiedel, Romal Thoppilan, Zi Yang, Apoorv Kulshreshtha, Gaurav Nemade, Yifeng Lu, and Quoc V. Le. 2020. Towards a human-like opendomain chatbot. CoRR, abs/2001.09977. Ricardo Baeza-Yates, Berthier Ribeiro-Neto, et al. 1999. Modern information retrieval, volume 463. ACM press New York. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: an automatic metric for MT evaluation with improved correlation with human judgments. In IEEvaluation@ACL, pages 65–72. Association for Computational Linguistics. 3661 Razvan C. Bunescu and Raymond J. Mooney. 2007. Learning to extract relations from the web using minimal supervision. In ACL. The Association for Computational Linguistics. Chun-Yen Chen, Dian Yu, Weiming Wen, Yi Mang Yang, Jiaping Zhang, Mingyang Zhou, Kevin Jesse, Austin Chau, Antara Bhowmick, Shreenath Iyer, et al. 2018. Gunrock: Building a human-like social bot by leveraging large scale real user data. Alexa Prize Proceedings. Jason Ingyu Choi, Ali Ahmadvand, and Eugene Agichtein. 2019. Offline and online satisfaction prediction in open-domain conversational systems. In CIKM, pages 1281–1290. ACM. Jacob Cohen. 1968. Weighted kappa: Nominal scale agreement provision for scaled disagreement or partial credit. Psychological bulletin, 70(4):213. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT (1), pages 4171–4186. Association for Computational Linguistics. Emily Dinan, Varvara Logacheva, Valentin Malykh, Alexander Miller, Kurt Shuster, Jack Urbanek, Douwe Kiela, Arthur Szlam, Iulian Serban, Ryan Lowe, et al. 2019. The second conversational intelligence challenge (convai2). arXiv preprint arXiv:1902.00098. Pradeep Dubey. 1975. On the uniqueness of the shapley value. International Journal of Game Theory, 4(3):131–139. Gabriel Forgues, Joelle Pineau, Jean-Marie Larchevêque, and Réal Tremblay. 2014. Bootstrapping dialog systems with word embeddings. In Nips, modern machine learning and natural language processing workshop, volume 2. Sarik Ghazarian, Ralph M. Weischedel, Aram Galstyan, and Nanyun Peng. 2020. Predictive engagement: An efficient metric for automatic evaluation of open-domain dialogue systems. In AAAI, pages 7789–7796. AAAI Press. Amirata Ghorbani and James Y. Zou. 2019. Data shapley: Equitable valuation of data for machine learning. In ICML, volume 97 of Proceedings of Machine Learning Research, pages 2242–2251. PMLR. Faruk Gul. 1989. Bargaining foundations of shapley value. Econometrica: Journal of the Econometric Society, pages 81–95. Fenfei Guo, Angeliki Metallinou, Chandra Khatri, Anirudh Raju, Anu Venkatesh, and Ashwin Ram. 2018. Topic-based evaluation for conversational bots. CoRR, abs/1801.03622. Prakhar Gupta, Shikib Mehri, Tiancheng Zhao, Amy Pavel, Maxine Eskénazi, and Jeffrey P. Bigham. 2019. Investigating evaluation of open-domain dialogue systems with human generated multiple references. CoRR, abs/1907.10568. Braden Hancock, Antoine Bordes, Pierre-Emmanuel Mazaré, and Jason Weston. 2019. Learning from dialogue after deployment: Feed yourself, chatbot! In ACL (1), pages 3667–3684. Association for Computational Linguistics. Braden Hancock, Paroma Varma, Stephanie Wang, Martin Bringmann, Percy Liang, and Christopher Ré. 2018. Training classifiers with natural language explanations. In ACL (1), pages 1884–1895. Association for Computational Linguistics. Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nezihe Merve Gürel, Bo Li, Ce Zhang, Costas J. Spanos, and Dawn Song. 2019a. Efficient taskspecific data valuation for nearest neighbor algorithms. PVLDB, 12(11):1610–1623. Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nick Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J. Spanos. 2019b. Towards efficient data valuation based on the shapley value. In AISTATS, volume 89 of Proceedings of Machine Learning Research, pages 1167–1176. PMLR. Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In ICML, volume 70 of Proceedings of Machine Learning Research, pages 1885–1894. PMLR. Ranjay Krishna, Vincent S. Chen, Paroma Varma, Michael Bernstein, Christopher Ré, and Fei-Fei Li. 2019. Scene graph prediction with limited labels. In ICCV, pages 2580–2590. IEEE. Margaret Li, Jason Weston, and Stephen Roller. 2019. ACUTE-EVAL: improved dialogue evaluation with optimized questions and multi-turn comparisons. CoRR, abs/1909.03087. Kaihui Liang, Austin Chau, Yu Li, Xueyuan Lu, Dian Yu, Mingyang Zhou, Ishan Jain, Sam Davidson, Josh Arnold, Minh Nguyen, et al. 2020a. Gunrock 2.0: A user adaptive social conversational system. arXiv preprint arXiv:2011.08906. Weixin Liang, Yanhao Jiang, and Zixuan Liu. 2021. GraghVQA: Language-guided graph neural networks for graph-based visual question answering. In MAI@NAACL-HLT. Association for Computational Linguistics. Weixin Liang, Feiyang Niu, Aishwarya N. Reganti, Govind Thattai, and Gökhan Tür. 2020b. LRTA: A transparent neural-symbolic reasoning framework with modular supervision for visual question answering. CoRR, abs/2011.10731. 3662 Weixin Liang, Youzhi Tian, Chengcai Chen, and Zhou Yu. 2020c. MOSS: end-to-end dialog system framework with modular supervision. In AAAI, pages 8327–8335. AAAI Press. Weixin Liang and James Zou. 2021. Neural group testing to accelerate deep learning. In IEEE International Symposium on Information Theory, ISIT 2021. IEEE. Weixin Liang, James Zou, and Zhou Yu. 2020d. ALICE: active learning with contrastive natural language explanations. In EMNLP (1), pages 4380– 4391. Association for Computational Linguistics. Weixin Liang, James Zou, and Zhou Yu. 2020e. Beyond user self-reported likert scale ratings: A comparison model for automatic dialog evaluation. In ACL, pages 1363–1374. Association for Computational Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Chia-Wei Liu, Ryan Lowe, Iulian Serban, Michael Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In EMNLP, pages 2122–2132. The Association for Computational Linguistics. Ryan Lowe, Michael Noseworthy, Iulian Vlad Serban, Nicolas Angelard-Gontier, Yoshua Bengio, and Joelle Pineau. 2017. Towards an automatic turing test: Learning to evaluate dialogue responses. In ACL (1), pages 1116–1126. Association for Computational Linguistics. Bingfeng Luo, Yansong Feng, Zheng Wang, Songfang Huang, Rui Yan, and Dongyan Zhao. 2018. Marrying up regular expressions with neural networks: A case study for spoken language understanding. arXiv preprint arXiv:1805.05588. Neil Mallinar, Abhishek Shah, Rajendra Ugrani, Ayush Gupta, Manikandan Gurusankar, Tin Kam Ho, Q. Vera Liao, Yunfeng Zhang, Rachel K. E. Bellamy, Robert Yates, Chris Desmarais, and Blake McGregor. 2019. Bootstrapping conversational agents with weak supervision. In AAAI, pages 9528–9533. AAAI Press. Shikib Mehri and Maxine Eskénazi. 2020. Unsupervised evaluation of interactive dialog with dialogpt. In SIGdial, pages 225–235. Association for Computational Linguistics. Mike Mintz, Steven Bills, Rion Snow, and Daniel Jurafsky. 2009. Distant supervision for relation extraction without labeled data. In ACL/IJCNLP, pages 1003–1011. The Association for Computer Linguistics. Jekaterina Novikova, Ondrej Dusek, Amanda Cercas Curry, and Verena Rieser. 2017. Why we need new evaluation metrics for NLG. In EMNLP, pages 2241–2252. Association for Computational Linguistics. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311– 318. ACL. Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. 2020. Estimating training data influence by tracing gradient descent. In NeurIPS. Ashwin Ram, Rohit Prasad, Chandra Khatri, Anu Venkatesh, Raefer Gabriel, Qing Liu, JeffNunn, Behnam Hedayatnia, Ming Cheng, Ashish Nagar, Eric King, Kate Bland, Amanda Wartick, Yi Pan, Han Song, Sk Jayadevan, Gene Hwang, and Art Pettigrue. 2018. Conversational AI: the science behind the alexa prize. CoRR, abs/1801.03604. Alexander Ratner, Stephen H. Bach, Henry R. Ehrenberg, Jason A. Fries, Sen Wu, and Christopher Ré. 2020. Snorkel: rapid training data creation with weak supervision. VLDB J., 29(2-3):709–730. Alexander J. Ratner, Christopher De Sa, Sen Wu, Daniel Selsam, and Christopher Ré. 2016. Data programming: Creating large training sets, quickly. In NIPS, pages 3567–3575. Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2020. Recipes for building an open-domain chatbot. CoRR, abs/2004.13637. Vasile Rus and Mihai C. Lintean. 2012. A comparison of greedy and optimal assessment of natural language student input using word-to-word similarity metrics. In BEA@NAACL-HLT, pages 157–162. The Association for Computer Linguistics. Chongyang Tao, Lili Mou, Dongyan Zhao, and Rui Yan. 2018. RUBER: an unsupervised method for automatic evaluation of open-domain dialog systems. In AAAI, pages 722–729. AAAI Press. Paroma Varma, Bryan D. He, Payal Bajaj, Nishith Khandwala, Imon Banerjee, Daniel L. Rubin, and Christopher Ré. 2017. Inferring generative model structure with static analysis. In NIPS, pages 240– 250. Anu Venkatesh, Chandra Khatri, Ashwin Ram, Fenfei Guo, Raefer Gabriel, Ashish Nagar, Rohit Prasad, Ming Cheng, Behnam Hedayatnia, Angeliki Metallinou, Rahul Goel, Shaohua Yang, and Anirudh Raju. 2018. On evaluating and comparing conversational agents. CoRR, abs/1801.03625. 3663 Huichuan Xia and Brian McKernan. 2020. Privacy in crowdsourcing: a review of the threats and challenges. Comput. Support. Cooperative Work., 29(3):263–301. Sanghyun Yi, Rahul Goel, Chandra Khatri, Tagyoung Chung, Behnam Hedayatnia, Anu Venkatesh, Raefer Gabriel, and Dilek Hakkani-Tür. 2019. Towards coherent and engaging spoken dialog response generation using automatic conversation evaluators. CoRR, abs/1904.13015. Chen Yu, Paul M. Aoki, and Allison Woodruff. 2004. Detecting user engagement in everyday conversations. In INTERSPEECH. ISCA. Dian Yu and Zhou Yu. 2021. Midas: A dialog act annotation scheme for open domain human machine spoken conversations. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, page 1103–1120. Zhou Yu, Leah Nicolich-Henkin, Alan W. Black, and Alexander I. Rudnicky. 2016. A wizard-of-oz study on A non-task-oriented dialog systems that reacts to user engagement. In SIGDIAL Conference, pages 55–63. The Association for Computer Linguistics. Zhou Yu, Vikram Ramanarayanan, Patrick L. Lange, and David Suendermann-Oeft. 2017. An opensource dialog system with real-time engagement tracking for job interview training applications. In IWSDS, volume 510 of Lecture Notes in Electrical Engineering, pages 199–207. Springer. 3664 A Appendix A.1 Implementation Details of HERALD We use K = 10 for the KNN Regressor. We load and fine-tune pre-trained BERT as the feature extractor φ. The details of extending BERT to encode multi-turn dialogs are as follows. Each dialog turn (along with its dialog context) is represented as a sequence of tokens in the following input format (Liang et al., 2020c): Starting with a special starting token [CLS ], we concatenate tokenized user and system utterances in chronological order with [S EP] as the separators for adjacent utterance. In other words, we represent each dialog as a sequence: [CLS ], S 1,1, S 1,2, ..., [S EP], U1,1, U1,2, ..., [S EP], S 2,1, S 2,2, ..., [S EP] where S i,j and Ui,j are the jth token of the system and user utterance in the ith turn. Following BERT, we also add a learned embedding to every token indicating whether it comes from user utterances or system utterances . In addition, since the disengaging class and the non-disengaging class are imbalanced, we up-sample the disengaging dialog turns for both the training set and the development set. Though it is also possible to handle the imbalanced classes by adding weights for two classes, we did not take this approach because we do not have a closedform solution for calculating the shapley value for weighted KNN in O(N log N) time. Improving the architecture of HERALD and extending HERALD to other machine learning tasks (Liang and Zou, 2021; Liang et al., 2020d,b, 2021) are interesting directions of future work. A.2 Reproducibility The source code of HERALD can be found in the supplementary materials. We run experiments on a server of eight GTX 1080 GPUs. The average runtime for all stages of HERALD is less than 10 minutes. The number of parameters is similar to BERT. We use the default hyperparameters of BERT. The public examples of Google Meena Dataset can be downloaded from https: //github.com/google-research/google-research/ blob/master/meena/meena.txt The public examples of Facebook Blender Dataset can be downloaded from https://parl.ai/projects/recipes/ chatlog_2.7B_render.html The public examples of ConvAI2 Dataset can be downloaded from http://convai.io/data/data_volunteers.json and http://convai.io/data/summer_wild_evaluation_ dialogs.json (a) Denoising with Shapley Value in Gunrock Movie Dataset (b) Denoising with Shapley Value in ConvAI2 Dataset Figure 6: Removing data points with low Shapley value improves the performance of the KNN classifier. Additional Shapley Value Analysis We also present addition analysis to show how Shapley denoising works as shown in Figure 6. We present the experiments on both Gunrock Movie Dataset and ConvAI2 Dataset. Figure 6 presents a quantitative analysis of Shapley value. According to the Shapley value, we remove data points one by one starting from the least valuable to the most valuable. Each time, after the data point is removed, we create new KNN classifier models on the remaining dialog turns and labels and evaluate them on the test set with expert annotations. As shown in Figure 6, removing training data with low Shapley values increases the performance to a certain point before convergence for K of all choices. We observe a similar trend when re-training a model on the remaining data. In contrast, removing data randomly or removing data from the most to least valuable data decreases the performance on the test set. This shows that low Shapley value data effectively capture outliers and corruptions, which further justifies our design choice of denoising with Shapely value. A.3 Addition Dialog Examples We show additional dialog examples. Figure 7 shows a full dialog example from ConvAI dataset. Figure 8 shows a full dialog example from Gunrock Movie dataset. 3665 Figure 7: A full example from ConvAI Dataset. Figure 8: A full example from Gunrock Movie Dataset.
|
2021
|
283
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3666–3681 August 1–6, 2021. ©2021 Association for Computational Linguistics 3666 Value-Agnostic Conversational Semantic Parsing Emmanouil Antonios Platanios, Adam Pauls, Subhro Roy, Yuchen Zhang, Alex Kyte, Alan Guo, Sam Thomson, Jayant Krishnamurthy, Jason Wolfe, Jacob Andreas, Dan Klein Microsoft Semantic Machines [email protected] Abstract Conversational semantic parsers map user utterances to executable programs given dialogue histories composed of previous utterances, programs, and system responses. Existing parsers typically condition on rich representations of history that include the complete set of values and computations previously discussed. We propose a model that abstracts over values to focus prediction on type- and function-level context. This approach provides a compact encoding of dialogue histories and predicted programs, improving generalization and computational efficiency. Our model incorporates several other components, including an atomic span copy operation and structural enforcement of well-formedness constraints on predicted programs, that are particularly advantageous in the low-data regime. Trained on the SMCALFLOW and TREEDST datasets, our model outperforms prior work by 7.3% and 10.6% respectively in terms of absolute accuracy. Trained on only a thousand examples from each dataset, it outperforms strong baselines by 12.4% and 6.4%. These results indicate that simple representations are key to effective generalization in conversational semantic parsing. 1 Introduction Conversational semantic parsers, which translate natural language utterances into executable programs while incorporating conversational context, play an increasingly central role in systems for interactive data analysis (Yu et al., 2019), instruction following (Guu et al., 2017), and task-oriented dialogue (Zettlemoyer and Collins, 2009). An example of this task is shown in Figure 1. Typical models are based on an autoregressive sequence prediction approach, in which a detailed representation of the dialogue history is concatenated to the input sequence, and predictors condition on this sequence and all previously generated components of the output (Suhr et al., 2018). While this approach can capture arbitrary dependencies between inputs and outputs, it comes at the cost of sample- and computational inefficiency. We propose a new “value-agnostic” approach to contextual semantic parsing driven by type-based representations of the dialogue history and functionbased representations of the generated programs. Types and functions have long served as a foundation for formal reasoning about programs, but their use in neural semantic parsing has been limited, e.g., to constraining the hypothesis space (Krishnamurthy et al., 2017), guiding data augmentation (Jia and Liang, 2016), and coarsening in coarse-to-fine models (Dong and Lapata, 2018). We show that representing conversation histories and partial programs via the types and functions they contain enables fast, accurate, and sample-efficient contextual semantic parsing. We propose a neural encoder– decoder contextual semantic parsing model which, in contrast to prior work: 1. uses a compact yet informative representation of discourse context in the encoder that considers only the types of salient entities that were predicted by the model in previous turns or that appeared in the execution results of the predicted programs, and 2. conditions the decoder state on the sequence of function invocations so far, without conditioning on any concrete values passed as arguments to the functions. Our model substantially improves upon the best published results on the SMCALFLOW (Semantic Machines et al., 2020) and TREEDST (Cheng et al., 2020) conversational semantic parsing datasets, improving model performance by 7.3% and 10.6%, respectively, in terms of absolute accuracy. In further experiments aimed at quantifying sample efficiency, 3667 ENTITY PROPOSERS Number(2) Month.May Propose entities from the current user utterance. PREVIOUS PROGRAM EXTRACTED DIALOGUE HISTORY TYPES [0] Constraint[Event]() [1] Constraint[Any]() [2] like(value = "shopping") [3] Time(hour = 2, meridiem = PM) [4] Constraint[Event](subject = [2], start = [3]) [5] revise(oldLoc = [0], rootLoc = [1], new = [4]) Function Argument Invocation Copy Entity Constant Reference (to the previous function invocation) Value PREDICTED PROGRAM LINEARIZED REPRESENTATION Can you delete my event called holiday shopping ? PREVIOUS USER UTTERANCE I can’t find an event with that name. PREVIOUS AGENT UTTERANCE CURRENT USER UTTERANCE Oh, it’s just called shopping. It may be at 2. parsing parsing execution delete( find( Constraint[Event]( subject = like("holiday shopping") ) ) ) revise( oldLoc = Constraint[Event](), rootLoc = Constraint[Any](), new = Constraint[Event]( subject = like("shopping"), start = Time( hour = 2, meridiem = PM ) ) ) Unit Constraint[String] Constraint[Event] Event String EventNotFoundError Set of salient types extracted from the dialogue history and used by the parser as a compact representation of the history to condition on. The last type comes from the program execution results. Representation predicted by the proposed model. PARSER Figure 1: Illustration of the conversational semantic parsing problem that we focus on and the representations that we use. The previous turn user utterance and the previous program are shown in blue on the top. The dialogue history representation extracted using our approach is shown on the top right. The current turn user utterance is shown in red on the bottom left. The current utterance, the set of proposed entities, and the extracted dialogue history representation form the input to our parser. Given this input, the parser predicts a program that is shown on the bottom right (in red rectangles). it improves accuracy by 12.4% and 6.4% respectively when trained on only a thousand examples from each dataset. Our model is also effective at non-contextual semantic parsing, matching state-ofthe-art results on the JOBS, GEOQUERY, and ATIS datasets (Dong and Lapata, 2016). This is achieved while also reducing the test time computational cost by a factor of 10 (from 80ms per utterance down to 8ms when running on the same machine; more details are provided in Appendix H), when compared to our fastest baseline, which makes it usable as part of a real-time conversational system. One conclusion from these experiments is that most semantic parses have structures that depend only weakly on the values that appear in the dialogue history or in the programs themselves. Our experiments find that hiding values alone results in a 2.6% accuracy improvement in the low-data regime. By treating types and functions, rather than values, as the main ingredients in learned representations for semantic parsing, we improve model accuracy and sample efficiency across a diverse set of language understanding problems, while also significantly reducing computational costs. 2 Proposed Model Our goal is to map natural language utterances to programs while incorporating context from dialogue histories (i.e., past utterances and their associated programs and execution results). We model a program as a sequenceo of function invocations, each consisting of a function and zero or more argument values, as illustrated at the lower right of Figure 1. The argument values can be either literal values or references to results of previous function invocations. The ability to reference previous elements of the sequence, sometimes called a target-side copy, allows us to construct programs that involve re-entrancies. Owing to this referential structure, a program can be equivalently represented as a directed acyclic graph (see e.g., Jones et al., 2012; Zhang et al., 2019). We propose a Transformer-based (Vaswani et al., 2017) encoder–decoder model that predicts programs by generating function invocations sequentially, where each invocation can draw its arguments from an inventory of values (§2.5)—possibly copied from the utterance—and the results of previous function invocations in the current program. The encoder (§2.2) transforms a natural language utterance and a dialogue history to a continuous representation. Subsequently, the decoder (§2.3) uses this representation to define an autoregressive distribution over function invocation sequences and chooses a high-probability sequence by performing beam search. As our experiments (§3) will show, a na¨ıve encoding of the complete dialogue history and program results in poor model accuracy. 3668 CURRENT PROGRAM with revision revise( oldLoc = Constraint[Event](), rootLoc = RoleConstraint(start), new = Time(hour = 3, meridiem = PM) ) PREVIOUS PROGRAM delete(find( Constraint[Event]( subject = like("holiday shopping"), start = Time(hour = 2, meridiem = PM), end = Time(hour = 5, meridiem = PM) ) )) CURRENT PROGRAM without revision delete(find( Constraint[Event]( subject = like("holiday shopping"), start = Time(hour = 3, meridiem = PM), end = Time(hour = 5, meridiem = PM) ) )) "It actually starts at 3pm." Contains information that is not mentioned in the current utterance Only contains information that is mentioned in the current utterance Figure 2: Illustration of the revise meta-computation operator (§2.1) used in our program representations. This operator can remove the need to copy program fragments from the dialogue history. 2.1 Preliminaries Our approach assumes that programs have type annotations on all values and function calls, similar to the setting of Krishnamurthy et al. (2017).1 Furthermore, we assume that program prediction is local in that it does not require program fragments to be copied from the dialogue history (but may still depend on history in other ways). Several formalisms, including the typed references of Zettlemoyer and Collins (2009) and the meta-computation operators of Semantic Machines et al. (2020), make it possible to produce local program annotations even for dialogues like the one depicted in Figure 2, which reuse past computations. We transformed the datasets in our experiments to use such metacomputation operators (see Appendix C). We also optionally make use of entity proposers, similar to Krishnamurthy et al. (2017), which annotate spans from the current utterance with typed values. For example, the span “one” in “Change it to one” might be annotated with the value 1 of type Number. These values are scored by the decoder along with other values that it considers (§2.5) when predicting argument values for function invocations. Using entity proposers aims to 1This requirement can be trivially satisfied by assigning all expressions the same type, but in practice defining a set of type declarations for the datasets in our experiments was not difficult (refer to Appendix C for details). help the model generalize better to previously unseen values that can be recognized in the utterance using hard-coded heuristics (e.g., regular expressions), auxiliary training data, or other runtime information (e.g., a contact list). In our experiments we make use of simple proposers that recognize numbers, months, holidays, and days of the week, but one could define proposers for arbitrary values (e.g., song titles). As described in §2.5, certain values can also be predicted directly without the use of an entity proposer. 2.2 Encoder The encoder, shown in Figure 3, maps a natural language utterance to a continuous representation. Like many neural sequence-to-sequence models, we produce a contextualized token representation of the utterance, Hutt ∈RU×henc, where U is the number of tokens and henc is the dimensionality of their embeddings. We use a Transformer encoder (Vaswani et al., 2017), optionally initialized using the BERT pretraining scheme (Devlin et al., 2019). Next, we need to encode the dialogue history and combine its representation with Hutt to produce history-contextualized utterance token embeddings. Prior work has incorporated history information by linearizing it and treating it as part of the input utterance (Cheng et al., 2018; Semantic Machines et al., 2020; Aghajanyan et al., 2020). While flexible and easy to implement, this approach presents a number of challenges. In complex dialogues, history encodings can grow extremely long relative to the user utterance, which: (i) increases the risk of overfitting, (ii) increases computational costs (because attentions have to be computed over long sequences), and (iii) necessitates using small batch sizes during training, making optimization difficult. Thanks to the predictive locality of our representations (§2.1), our decoder (§2.3) never needs to retrieve values or program fragments from the dialogue history. Instead, context enters into programs primarily when programs use referring expressions that point to past computations, or revision expressions that modify them. Even though this allows us to dramatically simplify the dialogue history representation, effective generation of referring expressions still requires knowing something about the past. For example, for the utterance “What’s next?” the model needs to determine what “What” refers to. Perhaps more interestingly, the presence of dates in recent 3669 DIALOGUE HISTORY TYPES Unit Constraint[String] Constraint[Event] Event String EventNotFoundError embed decoder UTTERANCE ENCODER DIALOGUE HISTORY ENCODER USER UTTERANCE "Oh, it's just called shopping. It may be at 2." attention K V Q Figure 3: Illustration of our encoder (§2.2), using the example of Figure 1. The utterance is processed by a Transformer-based (Vaswani et al., 2017) encoder and combined with information extracted from the set of dialogue history types using multi-head attention. turns (or values that have dates, such as meetings) should make the decoder more eager to generate referring calls that retrieve dates from the dialogue history; especially so if other words in the current utterance hint that dates may be useful and yet date values cannot be constructed directly from the current utterance. Subsequent steps of the decoder which are triggered by these other words can produce functions that consume the referred dates. We thus hypothesize that it suffices to strip the dialogue history down to its constituent types, hiding all other information.2 Specifically, we extract a set T of types that appear in the dialogue history up to m turns back, where m = 1 in our experiments.3 Our encoder then transforms Hutt into a sequence of history-contextualized embeddings Henc by allowing each token to attend over T . This is motivated by the fact that, in many cases, dialogue history is important for determining the meaning of specific tokens in the utterance, rather than the whole utterance. Specifically, we learn embeddings T ∈R|T |×htype for the extracted types, where htype is the embedding size, and use the attention mechanism of Vaswani et al. (2017) to contextualize Hutt: Henc ≜Hutt + MHA( Hutt |{z} Queries , T |{z} Keys , T |{z} Values ), (1) where “MHA” stands for multi-head attention, and each head applies a separate linear transformation to the queries, keys, and values. Intuitively, 2For the previous example, if the type List[Event] appeared in the history then we may infer that “What” probably refers to an Event. 3We experimented with different values of m and found that increasing it results in worse performance, presumably due to overfitting. [0] +( 1, 2) [1] +([0], 3) [2] +([1], 4) [3] +([2], 5) [0] +(Number, Number) [1] +(Number, Number) [2] +(Number, Number) Consider the following program representing the expression 1 + 2 + 3 + 4 + 5: While generating this invocation, the decoder only gets to condition on the following program prefix: Argument values are masked out! Figure 4: Illustration showing the way in which our decoder is value-agnostic. Specifically, it shows which part of the generated program prefix, our decoder conditions on while generating programs (§2.3). each utterance-contextualized token is further contextualized in (1) by adding to it a mixture of embeddings of elements in T , where the mixture coefficients depends only on that utterancecontextualized token. This encoder is illustrated in Figure 3. As we show in §3.1, using this mechanism performs better than the na¨ıve approach of appending a set-of-types vector to Hutt. 2.3 Decoder: Programs The decoder uses the history-contextualized representation Henc of the current utterance to predict a distribution over the program π that corresponds to that utterance. Each successive “line” πi of π invokes a function fi on an argument value tuple (vi1, vi2, . . . , viAi), where Ai is the number of (formal) arguments of fi. Applying fi to this ordered tuple results in the invocation fi(ai1 = vi1, ai2 = vi2, . . .), where (ai1, ai2, . . . , aiAi) name the formal arguments of fi. Each predicted value vij can be the result of a previous function invocation, a constant value, a value copied from the current utterance, or a proposed entity (§2.1), as illustrated in the lower right corner of Figure 1. These different argument sources are described in §2.5. Formally, the decoder defines a distribution of programs π: p(π | Henc) = P Y i=1 p(πi | f<i, Henc), (2) where P is the number of function invocations in the program, and f<i ≜{f1, . . . , fi−1}. Additionally, we assume that argument values are conditionally independent given fi and f<i, resulting in: p(πi | f<i) = p(fi |f<i) | {z } function scoring Ai Y j=1 p(vij |f<i, fi) | {z } argument value scoring , (3) where we have elided the conditioning on Henc. Here, functions depend only on previous functions 3670 FUNCTION EMBEDDER from: City NAME TYPE argument embedding FUNCTION SIGNATURE NAME TYPE TYPE ARGUMENT ARGUMENT Book[Flight](from: City, to: City): Booking[Flight] POOLING function embedding ARGUMENT EMBEDDER Figure 5: Illustration of our function encoder (§2.4), using a simplified example function signature. (not their argument values or results) and argument values depend only on their calling function (not on one another or any of the previous argument values).4 This is illustrated in Figure 4. In addition to providing an important inductive bias, these independence assumptions allow our inference procedure to efficiently score all possible function invocations at step i, given the ones at previous steps, at once (i.e., function and argument value assignments together), resulting in an efficient search algorithm (§2.6). Note that there is also a corresponding disadvantage (as in many machine translation models) that a meaningful phrase in the utterance could be independently selected for multiple arguments, or not selected at all, but we did not encounter this issue in our experiments; we rely on the model training to evade this problem through the dependence on Henc. 2.4 Decoder: Functions In Equation 3, the sequence of functions f1, f2, . . . in the current program is modeled by Q i p(fi |f<i, Henc). We use a standard autoregressive Transformer decoder that can also attend to the utterance encoding Henc (§2.2), as done by Vaswani et al. (2017). Our decoder generates sequences over the vocabulary of functions. This means that each function fi needs an embedding fi (used as both an input to the decoder and an output), which we construct compositionally. We assume that each unique function f has a type signature that specifies a name n, a list of type parameters {τ1, . . . , τT } (to support polymorphism),5 a list of argument names and types ((a1, t1), . . . , (aA, tA)), and a result type r. An 4We also tried defining a jointly normalized distribution over entire function invocations (Appendix A), but found that it results in a higher training cost for no accuracy benefits. 5The type parameters could themselves be parameterized, but we ignore this here for simplicity of exposition. example is shown in Figure 5. We encode the function and argument names using the utterance encoder of §2.2 and learn embeddings for the types, to obtain (n, r), {τ1, . . . , τT }, and {(a1, t1), . . . , (aA, tA)}. Then, we construct an embedding for each function as follows: a = Pool(a1 + t1, . . . , aA + tA), (4) f = n + Pool(τ1, . . . , τT ) + a + r, (5) where “Pool” is the max-pooling operation which is invariant to the arguments’ order. Our main motivation for this function embedding mechanism is the ability to take cues from the user utterance (e.g., due to a function being named similarly to a word appearing in the utterance). If the functions and their arguments have names that are semantically similar to corresponding utterance parts, then this approach enables zero-shot generalization.6 However, there is an additional potential benefit from parameter sharing due to the compositional structure of the embeddings (see e.g., Baroni, 2020). 2.5 Decoder: Argument Values This section describes the implementation of the argument predictor p(vij | f<i, fi). There are four different kinds of sources that can be used to fill each available argument slot: references to previous function invocations, constants from a static vocabulary, copies that copy string values from the utterance, and entities that come from entity proposers (§2.1). Many sources might propose the same value, including multiple sources of the same kind. For example, there may be multiple spans in the utterance that produce the same string value in a program, or an entity may be proposed that is also available as a constant. To address this, we marginalize over the sources of each value: p(vij | f<i, fi)= X s∈S(vij) p(vij, s|f<i, fi), (6) where vij represents a possible value for the argument named aij, and s ∈S(vij) ranges over the possible sources for that value. For example, given the utterance “Change that one to 1:30pm” and the value 1, the set S(1) may contain entities that correspond to both “one” and “1” from the utterance. 6The data may contain overloaded functions that have the same name but different type signatures (e.g., due to optional arguments). The overloads are given distinct identifiers f, but they often share argument names, resulting in at least partially shared embeddings. 3671 The argument scoring mechanism considers the last-layer decoder state hi dec that was used to predict fi via p(fi |f<i) ∝exp(f ⊤ i hi dec). We specialize this decoder state to argument aij as follows: hi,aij dec ≜ˆhi dec ⊙tanh(fi + aij), (7) where ⊙represents elementwise multiplication, fi is the embedding of the current function fi, aij is the encoding of argument aij as defined in §2.4, and ˆhdec is a projection of hdec to the necessary dimensionality. Intuitively, tanh(fi + aij) acts as a gating function over the decoder state, deciding what is relevant when scoring values for argument aij. This argument-specific decoder state is then combined with a value embedding to produce a probability for each (sourced) value assignment: p(v, s | f<i, fi) ∝ exp n ˜v⊤(hi,a dec + wkind(s) a ) + bkind(s) a o , (8) where a is the argument name aij, kind(s) ∈ {REFERENCE, CONSTANT, COPY, ENTITY}, ˜v is the embedding of (v, s) which is described next, and wk a and bk a are model parameters that are specific to a and the kind of the source s. References. References are pointers to the return values of previous function invocations. If the source s for the proposed value v is the result of the kth invocation (where k < i), we take its embedding ˜v to be a projection of hk dec that was used to predict that invocation’s function and arguments. Constants. Constants are values that are always proposed, so the decoder always has the option of generating them. If the source s for the proposed value v is a constant, we embed it by applying the utterance encoder on a string rendering of the value. The set of constants is automatically extracted from the training data (see Appendix B). Copies. Copies are string values that correspond to substrings of the user utterance (e.g., person names). String values can only enter the program through copying, as they are not in the set of constants (i.e., they cannot be “hallucinated” by the model; see Pasupat and Liang, 2015; Nie et al., 2019). One might try to construct an approach based on a standard token-based copy mechanism (e.g., Gu et al., 2016). However, this would allow copying non-contiguous spans and would also require marginalizing over identical tokens as opposed to spans, resulting in more ambiguity. Instead, we propose a mechanism that enables the decoder to copy contiguous spans directly from the utterance. Its goal is to produce a score for each of the U(U + 1)/2 possible utterance spans. Na¨ıvely, this would result in a computational cost that is quadratic in the utterance length U, and so we instead chose a simple scoring model that avoids it. Similar to Stern et al. (2017) and Kuribayashi et al. (2019), we assume that the score for a span factorizes, and define the embedding of each span value as the concatenation of the contextual embeddings of the first and last tokens of the span, ˜v = [hkstart utt ; hkend utt ]. To compute the copy scores we also concatenate hi,a dec with itself in Equation 8. Entities. Entities are treated the same way as copies, except that instead of scoring all spans of the input, we only score spans proposed by the external entity proposers discussed in §2.1. Specifically, the proposers provide the model with a list of candidate entities that are each described by an utterance span and an associated value. The candidates are scored using an identical mechanism to the one used for scoring copies. This means that, for example, the string “sept” could be linked to the value Month.September even though the string representations do not match perfectly. Type Checking. When scoring argument values for function fi, we know the argument types, as they are specified in the function’s signature. This enables us to use a type checking mechanism that allows the decoder to directly exclude values with mismatching types. For references, the value types can be obtained by looking up the result types of the corresponding function signatures. Additionally, the types are always pre-specified for constants and entities, and copies are only supported for a subset of types (e.g., String, PersonName; see Appendix B). The type checking mechanism sets p(vij | f<i, fi) = 0 whenever vij has a different type than the expected type for aij. Finally, because copies can correspond to multiple types, we also add a type matching term to the copy score. This term is defined as the inner product of the argument type embedding and a (learnable) linear projection of hkstart utt and hkend utt concatenated, where kstart and kend denote the span start and end indices. 2.6 Decoder: Search Similar to other sequence-to-sequence models, we employ beam search over the sequence of function invocations when decoding. However, in contrast to other models, our assumptions (§2.3) allow us to 3672 Dataset SMCALFLOW TREEDST V1.1 V2.0 Best Reported Result 66.5 68.2 62.2 Our Model 73.8 75.3 72.8 Table 1: Test set exact match accuracy comparing our model to the best reported results for SMCALFLOW (Seq2Seq model from the public leaderboard; Semantic Machines et al., 2020) and TREEDST (TED-PP model; Cheng et al., 2020). The evaluation on each dataset in prior work requires us to repeat some idiosyncrasies that we describe in Appendix D. efficiently implement beam search over complete function invocations, by leveraging the fact that: max πi p(πi)=max fi p(fi) Ai Y j=1 max vij p(vij |fi) , (9) where we have omitted the dependence on f<i. This computation is parallelizable and it also allows the decoder to avoid choosing a function if there are no high scoring assignments for its arguments (i.e., we are performing a kind of lookahead). This also means that the paths explored during the search are shorter for our model than for models where each step corresponds to a single decision, allowing for smaller beams and more efficient decoding. 3 Experiments We first report results on SMCALFLOW (Semantic Machines et al., 2020) and TREEDST (Cheng et al., 2020), two recently released large-scale conversational semantic parsing datasets. Our model makes use of type information in the programs, so we manually constructed a set of type declarations for each dataset and then used a variant of the HindleyMilner type inference algorithm (Damas and Milner, 1982) to annotate programs with types. As mentioned in §2.1, we also transformed TREEDST to introduce meta-computation operators for references and revisions (more details can be found in Appendix C).7 We also report results on nonconversational semantic parsing datasets in §3.2. We use the same hyperparameters across all experiments (see Appendix E), and we use BERTmedium (Turc et al., 2019) to initialize our encoder. 3.1 Conversational Semantic Parsing Test set results for SMCALFLOW and TREEDST are shown in Table 1. Our model significantly outperforms the best published numbers in each case. 7The transformed datasets are available at https: //github.com/microsoft/task_oriented_dialogue_ as_dataflow_synthesis/tree/master/datasets. Dataset SMCALFLOW TREEDST # Training Dialogues 1k 10k 33k 1k 10k 19k Seq2Seq 36.8 69.8 74.5 28.2 47.9 50.3 Seq2Tree 43.6 69.3 77.7 23.6 46.9 48.8 Seq2Tree++ 48.0 71.9 78.2 74.8 75.4 86.9 w/o BERT Our Model 53.8 73.2 78.5 78.6 87.6 88.5 Seq2Seq 44.6 64.1 67.8 28.6 40.2 47.2 Seq2Tree 50.8 74.6 78.6 30.9 50.6 51.6 w/ BERT Our Model 63.2 77.2 80.4 81.2 87.1 88.3 (a) Baseline comparison. Dataset SMCALFLOW TREEDST # Training Dialogues 1k 10k 33k 1k 10k 19k Our Model 63.2 77.2 80.4 81.2 87.1 88.3 Value Dependence 60.6 76.4 79.4 79.3 86.2 86.5 No Name Embedder 62.8 76.7 80.3 81.1 87.0 88.1 No Types 62.4 76.5 79.9 80.6 87.1 88.3 No Span Copy 60.2 76.2 79.8 79.0 86.7 87.4 No Entity Proposers 59.6 76.4 79.8 80.5 86.9 88.2 Parser All of the Above 58.9 75.8 77.3 72.9 80.2 80.6 No History 59.0 70.0 73.8 68.3 75.0 76.5 Previous Turn 61.3 75.9 77.4 80.5 86.9 87.4 History Linear Encoder 63.0 76.5 80.2 81.2 87.1 88.3 (b) Ablation study. Table 2: Validation set exact match accuracy across varying amounts of training data (each subset is sampled uniformly at random). The best results in each case are shown in bold red and are underlined. In order to further understand the performance characteristics of our model and quantify the impact of each modeling contribution, we also compare to a variety of other models and ablated versions of our model. We implemented the following baselines: – Seq2Seq: The OpenNMT (Klein et al., 2017) implementation of a pointer-generator network (See et al., 2017) that predicts linearized plans represented as S-expressions and is able to copy tokens from the utterance while decoding. This model is very similar to the model used by Semantic Machines et al. (2020) and represents the current state-of-the-art for SMCALFLOW.8 – Seq2Tree: The same as Seq2Seq, except that it generates invocations in a top-down, pre-order program traversal. Each invocation is embedded as a unique item in the output vocabulary. Note that SMCALFLOW contains re-entrant programs represented with LISP-style let bindings. Both the Seq2Tree and Seq2Seq are unaware of the special meaning of let and predict calls to let as any other function, and references to bound 8Semantic Machines et al. (2020) used linearized plans to represent the dialogue history, but our implementation uses previous user and agent utterances. We found no difference in performance. 3673 variables as any other literal. – Seq2Tree++: An enhanced version of the model by Krishnamurthy et al. (2017) that predicts typed programs in a top-down fashion. Unlike Seq2Seq and Seq2Tree, this model can only produce well-formed and well-typed programs. It also makes use of the same entity proposers (§2.1) similar to our model, and it can atomically copy spans of up to 15 tokens by treating them as additional proposed entities. Furthermore, it uses the linear history encoder that is described in the next paragraph. Like our model, re-entrancies are represented as references to previous outputs in the predicted sequence. We also implemented variants of Seq2Seq and Seq2Tree that use BERT-base9 (Devlin et al., 2019) as the encoder. Our results are shown in Table 2a. Our model outperforms all baselines on both datasets, showing particularly large gains in the low data regime, even when using BERT. Finally, we implemented the following ablations, with more details provided in Appendix G: – Value Dependence: Introduces a unique function for each value in the training data (except for copies) and transforms the data so that values are always produced by calls to these functions, allowing the model to condition on them. – No Name Embedder: Embeds functions and constants atomically instead of using the approach of §2.4 and the utterance encoder. – No Types: Collapses all types to a single type, which effectively disables type checking (§2.5). – No Span Copy: Breaks up span-level copies into token-level copies which are put together using a special concatenate function. Note that our model is value-agnostic and so this ablated model cannot condition on previously copied tokens when copying a span token-by-token. – No Entity Proposers: Removes the entity proposers, meaning that previously entity-linked values have to be generated as constants. – No History: Sets Henc = Hutt (§2.2). – Previous Turn: Replaces the type-based history encoding with the previous turn user and system utterances or linearized system actions. – Linear Encoder: Replaces the history attention 9We found that BERT-base worked best for these baselines, but was no better than the smaller BERT-medium when used with our model. Also, unfortunately, incorporating BERT in Seq2Tree++ turned out to be challenging due to the way that model was originally implemented. Method Dataset JOBS GEO ATIS Zettlemoyer and Collins (2007) — 86.1 84.6 Wang et al. (2014) 90.7 90.4 91.3 Zhao and Huang (2015) 85.0 88.9 84.2 Saparov et al. (2017) 81.4 83.9 — Dong and Lapata (2016) 90.0 87.1 84.6 Rabinovich et al. (2017) 92.9 87.1 85.9 Yin and Neubig (2018) — 88.2 86.2 Dong and Lapata (2018) — 88.2 87.7 Aghajanyan et al. (2020) — 89.3 — Our Model 91.4 91.4 90.2 Neural Methods ⌞No BERT 91.4 90.0 91.3 Table 3: Validation set exact match accuracy for singleturn semantic parsing datasets. Note that Aghajanyan et al. (2020) use BART (Lewis et al., 2020), a large pretrained encoder. The best results for each dataset are shown in bold red and are underlined. mechanism with a linear function over a multihot embedding of the history types. The results, shown in Table 2b, indicate that all of our features play a role in improving accuracy. Perhaps most importantly though, the “value dependence” ablation shows that our function-based program representations are indeed important, and the “previous turn” ablation shows that our typebased program representations are also important. Furthermore, the impact of both these modeling decisions grows larger in the low data regime, as does the impact of the span copy mechanism. 3.2 Non-Conversational Semantic Parsing Our main focus is on conversational semantic parsing, but we also ran experiments on nonconversational semantic parsing benchmarks to show that our model is a strong parser irrespective of context. Specifically, we manually annotated the JOBS, GEOQUERY, and ATIS datasets with typed declarations (Appendix C) and ran experiments comparing with multiple baseline and state-of-the-art methods. The results, shown in Table 3, indicate that our model meets or exceeds state-of-the-art performance in each case. 4 Related Work Our approach builds on top of a significant amount of prior work in neural semantic parsing and also context-dependent semantic parsing. Neural Semantic Parsing. While there was a brief period of interest in using unstructured sequence models for semantic parsing (e.g., Andreas 3674 et al., 2013; Dong and Lapata, 2016), most research on semantic parsing has used tree- or graph-shaped decoders that exploit program structure. Most such approaches use this structure as a constraint while decoding, filling in function arguments one-at-atime, in either a top-down fashion (e.g., Dong and Lapata, 2016; Krishnamurthy et al., 2017) or a bottom-up fashion (e.g., Misra and Artzi, 2016; Cheng et al., 2018). Both directions can suffer from exposure bias and search errors during decoding: in top-down when there’s no way to realize an argument of a given type in the current context, and in bottom-up when there are no functions in the programming language that combine the predicted arguments. To this end, there has been some work on global search with guarantees for neural semantic parsers (e.g., Lee et al., 2016) but it is expensive and makes certain strong assumptions. In contrast to this prior work, we use program structure not just as a decoder constraint but as a source of independence assumptions: the decoder explicitly decouples some decisions from others, resulting in good inductive biases and fast decoding algorithms. Perhaps closest to our work is that of Dong and Lapata (2018), which is also about decoupling decisions, but uses a dataset-specific notion of an abstracted program sketch along with different independence assumptions, and underperforms our model in comparable settings (§3.2). Also close are the models of Cheng et al. (2020) and Zhang et al. (2019). Our method differs in that our beam search uses larger steps that predict functions together with their arguments, rather than predicting the argument values serially in separate dependent steps. Similar to Zhang et al. (2019), we use a target-side copy mechanism for generating references to function invocation results. However, we extend this mechanism to also predict constants, copy spans from the user utterance, and link externally proposed entities. While our span copy mechanism is novel, it is inspired by prior attempts to copy spans instead of tokens (e.g., Singh et al., 2020). Finally, bottom-up models with similarities to ours include SMBOP (Rubin and Berant, 2020) and BUSTLE (Odena et al., 2020). Context-Dependent Semantic Parsing. Prior work on conversational semantic parsing mainly focuses on the decoder, with few efforts on incorporating the dialogue history information in the encoder. Recent work on context-dependent semantic parsing (e.g., Suhr et al., 2018; Yu et al., 2019) conditions on explicit representations of user utterances and programs with a neural encoder. While this results in highly expressive models, it also increases the risk of overfitting. Contrary to this, Zettlemoyer and Collins (2009), Lee et al. (2014) and Semantic Machines et al. (2020) do not use context to resolve references at all. They instead predict context-independent logical forms that are resolved in a separate step. Our approach occupies a middle ground: when combined with local program representations, types, even without any value information, provide enough information to resolve context-dependent meanings that cannot be derived from isolated sentences. The specific mechanism we use to do this “infuses” contextual type information into input sentence representations, in a manner reminiscent of attention flow models from the QA literature (e.g., Seo et al., 2016). 5 Conclusion We showed that abstracting away values while encoding the dialogue history and decoding programs significantly improves conversational semantic parsing accuracy. In summary, our goal in this work is to think about types in a new way. Similar to previous neural and non-neural methods, types are an important source of constraints on the behavior of the decoder. Here, for the first time, they are also the primary ingredient in the representation of both the parser actions and the dialogue history. Our approach, which is based on type-centric encodings of dialogue states and function-centric encodings of programs (§2), outperforms prior work by 7.3% and 10.6%, on SMCALFLOW and TREEDST, respectively (§3), while also being more computationally efficient than competing methods. Perhaps more importantly, it results in even more significant gains in the low-data regime. This indicates that choosing our representations carefully and making appropriate independence assumptions can result in increased accuracy and computational efficiency. 6 Acknowledgements We thank the anonymous reviewers for their helpful comments, Jason Eisner for his detailed feedback and suggestions on an early draft of the paper, Abulhair Saparov for helpful conversations and pointers about semantic parsing baselines and prior work, and Theo Lanman for his help in scaling up some of our experiments. 3675 References Armen Aghajanyan, Jean Maillard, Akshat Shrivastava, Keith Diedrick, Michael Haeger, Haoran Li, Yashar Mehdad, Veselin Stoyanov, Anuj Kumar, Mike Lewis, and Sonal Gupta. 2020. Conversational Semantic Parsing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5026–5035. Association for Computational Linguistics. Jacob Andreas, Andreas Vlachos, and Stephen Clark. 2013. Semantic Parsing as Machine Translation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 47–52, Sofia, Bulgaria. Association for Computational Linguistics. Marco Baroni. 2020. Linguistic Generalization and Compositionality in Modern Artificial Neural Networks. Philosophical Transactions of the Royal Society B, 375(1791):20190307. Jianpeng Cheng, Devang Agrawal, H´ector Mart´ınez Alonso, Shruti Bhargava, Joris Driesen, Federico Flego, Dain Kaplan, Dimitri Kartsaklis, Lin Li, Dhivya Piraviperumal, Jason D. Williams, Hong Yu, Diarmuid ´O S´eaghdha, and Anders Johannsen. 2020. Conversational Semantic Parsing for Dialog State Tracking. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8107–8117. Association for Computational Linguistics. Jianpeng Cheng, Siva Reddy, Vijay Saraswat, and Mirella Lapata. 2018. Learning an Executable Neural Semantic Parser. Computational Linguistics, 45(1):59–94. Publisher: MIT Press. Luis Damas and Robin Milner. 1982. Principal TypeSchemes for Functional Programs. In Proceedings of the 9th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’82, page 207–212, New York, NY, USA. Association for Computing Machinery. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Li Dong and Mirella Lapata. 2016. Language to Logical Form with Neural Attention. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 33–43, Berlin, Germany. Association for Computational Linguistics. Li Dong and Mirella Lapata. 2018. Coarse-to-Fine Decoding for Neural Semantic Parsing. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 731–742, Melbourne, Australia. Association for Computational Linguistics. Jiatao Gu, Zhengdong Lu, Hang Li, and Victor O.K. Li. 2016. Incorporating Copying Mechanism in Sequence-to-Sequence Learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1631–1640, Berlin, Germany. Association for Computational Linguistics. Kelvin Guu, Panupong Pasupat, E. Liu, and Percy Liang. 2017. From language to programs: Bridging reinforcement learning and maximum marginal likelihood. ArXiv, abs/1704.07926. Robin Jia and Percy Liang. 2016. Data recombination for neural semantic parsing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12–22, Berlin, Germany. Association for Computational Linguistics. Bevan Jones, Jacob Andreas, Daniel Bauer, Karl Moritz Hermann, and Kevin Knight. 2012. Semantics-Based Machine Translation with Hyperedge Replacement Grammars. In Proceedings of COLING 2012, pages 1359–1376, Mumbai, India. The COLING 2012 Organizing Committee. Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Optimization. arXiv:1412.6980 [cs.LG]. ArXiv: 1412.6980. Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. OpenNMT: Opensource toolkit for neural machine translation. In Proceedings of ACL 2017, System Demonstrations, pages 67–72, Vancouver, Canada. Association for Computational Linguistics. Jayant Krishnamurthy, Pradeep Dasigi, and Matt Gardner. 2017. Neural Semantic Parsing with Type Constraints for Semi-Structured Tables. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1516–1526, Copenhagen, Denmark. Association for Computational Linguistics. Tatsuki Kuribayashi, Hiroki Ouchi, Naoya Inoue, Paul Reisert, Toshinori Miyoshi, Jun Suzuki, and Kentaro Inui. 2019. An Empirical Study of Span Representations in Argumentation Structure Parsing. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4691– 4698, Florence, Italy. Association for Computational Linguistics. Kenton Lee, Yoav Artzi, Jesse Dodge, and Luke Zettlemoyer. 2014. Context-dependent Semantic Parsing for Time Expressions. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3676 1437–1447, Baltimore, Maryland. Association for Computational Linguistics. Kenton Lee, Mike Lewis, and Luke Zettlemoyer. 2016. Global Neural CCG Parsing with Optimality Guarantees. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2366–2376, Austin, Texas. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising Sequence-to-Sequence Pretraining for Natural Language Generation, Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880. Association for Computational Linguistics. Dipendra Kumar Misra and Yoav Artzi. 2016. Neural Shift-Reduce CCG Semantic Parsing. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1775–1786, Austin, Texas. Association for Computational Linguistics. Feng Nie, Jin-Ge Yao, Jinpeng Wang, Rong Pan, and Chin-Yew Lin. 2019. A Simple Recipe towards Reducing Hallucination in Neural Surface Realisation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2673–2679, Florence, Italy. Association for Computational Linguistics. Augustus Odena, Kensen Shi, David Bieber, Rishabh Singh, and Charles Sutton. 2020. BUSTLE: Bottomup Program Synthesis Through Learning-guided Exploration. arXiv:2007.14381 [cs, stat]. ArXiv: 2007.14381. Panupong Pasupat and Percy Liang. 2015. Compositional Semantic Parsing on Semi-Structured Tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1470–1480, Beijing, China. Association for Computational Linguistics. Maxim Rabinovich, Mitchell Stern, and Dan Klein. 2017. Abstract Syntax Networks for Code Generation and Semantic Parsing. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1139–1149, Vancouver, Canada. Association for Computational Linguistics. Ohad Rubin and Jonathan Berant. 2020. SmBoP: Semi-autoregressive Bottom-up Semantic Parsing. arXiv:2010.12412 [cs]. ArXiv: 2010.12412. Abulhair Saparov, Vijay Saraswat, and Tom Mitchell. 2017. Probabilistic Generative Grammar for Semantic Parsing. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 248–259, Vancouver, Canada. Association for Computational Linguistics. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073– 1083, Vancouver, Canada. Association for Computational Linguistics. Semantic Machines, Jacob Andreas, John Bufe, David Burkett, Charles Chen, Josh Clausman, Jean Crawford, Kate Crim, Jordan DeLoach, Leah Dorner, Jason Eisner, Hao Fang, Alan Guo, David Hall, Kristin Hayes, Kellie Hill, Diana Ho, Wendy Iwaszuk, Smriti Jha, Dan Klein, Jayant Krishnamurthy, Theo Lanman, Percy Liang, Christopher H. Lin, Ilya Lintsbakh, Andy McGovern, Aleksandr Nisnevich, Adam Pauls, Dmitrij Petters, Brent Read, Dan Roth, Subhro Roy, Jesse Rusak, Beth Short, Div Slomin, Ben Snyder, Stephon Striplin, Yu Su, Zachary Tellman, Sam Thomson, Andrei Vorobev, Izabela Witoszko, Jason Wolfe, Abby Wray, Yuchen Zhang, and Alexander Zotov. 2020. Task-Oriented Dialogue as Dataflow Synthesis. Transactions of the Association for Computational Linguistics, 8:556–571. Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. 2016. Bidirectional Attention Flow for Machine Comprehension. arXiv:1611.01603 [cs.CL]. ArXiv: 1611.01603. Abhinav Singh, Patrick Xia, Guanghui Qin, Mahsa Yarmohammadi, and Benjamin Van Durme. 2020. CopyNext: Explicit Span Copying and Alignment in Sequence to Sequence Models. In Proceedings of the Fourth Workshop on Structured Prediction for NLP, pages 11–16. Association for Computational Linguistics. Mitchell Stern, Jacob Andreas, and Dan Klein. 2017. A Minimal Span-Based Neural Constituency Parser. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 818–827, Vancouver, Canada. Association for Computational Linguistics. Alane Suhr, Srinivasan Iyer, and Yoav Artzi. 2018. Learning to Map Context-Dependent Sentences to Executable Formal Queries. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2238–2249, New Orleans, Louisiana. Association for Computational Linguistics. Iulia Turc, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Well-Read Students Learn Better: On the Importance of Pre-training Compact Models. arXiv:1908.08962 [cs.CL]. ArXiv: 1908.08962. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is All 3677 you Need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008. Curran Associates, Inc. Adrienne Wang, Tom Kwiatkowski, and Luke Zettlemoyer. 2014. Morpho-syntactic Lexical Generalization for CCG Semantic Parsing. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1284– 1295, Doha, Qatar. Association for Computational Linguistics. Pengcheng Yin and Graham Neubig. 2018. TRANX: A Transition-based Neural Abstract Syntax Parser for Semantic Parsing and Code Generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 7–12, Brussels, Belgium. Association for Computational Linguistics. Tao Yu, Rui Zhang, Michihiro Yasunaga, Yi Chern Tan, Xi Victoria Lin, Suyi Li, Heyang Er, Irene Li, Bo Pang, Tao Chen, Emily Ji, Shreya Dixit, David Proctor, Sungrok Shim, Jonathan Kraft, Vincent Zhang, Caiming Xiong, Richard Socher, and Dragomir Radev. 2019. SParC: Cross-Domain Semantic Parsing in Context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4511–4523, Florence, Italy. Association for Computational Linguistics. Luke Zettlemoyer and Michael Collins. 2007. Online Learning of Relaxed CCG Grammars for Parsing to Logical Form. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 678–687, Prague, Czech Republic. Association for Computational Linguistics. Luke Zettlemoyer and Michael Collins. 2009. Learning Context-Dependent Mappings from Sentences to Logical Form. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 976–984, Suntec, Singapore. Association for Computational Linguistics. Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR Parsing as Sequence-toGraph Transduction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 80–94, Florence, Italy. Association for Computational Linguistics. Kai Zhao and Liang Huang. 2015. Type-Driven Incremental Semantic Parsing with Polymorphism. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1416–1421, Denver, Colorado. Association for Computational Linguistics. 3678 A Invocation Joint Normalization Instead of the distribution in Equation 3, we can define a distribution over fi and {vij}Ai j=1 that factorizes in the same way but is also jointly normalized: p(πi | f<i) ∝h(fi) Ai Y j=1 g(fi, vij), (10) where h and g are defined as presented in §2.4 and §2.5, respectively, before normalization. This model has the same cost as the locally normalized model at test time but is significantly more expensive at training time as we need to score all possible function invocations, as opposed to always conditioning on the gold functions. It can in principle avoid some of the exposure bias problems of the locally normalized model, but we observed no accuracy improvements in our experiments. B Value Sources In our model, the type of a value determines what sources it can be generated from. We enforce that values of certain types can only be copied or entitylinked. Any values that do not fall under these constraints are added to a static vocabulary of constants, and the model is always permitted to generate them, as long as they pass type checking. Values that fall under these constraints are not added to this vocabulary so that they cannot be “hallucinated” by the model. The specific constraints that we use are described in the following paragraphs. Types that must be copied: Types for which the model is only allowed to construct values directly from string literals copied from the utterance. In §2.5 we noted that strings can be copied from the utterance to become string literals in the generated program. For certain types t, arguments of type t may also be willing to accept copied strings; in this case we generate a constructor call that constructs a t object from the string literal. For SMCALFLOW, these copyable types are String, PersonName, RespondComment, and LocationKeyphrase. For the other datasets it is just String. We declare training examples where a value of a copyable type appears in the program, but is not a substring of the corresponding utterance, as likely annotation errors and ignore them during training (but not during evaluation). Even though such examples are very rare for SMCALFLOW (∼0.5% of the examples), they turned out to be relatively frequent in TREEDST (∼6% of the examples), as we discuss in Appendix C. Types that must be entity-linked: Types for which argument values can only be picked from the set of proposed entities (§2.1) and cannot be otherwise hallucinated from the model, or directly copied from the utterance. The Number type is treated in a special way for all datasets, where numbers 0, 1, and 2 are allowed to be hallucinated, but all other numbers must be entity-linked. Furthermore, for SMCALFLOW the set of types that must be entity-linked also contains the Month, DayOfWeek, and Holiday types. Based on this, we can detect probable annotation errors. C Dataset Preparation We now describe how we processed the datasets to satisfy the requirements mentioned in §2.1. We have made the processed datasets available at https: //github.com/microsoft/task_oriented_dialogue_ as_dataflow_synthesis/tree/master/datasets. C.1 Type Declarations We manually specified the necessary type declarations by inspection of all functions in the training data. In some cases, we found it helpful to transform the data into an equivalent set of function calls that simplified the resulting programs, while maintaining a one-to-one mapping with the original representations. For example, SMCALFLOW contains a function called get that takes in an object of some type and a Path, which specifies a field of that object, and acts as an accessor. For example, the object could be an Event and the specified path may be "subject". We transform such invocations into invocations of functions that are instantiated separately for each unique combination of the object type and the provided path. For the aforementioned example, the corresponding new function would be defined as: def Event.subject(obj: Event): String All such transformations are invertible, so we can convert back to the original format after prediction. C.2 Meta-Computation Operators The meta-computation operators are only required for the conversational semantic parsing datasets, and SMCALFLOW already makes use of them. Therefore, we only had to convert TREEDST. To this end, we introduced two new operators: 3679 def refer[T](): T def revise[T, R]( root: Root[T], path: Path[R], revision: R => R, ): T refer goes through the programs and system actions in the dialogue history, starting at the most recent turn, finds the first sub-program that evaluates to type T, and replaces its invocation with that sub-program. Similarly, revise finds the first program whose root matches the specified root, walks down the tree along the specified path, and applies the provided revision on the sub-program rooted at the end of that path. It then replaces its invocation with this revised program. We performed an automated heuristic transformation of TREEDST so that it makes use of these meta-operators. We only applied the extracted transformations when executing them on the transformed programs using the gold dialogue history resulted in the original program (i.e., before applying any of our transformations). Therefore, when using the gold dialogue history, this transformation is also guaranteed to be invertible. We emphasize that we execute these meta-computation operators before computing accuracy so that our final evaluation results are comparable to prior work. C.3 Annotation Errors While preparing the datasets for our experiments using our automated transformations, we noticed that they contain some inconsistencies. For example, in TREEDST, the tree fragment: ...restaurant.book.restaurant.book... seemed to be interchangeable with: ...restaurant.object.equals... The annotation and checking mechanisms we employ impose certain regularity requirements on the data that are violated by such examples. Therefore, we had three choices for such examples: (i) we could add additional type declarations, (ii) we could discard them, or (iii) we could collapse the two annotations together, resulting in a lossy conversion. We used our best judgment when choosing among these options, preferring option (iii) where it was possible to do so automatically. We believe that all such cases are annotation errors, but we cannot know for certain without more information about how the TREEDST dataset was constructed. Overall, about 122 dialogues (0.4%) did not pass our checks for SMCALFLOW, and 585 dialogues (3.0%) for TREEDST. When converting back to the original format, we tally an error for each discarded example, and select the most frequent version of any lossily collapsed annotation. Our approach also provides two simple yet effective consistency checks for the training data: (i) running type inference using the provided type declarations to detect ill-typed examples, and (ii) using the constraints described Appendix B to detect other forms of annotation errors. We found that these two checks together caught 68 potential annotation errors (<0.5%) in SMCALFLOW and ∼1,000 potential errors (∼6%) in TREEDST. TREEDST was particularly interesting as we found a whole class of examples where user utterances were replaced with system utterances. Note that our model does not technically require any of these checks. It is possible to generate type signatures that permit arbitrary function/argument pairs based on observed data and to configure our model so that any observed value may be generated as a constant (i.e., not imposing the constraints described in Appendix B). In practice we found that constraining the space of programs provides useful sanity checks in addition to accuracy gains. C.4 Non-Conversational Semantic Parsing We obtained the JOBS, GEOQUERY, and ATIS datasets from the repository of Dong and Lapata (2016). For each dataset, we defined a library that specifies function and type declarations. D Evaluation Details To compare with prior work for SMCALFLOW (Semantic Machines et al., 2020) and TREEDST (Cheng et al., 2020), we replicated their setups. For SMCALFLOW, we predict plans always conditioning on the gold dialogue history for each utterance, but we consider any predicted plan wrong if the refer are correct flag is set to false. This flag is meant to summarize the accuracy of a hypothetical model for resolving calls to refer, but is not relevant to the problem of program prediction. We also canonicalize plans by sorting keyword arguments and normalizing numbers (so that 30.0 and 30 are considered equivalent, for example). For TREEDST, our model predicts programs that use the refer and revise operators, and we execute them against the dialogue history that consists of predicted programs and gold (oracle) system 3680 actions (following Cheng et al. (2020)) when converting back to the original tree representation. We canonicalize the resulting trees by lexicographically sorting the children of each node. For our baseline comparisons and ablations (shown in Tables 2a and 2b), we decided to ignore the refer are correct flag for SMCALFLOW because it assumes that refer is handled by some other model and for these experiments we are only interested in evaluating program prediction. Also, for TREEDST we use the gold plans for the dialogue history in order to focus on the semantic parsing problem, as opposed to the dialogue state tracking problem. For the non-conversational semantic parsing datasets we replicated the evaluation approach of Dong and Lapata (2016), and so we also canonicalize the predicted programs. E Model Hyperparameters We use the same hyperparameters for all of our conversational semantic parsing experiments. For the encoder, we use either BERT-medium (Turc et al., 2019) or a non-pretrained 2-layer Transformer (Vaswani et al., 2017) with a hidden size of 128, 4 heads, and a fully connected layer size of 512, for the non-BERT experiments. For the decoder we use a 2-layer Transformer with a hidden size of 128, 4 heads, and a fully connected layer size of 512, and set htype to 128, and harg to 512. For the non-conversational semantic parsing experiments we use a hidden size of 32 throughout the model as the corresponding datasets are very small. We also use a dropout of 0.2 for all experiments. For training, we use the Adam optimizer (Kingma and Ba, 2017), performing global gradient norm clipping with the maximum allowed norm set to 10. For batching, we bucket the training examples by utterance length and adapt the batch size so that the total number of tokens in each batch is 10,240. Finally, we average the log-likelihood function over each batch, instead of summing it. Experiments with BERT. We use a pre-training phase for 2,000 training steps, where we freeze the parameters of the utterance encoder and only train the dialogue history encoder and the decoder. Then, we train the whole model for another 8,000 steps. This because our model is not simply adding a linear layer on top of BERT, and so, unless initialized properly, we may end up losing some of the information contained in the pre-trained BERT model. During the pre-training phase, we linearly warm up the learning rate to 2 × 10−3 during the first 1,000 steps. We then decay it exponentially by a factor of 0.999 every 10 steps. During the full training phase, we linearly warm up the learning rate to 1 × 10−4 during the first 1,000 steps, and then decay it exponentially in the same fashion. Experiments without BERT. We use a single training phase for 30,000 steps, where we linearly warm up the learning rate to 5 × 10−3 during the first 1,000 steps, and then we decay it exponentially by a factor of 0.999 every 10 steps. We need a larger number of training steps in this case because none of the model components have been pre-trained. Also, the encoder is now much smaller, meaning that we can afford a higher learning rate. Even though these hyperparameters may seem very specific, we emphasize that our model is robust to the choice of hyperparameters and this setup was chosen once and shared across all experiments. F Baseline Models Seq2Seq. This model predicts linearized, tokenized S-expressions using the OpenNMT implementation of a Transformer-based (Vaswani et al., 2017) pointer-generator network (See et al., 2017). For example, the following program: +(length("some string"), 1) would correspond to the space-separated sequence: ( + ( length " some string " ) 1 ) In contrast to the model proposed in this paper, in this case tokens that belong to functions and values (i.e., that are outside of quotes) can also be copied directly from the utterance. Furthermore, there is no guarantee that this baseline will produce a well-formed program. Seq2Tree. This model uses the same underlying implementation as our Seq2Seq baseline—also with no guarantee that it will produce a well-formed program—but it predicts a different sequence. For example, the following program: +(+(1, 2), 3) would be predicted as the sequence: +(<NT>, 3) +(1, 2) Each item in the sequence receives a unique embedding in the output vocabulary and so, "+(1,2)" and "+(<NT>, 3)" share no parameters. <NT> is 3681 a special placeholder symbol that represents a substitution point when converting the linearized sequence back to a tree. Furthermore, copies are not inlined into invocations, but broken out into token sequences. For example, the following program: +(length("some string"), 1) would be predicted as the sequence: +(<NT>, 1) length(<NT>) " some string " Seq2Tree++. This is a re-implementation of Krishnamurthy et al. (2017) with some differences: (i) our implementation’s entity linking embeddings are computed over spans, including type information (as in the original paper) and a span embedding computed based on the LSTM hidden state at the start and end of each entity span, (ii) copies are treated as entities by proposing all spans up to length 15, and (iii) we use the linear dialogue history encoder described in §3.1. G Ablations The “value dependence” and the “no span copy” ablations are perhaps the most important in our experiments, and so we provide some more details about them in the following paragraphs. Value Dependence. The goal of this ablation is to quantify the impact of the dependency structure we propose in Equation 3. To this end, we first convert all functions to a curried form, where each argument is provided as part of a separate function invocation. For example, the following invocation: [0] event(subject = s0, start = t0, end = t1) is transformed to the following program fragment: [0] value(s0) [1] event_0(subject = [0]) [2] value(t0) [3] event_1(curried = [1], start = [2]) [4] value(t1) [5] event_2(curried = [3], end = [4]) When choosing a function, our decoder does not condition on the argument values of the previous invocations. In order to enable such conditioning without modifying the model implementation, we also transform the value function invocations whose underlying values are not copies, such that there exists a unique function for each unique value. This results in the following program: [0] value_s0(s0) [1] event_0(subject = [0]) [2] value_t0(t0) [3] event_1(curried = [1], start = [2]) [4] value_t1(t1) [5] event_2(curried = [3], end = [4]) Note that we keep the value s0, value t0, and value t1 function arguments because they allow the model to marginalize over multiple possible value sources (§2.5). The reason we do not transform the value functions that correspond to copies is because we attempted doing that on top of the span copy ablation, but it performed poorly and we decided that it may be a misrepresentation. Overall, this ablation offers us a way to obtain a bottom-up parser that maintains most properties of the proposed model, except for its dependency structure. No Span Copy. In order to ablate the proposed span copy mechanism we implemented a data transformation that replaces all copied values with references to the result of a copy function (for spans of length 1) or the result of a concatenate function called on the results of 2 or more calls to copy. For example, the function invocation: [0] event(subject = "water the plant") is converted to: [0] copy("water") [1] copy("the") [2] concatenate([0], [1]) [3] copy("plant") [4] concatenate([2], [3]) [5] event(subject = [4]) When applied on its own and not combined with other ablation, the single token copies are further inlined to produce the following program: [0] concatenate("water", "the") [1] concatenate([0], "plant") [2] event(subject = [1]) H Computational Efficiency For comparing model performance we computed the average utterance processing time across all of the SMCALFLOW validation set, using a single Nvidia V100 GPU. The fastest baseline required about 80ms per utterance, while our model only required about 8ms per utterance. This can be attributed to multiple reasons, such as the facts that: (i) our independence assumptions allow us to predict the argument value distributions in parallel, (ii) we avoid enumerating all possible utterance spans when computing the normalizing constant for the argument values, and (iii) we use ragged tensors to avoid unnecessary padding and computation.
|
2021
|
284
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3682–3692 August 1–6, 2021. ©2021 Association for Computational Linguistics 3682 MPC-BERT: A Pre-Trained Language Model for Multi-Party Conversation Understanding Jia-Chen Gu1∗, Chongyang Tao2, Zhen-Hua Ling1, Can Xu2, Xiubo Geng2, Daxin Jiang2† 1National Engineering Laboratory for Speech and Language Information Processing, University of Science and Technology of China, Hefei, China 2Microsoft, Beijing, China [email protected], [email protected], {chotao,caxu,xigeng,djiang}@microsoft.com Abstract Recently, various neural models for multiparty conversation (MPC) have achieved impressive improvements on a variety of tasks such as addressee recognition, speaker identification and response prediction. However, these existing methods on MPC usually represent interlocutors and utterances individually and ignore the inherent complicated structure in MPC which may provide crucial interlocutor and utterance semantics and would enhance the conversation understanding process. To this end, we present MPC-BERT, a pre-trained model for MPC understanding that considers learning who says what to whom in a unified model with several elaborated self-supervised tasks. Particularly, these tasks can be generally categorized into (1) interlocutor structure modeling including reply-to utterance recognition, identical speaker searching and pointer consistency distinction, and (2) utterance semantics modeling including masked shared utterance restoration and shared node detection. We evaluate MPCBERT on three downstream tasks including addressee recognition, speaker identification and response selection. Experimental results show that MPC-BERT outperforms previous methods by large margins and achieves new state-of-the-art performance on all three downstream tasks at two benchmarks. 1 Introduction Building a conversational agent with intelligence has drawn significant attention from both academia and industry. Most of existing methods have studied understanding conversations between two participants, aiming to return an appropriate response either in a generation-based (Shang et al., ∗Work done during the internship at Microsoft. †Corresponding author. Speaker Utterance Addressee I.1 How can I setup if I want add new server at xchat? I.2 From places, network servers, work I.1 group, his computer, and then I clicked on the shared folder. I.3 It did not allow you to see the files? I.2 I.2 It prompts for authentication and I I.3 don’t know what to put. I tried guest with no password. I.4 Put proper authentication in, then? I.2 I.3 I think you had kde on suse? I.2 Table 1: An MPC example in Ubuntu IRC channel. Here, “I.” is the abbreviation of “interlocutor”. 2015; Serban et al., 2016, 2017; Zhang et al., 2018b, 2020) or retrieval-based manner (Lowe et al., 2015; Wu et al., 2017; Zhou et al., 2018; Tao et al., 2019a,b; Gu et al., 2019a,b, 2020). Recently, researchers have paid more attention to a more practical and challenging scenario involving more than two participants, which is well known as multiparty conversation (MPC) (Ouchi and Tsuboi, 2016; Zhang et al., 2018a; Le et al., 2019; Hu et al., 2019). Table 1 shows an MPC example in the Ubuntu Internet Relay Chat (IRC) channel, which is composed of a sequence of (speaker, utterance, addressee) triples. In addition to returning an appropriate response, predicting who will be the next speaker (Meng et al., 2018) and who is the addressee of an utterance (Ouchi and Tsuboi, 2016; Zhang et al., 2018a; Le et al., 2019) are unique and important issues in MPC. An instance of MPC always contains complicated interactions between interlocutors, between utterances and between an interlocutor and an utterance. Therefore, it is challenging to model the conversation flow and fully understand the dialogue content. Existing studies on MPC learn the representations of interlocutors and utterances with neural networks, and their representation 3683 spaces are either separate (Ouchi and Tsuboi, 2016) or interactive (Zhang et al., 2018a). However, the semantics contained in the interlocutor and utterance representations may not be effectively captured as they are from two different representation spaces. Recently, to take advantage of the breakthrough in pre-training language models (PLMs) for natural language understanding, some studies proposed to integrate the speaker (Gu et al., 2020) or topic (Wang et al., 2020) information into PLMs. Despite of the performance improvement on response selection, these models still overlook the inherent relationships between utterances and interlocutors, such as “address-to”. Furthermore, most existing studies design models for each individual task in MPC (e.g., addressee recognition, speaker identification and response prediction) separately. Intuitively, these tasks are complementary among each other. Making use of these tasks simultaneously may produce better contextualized representations of interlocutors and utterances, and would enhance the conversation understanding, but is neglected in previous studies. On account of above issues, we propose MPCBERT which jointly learns who says what to whom in MPC by designing self-supervised tasks for PLMs, so as to improve the ability of PLMs on MPC understanding. Specifically, the five designed tasks includes reply-to utterance recognition, identical speaker searching, pointer consistency distinction, masked shared utterance restoration and shared node detection. The first three tasks are designed to model the interlocutor structure in MPC in a semantics-to-structure manner. In the output of MPC-BERT, an interlocutor is described through the encoded representations of the utterances it says. Thus, the representations of utterance semantics are utilized to construct the conversation structure in these three tasks. On the other hand, the last two tasks are designed to model the utterance semantics in a structure-to-semantics manner. Intuitively, the conversation structure influences the information flow in MPC. Thus, the structure information can also be used to strengthen the representations of utterance semantics in return. In general, these five self-supervised tasks are employed to jointly train the MPC-BERT in a multi-task learning framework, which helps the model to learn the complementary information among interlocutors and utterances, and that between structure and semantics. By this means, MPC-BERT can produce better interlocutor and utterance representations which can be effectively generalized to multiple downstream tasks of MPC. To measure the effectiveness of these selfsupervised tasks and to test the generalization ability of MPC-BERT, we evaluate it on three downstream tasks including addressee recognition, speaker identification and response selection, which are three core research issues of MPC. Two benchmarks based on Ubuntu IRC channel are employed for evaluation. One was released by Hu et al. (2019). The other was released by Ouchi and Tsuboi (2016) and has three experimental settings according to session lengths. Experimental results show that MPC-BERT outperforms the current state-of-the-art models by margins of 3.51%, 2.86%, 3.28% and 5.36% on the test sets of these two benchmarks respectively in terms of the session accuracy of addressee recognition, by margins of 7.66%, 2.60%, 3.38% and 4.24% respectively in terms of the utterance precision of speaker identification, and by margins of 3.82%, 2.71%, 2.55% and 3.22% respectively in terms of the response recall of response selection. In summary, our contributions in this paper are three-fold: (1) MPC-BERT, a PLM for MPC understanding, is proposed by designing five selfsupervised tasks based on the interactions among utterances and interlocutors. (2) Three downstream tasks are employed to comprehensively evaluate the effectiveness of our designed self-supervised tasks and the generalization ability of MPC-BERT. (3) Our proposed MPC-BERT achieves new state-ofthe-art performance on all three downstream tasks at two benchmarks. 2 Related Work Existing methods on building dialogue systems can be generally categorized into studying twoparty conversations and multi-party conversations (MPC). In this paper, we study MPC. In addition to predicting utterances, identifying the speaker and recognizing the addressee of an utterance are also important tasks for MPC. Ouchi and Tsuboi (2016) first proposed the task of addressee and response selection and created an MPC corpus for studying this task. Zhang et al. (2018a) proposed SI-RNN, which updated speaker embeddings role-sensitively for addressee and response selection. Meng et al. (2018) proposed a task of speaker classification as a surrogate task for speaker modeling. Le et al. 3684 (2019) proposed a who-to-whom (W2W) model to recognize the addressees of all utterances. Hu et al. (2019) proposed a graph-structured network (GSN) to model the graphical information flow for response generation. Wang et al. (2020) proposed to track the dynamic topic for response selection. Generally speaking, previous studies on MPC cannot unify the representations of interlocutors and utterances effectively. Also, they are limited to each individual task, ignoring the complementary information among different tasks. To the best of our knowledge, this paper makes the first attempt to design various self-supervised tasks for building PLMs aiming at MPC understanding, and to evaluate the performance of PLMs on three downstream tasks as comprehensively as possible. 3 MPC-BERT and Self-Supervised Tasks An MPC instance is composed of a sequence of (speaker, utterance, addressee) triples, denoted as {(sn, un, an)}N n=1, where N is the number of turns in the conversation. Our goal is to build a pre-trained language model for universal MPC understanding. Given a conversation, this model is expected to produce embedding vectors for all utterances which contain not only the semantic information of each utterance, but also the speaker and addressee structure of the whole conversation. Thus, it can be effectively adapted to various downstream tasks by fine-tuning model parameters. 3.1 Model Overview In this paper, BERT (Devlin et al., 2019) is chosen as the backbone of our PLM for MPC. Thus, we name it MPC-BERT. It is worth noting that our proposed self-supervised tasks for training MPCBERT can also be applied to other types of PLMs. We first give an overview of the input representations and the overall architectures of MPC-BERT. When constructing the input representations, in order to consider the speaker information of each utterance, speaker embeddings (Gu et al., 2020) are introduced as shown in Figure 1. Considering that the set of interlocutors are inconsistent in different conversations, a position-based interlocutor embedding table is initialized randomly at first and updated during pre-training, which means each interlocutor in a conversation is assigned with an embedding vector according to the order it appears in the conversation. Then, the speaker embeddings for each utterance can be derived by looking up this embedding table. The speaker embeddings are combined with standard token, position and segmentation embeddings and are then encoded by BERT. The output embeddings of BERT corresponding to different input tokens are utilized by different self-supervised tasks for further calculation. 3.2 Tasks of Interlocutor Structure Modeling The first three tasks follow the semantics-tostructure manner. In MPC-BERT, each interlocutor is described through the encoded representations of the utterances it says. Thus, the representations of utterance semantics are utilized to construct the conversation structure. Figure 1 shows the input representations and the model architectures of these three tasks. A [CLS] token is inserted at the start of each utterance, denoting its utterancelevel representation. Then, all utterances in a conversation are concatenated and a [SEP] token is inserted at the end of the whole sequence. It is notable that these three tasks share the same form of input data. Thus, the input only needs to be encoded once by BERT while the output can be fed into three tasks, which is computation-efficient. As shown in Figure 1, a task-dependent non-linear transformation layer is placed on top of BERT in order to adapt the output of BERT to different tasks. We will describe the details of these tasks as follows. 3.2.1 Reply-to Utterance Recognition To enable the model to recognize the addressee of each utterance, a self-supervised task named replyto utterance recognition (RUR) is proposed to learn which preceding utterance the current utterance replies to. After encoded by BERT, we extract the contextualized representations for each [CLS] token representing individual utterances. Next, a non-linear transformation followed by a layer normalization are performed to derive the utterance representations for this specific task {urur i }N i=1, where urur i ∈Rd and d = 768. Then, for a specific utterance Ui, its matching scores with all its preceding utterances are calculated as mij = softmax(urur⊤ i · Arur · urur j ), (1) where Arur ∈Rd×d is a linear transformation, mij denotes the matching degree of Uj being the replyto utterance of Ui, and 1 ≤j < i. We construct a set S by sampling a certain number of utterances 3685 Ui’ Ui UN [SEP] Input Token Embeddings Segment Embeddings Position Embeddings Speaker Embeddings ... ... ... ... ... ... ... ... Pre-trained Language Model (BERT) E[CLS] EU_i’ E[CLS] EU_i E[CLS] EU_N E[SEP] Output ... ... [CLS] [CLS] [CLS] (a) Reply-to Utterance Recognition Non-linear Transformation + Layer Normalization ui'rur uirur uNrur ... ... mij ... ... ... ... ... ... Uj’ Uj ... [CLS] [CLS] ... ... ... ... ... ... ... E[CLS] EU_j’ E[CLS] EU_j ... ... uj'rur ujrur ... ... (b) Identical Speaker Searching Non-linear Transformation + Layer Normalization ui'iss uiiss uNiss ... ... ... uj'iss ujiss ... ... (c) Pointer Consistency Distinction Non-linear Transformation + Layer Normalization ui'pcd uipcd uNpcd ... ... ... uj'pcd ujpcd ... ... Pointer Pointer Similarity Classifier ... Figure 1: Input representations and model architectures of the three self-supervised tasks for interlocutor structure modeling, including (a) reply-to utterance recognition, (b) identical speaker searching and (c) pointer consistency distinction. in a conversation and this recognition operation is performed for each utterance in S. Meanwhile, a dynamic sampling strategy is adopted so that models can see more samples. Finally, the pretraining objective of this self-supervised task is to minimize the cross-entropy loss as Lrur = − X i∈S i−1 X j=1 yij log(mij), (2) where yij = 1 if Uj is the reply-to utterance of Ui and yij = 0 otherwise. 3.2.2 Identical Speaker Searching Having knowledge of who is the speaker of an utterance is also important for MPC. The task of identical speaker searching (ISS) is designed by masking the speaker embedding of a specific utterance in the input representation, and aims to predict its speaker given the conversation. Since the set of interlocutors vary across conversations, the task of predicting the speaker of an utterance is reformulated as searching for the utterances sharing the identical speaker. First, for a specific utterance, its speaker embedding is masked with a special [Mask] interlocutor embedding to avoid information leakage. Given the utterance representations for this specific task {uiss i }N i=1 where uiss i ∈Rd, the matching scores of Ui with all its preceding utterances are calculated similarly with Eq. (1). Here, mij denotes the matching degree of Uj sharing the same speaker with Ui. For each instance in the dynamic sampling set S, there must be an utterance in previous turns sharing the same speaker. Otherwise, it is removed out of the set. Finally, the pre-training objective of this task is to minimize the cross-entropy loss similarly with Eq. (2). Here, yij = 1 if Uj shares the same speaker with Ui and yij = 0 otherwise. 3.2.3 Pointer Consistency Distinction We design a task named pointer consistency distinction (PCD) to jointly model speakers and addressees in MPC. In this task, a pair of utterances representing the “reply-to” relationship is defined as a speaker-to-addressee pointer. Here, we assume that the representations of two pointers directing from the same speaker to the same addressee should be consistent. As illustrated in Figure 2 (a), speaker Sm speaks Ui and Uj which reply to Ui′ and Uj′ from speaker Sn respectively. Thus, the utterance tuples (Ui, Ui′) and (Uj, Uj′) both represent the pointer of Sm-to-Sn and their pointer representations should be consistent.. Given the utterance representations for this specific task {upcd i }N i=1 where upcd i ∈Rd, we first capture the pointer information contained in each utterance tuple. The element-wise difference and multiplication between an utterance tuple (Ui, Ui′) are computed and are concatenated as pii′ = [upcd i −upcd i′ ; upcd i ⊙upcd i′ ], (3) 3686 Ui Ui ... Uj Uj Sn Sm ... ... : Speaker : Utterance : Utterance-to-utterance : Speaker-to-utterance (a) Pointer consistency distinction U1 U2 U3 U5 U8 U4 U6 U7 U9 (b) Shared node detection Figure 2: Illustrations of the self-supervised tasks of (a) pointer consistency distinction and (b) shared node detection. Rectangles denote utterances, circles denote interlocutors, a solid line denotes an utterance replying to an utterance, and a dashed line denotes an utterance from an interlocutor. where pii′ ∈R2d. Then, we compress pii′ and obtain the pointer representation ¯pii′ as ¯pii′ = ReLU(pii′ · Wpcd + bpcd), (4) where Wpcd ∈R2d×d and bpcd ∈Rd are parameters. Identically, a consistent pointer representations ¯pjj′ and an inconsistent one ¯pkk′ sampled from this conversation are obtained. The similarities between every two pointers are calculated as mij = sigmoid(¯p⊤ ii′ · Apcd · ¯pjj′), (5) where mij denotes the matching degree of pointer ¯pii′ being consistent with pointer ¯pjj′. mik can be derived accordingly. Finally, the pre-training objective of this task is to minimize the hinge loss which enforces mij to be larger than mik by at least a margin ∆as Lpcd = max{0, ∆−mij + mik}. (6) 3.3 Tasks of Utterance Semantics Modeling Intuitively, the conversation structure might influence the information flow, so that it can be used to strengthen the representations of utterance semantics. Thus, two self-supervised tasks following the structure-to-semantics manner are designed. 3.3.1 Masked Shared Utterance Restoration There are usually several utterances replying-to a shared utterance in MPC. Intuitively, a shared utterance is semantically relevant to more utterances in the context than non-shared ones. Based on this characteristic, we design a task named masked shared utterance restoration (MSUR). We first randomly sample an utterance from all shared utterances in a conversation and all tokens in this sampled utterance are masked with a [MASK] token. Then the model is enforced to restore the masked utterance given the rest conversation. Formally, assuming Ui as the masked shared utterance and li as the number of tokens in Ui. Given the token representations for this task {umsur i,t }li t=1 where umsur i,t ∈Rd, the probability distribution of each masked token can be calculated as pui,t = softmax(umsur i,t · Wmsur + bmsur), (7) where Wmsur ∈Rd×V is the token embedding table, V denotes the vocabulary size, and bmsur ∈ RV is a bias vector. Finally, the pre-training objective of this self-supervised task is to minimize the negative log-likelihood loss as Lmsur = −1 li li X t=1 log pui,t, (8) where pui,t is the element in pui,t corresponding to the original token. 3.3.2 Shared Node Detection A full MPC instance can be divided into several sub-conversations and we assume that the representations of sub-conversations under the same parent node tend to be similar. As illustrated in Figure 2 (b), two sub-conversations {U3, U5, U7, U8} and {U4, U6, U9} share the same parent node U2. Thus, they should be semantically relevant. Under this assumption, we design a self-supervised task named shared node detection (SND), which utilizes the conversation structure to strengthen the capability of models on measuring the semantic relevance of two sub-conversations. We first construct the pre-training samples for this task. Empirically, only the sub-conversations under the top shared node in a conversation are collected in order to filter out the sub-conversations with few utterances. Given a full MPC, the two sub-conversations with the most utterances form a positive pair. For each positive pair, we replace one of its elements with another sub-conversation randomly sampled from the training corpus to form a negative pair. Formally, given two sub-conversations ci and cj, utterances in each sub-conversation are first concatenated respectively to form two segments. Then, the two segments are concatenated with a [SEP] token and a [CLS] token is inserted at the beginning of the whole sequence. This sequence are encoded by BERT to derive the contextualized 3687 representation for the [CLS] token. A non-linear transformation with sigmoid activation is further applied to this representation for calculating the matching score mij, i.e., the probability of ci and cj sharing the same parent node. Finally, the pretraining objective of this task is to minimize the cross-entropy loss as Lsnd = −[yijlog(mij) + (1 −yij)log(1 −mij)], (9) where yij = 1 if ci and cj share the same parent node and yij = 0 otherwise. 3.4 Multi-task Learning In addition, we also adopt the tasks of masked language model (MLM) and next sentence prediction (NSP) in original BERT pre-training (Devlin et al., 2019), which have been proven effective for incorporating domain knowledge (Gu et al., 2020; Gururangan et al., 2020). Finally, MPCBERT is trained by performing multi-task learning that minimizes the sum of all loss functions as L = Lrur + Liss + Lpcd + Lmsur + Lsnd + Lmlm + Lnsp. (10) 4 Downstream Tasks 4.1 Addressee Recognition Given a multi-party conversation where part of the addressees are unknown, Ouchi and Tsuboi (2016) and Zhang et al. (2018a) recognized an addressee of the last utterance. Le et al. (2019) recognized addressees of all utterances in a conversation. In this paper, we follow the more challenging setting in Le et al. (2019). Formally, models are asked to predict {ˆan}N n=1 given {(sn, un, an)}N n=1\{an}N n=1, where ˆan is selected from the interlocutor set in this conversation and \ denotes exclusion. When applying MPC-BERT, this task is reformulated as finding a preceding utterance from the same addressee. Its RUR matching scores with all preceding utterances are calculated following Eq. (1). Then, the utterance with the highest score is selected and the speaker of the selected utterance is considered as the recognized addressee. Finally, the fine-tuning objective of this task is to minimize the crossentropy loss as Lar = − N X i=2 i−1 X j=1 yij log(mij), (11) where mij is defined in Eq. (1), yij = 1 if the speaker of Uj is the addressee of Ui and yij = 0 otherwise. 4.2 Speaker Identification This task aims to identify the speaker of the last utterance in a conversation. Formally, models are asked to predict ˆsN given {(sn, un, an)}N n=1\sN, where ˆsN is selected from the interlocutor set in this conversation. When applying MPC-BERT, this task is reformulated as identifying the utterances sharing the same speaker. For the last utterance UN, its speaker embedding is masked and its ISS matching scores mNj with all preceding utterances are calculated following Section 3.2.2. The finetuning objective of this task is to minimize the cross-entropy loss as Lsi = − N−1 X j=1 yNj log(mNj), (12) where yNj = 1 if Uj shares the same speaker with UN and yNj = 0 otherwise. 4.3 Response Selection This task asks models to select ˆuN from a set of response candidates given the conversation context {(sn, un, an)}N n=1\uN. The key is to measure the similarity between two segments of context and response. We concatenate each response candidate with the context and extract the contextualized representation e[CLS] for the first [CLS] token using MPC-BERT. Then, e[CLS] is fed into a nonlinear transformation with sigmoid activation to obtain the matching score between the context and the response. Finally, the fine-tuning objective of this task is to minimize the cross-entropy loss according to the true/false labels of responses in the training set as Lrs = −[ylog(mcr)+(1−y)log(1−mcr)], (13) where y = 1 if the response r is a proper one for the context c; otherwise y = 0. 5 Experiments 5.1 Datasets We evaluated our proposed methods on two Ubuntu IRC benchmarks. One was released by Hu et al. (2019), in which both speaker and addressee labels was provided for each utterance. The other benchmark was released by Ouchi and Tsuboi 3688 Datasets Train Valid Test Hu et al. (2019) 311,725 5,000 5,000 Ouchi and Tsuboi (2016) Len-5 461,120 28,570 32,668 Len-10 495,226 30,974 35,638 Len-15 489,812 30,815 35,385 Table 2: Statistics of the two benchmarks evaluated in this paper. (2016). Here, we adopted the version shared in Le et al. (2019) for fair comparison. The conversation sessions were separated into three categories according to the session length (Len5, Len-10 and Len-15) following the splitting strategy of previous studies (Ouchi and Tsuboi, 2016; Zhang et al., 2018a; Le et al., 2019). Table 2 presents the statistics of the two benchmarks evaluated in our experiments. 5.2 Baseline Models Non-pre-training-based models Ouchi and Tsuboi (2016) proposed a dynamic model DRNN which updated speaker embeddings with the conversation flow. Zhang et al. (2018a) improved DRNN to SI-RNN which updated speaker embeddings role-sensitively. Le et al. (2019) proposed W2W which jointly modeled interlocutors and utterances in a uniform framework, and predicted all addressees. Pre-training-based models BERT (Devlin et al., 2019) was pre-trained to learn general language representations with MLM and NSP tasks. SABERT (Gu et al., 2020) added speaker embeddings and further pre-trained BERT on a domain-specific corpus to incorporate domain knowledge. We re-implemented SA-BERT with the pre-training corpus used in this paper to ensure fair comparison. 5.3 Implementation Details The version of BERT-base-uncased was adopted for all our experiments. For pre-training, GELU (Hendrycks and Gimpel, 2016) was employed as the activation for all non-linear transformations. The Adam method (Kingma and Ba, 2015) was employed for optimization. The learning rate was initialized as 0.00005 and the warmup proportion was set to 0.1. We pre-trained BERT for 10 epochs. The training set of the dateset used in Hu et al. (2019) was employed for pre-training. The maximum utterance number was set to 7. The maximum sequence length was set to 230. The maximum sampling numbers for each example were set to 4 for RUR, 2 for ISS and 2 for PCD. ∆in Eq. (6) was set to 0.4, achieving the best performance out of {0.2, 0.4, 0.6, 0.8} on the validation set. The pre-training was performed using a GeForce RTX 2080 Ti GPU and the batch size was set to 4. For fine-tuning, some configurations were different according to the characteristics of these datasets. For Hu et al. (2019), the maximum utterance number was set to 7 and the maximum sequence length was set to 230. For the three experimental settings in Ouchi and Tsuboi (2016), the maximum utterance numbers were set to 5, 10 and 15, and the maximum sequence lengths were set to 120, 220 and 320. All parameters in PLMs were updated. The learning rate was initialized as 0.00002 and the warmup proportion was set to 0.1. For Hu et al. (2019), the fine-tuning process was performed for 10 epochs for addressee recognition, 10 epochs for speaker identification, and 5 epochs for response selection. For Ouchi and Tsuboi (2016), the fine-tuning epochs were set to 5, 5 and 3 respectively. The fine-tuning was also performed using a GeForce RTX 2080 Ti GPU. The batch sizes were set to 16 for Hu et al. (2019), and 40, 20, and 12 for the three experimental settings in Ouchi and Tsuboi (2016) respectively. The validation set was used to select the best model for testing. All codes were implemented in the TensorFlow framework (Abadi et al., 2016) and are published to help replicate our results. 1 5.4 Metrics and Results Addressee recognition We followed the metrics of previous work (Le et al., 2019) by employing precision@1 (P@1) to evaluate each utterance with ground truth. Also, a session is marked as positive if the addressees of all its utterances are correctly recognized, which is calculated as accuracy (Acc.). Table 3 presents the results of addressee recognition. It shows that MPC-BERT outperforms the best performing model, i.e., SA-BERT, by margins of 3.51%, 2.86%, 3.28% and 5.36% on these test sets respectively in terms of Acc., verifying the effectiveness of the proposed five selfsupervised tasks as a whole. To further illustrate the effectiveness of each task, ablation tests were performed as shown in the last five rows of Table 3. We can observe that all self-supervised tasks are useful as removing any of them causes performance 1https://github.com/JasonForJoy/MPC-BERT 3689 Hu et al. (2019) Ouchi and Tsuboi (2016) Len-5 Len-10 Len-15 P@1 Acc. P@1 Acc. P@1 Acc. P@1 Acc. Preceding (Le et al., 2019) 63.50 40.46 56.84 21.06 54.97 13.08 Subsequent (Le et al., 2019) 61.03 40.25 54.57 20.26 53.07 12.79 DRNN (Ouchi and Tsuboi, 2016) 72.75 58.18 65.58 34.47 62.60 22.58 SIRNN (Zhang et al., 2018a) 75.98 62.06 70.88 40.66 68.13 28.05 W2W (Le et al., 2019) 77.55 63.81 73.52 44.14 73.42 34.23 BERT (Devlin et al., 2019) 96.16 83.50 85.95 75.99 83.41 58.22 81.09 44.94 SA-BERT (Gu et al., 2020) 97.12 88.91 86.81 77.45 84.46 60.30 82.84 47.23 MPC-BERT 98.31 92.42 88.73 80.31 86.23 63.58 85.55 52.59 MPC-BERT w/o. RUR 97.75 89.98 87.51 78.42 85.63 62.26 84.78 50.83 MPC-BERT w/o. ISS 98.20 91.96 88.67 80.25 86.14 63.40 85.02 51.12 MPC-BERT w/o. PCD 98.20 91.90 88.51 80.06 85.92 62.84 85.21 51.17 MPC-BERT w/o. MSUR 98.08 91.32 88.70 80.26 86.21 63.46 85.28 51.23 MPC-BERT w/o. SND 98.25 92.18 88.68 80.25 86.14 63.41 85.29 51.39 Table 3: Evaluation results of addressee recognition on the test sets. Results except ours are cited from Le et al. (2019). Numbers in bold denote that the improvement over the best performing baseline is statistically significant (t-test with p-value < 0.05). Hu et al. (2019) Ouchi and Tsuboi (2016) Len-5 Len-10 Len-15 BERT (Devlin et al., 2019) 71.81 62.24 53.17 51.58 SA-BERT (Gu et al., 2020) 75.88 64.96 57.62 54.28 MPC-BERT 83.54 67.56 61.00 58.52 MPC-BERT w/o. RUR 82.48 66.88 60.12 57.33 MPC-BERT w/o. ISS 77.95 66.77 60.03 56.73 MPC-BERT w/o. PCD 83.39 67.12 60.62 58.00 MPC-BERT w/o. MSUR 83.51 67.21 60.76 58.03 MPC-BERT w/o. SND 83.47 67.04 60.44 58.12 Table 4: Evaluation results of speaker identification on the test sets in terms of P@1. Numbers in bold denote that the improvement over the best performing baseline is statistically significant (t-test with p-value < 0.05). drop. Among the five tasks, RUR plays the most important role, and the tasks focusing on modeling interlocutor structure contribute more than those for utterance semantics. Speaker identification Similarly, P@1 was employed as the evaluation metric of speaker identification for the last utterance of a conversation and the results are shown in Table 4. It shows that MPC-BERT outperforms SA-BERT by margins of 7.66%, 2.60%, 3.38% and 4.24% respectively in terms of P@1. Besides, from the ablation results we find that all tasks are useful for improving the performance of speaker identification and ISS and RUR contribute the most. In particular, removing PCD, MSUR and SND only leads to slight performance drop. The reason might be that the information conveyed by these tasks is redundant. Response selection The Rn@k metrics adopted by previous studies (Ouchi and Tsuboi, 2016; Zhang et al., 2018a) were used here. Each model was tasked with selecting k best-matched responses from n available candidates, and we calculated the recall as Rn@k. Two settings were followed in which k was set to 1 and n was set to 2 or 10. Table 5 presents the results of response selection. It shows that MPC-BERT outperforms SABERT by margins of 3.82%, 2.71%, 2.55% and 3.22% respectively in terms of R10@1. Ablation tests show that SND is the most useful task for response selection and the two tasks focusing on the utterance semantics contribute more than those 3690 Hu et al. (2019) Ouchi and Tsuboi (2016) Len-5 Len-10 Len-15 R2@1 R10@1 R2@1 R10@1 R2@1 R10@1 R2@1 R10@1 DRNN (Ouchi and Tsuboi, 2016) 76.07 33.62 78.16 36.14 78.64 36.93 SIRNN (Zhang et al., 2018a) 78.14 36.45 80.34 39.20 80.91 40.83 BERT (Devlin et al., 2019) 92.48 73.42 85.52 53.95 86.93 57.41 87.19 58.92 SA-BERT (Gu et al., 2020) 92.98 75.16 86.53 55.24 87.98 59.27 88.34 60.42 MPC-BERT 94.90 78.98 87.63 57.95 89.14 61.82 89.70 63.64 MPC-BERT w/o. RUR 94.48 78.16 87.20 57.56 88.96 61.47 89.07 63.24 MPC-BERT w/o. ISS 94.58 78.82 87.54 57.77 88.98 61.76 89.58 63.51 MPC-BERT w/o. PCD 94.66 78.70 87.50 57.51 88.75 61.62 89.45 63.46 MPC-BERT w/o. MSUR 94.36 78.22 87.11 57.58 88.59 61.05 89.25 63.20 MPC-BERT w/o. SND 93.92 76.96 87.30 57.54 88.77 61.54 89.27 63.34 Table 5: Evaluation results of response selection on the test sets. Results except ours are cited from Ouchi and Tsuboi (2016) and Zhang et al. (2018a). Numbers in bold denote that the improvement over the best performing baseline is statistically significant (t-test with p-value < 0.05). 5 10 15 Length 50 60 70 80 Session Accuracy BERT SA-BERT MPC-BERT (a) Addressee recognition 5 10 15 Length 55 60 65 Utterance Preision BERT SA-BERT MPC-BERT (b) Speaker identification 5 10 15 Length 54 56 58 60 62 64 Response Recall BERT SA-BERT MPC-BERT (c) Response selection Figure 3: Performance of models under different session lengths on the test sets of Ouchi and Tsuboi (2016) on the tasks of (a) addressee recognition, (b) speaker identification and (c) response selection. focusing on the interlocutor structures. 5.5 Discussions Figure 3 illustrates how the performance of BERT, SA-BERT and MPC-BERT changed with respect to different session lengths on the test sets of Ouchi and Tsuboi (2016). It can be seen that the performance of addressee recognition and speaker identification dropped as the session length increased. The reason might be that longer sessions always contain more interlocutors which increase the difficulties of predicting interlocutors. Meanwhile, the performance of response selection was significantly improved as the session length increased. It can be attributed to that longer sessions enrich the representations of contexts with more details which benefit response selection. Furthermore, as the session length increased, the performance of MPC-BERT dropped more slightly than that of SA-BERT on addressee recognition and speaker identification, and the R10@1 gap between MPC-BERT and SA-BERT on response selection enlarged from 2.71% to 3.22%. These results imply the superiority of MPC-BERT over SA-BERT on modeling long MPCs with complicated structures. 6 Conclusion In this paper, we present MPC-BERT, a pre-trained language model with five self-supervised tasks for MPC understanding. These tasks jointly learn who says what to whom in MPCs. Experimental results on three downstream tasks show that MPC-BERT outperforms previous methods by large margins and achieves new state-of-the-art performance on two benchmarks. Acknowledgments We thank anonymous reviewers for their valuable comments. 3691 References Mart´ın Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek Gordon Murray, Benoit Steiner, Paul A. Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2016. Tensorflow: A system for large-scale machine learning. In 12th USENIX Symposium on Operating Systems Design and Implementation, OSDI 2016, Savannah, GA, USA, November 2-4, 2016., pages 265–283. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Jia-Chen Gu, Tianda Li, Quan Liu, Zhen-Hua Ling, Zhiming Su, Si Wei, and Xiaodan Zhu. 2020. Speaker-aware BERT for multi-turn response selection in retrieval-based chatbots. In CIKM ’20: The 29th ACM International Conference on Information and Knowledge Management, Virtual Event, Ireland, October 19-23, 2020, pages 2041– 2044. Jia-Chen Gu, Zhen-Hua Ling, and Quan Liu. 2019a. Interactive matching network for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019, pages 2321–2324. Jia-Chen Gu, Zhen-Hua Ling, Xiaodan Zhu, and Quan Liu. 2019b. Dually interactive matching network for personalized response selection in retrieval-based chatbots. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 1845–1854. Association for Computational Linguistics. Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8342–8360. Dan Hendrycks and Kevin Gimpel. 2016. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. CoRR, abs/1606.08415. Wenpeng Hu, Zhangming Chan, Bing Liu, Dongyan Zhao, Jinwen Ma, and Rui Yan. 2019. GSN: A graph-structured network for multi-party dialogues. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 5010–5016. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Ran Le, Wenpeng Hu, Mingyue Shang, Zhenjun You, Lidong Bing, Dongyan Zhao, and Rui Yan. 2019. Who is speaking to whom? learning to identify utterance addressee in multi-party conversations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 1909– 1919. Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle Pineau. 2015. The ubuntu dialogue corpus: A large dataset for research in unstructured multi-turn dialogue systems. In Proceedings of the SIGDIAL 2015 Conference, The 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue, 2-4 September 2015, Prague, Czech Republic, pages 285–294. Zhao Meng, Lili Mou, and Zhi Jin. 2018. Towards neural speaker modeling in multi-party conversation: The task, dataset, and models. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, LREC 2018, Miyazaki, Japan, May 7-12, 2018. European Language Resources Association (ELRA). Hiroki Ouchi and Yuta Tsuboi. 2016. Addressee and response selection for multi-party conversation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 2133–2143. Iulian Vlad Serban, Alessandro Sordoni, Yoshua Bengio, Aaron C. Courville, and Joelle Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, February 12-17, 2016, Phoenix, Arizona, USA, pages 3776–3784. Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C. Courville, and Yoshua Bengio. 2017. A hierarchical latent variable encoder-decoder model for generating dialogues. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 3295– 3301. AAAI Press. 3692 Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. Neural responding machine for short-text conversation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 2631, 2015, Beijing, China, Volume 1: Long Papers, pages 1577–1586. Chongyang Tao, Wei Wu, Can Xu, Wenpeng Hu, Dongyan Zhao, and Rui Yan. 2019a. Multirepresentation fusion network for multi-turn response selection in retrieval-based chatbots. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM 2019, Melbourne, VIC, Australia, February 11-15, 2019, pages 267–275. ACM. Chongyang Tao, Wei Wu, Can Xu, Wenpeng Hu, Dongyan Zhao, and Rui Yan. 2019b. One time of interaction may not be enough: Go deep with an interaction-over-interaction network for response selection in dialogues. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28August 2, 2019, Volume 1: Long Papers, pages 1– 11. Weishi Wang, Steven C. H. Hoi, and Shafiq R. Joty. 2020. Response selection for multi-party conversations with dynamic topic tracking. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 6581– 6591. Yu Wu, Wei Wu, Chen Xing, Ming Zhou, and Zhoujun Li. 2017. Sequential matching network: A new architecture for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 496–505. Rui Zhang, Honglak Lee, Lazaros Polymenakos, and Dragomir R. Radev. 2018a. Addressee and response selection in multi-party conversations with speaker interaction rnns. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5690–5697. Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, and Bill Dolan. 2018b. Generating informative and diverse conversational responses via adversarial information maximization. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr´eal, Canada, pages 1815–1825. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT : Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, ACL 2020, Online, July 5-10, 2020, pages 270–278. Association for Computational Linguistics. Xiangyang Zhou, Lu Li, Daxiang Dong, Yi Liu, Ying Chen, Wayne Xin Zhao, Dianhai Yu, and Hua Wu. 2018. Multi-turn response selection for chatbots with deep attention matching network. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 1118–1127.
|
2021
|
285
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3693–3703 August 1–6, 2021. ©2021 Association for Computational Linguistics 3693 Best of Both Worlds: Making High Accuracy Non-incremental Transformer-based Disfluency Detection Incremental Morteza Rohanian and Julian Hough Cognitive Science Group School of Electronic Engineering and Computer Science Queen Mary University of London {m.rohanian, j.hough} @qmul.ac.uk Abstract While Transformer-based text classifiers pretrained on large volumes of text have yielded significant improvements on a wide range of computational linguistics tasks, their implementations have been unsuitable for live incremental processing thus far, operating only on the level of complete sentence inputs. We address the challenge of introducing methods for word-by-word left-to-right incremental processing to Transformers such as BERT, models without an intrinsic sense of linear order. We modify the training method and live decoding of non-incremental models to detect speech disfluencies with minimum latency and without pre-segmentation of dialogue acts. We experiment with several decoding methods to predict the rightward context of the word currently being processed using a GPT-2 language model and apply a BERT-based disfluency detector to sequences, including predicted words. We show our method of incrementalising Transformers maintains most of their high non-incremental performance while operating strictly incrementally. We also evaluate our models’ incremental performance to establish the trade-off between incremental performance and final performance, using different prediction strategies. We apply our system to incremental speech recognition results as they arrive into a live system and achieve state-of-the-art results in this setting. 1 Introduction Conversational systems provide a significant addition to the present approaches in mental health care delivery. Interactions with these conversational agents have been shown to contain observable indicators of cognitive states, such as the rate of filled pauses and different temporal and turnrelated features (Gratch et al., 2014). Alzheimer’s Disease (AD) patients, for example, have trouble performing tasks that leverage semantic information; they have difficulties with verbal fluency and object recognition. AD patients speak more slowly with long pauses and spend extra time looking for the correct word, which leads to speech disfluency (L´opez-de Ipi˜na et al., 2013; Nasreen et al., 2021). Disfluency markers can be key features for identifying certain cognitive disorders for application in conversational agents (Rohanian et al., 2020). Such conversational systems are primarily used for content processing, which is then analyzed offline. There is much work on detecting disfluencies for offline analysis of transcripts. However, given that these disfluency detection models do not work for live systems and depend on rich transcription data, including pre-segmentation of dialogue acts, to facilitate more cost-effective analysis of other data, we need systems capable of performing directly and incrementally off the speech signal, or at least from the results of automatic speech recognition (ASR) as they arrive in the system. As it receives word-by-word data, an incremental model must operate with minimum latency and do so without changing its initial assumptions and delivering its best decisions as early as possible following the principles outlined in (Hough and Purver, 2014). Here we design and evaluating models that work with online, incremental speech recognition output to detect disfluencies with varying levels of granularity. The best neural language encoders currently used in computational linguistics consider word sequences as a whole, and their implementations have been unsuitable for live incremental processing. Transformers (Vaswani et al., 2017), for instance, operate on representations that do not naturally have an organizing principle of linear word order. We analyze how these models work under incremental frameworks, where it is essential to present partial output relying on partial input pro3694 vided up to a certain time step that may occur in interactive healthcare systems. We explore whether we can adjust such models to function incrementally and how useful they are in terms of overall accuracy and incremental metrics. To further enhance the models’ incremental performance, we use two general strategies to adjust the training regime and the real-time procedure: incremental training (‘chunk-based’ training and add-M training) and incremental decoding (constant latency and prophecies). We employ three prominent decoding methods to predict the rightward context of the word currently being processed: beam search, top-k sampling, and top-p sampling. We also measure our models’ incremental performance to set the trade-off between incremental performance and final performance. 2 Related Work Although considerable work has been done on detecting disfluencies, much of this work uses transcripts as texts rather than live speech inputs, with the goal of ‘cleaning’ the disfluent content for post-processing purposes. They are almost exclusively conducted on pre-segmented utterances of the Switchboard corpus of telephone conversations (Godfrey et al., 1992). Several disfluency detection efforts involve sentence-based parsing and language models (Johnson and Charniak, 2004; Zwarts et al., 2010). Sequence labeling models with start-inside-outside (BIO) style tags have been used in recent neural sequence approaches to disfluency detection based on bi-directional Long Short Term Memory (BiLSTM) networks and Transformers, in which the sequences are available in full (Zayats et al., 2016; Lou and Johnson, 2020; Wang et al., 2020). Such offline methods are insufficient if we intend to infer meaning from repairs and edit words for disfluency detection in real-time, which is beneficial in a healthcare domain dialogue system that seeks to get a consistent and clear understanding of user statements and the user’s cognitive state. Methods based on strictly incremental operation have been rare. Hough and Purver (2014) used a line of classifiers and language model features in a strong incremental operating system without looking ahead. Incremental dependency parsing combined with the removal of disfluency was also studied (Rasooli and Tetreault, 2015). Some studies have used recurrent neural networks for live disfluency identification. Using a basic Elman Recurrent Neural Network (RNN), Hough and Schlangen (2015) investigated incremental processing, with an objective coupling detection accuracy with low latency. Language models have been used as an additional task for the identification of disfluencies, relying on the intuition that disfluencies can be detected by divergences from clean language models, with Johnson and Charniak (2004)’s noisy channel model beginning this effort. Shalyminov et al. (2018) made language modelling an auxiliary task to disfluency detection in a deep multi-task learning (MTL) set-up, gaining accuracy over a vanilla RNN tagger. POS tags have also been used as an input for detecting disfluencies, showing slight increases in disfluency detection over using word values alone (Purver et al., 2018). While the work above operates only on transcripts pre-segmented into utterances, recent research has been performed on combining disfluency detection with utterance segmentation. This was done in a joint tagset of disfluency, and utterance segmentation tags by (Hough and Schlangen, 2017), showing an improvement over the performance of the individual tasks, and (Rohanian and Hough, 2020) show an improvement in both tasks when framed as a multi-task learning (MTL) set-up with a Long Short-term Memory network (LSTM), also simultaneously doing POS-tagging and language modelling. The recent live incremental systems fall short of the same accuracies achievable on pre-segmented transcripts, so there is a natural interest in using the best non-incremental sequence models and adapting them for incrementality. Madureira and Schlangen (2020) take up this effort in several other sequence tagging and classification tasks, showing how bidirectional encoders and Transformers can be modified to work incrementally. To reduce the impact of the partiality of the input, the models predict future content and wait for more rightward context. Dalvi et al. (2018) also use truncated inputs during the training phase of live machine translation to address the partial input sentence decoding problem Bidirectional encoders face. Here, we seek to add to this growing effort to investigate the trade-off of incremental performance against the final output quality of deep neural network-based language processing, applied to incremental disfluency detection. 3695 | A uh flight [ to Boston + { uh I mean } to Denver ] on Friday | Thank you | Disfluency f e f f f e e e rpS−5 rpnSub f f f f Utterance segmentation .w- -w- -w-w- -w-w-w-w-w-w-w- -w. .w- -w. POS tags DT UH NN IN NNP UH P RP V B IN NNP IN NNP V B P RP Figure 1: An utterance with the disfluency tags (repair structures and edit terms) and the utterance segmentation tags and POS tags used for preprocessing. 3 Disfluency Detection Disfluencies are generally assumed to have a reparandum-interregnum-repair structure in their fullest form as speech repairs (Shriberg, 1994; Meteer et al., 1995). A reparandum is a stretch of speech later corrected by the speaker; the corrected expression is a repair, the beginning of which is referred to as repair onset. An interregnum word is a filler or a reference expression between the repair and reparandum, usually an interruption and hesitation step when the speaker expresses a repair, giving the structure as in (1). John [ likes | {z } reparandum + { uh } | {z } interregnum loves ] | {z } repair Mary (1) In the absence of reparandum and repair, the disfluency is reduced to an isolated edit term. A marked, lexicalised edit term such as a filled pause (“uh” or “um”) or more phrasal terms such as “I mean” and “you know” may occur. The identification of these elements and their structure is then the task of disfluency detection. The task of detecting incremental disfluencies adds to the difficulty of doing this in real-time, word-by-word, from left to right. Disfluency recognition is then treated as the same problem that a human processor faces with a disfluent expression: only when an interregnum is detected, or maybe even when a repair is initiated, does it become clear that the earlier content is now to be regarded as ‘to be repaired,’ i.e., to be classified as a reparandum. Therefore, the task cannot be defined as a simple sequence labeling task in which the tags for the reparandum, interregnum, and repair phases are assigned left-to-right over words as seen in the above example; in this case, it will require the assumption that “likes” would be repaired, at a time when there is no data to make it available. We use a tag set that encodes the start of the reparandum only at a time when it can be inferred, primarily when the repair starts – the disfluency detection task is to tag words as in the top line of tags in Fig. 1 as either fluent (f) an edit term (e), a repair onset word (rpS−N for the reparandum starting N words back) and a repair end word of the type repeat (rpnRep), substitution (rpnSub) or delete (rpnDel). 4 Model To incrementalise a Transformer-based model for word-by-word disfluency detection, we devise a model built on top of a pre-trained BERT architecture (Devlin et al., 2019) with a Conditional Random Field (CRF) output architecture to tag sequences with tags such as those in the top line of Fig. 1. We use a BERT-based encoder and try different strategies to incrementalise the system’s operation and output, using language models to predict future word sequences as described in Section 5 while maintaining BERT’s non-incremental quality. Utterance segmentation Our models are designed to work not only with pre-segmented data but also on raw transcripts and ASR results, where utterance segmentation is required to leverage the use of sentence-based linguistic knowledge in BERT. Utterance segmentation has a clear interdependence with and influence on the detection of disfluency as disfluent restarts and repairs may be incorrectly predicted at fluent utterance boundaries without segmentation. In this paper, rather than performing utterance segmentation in tandem with disfluency detection, we perform it on words as they arrive in the system as a live segmentation task before sending the current prefix of the utterance to the disfluency detection system. We use the word-by-word segmentation system from (Rohanian and Hough, 2020) where four output tags define ranges of transcribed words or word hypotheses using a BIES tag scheme (Beginning, Inside, End, and Single) to allow for the prediction of an utterance ending. The tagset allows information to be captured from the context of the word to decide whether this word continues a current utterance (the - prefix) or starts anew (the . prefix), and also allows live prediction of whether the next word will continue the current utterance (the - suffix) or 3696 whether the current word finishes the utterance (the . suffix). An example of the scheme is shown in the second line of Fig. 1. CRF We use a CRF output architecture to predict a tag for every token. Although this model generates predictions for the whole sequence, the labels are outputted individually. There are important dependencies between adjacent labels in disfluency detection, and explicit modeling of these relationships can help. The addition of the CRF enables the model to test for the most optimal path across all available label sequences. 4.1 Input Features In addition to the word values, we also experiment with two other inputs: Part-of-speech tags POS tags may enhance the identification of disfluencies on various settings. POS tagging helps detect disfluency structure as the parallelism between the reparandum and repair in substitutions, as shown in the repeated IN NNP sequences in Fig. 1. Word timings We also experiment with the duration from the ending of the previous word to the ending of the current word as it enters the system, either from ground truth word transcriptions or from ASR results. 5 Strategies for Incrementalising BERT Here we describe the different strategies we used to modify the training and live decoding methods of non-incremental models to detect speech disfluencies word-by-word incrementally. The general principle is to leverage high accuracy full sequence classification using BERT but deploying it on sequences, including future predictions for words up to the hypothesised end of the current utterance. 5.1 Modifying the Training Procedure Training is performed on full sentences/utterances, but the decoder produces outputs based on partial input data at the test time. This disparity between training and decoding can potentially affect our models’ performance. Based on (Dalvi et al., 2018), we present two methods to address this issue: chunk-based training and add-M training. Chunk-based training In chunk-based training, we change the training scheme by removing the ends of each sentence in the training set and simply break each training sentence into chunks of N tokens. Here we use 2 and 3 for N. Add-M training We begin with the first N words in training sentences in add-M training. The next training instances are then generated by N +M, N +2M, N +3M... words before the end of the sentence is reached. In our experiments, we found setting N=1 and M=1 worked best. 5.2 Modifying the Decoding Procedure Constant latency The technique of constant latency requires allowing certain ‘future’ words to be seen before a label to previous words is given. It is a form of look-ahead based on Baumann et al. (2011), in which before making the first decision with respect to previous time steps, the processor is required to wait for some correct context. We explore the one- or two-word contexts of our input. This suggests that the model generates the first label for word t after the word t + 1 is seen or the model observes words t + 1 and t + 2 before tagging word t. This has an inherent limit on the latency achievable, and we use this as a baseline incremental decoding system. Prophecy-based decoding For our other decoding strategies, we use a ‘prophecy’-based approach to predicting future word sequences, following the task of open-ended language generation, which, given an input text passage as context, is to produce text that constitutes a cohesive continuation (Holtzman et al., 2019). Inspired by (Madureira and Schlangen, 2020), using the GPT-2 language model (Radford et al., 2019), we first give each word as a left context and create a continuation until the end of an utterance to create a hypothetical complete context that satisfies the requirements of the models’ non-incremental structure. Formally, with m tokens x1...xm as our context, the task is to create the next n continuation tokens to achieve the completed sequence x1...xm+n. It is assumed that the models compute P(x1:m+n) using a standard left-to-right decomposition of the text probability as in (2). This process is used to build the utterance continuation token-by-token using a specific decoding technique. P(x1:m+n) = m+n Y i=1 P(xi|x1...xi−1) (2) Three of the most common decoding methods are used in this paper: Beam search, Top-k sampling, and Top-p sampling. Example word sequence prophecies from these decoding methods 3697 (a) (b) (c) Figure 2: Using a ‘prophecy’-based approach to predict future word sequences, following the task of openended language generation with three different decoding methods. (a) Beam search. (b) Top-k sampling. (c) Top-p sampling. are shown in Fig. 2. The right-most block shows the prediction of the continuation of the word sequences as each new word in the sequence “John likes uh loves Mary” is fed into the language model. Beam search Assuming that the model gives a greater likelihood to better quality text, we are looking for a sequence with the highest probability. During the search, a group of stacks is used to hold hypotheses. Beam size N is used to manage the search space by expanding the top N hypotheses in the existing stack. We used beam size 10 for all the models. Top-k sampling We define sampling as randomly choosing the next word based on its conditional probability distribution as in (3). xi ∼P(x|x1:i−1) (3) In the Top-k sampling, the most probable next k words are extracted and the probability mass is redistributed between only the following k words (Fan et al., 2018). Given a distribution P(x|x1:i−1), we extract its top-k vocabulary V (k) ⊂V as the set of size k which maximizes P x∈V (k) P(x|x1:i−1). After an initial investigation, we set k to 50 in all experiments. Top-p sampling Rather than selecting only the most probable K words, in Top-p sampling, we select the smallest possible range of words with their total likelihood exceeds the probability p (Holtzman et al., 2019). The probability mass is then redistributed between this set of words. With this method, the size of the word set will dynamically adjust based on the probability distribution of the next word. With the distribution P(x|x1:i−1), we consider its top-p sequence, with vocabulary V (p) ⊂V as the smallest set with P(x|x1:i−1) ≥p. We set p = 0.95. 6 Experimental Set-up We train on transcripts and test on both transcripts and ASR hypotheses. All models in testing have strictly word-by-word left to right input. In addition to using the latest word hypothesis as input, we train and evaluate the presented models with two kinds of additional inputs: time elapsed from the end of the previous word (hypothesis) to the current one and the POS tag of the current word. Results on the development set were used to find the best model to be evaluated on the test set. We used the data from (Hough and Schlangen, 2017) for ASR hypotheses – this was generated by a free trial version of IBM’s Watson SpeechTo-Text service for incremental ASR. The service offers good quality ASR on noisy data-on our selected held-out data on Switchboard, and the average WER is 26.5%. The Watson service, crucially for our task, does not filter out hesitation markers or disfluencies (Baumann et al., 2017). The service delivers results incrementally, so silence-based endpointing is not used. It also outputs word timings, which are close enough to the source timings to use as features in the live version of our system. The word embedding for LSTM was initialised with 50-dimensional embedding trained on Google News (Mikolov et al., 2013). The model has been implemented using Tensorflow 2.1. We train all models for a maximum of 50 epochs; otherwise, stop training if there is no improvement on the best score on the validation set after 7 epochs. A large version of the pre-trained BERT is used with 340M parameters (24-layer blocks, 16 self3698 Input Model Pre-segmented transcripts (per word) Transcripts (per word) ASR (per 10 second window) Frm FrpS Fe Frm FrpS Fe Frm FrpS Fe Words STIR (HS’15/ PHH’18) 0.741 / 0.749 -/0.827 0.880/RNN (HS’15) 0.689 0.873 LSTM 0.686 0.771 0.928 0.59 0.678 0.904 0.548 0.726 LSTM-MTL (RH’20) 0.737 0.799 0.938 0.629 0.743 0.917 0.573 0.757 BERT 0.758 0.851 0.960 0.659 0.782 0.947 0.524 0.603 0.812 Word + Timings LSTM 0.681 0.777 0.921 0.623 0.718 0.908 0.555 0.721 LSTM-MTL (RH’20) 0.741 0.812 0.929 0.629 0.741 0.922 0.559 0.751 BERT 0.752 0.842 0.958 0.678 0.791 0.939 0.502 0.594 0.793 Word + POS STIR (HP’14 / PHH’18) 0.779 / 0.768 -/0.833 0.937/RNN (HS’15 / PHH’18) 0.711 / 0.668 -/0.790 0.902/LSTM joint tagset (HS’17) 0.599 0.686 0.907 0.557 0.726 LSTM-MTL (SEL’18) 0.753 0.816 0.919 0.548 Words + Timings + POS LSTM joint tagset (HS’17) 0.601 0.719 0.918 0.555 0.727 LSTM 0.692 0.778 0.931 0.601 0.720 0.910 0.557 0.727 LSTM-MTL (RH’20) 0.743 0.811 0.932 0.633 0.743 0.931 0.571 0.757 BERT 0.757 0.853 0.958 0.676 0.802 0.944 0.522 0.605 0.809 Table 1: Final disfluency detection accuracy results on Switchboard data attention heads, and 1024 hidden-size) for the model. In our analysis, when fine-tuning BERT, we followed the hyper-parameters of (Devlin et al., 2019). Since the datasets we use are tokenized, and each token has a matching tag, we adopt the directions provided by (Devlin et al., 2019) to deal with the sub-tokenization of BERT: to determine its label, the scores of the first sub-token are used, and further sub-token scores are discarded. Data We use standard Switchboard training data (all conversation numbers starting sw2*,sw3 * in the Penn Treebank III release: 100k utterances, 650k words) and use standard held-out data (PTB III files sw4[5-9] *: 6.4k utterances, 49k words) as our validation set. We test on the standard test data (PTB III files 4[0-1] *) with partial words and punctuation stripped away from all files. We only choose a subset of the held-out and test data for the ASR results in assessment, whereby both channels achieve below 40 percent WER to ensure good separation- this left us with 18 dialogues in validation data and 17 dialogues for test data. 6.1 Evaluation Criteria We calculate F1 accuracy for repair onset detection FrpS and for edit term words Fe, which includes interregna and Frm for reparandum detection. Performing the task live, on hypotheses of speech recognition that may not be quite equivalent to the annotated gold-standard transcription involves the use of time-based local accuracy metrics in a time window (i.e., within this time frame, has a disfluency been detected, even if not on the identical words?)-we, therefore, measure the F1 score over 10-second windows of each speaker’s channel. For incremental performance, we measure latency and output stability over time. We use the first time to detection (FTD) metric of (Zwarts et al., 2010) for latency: the average latency (in number of words) before the first detection of a gold standard repair onset or edit term word. For stability, we evaluate the edit overhead (EO) of output labels (Baumann et al., 2011), the proportion of the unnecessary edits (insertions and deletions) required to achieve the final labels produced by the model, with perfect performance being 0%. 6.2 Competitor Baselines We compare our incrementalised BERT model against a number of existing baselines, largely from existing incremental disfluency detection systems trained and tested on the same data: STIR (HP’14/HS’15/PHH’18): Hough and Purver (2014)’s STrongly Incremental Repair detection (STIR) non-deep model using n-gram language model features in a pipeline of Random Forest classifiers. The reparandum is detected by a backward search, showing robustness for longer lengths of repair compared to deep sequence tagging models (Purver et al., 2018). A state-ofthe-art incremental model on pre-segmented transcripts. RNN (HS’15): (Hough and Schlangen, 2015)’s RNN-based model, the first deep learning-based 3699 Training Scheme Model Final output F1 Incrementality Frm FrpS Fe EO FTD Chunk LSTM .591 .674 .901 0.21 0.06 MTL .631 .739 .911 0.41 0.07 BERT .647 .780 .938 0.61 0.32 Add-M LSTM .598 .683 .909 0.20 0.03 MTL .628 .751 .921 0.38 0.10 BERT .664 .788 .949 0.60 0.31 Table 2: Final accuracy vs. incremental performance trade-off in the different models on un-segmented transcripts. incremental disfluency detection model using the same tagset as in our model. Results from Purver et al. (2018) are used, which reproduced the model with some degradation in the results. LSTM: An LSTM version of Hough and Schlangen (2015) on pre-segmented transcripts LSTM joint tagset (HS’17) Hough and Schlangen (2017)’s model, which simultaneously predicts utterance segmentation using a joint tag set of utterance segmentation tags and disfluency tags, the latter of which is the same as our own. This is the only other work to use word timing information and to be testable on ASR results. LSTM-MTL (SEL’18) Shalyminov et al. (2018)’s multi-task learning model, which tags according to our tag set but simultaneously does language modelling by predicting the probability of the current word given the history. Also adds ground-truth POS tags to input. LSTM-MTL (RH’20): Rohanian and Hough (2020)’s multi-task learning model, which simultaneously predicts utterance segmentation, POS tags and language model probabilities, exhibiting state-of-the-art results for a strictly incremental deep model. The model is used as described by the authors and also here with the addition of timing information and gold standard POS information (as opposed to simultaneously predicted POS tags). It is also applied to ASR results as it is a suitable model to do so. This same model provides the automatic live utterance segmentation in our own model. 7 Results The results in terms of the final output of our best performing incremental BERT system in the three testing regimes versus its competitors is shown in Model F1 Repeats Substitution Deletes With Standard Training LSTM 0.94 0.70 0.48 MTL 0.96 0.72 0.46 BERT 0.96 0.77 0.54 With Add-M Training LSTM 0.95 0.71 0.48 MTL 0.96 0.73 0.47 BERT 0.96 0.79 0.54 Table 3: Performance on different types of repair. Table 1.1 We found our best model was the add-M trained model, and the best decoding strategy was using top-p sampling for predicting future words. Disfluency detection on transcripts For repair detection, our system’s best FrpS score for detecting repair onsets on pre-segmented transcripts at 0.853 beats state-of-the-art incremental systems. This performance degrades using automatic segmentation to 0.802, a state-of-the-art result for this setting. Its Frm accuracy of 0.757 on reparandum words on pre-segmented transcripts is only beaten by HP’14/PHH’18 model using word and POS input, making it a state-of-the-art strictly incremental deep model. This performance degrades to 0.678 on raw transcripts but is a state-of-the-art result for this setting. In terms of edit term detection, stateof-the-art detection results of 0.960 and 0.944 are achieved on the pre-segmented and unsegmented settings, improving over the existing benchmarks of HP’14 and RH’20. These results suggest we have achieved the aim of a strictly incremental model achieving high final accuracies. Disfluency detection on ASR results Using the ASR results from HS’17 for comparison, a significant improvement can be seen over the previously reported results on FrpS and Fe per 10-second window, improving from 0.557 to 0.605 and from 0.727 to 0.809 respectively. Given the previously reported best system gave strong correlations in terms of real repair rates, this is encouraging that our system could be very useful in a live setting. 7.1 Incremental Performance The purpose of this paper was to adapt a highperforming, non-incremental model for incremental operation. As can be seen in Table 2 and in Fig. 3, while our BERT model with top-p sample utterance prediction outperforms the multi-task 1Experiments are reproducible from https://github. com/mortezaro/tr-disfluency 3700 (a) (b) Figure 3: Incremental results of first time to detection (FTD) metric for rpS and e and edit overhead (EO) for disfluency detection labels.(a) On unsegmented transcripts. (b) On ASR results. model and vanilla LSTM model in terms of final output accuracy, its incremental output stability is slightly below its competitors, with the best edit overhead of 63% unnecessary edits versus 25% (LSTM joint tagset (HS’17)) and 42% (LSTMMTL (RH’20)) on ASR results, meaning the output is slightly, though not severely, more jittery. Of the prophecy-based approaches, we found the top-p sampling method gave the most stable results (EO=61% with chunk training, EO=60% with add-M training) and beam search gave the least stable. As shown in Fig. 3, while the constant latency approaches offer large advantages in EO over prophecy-based models on transcripts, that advantage disappears on ASR results, where the prophecy models generally outperform them. As can be seen in Table 2, there is a slight improvement in stability across all systems using the add-M training regime for final output and incremental performance. In terms of latency, results are even more encouraging, with the best FTD for rpS of 0.31 words (versus 0.03 and 0.07) on transcripts, which shows a relatively short latency of detecting the repair for the first time– this suggests a responsive, sensitive system. 7.2 Error Analysis We conduct an error analysis in terms of performance on different repair types and in terms of repairs with different lengths. Table 3 shows the performance in terms of FrpS score on detecting repairs of the three different types: verbatim repeats, substitutions, and deletes (restarts). Our BERT model performs best, either jointly or uniquely, across all three types, with a gain of 0.06 over its nearest competitors for substitutions and deletes. Through large-scale training, the enhanced linguistic knowledge equips it to recognize the syntactic 3701 Model Reparandum length Reparandum length of nested disfluencies 1 2 3 4 5 6 1 2 3 4 5 6 With Standard Training LSTM .843 .675 .405 .311 .134 .131 .747 .586 .382 .320 .110 .104 MTL .856 .683 .431 .335 .134 .131 .763 .586 .405 .291 .110 .104 BERT .892 .716 .469 .379 .310 .187 .818 .623 .405 .320 .130 .140 With Add-M Training LSTM .843 .675 .434 .334 .134 .131 .741 .586 .382 .320 .110 .104 MTL .851 .709 .468 .335 .134 .131 .779 .586 .405 .291 .130 .104 BERT .892 .719 .472 .379 .310 .187 .833 .645 .405 .320 .130 .140 Table 4: F1 of models on repairs with reparanda of different length and lexical parallelism in more complex repairs while retaining high accuracy on repeats. Table 4 shows the degradation in performance in detecting repairs of different lengths. With Add-M training, the BERT model degrades less and performs (joint) best on all lengths and nested disfluencies. While the performance on length five repairs is considerably better than the other deep models, the 0.187 accuracy on length six repairs is what gives it a slight disadvantage compared to the HP’14 explicit backtracking system (reported as high as 0.500 in PHH’18), which likely accounts for the lower Frm score despite the superior FrpS score of our system. 8 Discussion and Conclusion Our incremental GPT-2 and BERT-driven system performs well at detecting repair disfluencies on pre-segmented and unsegmented transcripts, achieving state-of-the-art results for a strictly incremental repair onset detection. Our system is competitive at reparadnum word detection and achieves state-of-the-art results in edit term detection. The results on ASR transcripts are also state-of-the-art. The high sequence-final performance comes at the expense of marginally increased jitter in the word-by-word output, but with sensitive and fast repair detection, on average first detecting the repair under a third of a second after the end of the repair onset word. These results suggest it is beginning to enjoy the best of both worlds in leveraging the right-ward context which BERT uses for its high performance, while the continuation predictions from the GPT-2 model are good enough to allow good incremental performance before the true right-ward context is available. The linguistic knowledge in the BERT model allows it to recognize parallelism in reparandum and repair phases and the absence thereof to increase performance on detecting substitution and delete repairs. This improvement to existing deep disfluency detection models, and, with appropriate use of open-ended language generation techniques with a GPT-2 language model, its good incremental performance, is consistent with a growing body of work (Heeman and Allen, 1999; Johnson and Charniak, 2004; Zwarts et al., 2010; Hough and Purver, 2014; Shalyminov et al., 2018; Rohanian and Hough, 2020), showing good language modelling can lead to good disfluency detection, as they are inherently part of the same process. Our system still fails to detect longer repairs compared to an explicit backtracking mechanism like (Hough and Purver, 2014). While the vanishing gradient problem is partly overcome here, the strictly left-to-right constraint on decoding puts memory limitations on any repair detection system. In future, we will explore efficient ways to navigate this space whilst not filtering out rarer repair forms. The results on ASR results show our disfluency detection system is ready for use in a live setting with a good degree of accuracy, and work is currently underway to use it to help detect a variety of different cognitive conditions, including Alzheimer’s Disease, in a live diagnostic system. Acknowledgments We thank the anonymous ACL-IJCNLP reviewers for their helpful comments and Matthew Purver for his continuous support and supervision on the wider project. References Timo Baumann, Okko Buß, and David Schlangen. 2011. Evaluation and optimisation of incremental processors. Dialogue & Discourse, 2(1):113–141. Timo Baumann, Casey Kennington, Julian Hough, and David Schlangen. 2017. Recognising conversational speech: What an incremental asr should do for a dialogue system and how to get there. In Dialogues with social robots, pages 421–432. Springer. 3702 Fahim Dalvi, Nadir Durrani, Hassan Sajjad, and Stephan Vogel. 2018. Incremental decoding and training methods for simultaneous translation in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 493–499. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833. John J Godfrey, Edward C Holliman, and Jane McDaniel. 1992. Switchboard: Telephone speech corpus for research and development. In Acoustics, Speech, and Signal Processing, IEEE International Conference on, volume 1, pages 517–520. IEEE Computer Society. Jonathan Gratch, Ron Artstein, Gale M Lucas, Giota Stratou, Stefan Scherer, Angela Nazarian, Rachel Wood, Jill Boberg, David DeVault, Stacy Marsella, et al. 2014. The distress analysis interview corpus of human and computer interviews. In LREC, pages 3123–3128. Peter A Heeman and James Allen. 1999. Speech repains, intonational phrases, and discourse markers: modeling speakers’ utterances in spoken dialogue. Computational Linguistics, 25(4):527–572. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751. Julian Hough and Matthew Purver. 2014. Strongly incremental repair detection. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 78–89. Julian Hough and David Schlangen. 2015. Recurrent neural networks for incremental disfluency detection. In Sixteenth Annual Conference of the International Speech Communication Association. Julian Hough and David Schlangen. 2017. Joint, incremental disfluency detection and utterance segmentation from speech. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 326–336. Karmele L´opez-de Ipi˜na, Jesus-Bernardino Alonso, Carlos Manuel Travieso, Jordi Sol´e-Casals, Harkaitz Egiraun, Marcos Faundez-Zanuy, Aitzol Ezeiza, Nora Barroso, Miriam Ecay-Torres, Pablo MartinezLage, et al. 2013. On the selection of noninvasive methods based on speech analysis oriented to automatic alzheimer disease diagnosis. Sensors, 13(5):6730–6745. Mark Johnson and Eugene Charniak. 2004. A TAGbased noisy-channel model of speech repairs. In ACL, pages 33–39. Paria Jamshid Lou and Mark Johnson. 2020. Improving disfluency detection by self-training a selfattentive model. arXiv preprint arXiv:2004.05323. Brielen Madureira and David Schlangen. 2020. Incremental processing in the age of non-incremental encoders: An empirical assessment of bidirectional models for incremental NLU. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 357–374, Online. Association for Computational Linguistics. M. Meteer, A. Taylor, R. MacIntyre, and R. Iyer. 1995. Disfluency annotation stylebook for the switchboard corpus. ms. Technical report, Department of Computer and Information Science, University of Pennsylvania. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119. Shamila Nasreen, Morteza Rohanian, Matthew Purver, and Julian Hough. 2021. Alzheimer’s dementia recognition from spontaneous speech using disfluency and interactional features. Frontiers in Computer Science, 3:49. Matthew Purver, Julian Hough, and Christine Howes. 2018. Computational models of miscommunication phenomena. Topics in cognitive science, 10(2):425– 451. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. Mohammad Sadegh Rasooli and Joel R. Tetreault. 2015. Yara parser: A fast and accurate dependency parser. Computing Research Repository, arXiv:1503.06733. Version 2. Morteza Rohanian and Julian Hough. 2020. Reframing incremental deep language models for dialogue processing with multi-task learning. In Proceedings of the 28th International Conference on Computational Linguistics, pages 497–507, Barcelona, Spain (Online). International Committee on Computational Linguistics. Morteza Rohanian, Julian Hough, and Matthew Purver. 2020. Multi-modal fusion with gating using audio, 3703 lexical and disfluency features for alzheimer’s dementia recognition from spontaneous speech. In Proc. Interspeech, pages 2187–2191. Igor Shalyminov, Arash Eshghi, and Oliver Lemon. 2018. Multi-task learning for domain-general spoken disfluency detection in dialogue systems. In Proceedings of the 22nd SemDial Workshop on the Semantics and Pragmatics of Dialogue (AixDial), Aixen-Provence. Elizabeth Shriberg. 1994. Preliminaries to a Theory of Speech Disfluencies. Ph.D. thesis, University of California, Berkeley. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. arXiv preprint arXiv:1706.03762. Shaolei Wang, Wangxiang Che, Qi Liu, Pengda Qin, Ting Liu, and William Yang Wang. 2020. Multi-task self-supervised learning for disfluency detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9193–9200. Vicky Zayats, Mari Ostendorf, and Hannaneh Hajishirzi. 2016. Disfluency detection using a bidirectional lstm. arXiv preprint arXiv:1604.03209. Simon Zwarts, Mark Johnson, and Robert Dale. 2010. Detecting speech repairs incrementally using a noisy channel approach. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1371–1378.
|
2021
|
286
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3704–3717 August 1–6, 2021. ©2021 Association for Computational Linguistics 3704 NeuralWOZ: Learning to Collect Task-Oriented Dialogue via Model-Based Simulation Sungdong Kim1,2 Minsuk Chang1,2 Sang-Woo Lee1,2 NAVER AI Lab1 NAVER Clova2 {sungdong.kim, minsuk.chang, sang.woo.lee}@navercorp.com Abstract We propose NeuralWOZ, a novel dialogue collection framework that uses model-based dialogue simulation. NeuralWOZ has two pipelined models, Collector and Labeler. Collector generates dialogues from (1) user’s goal instructions, which are the user context and task constraints in natural language, and (2) system’s API call results, which is a list of possible query responses for user requests from the given knowledge base. Labeler annotates the generated dialogue by formulating the annotation as a multiple-choice problem, in which the candidate labels are extracted from goal instructions and API call results. We demonstrate the effectiveness of the proposed method in the zero-shot domain transfer learning for dialogue state tracking. In the evaluation, the synthetic dialogue corpus generated from NeuralWOZ achieves a new state-of-theart with improvements of 4.4% point joint goal accuracy on average across domains, and improvements of 5.7% point of zero-shot coverage against the MultiWOZ 2.1 dataset.1 1 Introduction For a task-oriented dialogue system to be scalable, the dialogue system needs to be able to quickly adapt and expand to new scenarios and domains. However, the cost and effort in collecting and annotating an expanding dataset is not only laborintensive but also proportional to the size and variety of the unseen scenarios. There are three types of dialogue system expansions. (1) The simplest expansion is the addition of new instances in the knowledge base (KB) under the identical schema. For example, the addition of newly opened restaurants in the KB of restaurant domain falls under this category. (2) A slightly more complicated expansion involves modifications to the KB schema, and possibly the related 1The code is available at github.com/naver-ai/neuralwoz. Figure 1: Overview of NeuralWOZ. The NeuralWOZ takes goal instruction for the user side (U) and API call results for the system side (S) to synthesize dialogue. First, it generates dialogue from the inputs and then labels dialogue state (Bt) and active domain (Domaint) by turn t on the dialogue. instances. For example, additions of new constraint types to access the KB due to the change in needs of the user often require a restructuring of the KB. If a dialogue system built with only restaurant search in mind observes user’s requests about not only “restaurant location” and but also “traffic information” for navigating, the system now needs a new knowledge base including the additional different domain. (3) The most complex expansion is the one that expands across multiple domains. For example, imagine an already built dialogue system 3705 supported restaurant and hotel reservation domains, but now needs to expand to points of interest or other domains. It is difficult to expand to new domain without collecting new data instances and building a new knowledge base, if the schema between the source (restaurant and hotel in this case) and target domain (point of interest) look different. To support development of scalable dialogue systems, we propose NeuralWOZ, a model-based dialogue collection framework. NeuralWOZ uses goal instructions and KB instances for synthetic dialogue generation. NeuralWOZ mimics the mechanism of a Wizard-of-Oz (Kelley, 1984; Dahlb¨ack et al., 1993) and Figure 1 illustrates our approach. NeuralWOZ has two neural components, Collector and Labeler. Collector generates a dialogue by using the given goal instruction and candidate relevant API call results from the KB as an input. Labeler annotates the generated dialogue with appropriate labels by using the schema structure of the dialogue domain as meta information. More specifically, Labeler selects the labels from candidate labels which can be obtained from the goal instruction and the API call results. As a result, NeuralWOZ is able to generate a dialogue corpus without training data of the target domain. We evaluate our method for zero-shot domain transfer task (Wu et al., 2019; Campagna et al., 2020) to demonstrate the ability to generate corpus for unseen domains, when no prior training data exists. In dialogue state tracking (DST) task with MultiWOZ 2.1 (Eric et al., 2019), the synthetic data generated with NeuralWOZ achieves 4.4% point higher joint goal accuracy and 5.7% point higher zero-shot coverage than the existing baseline. Additionally, we examine few-shot and full data augmentation tasks using both training data and synthetic data. We also illustrate how to collect synthetic data beyond MultiWOZ domains, and discuss the effectiveness of the proposed approach as a data collection strategy. Our contributions are as follows: • NeuralWOZ, a novel method for generating dialogue corpus using goal instruction and knowledge base information • New state-of-the-art performance on the zeroshot domain transfer task • Analysis results highlighting the potential synergy of using the data generated from NeuralWOZ together with human-annotated data 2 Related Works 2.1 Wizard-of-Oz Wizard-of-Oz (WOZ) is a widely used approach for constructing dialogue data (Henderson et al., 2014a,b; El Asri et al., 2017; Eric and Manning, 2017; Budzianowski et al., 2018). It works by facilitating a role play between two people. “User” utilizes a goal instruction that describes the context of the task and details of request and “system” has access to a knowledge base, and query results from the knowledge base. They take turns to converse, while the user makes requests one by one following the instructions, the system responds according to the knowledge base, and labels user’s utterances. 2.2 Synthetic Dialogue Generation Other studies on dialogue datasets use the user simulator-based data collection approaches (Schatzmann et al., 2007; Li et al., 2017; Bordes et al., 2017; Shah et al., 2018; Zhao and Eskenazi, 2018; Shah et al., 2018; Campagna et al., 2020). They define domain schema, rules, and dialogue templates to simulate user behavior under certain goals. The ingredients to the simulation are designed by developers and the dialogues are realized by predefined mapping rules or paraphrasing by crowdworkers. If a training corpus for the target domain exists, neural models that synthetically generates dialogues can augment the training corpus (Hou et al., 2018; Yoo et al., 2019). For example, Yoo et al. (2020) introduce Variational Hierarchical Dialog Autoencoder (VHDA), where hierarchical latent variables exist for speaker identity, user’s request, dialog state, and utterance. They show the effectiveness of their model on single-domain DST tasks. SimulatedChat (Mohapatra et al., 2020) also uses goal instruction for dialogue augmentation. Although it does not solve zero-shot learning task with domain expansion in mind, we run auxiliary experiments to compare with NeuralWOZ, and the results are in the Appendix D. 2.3 Zero-shot Domain Transfer In zero-shot domain transfer tasks, there is no data for target domain, but there exists plenty of data for other domains similar to target domain. Solving the problem of domain expansion of dialogue systems can be quite naturally reducted to solving zero-shot domain transfer. Wu et al. (2019) conduct a landmark study on the zero-shot DST. They 3706 Figure 2: Illustration of Collector and Labeler. Collector takes goal instruction G and API call results A as the input, and outputs dialogue DT which consists of T turns. The state candidate C is prepopulated from the G and A as a full set for labeling. Finally, Labeler takes its value’s subset OSi and question q for each slot type Si and dialogue context Dt from Collector, and chooses answer ˜o from the OSi. suggest a model, Transferable Dialogue State Generator (TRADE), which is robust to a new domain where few or no training data for the domain exists. Kumar et al. (2020) and Li et al. (2021) follow the same experimental setup, and we also compare NeuralWOZ in the same experiment setup. Abstract Transaction Dialogue Model (ATDM) (Campagna et al., 2020), another method for synthesizing dialogue data, is another baseline for zero-shot domain transfer tasks we adopt. They use rules, abstract state transition, and templates to synthesize the dialogue, which is then fed into a model-based zero-shot learner. They achieved state-of-the-art in the task using the synthetic data on SUMBT (Lee et al., 2019), a pretrained BERT (Devlin et al., 2019) based DST model. 3 NeuralWOZ In this section, we describe the components of NeuralWOZ in detail, and how they interact with each other. Figure 2 illustrates the input and output of two modules in NeuralWOZ. The synthetic corpus, which Collector and Labeler made, are used for the training of the DST baselines, TRADE (Wu et al., 2019) and SUMBT (Lee et al., 2019) in our experiments. 3.1 Problem Statement Domain Schema In task-oriented dialogues, there are two slot types; informable and requestable slots (Henderson et al., 2014a; Budzianowski et al., 2018). The informable slots are the task constraints to find relevant information from user requests, for example, “restaurantpricerange”, “restaurant-food”, “restaurant-name”, and “restaurant-book people” in Figure 1. The requestable slots are the additional details of user requests, like “reference number” and “address” in Figure 1. Each slot S can have its corresponding value V in a scenario. In multi-domain scenarios, each domain has a knowledge base KB, which consists of slot-value pairs corresponding to its domain schema. The API call results in Figure 1 are the examples of the KB instances of the restaurant domain. Goal Instruction The goal instruction, G, is a natural language text describing constraints of user behavior in the dialogue D including informable and requestable slots. The paragraph consists of four sentences at the top of Figure 1 is an example. We define a set of informable slot-value pairs that explicitly expressed on the G as CG, which we formally define as CG = {(SG i , V G i ) | 1 ≤i ≤|CG|, SG i ∈informable}. (“restaurantpricerange”, “expensive”) and (“restaurant-food”, “british”) are examples of the elements of CG (Figure 1). API Call Results The API call results, A, are corresponding query results of the CG from KB. We formally define A = {ai | 1 ≤i ≤|A|, ai ∈KB}. Each ai is associated with its domain, domainai, and with slot-value pairs, Cai = {(Sai k , V ai k ) | 1 ≤ k ≤|Cai|}. A slot Sai k can be either informable or requestable slot. For example, the restaurant instance, “graffiti” in Figure 1, is a query result from (“restaurant-pricerange”, “expensive”) and (“restaurant-food”, “british”) described in the goal instruction. State Candidate We define informable slot-value pairs that are not explicit in G but accessible by A in D as CA = {(SA i , V A i ) | 1 ≤i ≤|CA|, SA i ∈ informable}. It contains all informable slot-value pairs from Ca1 to Ca|A|. The elements of CA are 3707 likely to be uttered by summaries of current states or recommendations of KB instances by the system side in D. The system utterance of the second turn in Figure 1 is an example (“I recommend graffiti.”). In this case, the slot-value pair (“restaurant-name”, “graffiti”) can be obtained from the A, not from the G. Finally, state candidate C is the union of CG and CA. It is a full set of the dialogue state for the dialogue D from given G and A. Thus, it can be used as label candidates of dialogue state tracking annotation. 3.2 Collector Collector is a sequence-to-sequence model, which takes a goal instruction G and API call results A as the input and generates dialogue DT . The generated dialogue DT = (r1, u1, ..., rT , uT ) is the sequence of system response r and user utterance u. They are represented by N tokens (w1, ..., wN)2. p(DT |G, A) = N Y i=1 p(wi|w<i, G, A) We denote the input of Collector as <s> ⊕G ⊕ </s> ⊕A, where the ⊕is concatenate operation. The <s> and </s> are special tokens to indicate start and seperator respectively. The tokenized natural language description of G is directly used as the tokens. The A takes concatenation of each ai (a1 ⊕· · · ⊕a|A|)3. For each ai, we flatten the result to the token sequence, <domain>⊕domainai ⊕<slot>⊕Sai 1 ⊕V ai 1 ⊕ · · · ⊕<slot> ⊕Sai |Cai| ⊕V ai |Cai|. The <domain> and <slot> are other special tokens as separators. The objective function of Collector is LC = −1 MC MC X j=1 Nj X i=1 log p(wj i |wj <i, Gj, Aj). Our Collector model uses the transformer architecture (Vaswani et al., 2017) initialized with pretrained BART (Lewis et al., 2020). Collector is trained using negative log-likelihood loss, where MC is the number of training dataset for Collector and Nj is target length of the j-th instance. Following Lewis et al. (2020), label smoothing is used during the training with the smoothing parameter of 0.1. 2Following Hosseini-Asl et al. (2020), we also utilize rolespecific special tokens <system> and <user> for the r and u respectively. 3we limit the |A| to a maximum 3 3.3 Labeler We formulate labeling as a multiple-choice problem. Specifically, Labeler takes a dialogue context Dt = (r1, u1, ..., rt, ut), question q, and a set of answer options O = {o1, o2, ..., o|O|}, and selects one answer ˜o ∈O. Labeler encodes the inputs for each oi separately, and soi ∈R1 is the corresponding logit score from the encoding. Finally, the logit score is normalized via softmax function over the answer option set O. p(oi|Dt, q, O) = exp(soi) P|O| j exp(soj) , soi = Labeler(Dt, q, oi), ∀i. The input of Labeler is a concatenation of Dt, q, and oi, <s>⊕Dt⊕</s>⊕q⊕</s>⊕oi⊕</s>, with special tokens. For labeling dialogue states to Dt, we use the slot description for each corresponding slot type, Si, as the question, for example, “what is area or place of hotel?” for “hotel-area” in Figure 2. We populate corresponding answer options OSi = {Vj|(Sj, Vj) ∈C, Sj = Si} from the state candidate set C. There are two special values, Dontcare to indicate the user has no preference and None to indicate the user is yet to specify a value for this slot (Henderson et al., 2014a; Budzianowski et al., 2018). We include these values in the OSi. For labeling the active domain of Dt, which is the domain at t-th turn of Dt, we define domain question, for example “what is the domain or topic of current turn?”, for q and use predefined domain set Odomain as answer options. In MultiWOZ, Odomain = {“Attraction”, “Hotel”, “Restaurant”, “Taxi”, “Train”}. Our Labeler model employs a pretrained RoBERTa model (Liu et al., 2019) as the initial weight. Dialogue state and domain labeling are trained jointly based on the multiple choice setting. Preliminary result shows that the imbalanced class problem is significant in the dialogue state labels. Most of the ground-truth answers is None given question4. Therefore, we revise the negative loglikelihood objective to weight other (not-None) answers by multiplying a constant β to the loglikelihood when the answer of training instance is 4The number of None in the training data is about 10 times more than the number of others 3708 not None. The objective function of Labeler is LL = −1 ML ML X j=1 T X t=1 Nq X i=1 Lj t,i Lj t,i = ( β log p(˜oj t,i|Dj t, qj i , Oj i ), if ˜oj t,i ̸= None log p(˜oj t,i|Dj t, qj i , Oj i ), otherwise , where ˜oj t,i denotes the answer of i-th question for j-th training dialogue at turn t, the Nq is the number of questions, and ML is the number of training dialogues for Labeler. We empirically set β to a constant 5. 3.4 Synthesizing a Dialogue We first define goal template G.5 G is a delexicalized version of G by changing each value V G i expressed on the instruction to its slot SG i . For example, the “expensive” and “british” of goal instruction in Figure 1 are replaced with “restaurantpricerange” and “restaurant-food”, respectively. As a result, domain transitions in G becomes convenient. First, G is sampled from a pre-defined set of goal template. API call results A, which correspond to domain transitions in G, are randomly selected from the KB. Especially, we constrain the sampling space of A when the consecutive scenario among domains in G have shared slot values. For example, the sampled API call results for restaurant and hotel domain should share the value of “area” to support the following instruction “I am looking for a hotel nearby the restaurant”. G and A are aligned to become GA. In other words, each value for SG i in G is assigned using the corresponding values in A.6 Then, Collector generates dialogue D, of which the total turn number is T, given GA and A. More details are in Appendix A. Nucleus sampling (Holtzman et al., 2020) is used for the generation. We denote dialogue state and active domain at turn t as Bt and domaint respectively. The Bt, {(Sj, Vj,t) | 1 ≤j ≤J}, has J number of predefined slots and their values at turn t. It means Labeler is asked J (from slot descriptions) + 1 (from domain question) questions regarding dialogue context Dt from Collector. Finally, the out5In Budzianowski et al. (2018), they also use templates like ours when allocating goal instructions to the user in the Wizard-of-Oz setup. 6Booking-related slots, e.g., the number of people, time, day, and etc., are randomly sampled for their values since they are independent of the A. put of Labeler is a set of dialogue context, dialogue state, and active domain at turn t triples {(D1, B1, domain1), ..., (DT , BT , domainT )}. 4 Experimental Setups 4.1 Dataset We use MultiWOZ 2.1 (Eric et al., 2019) dataset7 for our experiments. It is one of the largest publicly available multi-domain dialogue data and it contains 7 domains related to travel (attraction, hotel, restaurant, taxi, train, police, hospital), including about 10,000 dialogues. The MultiWOZ data is created using WOZ so it includes goal instruction per each dialogue and domain-related knowledge base as well. We train our NeuralWOZ using the goal instructions and the knowledge bases first. Then we evaluate our method on dialogue state tracking with and without synthesized data from the NeuralWOZ using five domains (attraction, restaurant, hotel, taxi, train) in our baseline, and follow the same preprocessing steps of Wu et al. (2019); Campagna et al. (2020). 4.2 Training NeuralWOZ We use the pretrained BART-Large (Lewis et al., 2020) for Collector and RoBERTa-Base (Liu et al., 2019) for Labeler. They share the same byte-level BPE vocab (Sennrich et al., 2016) introduced by Radford et al. (2019). We train the pipelined models using Adam optimizer (Kingma and Ba, 2017) with learning rate 1e-5, warming up steps 1,000, and batch size 32. The number of training epoch is set to 30 and 10 for Collector and Labeler respectively. For the training phase of Labeler, we use a state candidate set from ground truth dialogue states B1:T for each dialogue, not like the synthesizing phase where the options are obtained from goal instruction and API call results. We also evaluate the performance of Labeler itself like the training phase with validation data (Table 5). Before training Labeler on the MultiWOZ 2.1 dataset, we pretrain Labeler on DREAM8 (Sun et al., 2019) to boost Labeler’s performance. This is similar to coarse-tuning in Jin et al. (2019). The same hyper parameter setting is used for the pretraining. For the zero-shot domain transfer task, we exclude dialogues which contains target domain from 7https://github.com/budzianowski/multiwoz 8The DREAM is a multiple-choice question answering dataset in dialogue and includes about 84% of non-extractive answers. 3709 Model Training Hotel Restaurant Attraction Train Taxi Average TRADE Full dataset 50.5 / 91.4 61.8 / 92.7 67.3 / 87.6 74.0 / 94.0 72.7 / 88.9 65.3 / 89.8 Zero-shot (Wu) 13.7 / 65.6 13.4 / 54.5 20.5 / 55.5 21.0 / 48.9 60.2 / 73.5 25.8 / 59.6 Zero-shot (Campagna) 19.5 / 62.6 16.4 / 51.5 22.8 / 50.0 22.9 / 48.0 59.2 / 72.0 28.2 / 56.8 Zero-shot + ATDM 28.3 / 74.5 35.9 / 75.6 34.9 / 62.2 37.4 / 74.5 65.0 / 79.9 40.3 / 73.3 Zero-shot + NeuralWOZ 26.5 / 75.1 42.0 / 84.2 39.8 / 65.7 48.1 / 83.9 65.4 / 79.9 44.4 / 77.8 Zero-shot Coverage 52.5 / 82.2 68.0 / 90.8 59.1 / 75.0 65.0 / 89.3 90.0 / 89.9 66.9 / 85.4 SUMBT Full dataset 51.8 / 92.2 64.2 / 93.1 71.1 / 89.1 77.0 / 95.0 68.2 / 86.0 66.5 / 91.1 Zero-shot 19.8 / 63.3 16.5 / 52.1 22.6 / 51.5 22.5 / 49.2 59.5 / 74.9 28.2 / 58.2 Zero-shot + ATDM 36.3 / 83.7 45.3 / 82.8 52.8 / 78.9 46.7 / 84.2 62.6 / 79.4 48.7 / 81.8 Zero-shot + NeuralWOZ 31.3 / 81.7 48.9 / 88.4 53.0 / 79.0 66.9 / 92.4 66.7 / 83.9 53.4 / 85.1 Zero-shot Coverage 60.4 / 88.6 76.2 / 95.0 74.5 / 88.7 86.9 / 97.3 97.8 / 97.6 79.2 / 93.4 Table 1: Experimental results of zero-shot domain transfer on the test set of MultiWOZ 2.1. Joint goal accuracy / slot accuracy are reported. The Wu indicates original zero-shot scheme of the TRADE suggested by Wu et al. (2019) and reproduced by Campagna et al. (2020). The Campagna indicates a revised version of the original by Campagna et al. (2020). The + indicates the synthesized dialogue is used together for the training. the training data for both Collector and Labeler. This means we train our pipelines for every target domain separately. We use the same seed data for training as Campagna et al. (2020) did in the fewshot setting. All our implementations are conducted on NAVER Smart Machine Learning (NSML) platform (Sung et al., 2017; Kim et al., 2018) using huggingface’s transformers library (Wolf et al., 2020). The best performing models, Collector and Labeler, are selected by evaluation results from the validation set. 4.3 Synthetic Data Generation We synthesize 5,000 dialogues for every target domain for both zero-shot and few-shot experiments9, and 1,000 dialogues for full data augmentation. For zero-shot experiment, since the training data are unavailable for a target domain, we only use goal templates that contain the target domain scenario in the validation set similar to Campagna et al. (2020). We use nucleus sampling in Collector with parameters top p ratio in the range {0.92, 0.98} and temperature in the range {0.7, 0.9, 1.0}. It takes about two hours to synthesize 5,000 dialogues using one V100 GPU. More statistics is in Appendix B. 4.4 Baselines We compare NeuralWOZ with baseline methods both zero-shot learning and data augmentation using MultiWOZ 2.1 in our experiments. We use a baseline zero-shot learning scheme which does not 9In Campagna et al. (2020), the average number of synthesized dialogue over domains is 10,140. use synthetic data (Wu et al., 2019). For data augmentation, we use ATDM and VHDA. ATDM refers to a rule-based synthetic data augmentation method for zero-shot learning suggested by Campagna et al. (2020). It defines rules including state transitions and templates for simulating dialogues and creates about 10,000 synthetic dialogues per five domains in the MultiWOZ dataset. Campagna et al. (2020) feed the synthetic dialogues into zero-shot learner models to perform zero-shot transfer task for dialogue state tracking. We also employ TRADE (Wu et al., 2019) and SUMBT (Lee et al., 2019) as baseline zero-shot learners for fair comparisons with the ATDM. VHDA refers to model-based generation method using hierarchical variational autoencoder (Yoo et al., 2020). It generates dialogues incorporating information of speaker, goal of the speaker, turnlevel dialogue acts, and utterance sequentially. Yoo et al. (2020) augment about 1,000 dialogues for restaurant and hotel domains in the MultiWOZ dataset. For a fair comparison, we use TRADE as the baseline model for the full data augmentation experiments. Also, we compare ours with the VHDA on the single-domain augmentation setting following their report. 5 Experimental Results We use both joint goal accuracy (JGA) and slot accuracy (SA) as the performance measurement. The JGA is an accuracy which checks whether all slot values predicted at each turn exactly match the ground truth values, and the SA is the slotwise accuracy of partial match against the grouth 3710 Synthetic TRADE SUMBT no syn 44.2 / 96.5 46.7 / 96.7 ATDM 43.0 / 96.4 46.9 / 96.6 NeuralWOZ 45.8 / 96.7 47.1 / 96.8 Table 2: Full data augmentation on multi-domain DST. Joint goal accuracy / slot accuracy are reported. truth values. Especially for zero and few-shot setting, we follow the previous setup (Wu et al., 2019; Campagna et al., 2020). Following Campagna et al. (2020), the zero-shot learner model should be trained on data excluding the target domain, and tested on the target domain. We also add synthesized data from our NeuralWOZ which is trained in the same way, i.e., leave-one-out setup, to the training data in the experiment. 5.1 Zero-Shot Domain Transfer Learning Our method achieves new state-of-the-art of zeroshot domain transfer learning for dialogue state tracking on the MultiWOZ 2.1 dataset (Table 1). Except for the hotel domain, the performance over all target domains is significantly better than the previous sota method. We discuss the lower performance in hotel domain in the analysis section. Following the work of Campagna et al. (2020), we also measure zero-shot coverage, which refers to the accuracy ratio between zero-shot learning over target domain, and fully trained model including the target domain. Our NeuralWOZ achieves 66.9% and 79.2% zero-shot coverage on TRADE and SUMBT, respectively, outperforming previous state-of-the-art, ATDM, which achieves 61.2% and 73.5%, respectively. 5.2 Data Augmentation on Full Data Setting For full data augmentation, our synthesized data come from fully trained model including all five domains in this setting. Table 2 shows that our model still consistently outperforms in full data augmentation of multi-domain dialogue state tracking. Specifically, our NeuralWOZ performs 2.8% point better on the joint goal accuracy of TRADE than ATDM. Our augmentation improves the performance by a 1.6% point while ATDM degrades. We also compare NeuralWOZ with VHDA, a previous model-based data augmentation method for dialogue state tracking (Yoo et al., 2020). Since the VHDA only considers single-domain simulation, we use single-domain dialogue in hotel Synthetic Restaurant Hotel no syn 64.1 / 93.1 52.3 / 91.9 VHDA 64.9 / 93.4 52.7 / 92.0 NeuralWOZ 65.8 / 93.6 53.5 / 92.1 Table 3: Full data augmentation on single-domain DST. Joint goal accuracy / slot accuracy are reported. TRADE is used for evaluation. Domain Collector ↓ Labeler ↑ Full 5.0 86.8 w/o Hotel 5.4 79.2 w/o Restaurant 5.3 81.3 w/o Attraction 5.3 83.4 w/o Train 5.6 83.2 w/o Taxi 5.2 83.1 Table 4: Intrinsic evaluation results of NeuralWOZ on the validation set of MultiWOZ 2.1. Perplexity and joint goal accuracy are used for measurement respectively. The “w/o” means the domain is excluded from the full data. Different from the zero-shot experiments, the joint goal accuracy is computed by regarding all five domains. and restaurant domains for the evaluation. Table 3 shows that our method still performs better than the VHDA in this setting. NeuralWOZ has more than twice better joint goal accuracy gain than that of VHDA. 5.3 Intrinsic Evaluation of NeuralWOZ Table 4 shows the intrinsic evaluation results from two components (Collector and Labeler) of the NeuralWOZ on the validation set of MultiWOZ 2.1. We evaluate each component using perplexity for Collector and joint goal accuracy for Labeler, respectively. Note that the joint goal accuracy is achieved by using state candidate set, prepopulated as the multiple-choice options from the ground truth, B1:T , as the training time of Labeler. It can be seen as using meta information since its purpose is accurate annotation but not the dialogue state tracking itself. We also report the results by excluding target domain from full dataset to simulate zero-shot environment. Surprisingly, synthesized data from ours performs effectively even though the annotation by Labeler is not perfect. We conduct further analysis, the responsibility of each model, in the following section. 3711 Figure 3: Breakdown of accuracy by slot of hotel domain in the zero-shot experiments when using synthetic data. The analysis is conducted based on TRADE. 6 Analysis 6.1 Error Analysis Figure 3 shows the slot accuracy for each slot type in the hotel domain, which is the weakest domain from ours. Different from other four domains, only the hotel domain has two boolean type slots, “parking” and “internet”, which can have only “yes” or “no” as their value. Since they have abstract property for the tracking, Labeler’s labeling performance tends to be limited to this domain. However, it is noticeable that our accuracy of booking related slots (book stay, book people, book day) are much higher than the ATDM’s. Moreover, the model using synthetic data from the ATDM totally fails to track the “book stay” slot. In the synthesizing procedures of Campagna et al. (2020), they create the data with a simple substitution of a domain noun phrase when the two domains have similar slots. For example, “find me a restaurant in the city center” can be replaced with “find me a hotel in the city center” since the restaurant and hotel domains share “area” slot. We presume it is why they outperform over slots like “pricerange” and “area”. 6.2 Few-shot Learning We further investigate how our method is complementary with human-annotated data. Figure 4 illustrates our NeuralWOZ shows a consistent gain in the few-shot domain transfer setting. Unlike the performance with ATDM is saturated as few-shot ratio increases, the performance using our NeuralWOZ is improved continuously. We get about 5.8% point improvement from the case which does not use synthetic data when using 10% of humanannotated data for the target domain. It implies our method could be used more effectively with the Figure 4: Few-shot learning result in MultiWOZ 2.1. The score indicates average across domain. TRADE is used for the baseline model. Collector Labeler Hotel’s JGA Full Full 53.5 Full w/o Hotel 30.8 w/o Hotel Full 27.3 w/o Hotel w/o Hotel 26.5 Table 5: Result of responsibility analysis. We compare the performances of each model with and without the hotel domain in the training data. human-annotated data in a real scenario. 6.3 Ablation Study We discover whether Collector and Labeler are more responsible for the quality of synthesizing. Table 5 shows ablation results where each model of NeuralWOZ is trained the data including or withholding the hotel domain. Except for the training data for each model, the pipelined models are trained and dialogues are synthesized in the same way. Then, we train TRADE model using the synthesized data and evaluate it on hotel domain like the zero-shot setting. The performance gain from Collector which is trained including the target domain is 4.3% point, whereas the gain from Labeler is only 0.8% point. It implies the generation quality from Collector is more responsible for the performance of the zero-shot learner than accurate annotation of Labeler. 6.4 Qualitative Analysis Figure 5 is an qualitative example generated by NeuralWOZ. It shows the NeuralWOZ can generate an unseen movie domain which has a different schema from the traveling, the meta domain of the MultiWOZ dataset, even if it is trained on only the 3712 Figure 5: Unseen domain dialogue generation from NeuralWOZ. The movie domain is an example. It has very different domain schema from the domains in MultiWOZ dataset. dataset. It is harder to generalize when the schema structure of the target domain is different from the source domain. Other examples can be found in Appendix C. We would like to extend the NeuralWOZ to more challenging expansion scenario like these in future work. 6.5 Comparison on End-to-End Task To show that our framework can be used for other dialogue tasks, we test our data augmentation method on end-to-end task in MultiWOZ 2.1. We describe the result in Appendix D with discussion. In full data setting, Our method achieves 17.46 BLUE, 75.1 Inform rate, 64.6 Success rate, and 87.31 Combine rate, showing performance gain using the synthetic data. Appendix D also includes the comparison and discussion on SimulatedChat (Mohapatra et al., 2020). 7 Conclusion We propose NeuralWOZ, a novel dialogue collection framework, and we show our method achieves state-of-the-art performance on zero-shot domain transfer task. We find the dialogue corpus from NeuralWOZ is synergetic with human-annotated data. Finally, further analysis shows that NeuralWOZ can be applied for scaling dialogue system. We believe NeuralWOZ will spark further research into dialogue system environments where expansion target domains are distant from the source domains. Acknowledgments We thank Sohee Yang, Gyuwan Kim, Jung-Woo Ha, and other members of NAVER AI for their valuable comments. We also thank participants who helped our preliminary experiments for building data collection protocol. References Antoine Bordes, Y-Lan Boureau, and Jason Weston. 2017. Learning end-to-end goal-oriented dialog. Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, I˜nigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gaˇsi´c. 2018. MultiWOZ - a large-scale multi-domain Wizard-of-Oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016–5026, Brussels, Belgium. Association for Computational Linguistics. Giovanni Campagna, Agata Foryciarz, Mehrad Moradshahi, and Monica Lam. 2020. Zero-shot transfer learning with synthesized data for multi-domain dialogue state tracking. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 122–132, Online. Association for Computational Linguistics. Nils Dahlb¨ack, Arne J¨onsson, and Lars Ahrenberg. 1993. Wizard of oz studies: why and how. In Proceedings of the 1st international conference on Intelligent user interfaces, pages 193–200. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Layla El Asri, Hannes Schulz, Shikhar Sharma, Jeremie Zumer, Justin Harris, Emery Fine, Rahul Mehrotra, and Kaheer Suleman. 2017. Frames: a corpus for adding memory to goal-oriented dialogue systems. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 207– 219, Saarbr¨ucken, Germany. Association for Computational Linguistics. Mihail Eric, Rahul Goel, Shachi Paul, Adarsh Kumar, Abhishek Sethi, Peter Ku, Anuj Kumar Goyal, Sanchit Agarwal, Shuyang Gao, and Dilek Hakkani-Tur. 2019. Multiwoz 2.1: A consolidated multi-domain dialogue dataset with state corrections and state tracking baselines. arXiv preprint arXiv:1907.01669. 3713 Mihail Eric and Christopher D. Manning. 2017. Keyvalue retrieval networks for task-oriented dialogue. Matthew Henderson, Blaise Thomson, and Jason D. Williams. 2014a. The second dialog state tracking challenge. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pages 263–272, Philadelphia, PA, U.S.A. Association for Computational Linguistics. Matthew Henderson, Blaise Thomson, and Jason D Williams. 2014b. The third dialog state tracking challenge. In 2014 IEEE Spoken Language Technology Workshop (SLT), pages 324–329. IEEE. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. Yutai Hou, Yijia Liu, Wanxiang Che, and Ting Liu. 2018. Sequence-to-sequence data augmentation for dialogue language understanding. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1234–1245, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Di Jin, Shuyang Gao, Jiun-Yu Kao, Tagyoung Chung, and Dilek Hakkani-tur. 2019. Mmm: Multi-stage multi-task learning for multi-choice reading comprehension. John F Kelley. 1984. An iterative design methodology for user-friendly natural language office information applications. ACM Transactions on Information Systems (TOIS), 2(1):26–41. Hanjoo Kim, Minkyu Kim, Dongjoo Seo, Jinwoong Kim, Heungseok Park, Soeun Park, Hyunwoo Jo, KyungHyun Kim, Youngil Yang, Youngkwan Kim, et al. 2018. Nsml: Meet the mlaas platform with a real-world case study. arXiv preprint arXiv:1810.09957. Diederik P. Kingma and Jimmy Ba. 2017. Adam: A method for stochastic optimization. Adarsh Kumar, Peter Ku, Anuj Kumar Goyal, Angeliki Metallinou, and Dilek Hakkani-Tur. 2020. Ma-dst: Multi-attention based scalable dialog state tracking. Hwaran Lee, Jinsik Lee, and Tae-Yoon Kim. 2019. SUMBT: Slot-utterance matching for universal and scalable belief tracking. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5478–5483, Florence, Italy. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Shuyang Li, Jin Cao, Mukund Sridhar, Henghui Zhu, Shang-Wen Li, Wael Hamza, and Julian McAuley. 2021. Zero-shot generalization in dialog state tracking through generative question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1063–1074, Online. Association for Computational Linguistics. Xiujun Li, Zachary C. Lipton, Bhuwan Dhingra, Lihong Li, Jianfeng Gao, and Yun-Nung Chen. 2017. A user simulator for task-completion dialogues. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. Biswesh Mohapatra, Gaurav Pandey, Danish Contractor, and Sachindra Joshi. 2020. Simulated chats for task-oriented dialog: Learning to generate conversations from instructions. arXiv preprint arXiv:2010.10216. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Jost Schatzmann, Blaise Thomson, Karl Weilhammer, Hui Ye, and Steve Young. 2007. Agenda-based user simulation for bootstrapping a POMDP dialogue system. In Human Language Technologies 2007: The Conference of the North American Chapter of the Association for Computational Linguistics; Companion Volume, Short Papers, pages 149– 152, Rochester, New York. Association for Computational Linguistics. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725, Berlin, Germany. Association for Computational Linguistics. Pararth Shah, Dilek Hakkani-T¨ur, Gokhan T¨ur, Abhinav Rastogi, Ankur Bapna, Neha Nayak, and Larry Heck. 2018. Building a conversational agent overnight with dialogue self-play. arXiv preprint arXiv:1801.04871. Kai Sun, Dian Yu, Jianshu Chen, Dong Yu, Yejin Choi, and Claire Cardie. 2019. Dream: A challenge dataset and models for dialogue-based reading comprehension. 3714 Nako Sung, Minkyu Kim, Hyunwoo Jo, Youngil Yang, Jingwoong Kim, Leonard Lausen, Youngkwan Kim, Gayoung Lee, Donghyun Kwak, Jung-Woo Ha, et al. 2017. Nsml: A machine learning platform that enables you to focus on your models. arXiv preprint arXiv:1712.05902. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Chien-Sheng Wu, Andrea Madotto, Ehsan HosseiniAsl, Caiming Xiong, Richard Socher, and Pascale Fung. 2019. Transferable multi-domain state generator for task-oriented dialogue systems. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 808–819, Florence, Italy. Association for Computational Linguistics. Kang Min Yoo, Hanbit Lee, Franck Dernoncourt, Trung Bui, Walter Chang, and Sang-goo Lee. 2020. Variational hierarchical dialog autoencoder for dialog state tracking data augmentation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3406–3425, Online. Association for Computational Linguistics. Kang Min Yoo, Youhyun Shin, and Sang-goo Lee. 2019. Data augmentation for spoken language understanding via joint variational generation. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 7402–7409. Yichi Zhang, Zhijian Ou, and Zhou Yu. 2020. Taskoriented dialog systems that consider multiple appropriate responses under the same context. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9604–9611. Tiancheng Zhao and Maxine Eskenazi. 2018. Zeroshot dialog generation with cross-domain latent actions. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue, pages 1–10. 3715 A Goal Instruction Sampling for Synthesizing in NeuralWOZ Figure 6: An example of sampling goal instruction GA using goal template G and randomly selected API call results A. B Data Statistics # of Dialogues # of Turns Domain Slots Train Valid Test Train Valid Test Attraction area, name, type 2,717 401 395 8,073 1,220 1,256 Hotel price range, type, parking, book stay, book day, book people, area, stars, internet, name 3,381 416 394 14,793 1,781 1,756 Restaurant food, price range, area, name, book time, book day, book people 3,813 438 437 15,367 1,708 1,726 Taxi leave at, destination, departure, arrive by 1,654 207 195 4,618 690 654 Train destination, day, departure, arrive by, book people, leave at 3,103 484 494 12,133 1,972 1,976 Table 6: Data Statistics of MultiWOZ 2.1. C Additional Qualitative Examples Figure 7 shows other examples from our NeuralWOZ. The left subfigure shows an example of synthesized dialogue from NeuralWOZ in a restaurant, which is seen domain and has the same schema from the 3716 Attraction Hotel Restaurant Taxi Train Full # goal template 411 428 455 215 482 1,000 # synthesized dialogues 5,000 5,000 5,000 5,000 5,000 1,000 # synthesized turns 38,655 38,112 37,230 45,542 37,863 35,053 # synthesized tokens 947,791 950,272 918,065 1,098,917 873,671 856,581 Table 7: Statistics of the synthesized data used in NeuralWOZ using for zero-shot and full augmentation experiments. Figure 7: Qualitative examples of synthesized dialogues from NeuralWOZ in the restaurant domain. Model Belief State BLEU Inform Success Combined DAMD (Zhang et al., 2020) Oracle 17.3 80.3 65.1 90 SimpleTOD (Hosseini-Asl et al., 2020) Oracle 16.22 85.1 73.5 95.52 GPT2 (Mohapatra et al., 2020) Oracle 15.95 72.8 63.7 84.2 GPT2 + SimulatedChat (Mohapatra et al., 2020) Oracle 15.06 80.4 62.2 86.36 GPT2 (ours) Oracle 17.27 77.1 67.8 89.72 GPT2 + NeuralWOZ (ours) Oracle 17.69 78.1 67.6 90.54 DAMD (Zhang et al., 2020) Generated 18.0 72.4 57.7 83.05 SimpleTOD (Hosseini-Asl et al., 2020) Generated 14.99 83.4 67.1 90.24 GPT2 (Mohapatra et al., 2020) Generated 15.94 66.2 55.4 76.74 GPT2 + SimulatedChat (Mohapatra et al., 2020) Generated 14.62 72.5 53.7 77.72 GPT2 (ours) Generated 17.38 74.6 64.4 86.88 GPT2 + NeuralWOZ (ours) Generated 17.46 75.1 64.6 87.31 Table 8: Performance of the end-to-end task model. restaurant domain in MultiWOZ dataset. However, the “spicy club” is an unseen instance which is newly added to the schema for the synthesizing. The right subfigure shows other synthetic dialogue in restaurant, which is a seen domain but has different schema from restaurant domain in MultiWOZ dataset. It describes navigation in-car scenario which is borrowed from KVret dataset (Eric and Manning, 2017). It is a non-trivial problem to adapt to unseen scenario, even if it is in the same domain. D Additional Explanation on Comparison in End-to-End Task To compare our model with the model of (Mohapatra et al., 2020), we conduct end-to-end task experiments the previous work did. Table 8 illustrates the result. Though the performance of baseline implementation 3717 is different, we can see that the trend of performance improvement is comparable to the report of SimulatedChat. Two studies are also different in terms of modeling. In our method, all utterances in the dialogue are first collected based on goal instruction and KB information by Collector. After that, Labeler selects annotations from candidate labels, which can be inducted from goal instruction and KB information. On the other hand, SimulatedChat creates utterance and label sequentially with knowledge base access, for each turn. Thus, each generation of utterance is affected by the generated utterance of labels of the previous turn. In detail, the two methods also differ in terms of complexity. SimulatedChat creates a model for each domain separately, and for each domain, it creates five neural modules: user response generation, user response selector, agent query generator, agent response generator, and agent response selector. This results 25 neural models for data augmentation in the MultiWOZ experiments. On the contrary, NeuralWOZ only needs two neural models for data augmentation: Collector and Labeler. Another notable difference is that SimulatedChat does not generate multi-domain data in a natural way. The strategy of creating a model for each domain not only makes it difficult to transfer the knowledge to a new domain, but also makes it difficult to create multi-domain data. In SimulatedChat, the dialogue is created for each domain and then concatenated. Our model can properly reflect the information of all domains included in the goal instruction to generate synthetic dialogues, regardless of the number of domains. E Other Experiment Details The number of parameters of our models is 406M for Collector and 124M for Labeler, respectively. Both models are trained on two V100 GPUs with mixed precision floating point arithmetic. It takes about 4 (10 epochs) and 24 hours (30 epochs) for the training, respectively. We optimize hyperparameters of each model, learning rate {1e-5, 2e-5, 3e-5} and batch size {16, 32, 64}, based on greedy search. We set the maximum sequence length of Collector to 768 and the Labeler to 512. For the main experiments, we fix hyperparameter settings of TRADE (learning rate 1e-4 and batch size 32) and SUMBT (learning rate 5e-5 and batch size 4) same with previous works. We use the script of Campagna et al. (2020) for converting the TRADE’s data format to the SUMBT’s. For GPT2 (Radford et al., 2019) based model for the end2end task, we re-implement the model similar with SimpleTOD (Hosseini-Asl et al., 2020) but not using action. Thus, it generates dialogue context, dialogue state, database results, and system response in an autoregressive manner. We also use special tokens in the SimpleTOD (without special tokens for the action). We follow preprocessing procedure for the end2end task, including delexicalization suggested by (Budzianowski et al., 2018). We use 8 for batch size and 5e-5 for learning rate. Note that we also train our NeuralWOZ using 30% of training data and synthesize 5000 dialogues for the end2end experiments. However, we could not find detailed experiments setup of Mohapatra et al. (2020) including hyperparameter, the seed of each portion of training data, and evaluation, so it is not a fair comparison.
|
2021
|
287
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3718–3734 August 1–6, 2021. ©2021 Association for Computational Linguistics 3718 CDRNN: Discovering Complex Dynamics in Human Language Processing Cory Shain The Ohio State University [email protected] Abstract The human mind is a dynamical system, yet many analysis techniques used to study it are limited in their ability to capture the complex dynamics that may characterize mental processes. This study proposes the continuoustime deconvolutional regressive neural network (CDRNN), a deep neural extension of continuous-time deconvolutional regression (CDR, Shain and Schuler, 2021) that jointly captures time-varying, non-linear, and delayed influences of predictors (e.g. word surprisal) on the response (e.g. reading time). Despite this flexibility, CDRNN is interpretable and able to illuminate patterns in human cognition that are otherwise difficult to study. Behavioral and fMRI experiments reveal detailed and plausible estimates of human language processing dynamics that generalize better than CDR and other baselines, supporting a potential role for CDRNN in studying human language processing. 1 Introduction Central questions in psycholinguistics concern the mental processes involved in incremental human language understanding: which representations are computed when, by what mental algorithms (Frazier and Fodor, 1978; Just and Carpenter, 1980; Abney and Johnson, 1991; Tanenhaus et al., 1995; Almor, 1999; Gibson, 2000; Coltheart et al., 2001; Hale, 2001; Lewis and Vasishth, 2005; Levy, 2008, inter alia)? Such questions are often studied by caching out a theory of language processing in an experimental stimulus, collecting human responses, and fitting a regression model to test whether measures show the expected effects (e.g. Grodner and Gibson, 2005). Regression techniques have grown in sophistication, from ANOVA (e.g. Pickering and Branigan, 1998) to newer linear mixed-effects approaches (LME, Bates et al., 2015) that enable direct word-by-word analysis of effects in naturalistic human language processing (e.g. Demberg and Keller, 2008; Frank and Bod, 2011). However, these methods struggle to account for delayed effects. Because the human mind operates in real time and experiences computational bottlenecks of various kinds (Bouma and De Voogd, 1974; Just and Carpenter, 1980; Ehrlich and Rayner, 1981; Mollica and Piantadosi, 2017), delayed effects may be pervasive, and, if left uncontrolled, can yield misleading results (Shain and Schuler, 2018). Continuous-time deconvolutional regression (CDR) is a recently proposed technique to address delayed effects in measures of human cognition (Shain and Schuler, 2018, 2021). CDR fits parametric continuous-time impulse response functions (IRFs) that mediate between word features and response measures. An IRF maps the time elapsed between a stimulus and a response to a weight describing the expected influence of the stimulus on the response. CDR models the response as an IRF-weighted sum of preceding stimuli, thus directly accounting for effect latencies. Empirically, CDR reveals fine-grained processing dynamics and generalizes better to human reading and fMRI responses than established alternatives. However, CDR retains a number of simplifying assumptions (e.g. that the IRF is fixed over time) that may not hold of the human language processing system. Deep neural networks (DNNs), widely used in natural language processing (NLP), can relax these strict assumptions. Indeed, psycholinguistic regression analyses and NLP systems share a common structure: both fit a function from word features to some quantity of interest. However, psycholinguistic regression models face an additional constraint: they must be interpretable enough to allow researchers to study relationships between variables in the model. This requirement may be one reason why black box DNNs are not generally 3719 used to analyze psycholinguistic data, despite the tremendous gains DNNs have enabled in natural language tasks (Peters et al., 2018; Devlin et al., 2019; Radford et al., 2019; Brown et al., 2020, inter alia), in part by better approximating the complex dynamics of human cognition as encoded in natural language (Linzen et al., 2016; Gulordava et al., 2018; Tenney et al., 2019; Hewitt and Manning, 2019; Wilcox et al., 2019; Schrimpf et al., 2020). This study proposes an attempt to leverage the flexibility of DNNs for psycholinguistic data analysis. The continuous-time deconvolutional regressive neural network (CDRNN) is an extension of CDR that reimplements the impulse response function as a DNN describing the expected influence of preceding events (e.g. words) on future responses (e.g. reading times) as a function of their properties and timing. CDRNN retains the deconvolutional design of CDR while relaxing many of its simplifying assumptions (linearity, additivity, homosketasticity, stationarity, and context-independence, see Section 2), resulting in a highly flexible model. Nevertheless, CDRNN is interpretable and can shed light on the underlying data generating process. Results on reading and fMRI measures show substantial generalization improvements from CDRNN over baselines, along with detailed insights about the underlying dynamics that cannot easily be obtained from existing methods.1 2 Background Psycholinguists have been aware for decades that processing effects may lag behind the words that trigger them (Morton, 1964; Bouma and De Voogd, 1974; Rayner, 1977; Erlich and Rayner, 1983; Mitchell, 1984; Rayner, 1998; Vasishth and Lewis, 2006; Smith and Levy, 2013), possibly because cognitive “buffers” may exist to allow higher-level information processing to catch up with the input (Bouma and De Voogd, 1974; Baddeley et al., 1975; Just and Carpenter, 1980; Ehrlich and Rayner, 1981; Mollica and Piantadosi, 2017). They have also recognized the potential for non-linear, interactive, and/or time-varying relationships between word features and language processing (Smith and Levy, 2013; Baayen et al., 2017, 2018). No prior regression method can jointly address these 1Because of page constraints, additional replication details and synthetic results are provided in an external supplement, available here: https://osf.io/z89vn/. concerns in non-uniform time series (e.g. words with variable duration) like naturalistic psycholinguistic experiments. Discrete-time methods (e.g. lagged/spillover regression, Sims, 1971; Erlich and Rayner, 1983; Mitchell, 1984) ignore potentially meaningful variation in event duration, even if some (e.g. generalized additive models, or GAMs, Hastie and Tibshirani, 1986; Wood, 2006) permit non-linear and non-stationary (time-varying) feature interactions (Baayen et al., 2017). CDR (Shain and Schuler, 2018, 2021) addresses this limitation by fitting continuous-time IRFs, but assumes that the IRF is stationary (time invariant), that features scale linearly and combine additively, and that the response variance is constant (homoskedastic). By implementing the IRF as a time-varying neural network, CDRNN relaxes all of these assumptions, incorporating the featural flexibility of GAMs while retaining the temporal flexibility of CDR. Previous studies have investigated latency and non-linearity in human sentence processing. For example, Smith and Levy (2013) attach theoretical significance to the functional form of the relationship between word surprisal and processing cost, using GAMs to show that this relationship is linear and arguing on this basis that language processing is highly incremental. This claim is under active debate (Brothers and Kuperberg, 2021), underlining the importance of methods that can investigate questions of functional form. Smith and Levy (2013) also investigate the timecourse of surprisal effects using spillover and find a more delayed surprisal response in self-paced reading (SPR) than in eye-tracking. Shain and Schuler (2021) support the latter finding using CDR, and in addition show evidence of strong inertia effects in SPR, such that participants who have been reading quickly in the recent past also read more quickly now. However, this outcome may be an artifact of the stationarity assumption: CDR may be exploiting its estimates of rate effects in order to capture broad non-linear negative trends (e.g. task adaptation, Prasad and Linzen, 2019) in a stationary model. Similarly, the generally null word frequency estimates reported in Shain and Schuler (2021) may be due in part to the assumption of additive effects: word frequency and surprisal are related, and they may coordinate interactively to determine processing costs (Norris, 2006). Thus, in general, prior findings on the timecourse and functional form of effects in human sentence processing may be influenced by method3720 h(0) = x t RNN h τ IRF convolution P s RNN + in ∼ g(τ) x 1 RNN + in ∼ g(τ) x 1 RNN + in ∼ g(τ) x 1 RNN + in ∼ g(τ) x 1 Figure 1: CDRNN model. Subscripts omitted to reduce clutter. The IRF g(τ) at an event computes the expected contribution of each feature of the event vector h(0) to each element of the parameter vector s of the predictive distribution for a particular response value. The first layer of the IRF depends non-linearly on the properties of the event via hin and (optionally) on context via hRNN, which requires the recurrent connections in gray. Elements with random effects have dotted outlines. For variable definitions, see Appendix A. ological limitations: the GAM models of Smith and Levy (2013) ignore variable event duration, the CDR models of Shain and Schuler (2021) ignore non-linearity, and both approaches assume stationarity, context-independence, constant variance, and additive effects. By jointly relaxing these potentially problematic assumptions, CDRNN stands to support more reliable conclusions about human language comprehension, while also possibly enabling new insights into cognitive dynamics. 3 Model 3.1 Architecture This section presents a high-level description of the model design (for formal definition, see Appendix A). The CDRNN architecture is represented schematically in Figure 1. The primary goal of estimation is to identify the deep neural IRF g(τ) (top) that computes the influence of a preceding event on the predictive distribution over a subsequent response as a function of their distance in time τ. As shown, the IRF is a feedforward projection of τ into a matrix that defines a weighted sum over the values of input vector x, which is concatenated with a bias to capture general effects of stimulus timing (rate). This matrix multiplication determines the contribution of the stimulus event to the parameters of the predictive distribution (e.g. the mean and variance parameters of a Gaussian predictive distribution). Defining the IRF as a function of τ ensures that the model has a continuous-time definition. To capture non-linear effects of stimulus features, the IRF projection is itself parameterized by a projection of a hidden state h. The dependence on h permits non-linear influences of the properties of the stimulus sequence on the IRF itself. To generate h, the predictors x are concatenated with their timestamps t and submitted to the model as input. Inputs are cast to a hidden state for each preceding event as the sum of three quantities: a feedforward projection hin of each input, a forwarddirectional RNN projection hRNN of the events up to and including each input, and random effects hZ containing offsets for the relevant random effects level(s) (e.g. for each participant in an experiment). In this study, the recurrent component is treated as optional (gray arrows). Without the RNN, the model is non-stationary (via input t) but cannot capture contextual influences on the IRF. The summation over IRF outputs at the top of the figure ensures that the model is deconvolutional: each preceding input contributes to the response in some proportion, with that proportion determined by the features, context, and relative timing of that input. Because the IRF depends on a deep neural projection of the current stimulus as well as (optionally) the entire sequence of preceding stimuli, it implicitly estimates all interactions between these variables in governing the response. Predictors may thus coordinate in a non-linear, non-additive, and time-varying manner. The CDRNN IRF describes the influence over time of predictors on all parameters of the predictive distribution (in these experiments, the mean and variance parameters of a Gaussian predictive distribution). Such a design (i.e. modeling dependencies on the predictors of all parameters of the predictive distribution) has previously been termed distributional regression (B¨urkner, 2018). Despite their flexibility and task performance (Section 5), CDRNN models used in this study have few parameters (Table A1) by current deep learning standards because they are relatively shallow and small (Supplement S1). 3.2 Objective and Regularization Given (1) an input configuration C containing predictors X, input timestamps t, and response timestamps t′, (2) CDRNN parameter vector w, (3) output distribution p, (4) random effects vector z, and (5) response vector y, the model uses gradient de3721 scent to minimize the following objective: L (y | C; w, z) def = −log p (y | C; w, z) + (1) λz||z||2 2 + Lreg In addition to random effects shrinkage governed by λz and any arbitrary additional regularization penalties Lreg (see Supplement S1), models are regularized using dropout (Srivastava et al., 2014) with drop rate dh at the outputs of all feedforward hidden layers. Random effects are also dropped at rate dz, which is intended to encourage the model to find population-level estimates that accurately reflect central tendency. Finally, the recurrent contribution to the CDRNN hidden state (hRNN above) is dropped at rate dr, which is intended to encourage accurate IRF estimation even when context is unavailable. 3.3 Effect Estimation Because it is a DNN, CDRNN lacks parameters that selectively describe the size and shape of the response to a specific predictor (unlike CDR), and indeed individual parameters (e.g. individual biases or connection weights) are not readily interpretable. Thus, from a scientific perspective, the quantity of general interest is not a distribution over parameters, but rather over the effect of a predictor on the response. The current study proposes to accomplish this using perturbation analysis (e.g. Ribeiro et al., 2016; Petsiuk et al., 2018), manipulating the input configuration and quantifying the influence of this manipulation on the predicted response.2 For example, to obtain an estimate of rate effects (i.e. the base response or “deconvolutional intercept,” see Shain and Schuler, 2021), a reference stimulus can be constructed, and the response to it can be queried at each timepoint over some interval of interest. To obtain CDR-like estimates of predictor-wise IRFs, the reference stimulus can be increased by 1 in the predictor dimension of interest (e.g. word surprisal) and requeried, taking the difference between the obtained response and the reference response to reveal the influence of an extra unit of the predictor.3 This study uses the 2Perturbation analyses is one of a growing suite of tools for black box interpretation. It is used here because it straightforwardly links properties of the input to changes in the estimated response, providing a highly general method for querying aspects of the the non-linear, non-stationary, non-additive IRF defined by the CDRNN equations. 3Note that 1 is used here to maintain comparability of effect estimates to those generated by methods that assume training set mean of x and t as a reference, since this represents the response of the system to an average stimulus. The model also supports arbitrary additional kinds of queries, including of the curvature of an effect in the IRF over time and of the interaction between two effects at a point in time. Indeed, the IRF can be queried with respect to any combination of values for predictors, t, and τ, yielding an open-ended space of queries that can be constructed as needed by the researcher. Because the estimates of interest all derive from the model’s predictive distribution, uncertainty about them can be measured with Monte Carlo techniques as long as training involves a stochastic component, such as dropout (Srivastava et al., 2014) or batch normalization (Ioffe and Szegedy, 2015). This study estimates uncertainty using Monte Carlo dropout (Gal and Ghahramani, 2016), which recasts training neural networks with dropout as variational Bayesian approximation of deep Gaussian process models (Damianou and Lawrence, 2013). At inference time, an empirical distribution over responses to an input is constructed by resampling the model (i.e. sampling different dropout masks).4 As argued by Shain and Schuler (2021) for CDR, in addition to intervals-based tests, common hypothesis tests (e.g. for the presence of an effect) can be performed in a CDRNN framework via bootstrap model comparison on held out data (e.g. of models with and without the effect of interest). 4 Methods Following Shain and Schuler (2021), CDRNN is applied to naturalistic human language processing data from three experimental modalities: the Natural Stories self-paced reading corpus (∼1M instances, Futrell et al., 2020), the Dundee eye-tracking corpus (∼200K instances, Kennedy linearity of effects (especially CDR), but that 1 has no special meaning in the non-linear setting of CDRNN modeling, and effects can be queried at any offset from any reference. Results here show that deflections move relatively smoothly away from the reference, even at smaller steps than 1, and that IRFs queried at 1 are similar to those obtained from (linear) CDR, indicating that this method of effect estimation is reliable. Note finally that because predictors are underlyingly rescaled by their training set standard deviations (though plotted at the original scale for clarity), 1 here corresponds to 1 standard unit, as was the case with the CDR estimates discussed in Shain and Schuler (2021). 4Initial experiments also explored uncertainty quantification by implemententing CDRNN as a variational Bayesian DNN. Compared to the methods advocated here, the variational approach was more prone to instability, achieved worse fit, and yielded implausibly narrow credible intervals. 3722 Natural Stories (SPR) Dundee ms log-ms ms log-ms Model Train Dev Test Train Dev Test Train Dev Test Train Dev Test LME 19980† 20471† 20230† 0.0789† 0.0807† 0.0803† 13112† 14162† 14024† 0.1507† 0.1532† 0.1526† GAM 19873 20349 20109 0.0784 0.0802 0.0799 12882 13948 13771 0.1491 0.1518 0.1508 CDR 18118 18373 18212 0.0646 0.0652 0.0654 13073 14106 13960 0.1505 0.1539 0.1520 CDRNN-FF 18338 18677 18401 0.0644 0.0651 0.0650 12760 13863 13678 0.1479 0.1507 0.1498 CDRNN-RNN 18217 18624 18430 0.0636 0.0647 0.0642 12791 13897 13717 0.1476 0.1507 0.1495 Table 1: Reading. Mean squared error by model. Baselines as reported in Shain and Schuler (2021). Daggers (†) indicate convergence failures. et al., 2003), and the Natural Stories fMRI corpus (∼200K instances, Shain et al., 2020), using the train/dev/test splits for these corpora defined in Shain and Schuler (2021). Further details about datasets and preprocessing are given in Supplement S2. For reading data, CDRNN is compared to CDR as well as lagged LME and GAM baselines equipped with four spillover positions for each predictor (values from the current word, plus three preceding words), since LME and GAM are well established analysis methods in psycholinguistics (e.g. Baayen et al., 2007; Demberg and Keller, 2008; Frank and Bod, 2011; Smith and Levy, 2013; Baayen et al., 2017; Goodkind and Bicknell, 2018, inter alia). Because the distribution of reading times is heavy-tailed (Frank et al., 2013), following Shain and Schuler (2021) models are fitted to both raw and log-transformed reading times. For fMRI data, CDRNN is compared to CDR as well as four existing techniques for analyzing naturalistic fMRI data: pre-convolution with the canonical hemodynamic response function (HRF, Brennan et al., 2012; Willems et al., 2015; Henderson et al., 2015, 2016; Lopopolo et al., 2017), linear interpolation (Shain and Schuler, 2021), binning (Wehbe et al., 2020), and Lanczos interpolation (Huth et al., 2016). Statistical model comparisons use paired permutation tests of test set error (Demˇsar, 2006). Models use predictors established by prior psycholinguistic research (e.g. Rayner, 1998; Demberg and Keller, 2008; van Schijndel and Schuler, 2013; Staub, 2015; Shain and Schuler, 2018, inter alia): unigram and 5-gram surprisal, word length (reading only), saccade length (eye-tracking only), and previous was fixated (eye-tracking only). Predictor definitions are given in Appendix C. The deconvolutional intercept term rate (Shain and Schuler, 2018, 2021), an estimate of the general influence of observing a stimulus at a point in time, independently of its properties, is implicit in CDRNN (unlike CDR) and is therefore reported in all results. Reading models include random effects by subject, while fMRI models include random effects by subject and by functional region of interest (fROI). Unlike LME, where random effects capture linear differences in effect size between e.g. subjects, random effects in CDRNN capture differences in overall dynamics between subjects, including differences in size, IRF shape, functional form (e.g. linearity), contextual influences on the IRF, and interactions with other effects. Two CDRNN variants are considered in all experiments: the full model (CDRNN-RNN) containing an RNN over the predictor sequence, and a feedforward only model (CDRNN-FF) with the RNN ablated (gray arrows removed in Figure 1). This manipulation is of interest because CDRNN-FF is both more parsimonious (fewer parameters) and faster to train, and may therefore be preferred in the absence of prior expectation that the IRF is sensitive to context. All plots show means and 95% credible intervals. Code and documentation are available at https://github.com/coryshain/cdr. 5 Results Since CDRNN is designed for scientific modeling, the principal output of interest is the IRF itself and the light it might shed on questions of cognitive dynamics, rather than on performance in some task (predicting reading latencies or fMRI measures are not widely targeted engineering goals). However, predictive performance can help establish the trustworthiness of the IRF estimates. To this end, as a sanity check, this section first evaluates predictive performance on human data relative to existing regression techniques. While results may resemble “bake-off” comparisons familiar from machine learning (and indeed CDRNN does outperform all baselines), their primary purpose is to establish that the CDRNN estimates are trustworthy, since they describe the phenomenon of interest in a way that generalizes accurately to an unseen sample. Baseline models, including CDR, are as reported 3723 Model Train Expl Test Canonical HRF 11.3548† 11.8263† 11.5661† Linearly interpolated 11.4236† 11.9888† 11.6654† Averaged 11.3478† 11.9280† 11.6090† Lanczos interpolated 11.3536† 11.9059† 11.5871† CDR 11.2774 11.6928 11.5369 CDRNN-FF 10.5648 11.3602 11.3042 CDRNN-RNN 10.8736 11.5631 11.3914 Table 2: fMRI. Mean squared error by model. Baselines as reported in Shain and Schuler (2021). Daggers (†) indicate convergence failures. in Shain and Schuler (2021).5 5.1 Model Validation: Baseline Comparisons Table 1 gives mean squared error by dataset of CDRNN vs. baseline models on reading times from both Natural Stories and Dundee. Both versions of CDRNN outperform all baselines on the dev partition of all datasets except for raw (ms) latencies in Natural Stories (SPR), where CDRNN is edged out by CDR6 but still substantially outperforms the non-CDR baselines. Nonetheless, results indicate that CDRNN estimates of Natural Stories (ms) are similarly reliable to those of CDR, and, as discussed in Section 5.2, CDRNN largely replicates the CDR estimates on Natural Stories while offering advantages for analysis. Although CDR struggles against GAM baselines on Dundee, CDRNN has closed the gap. This is noteworthy in light of speculation in Shain and Schuler (2021) that CDR’s poorer performance on Dundee might be due in part to non-linear effects, which GAM can estimate but CDR cannot. CDRNN performance supports this conjecture: once the model can account for non-linearities, it overtakes GAMs. Results from fMRI are shown in Table 2, where both CDRNN variants yield substantial improvements to training, dev, and test set error. These results indicate that the relaxed assumptions afforded by CDRNN are beneficial for describing the fMRI response, which is known to saturate over time (Friston et al., 2000; Wager et al., 2005; Vazquez et al., 2006; Lindquist et al., 2009). Following Shain and Schuler (2021), model error is statistically compared using a paired permu5For all datasets, the CDR baseline used here is the variant that was deployed on the test set in Shain and Schuler (2021). 6Note that a major advantage of CDRNN is its ability to model dynamics in response variance, which are not reflected in squared error. For example, although CDRNN-FF achieves worse test set error than CDR on the Natural Stories (ms) task, it affords a 31,040 point log likelihood improvement. CDRNN FF RNN Baseline Modality p p LME Reading 0.0001*** 0.0001*** GAM Reading 0.0001*** 0.0001*** Canonical HRF fMRI 0.0001*** 0.0001*** Interpolated fMRI 0.0001*** 0.0001*** Averaged fMRI 0.0001*** 0.0001*** Lanczos fMRI 0.0001*** 0.0001*** CDR Both 0.0001*** 0.0001*** CDRNN-FF Both — 0.0048** Table 3: Permutation test of overall test set performance improvement from CDRNN variants over each baseline. tation test that pools across all datasets covered by a given baseline (reading data for LME and GAM, fMRI data for canonical HRF, linearly interpolated, averaged, and Lanczos interpolated, and both for CDR).7 Results are given in Table 3. As shown, both variants of CDRNN significantly improve over all baselines, and CDRNN-RNN significantly improves over CDRNN-FF. Notwithstanding, CDRNN-FF may be preferred in applications: simpler, faster to train, better at recovering synthetic models (Supplement S3), more reliable in noisy domains like fMRI, and close in performance to CDRNN-RNN. Results overall support the reliability of patterns revealed by CDRNN’s estimated IRF, which is now used to explore and visualize sentence processing dynamics. 5.2 Effect Latencies in CDRNN vs. CDR CDR-like IRF estimates can be obtained by increasing a predictor by 1 (standard deviation) relative to the reference and observing the change in the response over time. Visualizations using this approach are presented in Figure 2 alongside CDR estimates from Shain and Schuler (2021). In general, CDRNN finds similar patterns to CDR. This suggests both (1) that CDRNN is capable of recovering estimates from a preceding state-of-the-art deconvolutional model for these domains, and (2) that CDR estimates in these domains are not driven by artifacts introduced by its simplifying assumptions, since a model that lacks those assumptions and has a qualitatively different architecture largely recovers them. Nonetheless there are differences. For example, Dundee estimates decay more quickly over time in CDRNN than in CDR, indicating an even less pronounced influence of temporal diffusion in 7The comparison rescales each pair of error vectors by their joint standard deviation in order to enable comparability across datasets with different error variances. 3724 CDR NatStor (SPR) log-ms CDRNN-FF CDRNN-RNN Dundee log-ms NatStor (fMRI) BOLD Delay (s) • rate saccade length previous was fixated word length/sound power unigram surprisal 5-gram surprisal Figure 2: CDRNN-estimated IRFs across datasets, with CDR estimates from Shain and Schuler (2021) for reference. Sound power omitted from CDRNN fMRI models (see Appendix C for justification). eye-tracking than CDR had previously suggested. Estimates from CDRNN-FF and CDRNN-RNN roughly agree, except that CDRNN-RNN estimates for fMRI are more attenuated. CDR shows little uncertainty in the fMRI domain despite its inherent noise (Shain et al., 2020), while CDRNN more plausibly shows more uncertainty in its estimates for the noisier fMRI data. As noted in Section 2, Shain and Schuler (2021) report negative rate effects in reading — i.e., a local decrease in subsequent reading time at each word, especially in SPR. This was interpreted as an inertia effect (faster recent reading engenders faster current reading), but it might also be an artifact of non-linear decreases in latency over time (due to task habituation, e.g. Baayen et al., 2017; Harrington Stack et al., 2018; Prasad and Linzen, 2019) that CDR cannot model. CDRNN estimates nonetheless thus support the prior interpretation of rate effects as inertia, at least in SPR: a model that can flexibly adapt to non-linear habituation trends finds SPR rate estimates that are similar in shape and magnitude to those estimated by CDR. In addition, CDRNN finds a slower response to word surprisal in self-paced reading than in eye-tracking. This result converges with worddiscretized timecourses reported in Smith and Levy (2013), who find more extensive spillover of surprisal effects in SPR than in eye-tracking. Results thus reveal important hidden dynamics in the reading response (inertia effects), continuous-time delays in processing effects, and influences of modality the continuous dynamics of sentence processing, all of which are difficult to estimate using existing regression techniques. Greater response latency and more pronounced inertia effects in self-paced reading may be due to the fact that a gross motor task (paging via button presses) is overlaid on the sentence comprehension task. While the motor task is not generally of interest to psycholinguistic theories, controlling for its effects is crucial when using self-paced reading to study sentence comprehension (Mitchell, 1984). 5.3 Linearity of Surprisal Effects CDRNN also allows the analyst to explore other aspects of the IRF, such as functional curvature at a point in time. For example, in the context of reading, Smith and Levy (2013) argue for a linear increase in processing cost as a function of word surprisal. The present study allows this claim to be assessed across modalities by checking the curva3725 NatStor (SPR) Instantaneous Dundee NatStor (fMRI) Over Time Figure 3: CDRNN-FF-estimated functional curvature of the 5-gram surprisal response. In 3D plots, 95% credible intervals shown as vertical gray bars. ture of the 5-gram surprisal response (in raw ms) at a timepoint of interest (0ms for reading and ∼5s for fMRI). As shown in the top row of Figure 3, reading estimates are consistent with a linear response (the credible interval contains a straight line), as predicted, but are highly non-linear in fMRI, with a rapid peak above the mean (zero-crossing) followed by a sharp dip and plateau, and even an estimated increased response at values below the mean (though estimates at the extremes have high uncertainty). This may be due in part to ceiling effects: blood oxygen levels measured by fMRI are bounded, but reading times are not. While this is again a property of experimental modality rather than sentence comprehension itself, understanding such influences is important for drawing scientific conclusions from experimental data. For example, due to the possibility of saturation, fMRI may not be an ideal modality for testing scientific claims about the functional form of effects, and the linearity assumptions of e.g. CDR and LME may be particularly constraining. The curvature of effects can also be queried over time. If an effect is temporally diffuse but linear, its curvature should be roughly linear at any delay of interest. The second row of Figure 3 shows visualizations to this effect. These plots in fact subsume the kinds of univariate plots shown above: univariate IRFs to 5-gram surprisal like those plotted in Figure 2 are simply slices taken at a predictor value (1 sample standard deviation above the mean), whereas curvature estimates in the first row of Figure 3 are simply slices taken at a time value (0s for reading and 5s for fMRI). Plots are consistent with the linearity hypothesis for reading, but again show strong non-linearities in the fMRI domain that are consistent with saturation effects Delay (s) • rate sound power unigram surprisal 5-gram surprisal PCFG surprisal Figure 4: Effect interactions in a CDRNN-FF replication of Shain et al. (2020). 95% credible intervals shown as vertical gray bars. as discussed above. 5.4 Effect Interactions In addition to exploring multivariate relationships of a predictor with time, relationships between predictors can also be studied. Such relationships constitute “interactions” in a CDRNN model, though they are not constrained (cf. interactions in linear models) to be strictly multiplicative — indeed, a major advantage of CDRNN is that interactions come “for free”, along with estimates of their functional form. To explore effect interactions, a CDRNN-FF version of the full model in Shain et al. (2020) is fitted to the fMRI dataset. The model contains more predictors to explore than models considered above, including surprisal computed from a probabilistic context-free grammar (PCFG surprisal, see Appendix C for details). Univariate IRFs are shown in the top left panel of Figure 4, and pairwise interaction surfaces at a delay of 5s (near the peak response) are shown in the remaining panels. Plots show that the response at any value of the other predictors is roughly flat as a function of sound power (i.e. signal power of the auditory stimulus, middle row). This accords with prior arguments that the cortical language system, whose activity is measured here, does not strongly register low-level perceptual effects (Fedorenko et al., 2010; Braze et al., 2011). 3726 NatStor (SPR) Dundee NatStor (fMRI) Delay (s) • rate saccade length previous was fixated word length unigram surprisal 5-gram surprisal Figure 5: CDRNN-FF-estimated IRFs of the variance of the response by dataset. The estimate for unigram surprisal (middle left) shows an unexpected non-linearity: although activity increases with higher surprisal (lower frequency words), it also increases at lower surprisal (higher frequency words), suggesting the existence of high frequency items that nonetheless engender a large response. The interaction between PCFG surprisal and unigram surprisal possibly sheds light on this outcome, since it shows a sharper increase in the PCFG surprisal response in higher frequency (lower unigram surprisal) regions. This may be because the most frequent words in English tend to be function words that play an outsized role in syntactic structure building (e.g. prepositional phrase attachment decisions). In addition, 5-gram surprisal interacts with PCFG surprisal, showing a non-linear increase in response for words that are high on both measures. This is consistent with a unitary predictive mechanism that experiences strong error signals when both string-level (5-gram) and structural (PCFG) cues are poor. All these interactions should be interpreted with caution, since the uncertainty interval covers much weaker degrees of interaction. 5.5 IRFs of the Response Variance As discussed in Section 3, CDRNN implements distributional regression and thus also contains an IRF describing the influence of predictors on the variance of the predictive distribution as a function of time. IRFs of the variance can be visualized identically to IRFs of the mean. For example, Figure 5 shows the estimated change in the standard deviation of the predictive distribution over time from observing a stimulus.8 Estimates show stimulus-dependent changes 8Because standard deviation is a bounded variable and the IRF applies before the constraint function (softplus), the relationship between the standard deviation and the y axis of the plots is not straightforward. Estimates nonetheless clearly indicate the shape and relative contribution to the response in variance across datasets whose shapes are not straightforwardly related to that of the IRFs of the mean (Figure 2). For example, both reading datasets (left and center) generally show mean and standard deviation traveling together, with increases in the mean corresponding to increases in standard deviation. In Dundee, the shapes of these changes resemble each other strongly, whereas in Natural Stories the IRFs of the standard deviation (especially rate) differ substantially from the IRFs of the mean. By contrast, in fMRI (right), the IRFs of the standard deviation look roughly like inverted HRFs (especially for rate and 5-gram surprisal), indicating that BOLD variance tends to decrease with larger values of the predictors. While detailed interpretation of these patterns is left to future work, these results demonstrate the utility of CDRNN for analyzing a range of links between predictors and response that are otherwise difficult to study. 6 Conclusion This study proposed and evaluated CDRNN, a deep neural extension of continuous-time deconvolutional regression that relaxes implausible simplifying assumptions made by widely used regression techniques in psycholinguistics. In so doing, CDRNN provides detailed estimates of human language processing dynamics that are difficult to obtain using other measures. Results showed plausible estimates from human data that generalize better than alternatives and can illuminate hitherto understudied properties of the human sentence processing response. This outcome suggests that CDRNN may play a valuable role in analyzing human experimental data. References Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Man´e, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Vi´egas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. 2015. TensorFlow: Large-Scale Machine Learning on Heterogeneous Distributed Systems. variance of the stimulus features. 3727 Steven P Abney and Mark Johnson. 1991. Memory Requirements and Local Ambiguities of Parsing Strategies. J.\ Psycholinguistic Research, 20(3):233–250. Amit Almor. 1999. Noun-Phrase Anaphora and Focus: The Informational Load Hypothesis. Psychological Review, 106(4):748–765. Harald Baayen, Shravan Vasishth, Reinhold Kliegl, and Douglas Bates. 2017. The cave of shadows: Addressing the human factor with generalized additive mixed models. Journal of Memory and Language, 94(Supplement C):206–234. R Harald Baayen, Doug J Davidson, and Douglas M Bates. 2007. Mixed effects modelling with crossed random effects for subjects and items. manuscript. R Harald Baayen, Jacolien van Rij, Cecile de Cat, and Simon Wood. 2018. Autocorrelated errors in experimental data in the language sciences: Some solutions offered by Generalized Additive Mixed Models. In Dirk Speelman, Kris Heylen, and Dirk Geeraerts, editors, Mixed Effects Regression Models in Linguistics. Springer, Berlin. Alan D Baddeley, Neil Thomson, and Mary Buchanan. 1975. Word length and the structure of short term memory. Journal of Verbal Learning and Verbal Behavior, 15(6):575–589. Douglas Bates, Martin M¨achler, Ben Bolker, and Steve Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1):1– 48. H Bouma and A H De Voogd. 1974. On the control of eye saccades in reading. Vision Research, 14(4):273–284. David Braze, W Einar Mencl, Whitney Tabor, Kenneth R Pugh, R Todd Constable, Robert K Fulbright, James S Magnuson, Julie A Van Dyke, and Donald P Shankweiler. 2011. Unification of sentence processing via ear and eye: An fMRI study. cortex, 47(4):416–431. Jonathan Brennan, Yuval Nir, Uri Hasson, Rafael Malach, David J Heeger, and Liina Pylkk¨anen. 2012. Syntactic structure building in the anterior temporal lobe during natural story listening. Brain and Language, 120(2):163–173. Trevor Brothers and Gina R Kuperberg. 2021. Word predictability effects are linear, not logarithmic: Implications for probabilistic models of sentence comprehension. Journal of Memory and Language, 116:104174. Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, and Dario Amodei. 2020. Language models are few-shot learners. In Proceedings of Advances in Neural Information Processing Systems 33. Paul-Christian B¨urkner. 2018. Advanced Bayesian Multilevel Modeling with the R Package brms. R Journal, 10(1). Max Coltheart, Kathleen Rastle, Conrad Perry, Robyn Langdon, and Johannes Ziegler. 2001. DRC: a dual route cascaded model of visual word recognition and reading aloud. Psychological review, 108(1):204. Andreas Damianou and Neil D Lawrence. 2013. Deep gaussian processes. In Artificial intelligence and statistics, pages 207–215. PMLR. Vera Demberg and Frank Keller. 2008. Data from eyetracking corpora as evidence for theories of syntactic processing complexity. Cognition, 109(2):193–210. Janez Demˇsar. 2006. Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research, 7(Jan):1–30. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. NAACL19. Susan F Ehrlich and Keith Rayner. 1981. Contextual effects on word perception and eye movements during reading. Journal of verbal learning and verbal behavior, 20(6):641–655. Kate Erlich and Keith Rayner. 1983. Pronoun assignment and semantic integration during reading: Eye movements and immediacy of processing. Journal of Verbal Learning & Verbal Behavior, 22:75–87. Evelina Fedorenko, Po-Jang Hsieh, Alfonso NietoCasta˜n´on, Susan Whitfield-Gabrieli, and Nancy Kanwisher. 2010. New method for fMRI investigations of language: defining ROIs functionally in individual subjects. Journal of Neurophysiology, 104(2):1177–1194. Victoria Fossum and Roger Levy. 2012. Sequential vs. Hierarchical Syntactic Models of Human Incremental Sentence Processing. In Proceedings of {{CMCL}} 2012. Association for Computational Linguistics. Stefan Frank and Rens Bod. 2011. Insensitivity of the human sentence-processing system to hierarchical structure. Psychological Science. Stefan L Frank, Irene Fernandez Monsalve, Robin L Thompson, and Gabriella Vigliocco. 2013. Reading time data for evaluating broad-coverage models of English sentence processing. Behavior Research Methods, 45(4):1182–1190. 3728 Lyn Frazier and Jerry D Fodor. 1978. The sausage machine: a new two-stage parsing model. Cognition, 6:291–325. Karl J Friston, Andrea Mechelli, Robert Turner, and Cathy J Price. 2000. Nonlinear responses in fMRI: The Balloon model, Volterra kernels, and other hemodynamics. NeuroImage, 12(4):466–477. Richard Futrell, Edward Gibson, Harry J Tily, Idan Blank, Anastasia Vishnevetsky, Steven T Piantadosi, and Evelina Fedorenko. 2020. The Natural Stories corpus: a reading-time corpus of English texts containing rare syntactic constructions. Language Resources and Evaluation, pages 1–15. Yarin Gal and Zoubin Ghahramani. 2016. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In international conference on machine learning, pages 1050–1059. PMLR. Edward Gibson. 2000. The Dependency Locality Theory: A distance-based theory of linguistic complexity. In Alec Marantz, Yasushi Miyashita, and Wayne O’Neil, editors, Image, language, brain, pages 95– 106. MIT Press, Cambridge. Adam Goodkind and Klinton Bicknell. 2018. Predictive power of word surprisal for reading times is a linear function of language model quality. In Proceedings of the 8th Workshop on Cognitive Modeling and Computational Linguistics (CMCL 2018), pages 10–18. David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2007. English Gigaword Third Edition LDC2007T07. Daniel J Grodner and Edward Gibson. 2005. Consequences of the serial nature of linguistic input. Cognitive Science, 29:261–291. Kristina Gulordava, Piotr Bojanowski, ´Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless Green Recurrent Networks Dream Hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195–1205. John Hale. 2001. A Probabilistic Earley Parser as a Psycholinguistic Model. In Proceedings of the second meeting of the North American chapter of the Association for Computational Linguistics, pages 159– 166, Pittsburgh, PA. Caoimhe M Harrington Stack, Ariel N James, and Duane G Watson. 2018. A failure to replicate rapid syntactic adaptation in comprehension. Memory & cognition, 46(6):864–877. Trevor Hastie and Robert Tibshirani. 1986. Generalized additive models. Statist. Sci., 1(3):297–310. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778. Kenneth Heafield, Ivan Pouzyrevsky, Jonathan H Clark, and Philipp Koehn. 2013. Scalable modified KneserNey language model estimation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 690–696, Sofia, Bulgaria. John M Henderson, Wonil Choi, Matthew W Lowder, and Fernanda Ferreira. 2016. Language structure in the brain: A fixation-related fMRI study of syntactic surprisal in reading. Neuroimage, 132:293–300. John M Henderson, Wonil Choi, Steven G Luke, and Rutvik H Desai. 2015. Neural correlates of fixation duration in natural reading: evidence from fixationrelated fMRI. NeuroImage, 119:390–397. Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (GELUs). arXiv preprint arXiv:1606.08415. John Hewitt and Christopher D Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long Short-Term Memory. Neural Comput., 9(8):1735– 1780. Alexander G Huth, Wendy A de Heer, Thomas L Griffiths, Fr´ed´eric E Theunissen, and Jack L Gallant. 2016. Natural speech reveals the semantic maps that tile human cerebral cortex. Nature, 532(7600):453. Sergey Ioffe and Christian Szegedy. 2015. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. In International Conference on Machine Learning, pages 448–456. Marcel Adam Just and Patricia A Carpenter. 1980. A theory of reading: From eye fixations to comprehension. Psychological Review, 87(4):329–354. Alan Kennedy, James Pynte, and Robin Hill. 2003. The Dundee corpus. In Proceedings of the 12th European conference on eye movement. Diederik P Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6. Roger Levy. 2008. Expectation-based syntactic comprehension. Cognition, 106(3):1126–1177. 3729 Richard L Lewis and Shravan Vasishth. 2005. An activation-based model of sentence processing as skilled memory retrieval. Cognitive Science, 29(3):375–419. Martin A Lindquist, Ji Meng Loh, Lauren Y Atlas, and Tor D Wager. 2009. Modeling the hemodynamic response function in fMRI: Efficiency, bias and mismodeling. NeuroImage, 45(1, Supplement 1):S187 – S198. Tal Linzen, Emmanuel Dupoux, and Yoav Goldberg. 2016. Assessing the ability of LSTMs to learn syntax-sensitive dependencies. Transactions of the Association for Computational Linguistics, 4:521– 535. Alessandro Lopopolo, Stefan L Frank, Antal den Bosch, and Roel M Willems. 2017. Using stochastic language models (SLM) to map lexical, syntactic, and phonological information processing in the brain. PloS one, 12(5):e0177794. Mitchell P Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: the Penn Treebank. Computational Linguistics, 19(2):313–330. Don C Mitchell. 1984. An evaluation of subject-paced reading tasks and other methods for investigating immediate processes in reading. New methods in reading comprehension research, pages 69–89. Francis Mollica and Steve Piantadosi. 2017. An incremental information-theoretic buffer supports sentence processing. In Proceedings of the 39th Annual Cognitive Science Society Meeting. John Morton. 1964. The effects of context upon speed of reading, eye movements and eye-voice span. Quarterly Journal of Experimental Psychology, 16(4):340–354. Luan Nguyen, Marten van Schijndel, and William Schuler. 2012. Accurate Unbounded Dependency Recovery using Generalized Categorial Grammars. In Proceedings of COLING 2012. Dennis Norris. 2006. The Bayesian Reader: Explaining word recognition as an optimal Bayesian decision process. Psychological review, 113(2):327. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. arXiv preprint arXiv:1802.05365. Vitali Petsiuk, Abir Das, and Kate Saenko. 2018. RISE: Randomized Input Sampling for Explanation of Black-box Models. In Proceedings of the British Machine Vision Conference (BMVC). Martin J Pickering and Holly P Branigan. 1998. The representation of verbs: Evidence from syntactic priming in language production. Journal of Memory and language, 39(4):633–651. Boris T Polyak and Anatoli B Juditsky. 1992. Acceleration of stochastic approximation by averaging. SIAM Journal on Control and Optimization, 30(4):838–855. Grusha Prasad and Tal Linzen. 2019. Rapid syntactic adaptation in self-paced reading: detectable, but requires many participants. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog, 1(8):9. Keith Rayner. 1977. Visual attention in reading: Eye movements reflect cognitive processes. Memory \& Cognition, 5(4):443–448. Keith Rayner. 1998. Eye Movements in Reading and Information Processing: 20 Years of Research. Psychological Bulletin, 124(3):372–422. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. ”Why should I trust you?” Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144. Marten van Schijndel, Andy Exley, and William Schuler. 2013. A model of language processing as hierarchic sequential prediction. Topics in Cognitive Science, 5(3):522–540. Marten van Schijndel and William Schuler. 2013. An Analysis of Frequency- and Memory-Based Processing Costs. In Proceedings of NAACL-HLT 2013. Association for Computational Linguistics. Marten van Schijndel and William Schuler. 2015. Hierarchic syntax improves reading time prediction. In Proceedings of NAACL-HLT 2015. Association for Computational Linguistics. Martin Schrimpf, Idan A Blank, Greta Tuckute, Carina Kauf, Eghbal A Hosseini, Nancy G Kanwisher, Joshua B Tenenbaum, and Evelina Fedorenko. 2020. Artificial Neural Networks Accurately Predict Language Processing in the Brain. BioRxiv. Cory Shain, Idan Blank, Marten van Schijndel, William Schuler, and Evelina Fedorenko. 2020. fMRI reveals language-specific predictive coding during naturalistic sentence comprehension. Neuropsychologia, 138. Cory Shain and William Schuler. 2018. Deconvolutional time series regression: A technique for modeling temporally diffuse effects. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Cory Shain and William Schuler. 2021. ContinuousTime Deconvolutional Regression for Psycholinguistic Modeling. Cognition. 3730 Christopher A Sims. 1971. Discrete approximations to continuous time distributed lags in econometrics. Econometrica: Journal of the Econometric Society, pages 545–563. Nathaniel J Smith and Roger Levy. 2013. The effect of word predictability on reading time is logarithmic. Cognition, 128:302–319. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958. Adrian Staub. 2015. The effect of lexical predictability on eye movements in reading: Critical review and theoretical interpretation. Language and Linguistics Compass, 9(8):311–327. Michael K Tanenhaus, Michael J Spivey-Knowlton, Kathleen M Eberhard, and Julie C E Sedivy. 1995. Integration of visual and linguistic information in spoken language comprehension. Science, 268:1632–1634. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. ACL19. Shravan Vasishth and Richard L Lewis. 2006. Argument-head distance and processing complexity: Explaining both locality and antilocality effects. Language, 82(4):767–794. Alberto L Vazquez, Eric R Cohen, Vikas Gulani, Luis Hernandez-Garcia, Ying Zheng, Gregory R Lee, Seong-Gi Kim, James B Grotberg, and Douglas C Noll. 2006. Vascular dynamics and BOLD fMRI: CBF level effects and analysis considerations. Neuroimage, 32(4):1642–1655. Tor D Wager, Alberto Vazquez, Luis Hernandez, and Douglas C Noll. 2005. Accounting for nonlinear BOLD effects in fMRI: parameter estimates and a model for prediction in rapid event-related studies. NeuroImage, 25(1):206–218. Leila Wehbe, Idan A Blank, Cory Shain, Richard Futrell, Roger Levy, Titus von der Malsburg, Nathaniel Smith, Edward Gibson, and Evelina Fedorenko. 2020. Incremental language comprehension difficulty predicts activity in the language network but not the multiple demand network. bioRxiv. Ethan Wilcox, Roger Levy, and Richard Futrell. 2019. Hierarchical Representation in Neural Language Models: Suppression and Recovery of Expectations. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 181–190. Roel M Willems, Stefan L Frank, Annabel D Nijhof, Peter Hagoort, and Antal den Bosch. 2015. Prediction during natural language comprehension. Cerebral Cortex, 26(6):2506–2516. Dataset CDR CDRNN-FF CDRNN-RNN Synth 662 7,330 17,058 NatStor (SPR) 21,845 22,546 40,408 Dundee 2,080 6,870 14,838 NatStor (fMRI) 331 13,834 26,058 Table A1: Number of trainable parameters by model and dataset. Simon N Wood. 2006. Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC, Boca Raton. A Mathematical Definition This appendix formally defines the CDRNN model. CDRNN assumes the following quantities as input:9 • X ∈N: Number of predictor observations (e.g. word exposures) • Y ∈N: Number of response observations (e.g. fMRI scans) • Z ∈N: Number of random grouping factor levels (e.g. distinct participants) • K ∈N: Number of predictors • X ∈RX×K: Design matrix of X predictor observations of K dimensions each. • y ∈RY : Vector of Y response observations • Z ∈{0, 1}Y ×Z: Boolean matrix indicating random grouping factor levels associated with each response observation • t ∈RX: Vector of timestamps associated with each observation in X • t′ ∈RY : Vectors of timestamps associated with each observation in y • S ∈N: Number of parameters in predictive distribution (e.g. 2 for a normal distribution: mean and variance) For simplicity of exposition, X and y are assumed to contain data from a single time series (e.g. a single participant performing a single experiment). 9Throughout these definitions, vectors and matrices are notated in bold lowercase and uppercase, respectively (e.g. u, U). Objects with indexed names are designated using subscripts (e.g. vr). Vector and matrix indexing operations are notated using subscript square brackets, and slice operations are notated using ∗(e.g. X[∗,k] denotes the kth column of matrix X). Hadamard (pointwise) products are notated using ⊙. The notations 0 and 1 designate conformable column vectors of 0’s and 1’s, respectively. Superscripts are used for indexation and do not denote exponentiation. 3731 The definition below can be applied without loss of generality to data containing multiple time series by concatenating the output of the model as applied to multiple X, y pairs. X, y and their associated satellite data Z, t, t′ must be temporally sorted. Given these inputs, CDRNN estimates a latent impulse response function that relates timestamped predictors to all parameters of the assumed predictive distribution. For example, assuming a univariate normally distributed response, CDRNN learns an IRF with two output dimensions, one for the predictive mean, and one for the predictive variance. Regressing all parameters of the predictive distribution in this way has previously been called distributional regression (B¨urkner, 2018). CDRNN contains a recurrent neural network (RNN), neural projections that map inputs and RNN states to a hidden state for each preceding event, and neural projections that map the hidden states to predictions about (1) the influence of each event on the response (IRF) and (2) the parameter(s) of the error distribution (e.g. the variance of a Gaussian error). The definition assumes the following quantities: • Lin, LRNN, LIRF ∈N: Number of layers in the input projection, RNN, and IRF, respectively • Din(ℓ), DRNN(ℓ), Dh, DIRF(ℓ) ∈N: Number of output dimensions in the ℓth layer of the input projection, RNN, hidden state, and IRF, respectively The following values are deterministically assigned: • DIRF(LIRF) = S(K + 1) (the IRF generates a convolution weight for every predictor dimension, plus the timestamp, for each parameter of the predictive distribution) • Din(0) = K + 1 (input is predictors + time) • Din(Lin) = Dh In these definitions, integers x, y respectively refer to row indices of X, y. Let zy be the vector Z[y,∗] ⊤of random effects associated with the response at y. Let Wh,Z ∈RDh×Z, WIRF(1),Z ∈ R2DIRF(1)×Z, and Ws,Z ∈RS×Z be an embedding matrix for zy. Random effects offsets at response step y for the hidden state (hZ y), the weights and biases of the first layer of the IRF (wIRF(1),Z y , bIRF(1),Z y ), and the parameters of the predictive distribution (eZ y, i.e. random intercepts and variance parameters) are generated as follows: hZ y def = Wh,Zzy (2) " wIRF(1),Z y bIRF(1),Z y # def = WIRF(1),Zzy (3) sZ y def = Ws,Zzy (4) Following prior work in mixed effects models (Bates et al., 2015), to ensure that population-level estimates reliably encode central tendency, each output dimension of Wh,Z, WIRF(1),Z, and Ws,Z is constrained to have mean 0 across the levels of each random grouping factor (e.g. across participants in the study). The neural IRF is applied to a temporal offset τ representing the delay at which to query the response to an input (e.g. τ = 1 queries the response to an input 1s after the input occurred). The output of the neural IRF gℓ x,y(τ) ∈RDIRF(ℓ) applied to τ at layer ℓis defined as: g(1) x,y(τ) def = sIRF(1) wIRF(1) x,y τ + bIRF(1) x,y (5) g(ℓ) x,y(τ) def = sIRF(ℓ) WIRF(ℓ)g(ℓ−1) x,y (τ) + bIRF(ℓ) , (6) ℓ> 1 wIRF(1) x,y def = wIRF(1) + wIRF(1),Z y + WIRF(1) ∆ hx,y (7) bIRF(1) x,y def = bIRF(1) + bIRF(1),Z y + BIRF(1) ∆ hx,y (8) WIRF(ℓ) x,y , bIRF(ℓ) x,y , and sIRF(ℓ) are respectively the ℓth IRF layer’s weight matrix at predictor timestep x and response timestep y, bias vector at time x, y, and squashing function, and g(0) x,y(τ) = τ. wIRF(1), bIRF(1) are respectively globally applied initial weight and bias vectors for the first layer of the IRF, which transforms scalar τ, each of which is shifted by its corresponding random effects. WIRF(1) ∆ , BIRF(1) ∆ are respectively weight matrices used to compute additive modifications to WIRF(1) from CDRNN hidden state hx,y, similar in spirit to a residual network (He et al., 2016). Non-initial IRF layers are treated as stationary (i.e. their parameters are independent of x, y). The final output of the IRF is given by: gx,y(τ) def = reshape g(LIRF) x,y (τ), (S, K + 1) (9) 3732 The hidden state hx,y is computed as the squashed sum of several quantities: a global bias hbias, random effects hZ, a neural projection hin x,y of the inputs at x, y, and a neural projection hRNN x,y of the hidden state of an RNN over the sequence of predictors up to and including timestep x: hx,y def =sh hbias + hZ y + hin x,y + hRNN x,y (10) The IRF gx,y is therefore feature-dependent via the neural projection hin x,y of the input at x, y and context-dependent via the neural projection hRNN x,y of an RNN over the input up to x for the response at y. This design relaxes stationarity assumptions while also sharing structure across timepoints. The definitions of hin x,y and hRNN x,y are given below. Let tx be the element t[x] and xx be the xth predictor vector X[x,∗] ⊤. The inputs h(0) x,y to the CDRNN model are defined as the vertical concatenation of the predictors xx and the event timestamp tx: h(0) x,y def = xx tx (11) The output of the input projection at layer l and time x, y is defined as: hin(ℓ) x,y def = sin(ℓ) Win(ℓ)hin(ℓ−1) x,y + bin(ℓ) (12) where hin(0) x,y def = h(0) x,y. At the final layer, sin(Lin) is identity and bin(Lin) = 0, since hx,y already has a bias. The final output of the input projection is given by: hin x,y def = hin(Lin) x,y (13) Note that hin x,y is already non-stationary by virtue of its dependence on the event timestamp t[x], which allows the IRF to differ between timepoints (see e.g. Baayen et al., 2017, for development of a similar idea using generalized additive models). While this model of non-stationarity can be complex and non-linear, it is still limited by contextindependence. That is, the change in the IRF over time depends only on the amount of time elapsed since the start of the time series, independently of which events preceded. However, it is possible that the contents of the events in a time series may influence the IRF, above any deterministic change in response over time (for example, if several difficult preceding words have already taxed the processing buffer, additional processing costs may become larger). To account for this possibility, an RNN is built into the CDRNN design.10 Any variant of RNN can be used (this study uses a long shortterm memory network, or LSTM, Hochreiter and Schmidhuber, 1997). The ℓth RNN hidden state at x, y is designated by hRNN(ℓ) x,y . To account for the possibility of random variation in sensitivity to context, the initial hidden and cell states hRNN(ℓ) 0,y , cRNN(ℓ) 0,y depend on the random effects: hRNN(ℓ) 0,y def = hRNN(ℓ) 0 + WRNNh(ℓ) Z zy (14) cRNN(ℓ) 0,y def = cRNN(ℓ) 0 + WRNNc(ℓ) Z zy (15) where hRNN(ℓ) 0 , cRNN(ℓ) 0 are global biases and WRNNh(ℓ) Z , WRNNc(ℓ) Z are constrained to have mean 0 within each random grouping factor. Non-initial RNN states are computed via a standard LSTM update: h hRNN(ℓ) x,y , cRNN(ℓ) x,y i def =LSTM hRNN(ℓ) x−1,y , (16) cRNN(ℓ) x−1,y , hRNN(ℓ−1) x,y The hidden state of the final RNN layer is linearly projected to the dimensionality of the CDRNN hidden state: hRNN x,y def = WRNNprojhRNN(LRNN) x,y (17) To apply the CDRNN model to data, a mask F ∈{0, 1}Y ×X admits only those observations in X that precede each y[y]: F[y,x] def = ( 1 t[x] ≤t′[y] 0 otherwise (18) Letting τx,y denote the temporal offset between the predictors at x and the response at y, i.e. τx,y def = t′[y] −t[x]. A total of S(K +1) sparse convolution matrices Gs,k ∈RY ×X are defined to contain the predicted response to each preceding event for the kth dimension of h(0) x,y and the sth parameter of the predictive distribution, masked by F: Gs,k def = g1,1(τ1,1)[s,k] · · · gX,1(τX,1)[s,k] ... ... ... g1,Y (τ1,Y )[s,k] · · · gX,Y (τX,Y )[s,k] ⊙F (19) 10The experiments in this study also consider a variant without the RNN component, which is mathematically equivalent to setting hRNN x,y = 0. 3733 The convolved design matrix X′(s) ∈RY ×(K+1) for the sth parameter of the predictive distribution is then computed as: X′(s) [∗,k] def = Gs,k [X, t][∗,k] (20) Vector s ∈RS contains global, population-level estimates of the parameters of the predictive distribution. Under the univariate normal predictive distribution assumed in this study, s contains the predictive mean (µ, i.e. the intercept) and variance (σ2): s def = µ σ2 (21) Matrix SZ contains random predictive distribution parameter estimates for the yth response sZ y: SZ def = sZ 1 ⊤ ... sZ Y ⊤ (22) The vector of values for each response y for the sth predictive distribution parameter is given by summing the population value, random effects values, and convolved response values: S[∗,s] def = fconstraint(s) X′(s)1 + SZ [∗,s] + s[s] (23) where fconstraint(s) enforces any required constraints on the sth parameter of the predictive distribution. In the Gaussian predictive distribution assumed here, fconstraint(1) (the constraint function for the mean) is identity and fconstraint(2) (the constraint function for the variance) is the softplus bijection: softplus(x) def = ln(ex + 1) (24) Given an assumed distributional family F (here assumed to be univariate normal), the response in the CDRNN model is distributed as: y ∼F S[∗,1], . . . , S[∗,S] (25) B Asynchronously Measured Predictor Dimensions As discussed in Shain and Schuler (2018, 2021), CDR applies straightforwardly to time series with asynchronous predictor vectors and response values (i.e. measured at different times, such as word onsets that do not align with fMRI scan times). The CDR implementation of Shain and Schuler (2021) also supports asynchronously measured dimensions of the predictor matrix, simply by providing each predictor dimension with its own vector of timestamps. This allows e.g. Shain et al. (2020) to regress linguistic features (which are word-aligned) and sound power (which in their definition is measured at regular 100ms intervals) in the same model. Supporting asynchronously measured predictor dimensions is more challenging in CDRNN, especially if the RNN component is used. The solution used in CDR is not available because input dimensions that do not align in time are (1) arbitrarily grouped together and (2) erroneously treated as steps in the RNN input sequence. A more principled solution is to interleave the predictors in time order and pad irrelevant dimensions with zeros. For example, in a model with predictor A and predictor B that are sampled at different times, the values of A and B are temporally sorted together into a single time series, with the B value of A events set to zero and the A value of B events set to zero. This approach carries a computational cost: unlike CDR, the number of inputs to the convolution scales linearly on the number of asynchronously measured sets of predictors in the model. C Predictors The following predictors are common to all models presented here: • Rate (CDR/NN only): The deconvolutional intercept, i.e. the base response to a stimulus, independent of its features. In CDR, rate is estimated explicitly by fitting an IRF to intercept vector (Shain and Schuler, 2021) (i.e., implicitly, the response when all predictors are 0). In CDRNN, rate is a reference response, computed by taking the response to an average stimulus (since the zero vector may unlikely for a given input distribution, using it as a reference may not reliably reflect the model’s domain knowledge). In this study, all other IRF queries subtract out rate in order to show deviation from the reference. • Unigram surprisal: The negative log of the smoothed context-independent probability of a word according to a unigram KenLM model (Heafield et al., 2013) trained on Gigaword 3 (Graff et al., 2007). While this quantity is typically treated on a frequency or log probability scale in psycholinguistics, it is treated here on 3734 a surprisal (negative log prob) scale simply for easy of comparison with 5-gram surprisal (below), even though it is not a good estimate of the quantity typically targeted by surprisal (contextual predictability), since context is ignored. • 5-gram surprisal: The negative log of the smoothed probability of a word given the four preceding words according to a 5-gram KenLM model (Heafield et al., 2013) trained on Gigaword 3 (Graff et al., 2007). The following predictor is used in all reading models: • Word length: The length of the word in characters. The following predictors are used in eye-tracking models: • Saccade length: The length in words of the incoming saccade (eye movement), including the current word. • Previous was fixated: Indicator for whether the most recent fixation was to the immediately preceding word. Replications of Shain et al. (2020) use the following additional predictors: • PCFG surprisal: Lexicalized probabilistic context-free grammar surprisal computed using the incremental left-corner parser of van Schijndel et al. (2013) trained on a generalized categorial grammar (Nguyen et al., 2012) reannotation of Wall Street Journal sections 2 through 21 of the Penn Treebank (Marcus et al., 1993). • Sound power: Stimulus sound power (root mean squared energy), averaged over 250ms intervals. This implementation differs slightly from that of Shain et al. (2020), who sampled the measure every 100ms. The longer interval is designed to provide coverage over the extent of the HRF in this study, which uses a shorter history window for computational reasons (128 timesteps instead of 256). Both for computational reasons, especially under CDRNN-RNN (Appendix B) and because prior sound power estimates in this dataset have been weak (Shain et al., 2020), sound power is omitted from models used in the main comparison.
|
2021
|
288
|
Structural Guidance for Transformer Language Models Peng Qian1 Tahira Naseem2 Roger Levy1 Ram´on Fernandez Astudillo2 1 Department of Brain and Cognitive Sciences, MIT 2 IBM Research [email protected] [email protected] [email protected] [email protected] Abstract Transformer-based language models pretrained on large amounts of text data have proven remarkably successful in learning generic transferable linguistic representations. Here we study whether structural guidance leads to more human-like systematic linguistic generalization in Transformer language models without resorting to pre-training on very large amounts of data. We explore two general ideas. The “Generative Parsing” idea jointly models the incremental parse and word sequence as part of the same sequence modeling task. The “Structural Scaffold” idea guides the language model’s representation via additional structure loss that separately predicts the incremental constituency parse. We train the proposed models along with a vanilla Transformer language model baseline on a 14 million-token and a 46 million-token subset of the BLLIP dataset, and evaluate models’ syntactic generalization performances on SG Test Suites and sized BLiMP. Experiment results across two benchmarks suggest converging evidence that generative structural supervisions can induce more robust and humanlike linguistic generalization in Transformer language models without the need for data intensive pre-training. 1 Introduction Pre-trained Transformer architectures have led to huge progress in building more human-like language processing systems (Radford et al.; Devlin et al., 2019; Brown et al., 2020, among others). These models achieve impressive perplexity results on language modelling datasets, perform well on grammatical judgments (Warstadt et al., 2020), and provide useful linguistic representations that benefit a wide range of downstream tasks. Probing analyses also suggest that these models learn to implicitly encode syntactic information (Hewitt and Manning, 2019; Clark et al., 2019) that may support better linguistic generalization than recurrent neural network architectures (RNNs). However, the Transformer architecture (Vaswani et al., 2017) is an interesting subject of study beyond its success in transfer-learning settings. Transformer models lack the inductive biases of RNNs. Rather than maintaining vector-valued state and updating it in a recurrent manner, auto-regressive Transformer models encode all past decisions simultaneously at each inference step, thanks to a self-attention mechanism. The only notion of sequence order is also given by position embeddings summed to content embeddings in both input and auto-regressive signals. Previous works have shown the advantage of structural supervision in RNNs in learning to maintain syntactic states and non-local dependencies (Kuncoro et al., 2018; Wilcox et al., 2019; Futrell et al., 2019). It remains an open question whether Transformer language models can similarly benefit from generative structural supervision, and what form of structural supervision would more effectively induce human-like syntactic generalization. This work hypothesizes that the Transformer language model may benefit from explicit generative structural supervision to systematically generalize syntactic knowledge. Here we explore two major classes of structural guidance for Transformer language models based on joint modeling of language and constituency parses. The “generative parsing as language modeling” approach builds a Transformer-parameterized model to learn to predict actions that incrementally build constituency trees along with terminal words, following prior work on RNNs (Dyer et al., 2016; Choe and Charniak, 2016). The “structural scaffolding” approach follows the general idea of regularizing hidden representation through multi-task learning objective, with prior success in various NLP tasks (Zhang S NP The birds VP sang ADVP ⟨BOS⟩ NT(S) NT(NP) The birds REDUCE NT(VP) sang NT(ADVP) · · · w0 w1 w2 w3 y0:1 y1:2 y2:3 y3:4 w1 w2 w3 w0 w1 w2 (a) Vanilla language model NT(S) NT(NP) The birds REDUCE ⟨BOS⟩ NT(S) NT(NP) The birds (b) Parsing as Language Modelling w1 w2 w3 w0 w1 w2 y0:1 y1:2 y2:3 w1 w2 w3 w0 w1 w2 y0:1 y1:2 ⟨PAD⟩ (c) Language models with Structural Scaffold Figure 1: Top: Illustration of a partial constituency tree and corresponding transitions. Bottom: unidirectional transformer language model (a) without explicit structural supervision, (b) for modelling generative action parsing sequence, and (c) with structural scaffold for predicting the local incremental parsing state. and Weiss, 2016; Søgaard and Goldberg, 2016; Swayamdipta et al., 2018). We test these two approaches on two subsets of the BLLIP dataset (Charniak et al., 2000) and evaluate models’ syntactic generalization performances on SG Test Suites (Hu et al., 2020) and a sampled subset of the BLiMP Benchmark (Warstadt et al., 2020). We show evidence that generative structural supervision indeed induces more robust and human-like linguistic generalization in Transformer language models and explore the different trade-offs involved in the presented methods. 2 Models Here we explore joint modelling of structures and words parametrized with Transformers by considering both a sentence W and its constituency parse Y and modeling the joint distribution P(W, Y ). 2.1 Generative Parsing as Language Modeling A language model can be described formally as a probability distribution over strings of a language w1, · · · , wT , usually left-to-right factored. p(W) = p(w1, · · · , wT ) = T Y t=1 p(wt | w<t) (1) There are many possible approaches that can combine both language modeling and syntax modeling tasks. As long as both tasks share some of the parameters they can be considered a case of multi-task learning (Caruana, 1997). Of interest here is the model proposed in Recurrent Neural Network Grammars (RNNGs; Dyer et al., 2016) and parsing as language model (LSTM-LM; Choe and Charniak, 2016). Both approaches model the joint distribution of words W and constituency tree components Y as p(Y, W) = p(a1, · · · , aR) = R Y t=1 p(at | a<t) (2) where at are transitions of a state machine that generates both the sentence and the tree. These transitions are similar to the well-established transition sets used for transition-based parsing (Earley, 1970) but adapted to generate both text and parse simultaneously. For the reminder of this work, we will consider each at to be integer valued and indexing a dictionary of transitions. A transition a can be a word w or a transition action that generates a component of the constituency tree y. The actions include non-terminal symbols that open and label a new constituent with the label x, indicated as NT(x), or a REDUCE action closing the closest open constituent. An example of a partial parse tree and transitions can be found at the top of Figure 1. RNNG and LSTM-LM parametrize the same factorization in Equation 2 in different ways. RNNG utilizes stack-LSTMs, which allow it to dynamically create representations for partial tree components by composition. The LSTM-LM, however, uses a flat parametrization treating the transitions as a sequence in a conventional language model learnt with an LSTM (Hochreiter and Schmidhuber, 1997). It should also be noted that the LSTMLM is designed as a parser, while RNNG is also used as a language model. In order to derive a language model from a joint model, it is is necessary to marginalize over all possible parse trees p(W) = X Y ∈Y(W) p(Y, W) (3) which is an intractable problem since there is an exponentially large number of possible trees. The original RNNG work (Dyer et al., 2016) proposes an approximate solution based on importance sampling. In this work we use the word-synchronous beam search approximation introduced in Stern et al. (2017). The marginalized likelihood language model in Equation 3 is desirable because it makes no statistical independence assumption between language and syntax and shares all parameters across both tasks, with the exception of action specific embeddings. Particularly relevant for this work is the fact that both word and non-word transitions are predicted as language model output indiscriminately and are available at each prediction step through its history a<t. In this work we propose to parametrize Eq 2 with a Transformer language model (Vaswani et al., 2017). This is equivalent to the flat parametrization of the LSTM-LM but using a Transformer language model instead. Unlike LSTM-LM, which is a parsing model, we derive from it a language model by marginalization as in the RNNG. A Transformer language model can be succinctly described as a neural network of vertically stacked layers where the m-th layer is given by hm <t = FFm O · Am 1 (hm−1 <t ) Am 2 (hm−1 <t ) · · · Am N(hm−1 <t ) . (4) Here hm−1 <t ∈RH×t is the output of the previous decoder layer for all previous predictions of the model at time step t and H is the size of the hidden vector. The input to the first layer i.e. h0 <t are the embeddings of all previous transitions a<t concatenated with a start symbol. Each embedding is the sum of both a content embedding, dictionary vector that is being indexed, and a position embedding that encodes the absolute or relative position of each action in the sequence. FFm() is a feed-forward layer, Am 1 () · · · AM N () are multiple self-attention heads and O ∈RH×H is a matrix multiplication performed on the concatenated output of the attention heads. Both the feed-forward and the projection of N attention heads through O are wrapped around with residual, dropout and layer normalization operations that are here removed for clarity. Each attention head comprises a simple inner product attention mechanism Am n (hm−1 <t ) = V m n · hm−1 <t · softmax (Km n · hm−1 <t )T ·Qm n · hm−1 <t + M (5) where V m n , Km n , Qm n ∈RH/N×H are value, key and query projection matrices respectively and the softmax operation is normalized over columns to sum to one. The matrix M ∈{−∞, 0}t×t is used to prevent the model from attending to future states during training, enabling efficient parallelization. It is displayed here due to its relevance for the next section. Similarly to other models, to derive a distribution over all possible transitions, including words, nonterminal symbols and the REDUCE operation, we can use a softmax together with an inner product p(at | a<t) = softmax(EW∪Y · hm <t)at (6) where EW∪Y are the embeddings for the joint vocabulary of words, non-terminals and REDUCE transitions. Henceforth, we refer to this model as Parsing as Language Model, or PLM for short. Unlike LSTMs or the RNNG, the Transformer has direct access to all past decisions through selfattention and relies on position embeddings to encode word order. Thus, in principle, there is no structural bias for the model to favor past decisions that are close in time to inform current prediction. On one hand, this potential ability to use long distance information can enable a less local, more human like processing of language, but on the other hand, it can also result in an additional learning burden, especially if there is not sufficient learning data available. Also worth noting for the experiments proposed here is that the total number of parameters of a typical Transformer greatly exceeds that of an LSTM or a RNNG model. 2.2 Incorporating RNNG-like characteristics As previously mentioned, unlike any of the other models, the RNNG is able to create partial tree representations by composition using stack-LSTMs. S NP The birds ⟨BOS⟩ NT(S) NT(NP) The birds BUFFER head ⟨BOS⟩ NT(S) NT(NP) The birds STACK head Figure 2: Illustration of how the generated incremental constituency parse is used to inform attention patterns in the structure-guided attention heads. This changes the RNNG model structure dynamically as a function of the partial parse, a very desirable property to derive syntax-aware representations. Moreover, the fact that Recurrent Neural Networks such as LSTMs summarize all information about previous time steps on two hidden vectors, creates a bottleneck that forces the model to focus on the local state. This is a situation where a syntax-aware representation can provide additional value by enabling the local state to better encompass past structures. We conjecture that a similarly constrained local state might benefit Transformer models in learning linguistic regularities, especially in a limited training data scenario. In an attempt to capture a similar effect in the Transformer, we explore here the idea of masking some attention heads to reflect the parser state as in the stack-Transformer (Astudillo et al., 2020). In the stack-Transformer, two attention heads are specialized to attend only to the contents of buffer and stack respectively for dependency and semantic parsing tasks. Here we choose to specialize two heads as well for each layer in Equation 4, as depicted in Fig. 2. One attention head attends to the contents of the last open constituent whereas another head attends all other past decisions not involving that constituent. The rest of the heads are left free as in the original Transformer architecture. To constrain the attention heads, we only need to alter the mask M in Equation 5 to depend on head index n and past actions Mn(a<t), which results in a negligible computation overhead. This hard masking makes the model structure change dynamically depending on the partial parse and it forces some heads to focus on the local syntactic state. Nevertheless, unlike the RNNG, it does not create new representations of partial parses that can be composed in a recurrent manner at each time step, and some attention heads can still operate unrestricted. We hypothesize that structure-aware attention mechanism may still help the model achieve better generalization. The symbolic representation induces a strong inductive bias to how the model should use the structure that it generates on the fly. We henceforth refer to this model PLM-mask. 2.3 Scaffolding by Learning to Predict Local Parse States Given the strong coupling between the tasks, the marginal likelihood Transformer language model of the previous section can be expected to be strongly influenced by the additional syntax prediction task. This comes however at a big cost. First, sequences combine both words and non-terminal and reduce transitions, yielding longer sentences than those of a normal language model R > T. Furthermore the approximated marginalization is computationally intensive and also introduces an approximation error. One well-established regime that allows joint modeling of tasks at a low complexity is that of the syntactic scaffold (Zhang and Weiss, 2016; Søgaard and Goldberg, 2016; Swayamdipta et al., 2018). Scaffolding adds an additional structure prediction task at one of the layers of the model as a separate layer and only during training. This is a minimally intrusive change since it just branches some hidden vector of the network and computes an additional loss. It also has no influence on test runtime and avoids expensive steps such as marginalization. However, applying the idea of syntactic scaffolding to our present scenario poses one difficulty. If we use a standard language model predicting words w and predict the non-word symbols y separately, we face the problem that the two sequences have different lengths. To overcome this in a straightforward way, we predict the n-gram of non-word actions yt:t+n(t) corresponding to the partial parse synchronous with step t when we predict word wt. We use a secondary softmax layer for this action n-gram prediction. p(yt:t+n | y<t) = softmax(EY ∗· hm <t)yt:t+n (7) Here EY ∗is the vocabulary of all transition ngrams excluding words found in the train corpus plus a blank symbol. Note that since Scaffolding operates only at train time, we do not need to worry about generalization of these n-grams to test time. The models are thus trained to minimize the loss function −log p(Y, W) where p(Y, W) = QT t=1 p(wt | w<t) + QT t=1 p(yt:t+n(t) | w<t) (8) The scaffold can be set so that the synchronous non-word action n-grams yt:t+n(t) are predicted either before (Figure 1c, left) or after (Figure 1c, right) producing wt. We considered both variants in our experiments to empirically assess their impact on performance. We refer to this model as Transformer Language Model with Syntactic Scaffold, or ScLM in short, and its two versions ScLM-past and ScLM-next, for past and next ngram prediction. 3 Experiments 3.1 Model Training All models, including the baseline vanilla language models (LM in short), the syntactic scaffold models, and the generative parsing models, are based on the same architecture of GPT-2 small (Radford et al.) (117M parameters, 12 layers, H = 768) and use the same BPE tokenizer, but with randomly initialized weights. We believe this would give us a fair comparison to pretrained GPT-2 as well, in order to evaluate whether structural guidance helps improve sample efficiency. We implemented all the proposed models using Huggingface’s Transformer package (Wolf et al., 2020)1. As our goal here is to study whether structural guidance helps models learn robust humanlike generalization of syntactic knowledge, we train our model on the BLLIP dataset (Charniak et al., 2000), an English newswire style corpus used in Hu et al. (2020). This makes the results here more comparable to the results reported in previous work, especially with RNNGs. We train the proposed models and the baseline vanilla Transformer language models on BLLIP-MD, a 14 million-token corpus, and BLLIP-LG, a 46 million-token corpus, both of which are auto-parsed using a state-of-theart constituency parser (Kitaev and Klein, 2018). We used the parsed sentences to generate oracle parsing action sequence for PLM and PLM-mask. We collected a list of word-synchronous parsing 1Code available at https://github.com/IBM/ transformers-struct-guidance action sequences from the train and development oracle of BLLIP-LG and use it to parametrize the action n-gram vocabulary of ScLMs trained on both BLLIP-MD and BLLIP-LG. There are 3756 action n-gram types from the corpora, including one padding token and one blank token. All models were trained with learning rate 10−5, AdamW optimizer, and minibatch of size 5. We trained the models with multiple seeds within the capacity of our resources, in order to accommodate potential variance. In total, there are three seeds of LM, four of ScLM-past, four of ScLM-next, three of PLM, and three of PLM-mask for BLLIP-MD, and the same number of seeds of each model type for BLLIP-LG. Models were trained until convergence, as suggested by the loss of the development set during training. 3.2 Targeted Syntactic Evaluation To assess whether a trained model systematically generalizes its syntactic knowledge, we employ targeted syntactic evaluation paradigm (Marvin and Linzen, 2018). Specifically, we measure models’ performance on two held-out test datasets, a collection of syntactic generalization test suites from Hu et al. (2020) and BLiMP Benchmark from Warstadt et al. (2020). These two datasets cover a wide range of English syntactic phenomena. Tests from Hu et al. (2020), which we refer as SG Test Suites, consist of hand-designed test suites for evaluating fine-grained syntactic generalization in incremental processing of a linguistic input. The general method is to compare models’ surprisals p(continuation|prefix) of grammatical and ungrammatical continuations given certain sentence prefixes. We report the accuracy averaged across SG test suites. BLiMP Benchmark features minimal pairs of a grammatical sentence W and an ungrammatical counterpart W ∗. To evaluate a model on these minimal pairs, one simply compares the likelihood of W and W ∗assigned by the model. As is implied by the evaluation methods, we need to marginalize out the structure variables for PLM or PLM-mask models in order to estimate the surprisal of a continuation, given a sentence prefix or the likelihood of a complete sentence. We follow similar setup as in Futrell et al. (2019); Wilcox et al. (2019) applying word-synchronous beam search (Stern et al., 2017) to find a list Yk of k incremental parses given a sentence prefix w<t. BLLIP-MD BLLIP-LG 0.55 0.60 0.65 0.70 0.75 0.80 Accuracy Model Performance on SG Test Suites RNNG LM ScLM-past ScLM-next PLM PLM-mask GPT-2 BLLIP-MD BLLIP-LG 0.55 0.60 0.65 0.70 0.75 0.80 Accuracy Model Performance on BLiMP-10% Test Suites LM ScLM-past ScLM-next PLM PLM-mask GPT-2 Figure 3: Comparing models’ overall accuracy across test suites from SG Test Suites (top) and BLiMP-10% (bottom). RNNG performances are from Hu et al. (2020). We then sum the joint probability p(w<t, y<t) over the list of incremental parses given by the model to approximate the likelihood of p(w<t). We set the parse beam size to 100, word-synchronous beam size k as 10, and fast track size of 5. Since the search process can be computationally intensive, the large number of items in BLiMP benchmark poses a computational challenge. We therefore select the first 10% out of the 1000 items in each of the 67 tests of BLiMP Benchmark. We report the accuracy over the 100 items and refer to this down-sized BLiMP Benchmark as BLiMP-10%. We compare models’ performance on the SG Test Suites and BLiMP-10% in Figure 3. Each bar shows a model’s performance averaged across multiple seeds on a given benchmark, with each dot plotting the accuracy of a specific seed. Overall, syntactic generalization performance improves as the training data size increases from BLLIP-MD (14 million tokens) to BLLIP-LG (42 million tokens). Models with structural guidance achieve higher accuracy than the vanilla Transformer language model trained on the same set of raw text data without explicit structural information. We also include the results for the RNNGs taken from Hu et al. (2020). RNNG lags behind all Transformer models by a large margin in average scores. We also notice that among different forms of structural guidance, generative parsing as language modeling is the most effective in improving syntactic generalization performance against the baseline transformer language models. We didn’t observe consistent benefits of adding dynamic masking mechanism to PLM. While scaffolding approach slightly improves vanilla Transformer language models, it still falls behind the best performance of the model trained with generative parsing. We hypothesize that our scaffold did not fully exploit the compositional structure in the local parses by modelling each action n-gram as a distinct type, while the generative parsing models only predict actions in a relatively small set of non-terminal action space, which might make it easier for PLM and PLM-mask to learn compositional generalization. We leave it for future work to design new scaffolds that can take advantage of the combinatorial nature of syntactic structure. For completeness, we also ran the pre-trained GPT-2 model on the syntactic suites. This yielded a score of 0.808 on the SG Test Suites and 0.827 on BLiMP-10% for the small version of pre-trained GPT-2. Among models trained on BLLIP-LG, the average accuracy score on the SG Test Suites is 0.723 for PLMs, 0.748 for PLM-masks, and 0.665 for LMs. Similar trend is observed on BLiMP-10% as well, where among models trained on BLLIPLG the average accuracy is 0.751 for PLMs, 0.753 for PLM-masks, and 0.708 for LMs. The proposed PLM method is able to close the gap between GPT-2 small and the same model trained with BLLIP-LG by about half, while the improvement for BLiMP is more modest but still significative. It remains an open question whether scaling syntactic supervision to a larger dataset than BLLIP-LG would bring the generalization performance of PLM models closer to that of the pretrained GPT-2 model. 3.2.1 Relationship between Perplexity and Syntactic Generalization Performance We compare perplexity on the BLLIP held-out test set against syntactic generalization performance in Figure 4. Perplexities of PLM and PLM-mask models are computed setting the parse tree equal to the gold parse in Equation 3 to approximate the likelihood. Note that, unlike Hu et al. (2020), all 50 60 Word-level Perplexity 0.625 0.650 0.675 0.700 0.725 0.750 SG Accuracy Model LM ScLM-past ScLM-next PLM PLM-mask Corpus BLLIP-MD BLLIP-LG Corpus BLLIP-MD BLLIP-LG 50 60 Word-level Perplexity 0.68 0.70 0.72 0.74 0.76 BLiMP-10% Accuracy Model LM ScLM-past ScLM-next PLM PLM-mask Corpus BLLIP-MD BLLIP-LG Corpus BLLIP-MD BLLIP-LG Figure 4: Comparison between model perplexity on BLLIP test data and syntactic generalization performance on SG Test Suites (top) and BLiMP-10% (bottom). our models use the same BPE vocabulary and word tokenization from GPT-2. The only exception are the additional parsing actions in the vocabulary y. From Figure 4, both perplexity and syntactic generalization performance improve with dataset size. However, for both training dataset sizes, we see that structural guidance can improve syntactic generalization. PLM models consistently perform better than vanilla models. While all models achieve very similar perplexity results after being trained on a specific dataset, their syntactic generalization performances differ dramatically. 3.2.2 Effect of Structural Guidance on Learning Specific Syntactic Structures In addition to comparing model’s aggregated performances, we also compare their generalization performances in the clustered subsets of tests in SG Test Suites and BLiMP-10%. These subsets consist of several related tests that target specific type of syntactic phenomenon, such as NPI licensing, subject-verb agreement, filler-gap dependencies, etc. We also include the results for the RNNGs taken from Hu et al. (2020). Results in Figure 5 show converging evidence that structural guidance in the form of generative parsing can robustly improve learning of subjectverb agreement and NPI licensing, and helps the model to better capture incremental processing phenomenon such as garden-path effects, but seems to slightly hurt the performance on gross syntactic state. While overall the RNNG shows a poor performance this is mostly due to its very low scores for licensing suites. Excluding these suites only the RNNG shows a performance close to the PLM model, even outperforming it clearly for the gross syntactic state suites. In this category and binding PLM variants seem inferior to all other models. 4 Related Work Multitask learning (Caruana, 1997) has been applied to a variety of NLP tasks with traditional modeling approaches (Miller et al., 2000; Sutton and McCallum, 2005; Sutton et al., 2007) as well as more recent neural models (Collobert et al., 2011; Li et al., 2020a). A recurring theme has been the use of structure in the form of syntactic trees to benefit other NLP tasks. Among the early works exploring this direction, Punyakanok et al. (2008) showed that syntactic parses can benefit Semantic Role Labeling (SRL). Poon and Domingos (2009) extended this idea to induce first-order logic representation in a unsupervised fashion, by clustering the dependency structures. In both cases syntax forms part of a pipeline and is not strictly supervision for the end task. This trend continued with the rise of neural models. Collobert et al. (2011) improved deep convolution neural network for syntactic chunking models with additional POS supervision. Zhang and Weiss (2016); Søgaard and Goldberg (2016) observe the benefits of POS supervision at different depths of a neural network model with impact on dependency parsing, tagging and CCG super tagging performance. He et al. (2019) perform a syntax-based pruning of semantic roles, showing benefits in a multilingual setting. More recently, Sachan et al. (2020) incorporate a syntactic graph recurrent neural network into BERT models for better semantic role labeling. However, their method shows little or no benefit of syntax modeling for Named Entity Recognition and relation linking task. Neural machine translation (Chen et al., 2018) and text generation (Li et al., 2020a) have also been shown to benefit from syntactic modeling. In a recent work, Li et al. (2020b) use syntactic modeling in BERT based transformers to achieve performance gains 0.05 0.15 0.25 0.35 0.45 0.55 0.65 Accuracy Licensing (10 suites) 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 Long-Distance Dependencies (8 suites) 0.3 0.4 0.5 0.6 0.7 0.8 Agreement (3 suites) BLLIP-MD BLLIP-LG 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Accuracy Garden-Path Effects (6 suites) BLLIP-MD BLLIP-LG 0.70 0.75 0.80 0.85 0.90 0.95 Gross Syntactic State (4 suites) BLLIP-MD BLLIP-LG 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Center Embedding (2 suites) Model Performance on Specific Clusters of SG Test Suites RNNG LM ScLM-past ScLM-next PLM PLM-mask 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Accuracy Anaphor Agreement (2 suites) 0.55 0.60 0.65 0.70 Argument Structure (9 suites) 0.60 0.65 0.70 0.75 Binding (7 suites) 0.60 0.65 0.70 0.75 Control/Raising (5 suites) 0.70 0.75 0.80 0.85 Accuracy Determiner-Noun Agreement (8 suites) 0.50 0.55 0.60 0.65 0.70 0.75 0.80 Ellipsis (2 suites) 0.60 0.65 0.70 0.75 Filler Gap (7 suites) 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Irregular Forms (2 suites) BLLIP-MD BLLIP-LG 0.40 0.45 0.50 0.55 0.60 Accuracy Island Effects (8 suites) BLLIP-MD BLLIP-LG 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 NPI Licensing (7 suites) BLLIP-MD BLLIP-LG 0.55 0.60 0.65 0.70 0.75 Quantifiers (4 suites) BLLIP-MD BLLIP-LG 0.70 0.75 0.80 0.85 Subject-Verb Agreement (6 suites) Model Performance on Specific Clusters of BLiMP-10% Test Suites LM ScLM-past ScLM-next PLM PLM-mask Figure 5: Model performance comparison by specific linguistic phenomena clustered in SG Test Suites (top) and BLiMP-10% (bottom). RNNG performances are from Hu et al. (2020). on several text classification benchmarks. Other works have found that structural supervision in the form of intermediate fine-tuning (e.g., on CCG super tagging) is not helpful or even harmful (Pruksachatkun et al., 2020; Warstadt et al., 2019). The focus of our work is on gauging the impact of joint modeling on syntactic generalization performance. In this direction, the work of Swayamdipta et al. (2018) is close to the scaffolding version of our model. They predict multiple labels, extracted from syntactic information, as auxiliary task and show positive effects on shallow semantic parsing and co-reference resolution. We use however a single feature, constituency parsing n-gram, which is closer to prior work relying on Part-of-Speech information. In addition, we explore impact of using preceding structure as feature vs postceding structure, which as shown plays a role in the learning process. In terms of modeling objective and syntactic representations, our method is closest to the works of Choe and Charniak (2016); Dyer et al. (2016) that jointly model syntax and language. A more recent work from Peng et al. (2019) uses Rational Neural Networks language model that can derive binary unlabeled constituents from attention weights and can supervise the attention to attain a structural inductive bias. The proposed models show lower language modeling perplexity compared to their structure agnostic counterparts. We also extend here the idea of syntax-aware language modeling to transformer-based language models. Finally, our approach relates to the other works that propose ways of incorporating structural information into Transformer-based models. This includes the use of dependency or tree structure for constraining self-attention patterns (Strubell et al., 2018; Wang et al., 2019; Zhang et al., 2020), guiding cross-attention (Chen et al., 2018; Astudillo et al., 2020), modelling syntactic distance (Du et al., 2020), using syntactic information to guide the computation flow in the model (Shen et al., 2021), or through knowledge distillation (Kuncoro et al., 2020). Our structured masking in parsing as language modeling approach is close in spirit to the methods that modify attention mechanism according to syntactic connections (Astudillo et al., 2020); This work, however, primarily aims to study the impact of structural guidance on syntactic generalization. Therefore, we resort to simpler methods of incorporating structure to minimize the impact of modeling intricacies. 5 Conclusion Our work explores two forms of syntactic supervision as structural guidance for Transformer language models. Experiments suggest that generative parsing approach can effectively improve systematic generalization of learned syntactic knowledge in small training data regime, while a naive syntactic scaffold approach does not improve the baseline to the same extent despite reduced computation cost at inference time. Future work may explore alternative structural guidance strategies that combine the best of both approaches. Acknowledgments The authors would like to thank the anonymous reviewers for their helpful comments. This work was supported by the MIT-IBM Watson AI Lab. References Ram´on Fernandez Astudillo, Miguel Ballesteros, Tahira Naseem, Austin Blodgett, and Radu Florian. 2020. Transition-based parsing with stacktransformers. page 1001–1007. Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Rich Caruana. 1997. Multitask learning. Machine learning, 28(1):41–75. Eugene Charniak, Don Blaheta, Niyu Ge, Keith Hall, John Hale, and Mark Johnson. 2000. Bllip 1987-89 wsj corpus release 1. Linguistic Data Consortium, Philadelphia, 36. Kehai Chen, Rui Wang, Masao Utiyama, Eiichiro Sumita, and Tiejun Zhao. 2018. Syntax-directed attention for neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32. Do Kook Choe and Eugene Charniak. 2016. Parsing as language modeling. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2331–2336, Austin, Texas. Association for Computational Linguistics. Kevin Clark, Urvashi Khandelwal, Omer Levy, and Christopher D. Manning. 2019. What does BERT look at? an analysis of BERT’s attention. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 276–286, Florence, Italy. Association for Computational Linguistics. Ronan Collobert, Jason Weston, L´eon Bottou, Michael Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011. Natural language processing (almost) from scratch. Journal of machine learning research, 12(ARTICLE):2493–2537. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Wenyu Du, Zhouhan Lin, Yikang Shen, Timothy J. O’Donnell, Yoshua Bengio, and Yue Zhang. 2020. Exploiting syntactic structure for better language modeling: A syntactic distance approach. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6611– 6628, Online. Association for Computational Linguistics. Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. Recurrent neural network grammars. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 199–209, San Diego, California. Association for Computational Linguistics. Jay Earley. 1970. An efficient context-free parsing algorithm. Communications of the ACM, 13(2):94– 102. Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 32–42, Minneapolis, Minnesota. Association for Computational Linguistics. Shexia He, Zuchao Li, and Hai Zhao. 2019. Syntaxaware multilingual semantic role labeling. arXiv preprint arXiv:1909.00310. John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138, Minneapolis, Minnesota. Association for Computational Linguistics. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725–1744, Online. Association for Computational Linguistics. Nikita Kitaev and Dan Klein. 2018. Constituency parsing with a self-attentive encoder. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics. Adhiguna Kuncoro, Chris Dyer, John Hale, Dani Yogatama, Stephen Clark, and Phil Blunsom. 2018. LSTMs can learn syntax-sensitive dependencies well, but modeling structure makes them better. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1426–1436, Melbourne, Australia. Association for Computational Linguistics. Adhiguna Kuncoro, Lingpeng Kong, Daniel Fried, Dani Yogatama, Laura Rimell, Chris Dyer, and Phil Blunsom. 2020. Syntactic structure distillation pretraining for bidirectional encoders. Transactions of the Association for Computational Linguistics, 8:776–794. Yinghao Li, Rui Feng, Isaac Rehg, and Chao Zhang. 2020a. Transformer-based neural text generation with syntactic guidance. arXiv preprint arXiv:2010.01737. Zhongli Li, Qingyu Zhou, Chao Li, Ke Xu, and Yunbo Cao. 2020b. Improving bert with syntax-aware local attention. arXiv preprint arXiv:2012.15150. Rebecca Marvin and Tal Linzen. 2018. Targeted syntactic evaluation of language models. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1192–1202, Brussels, Belgium. Association for Computational Linguistics. Scott Miller, Heidi Fox, Lance Ramshaw, and Ralph Weischedel. 2000. A novel use of statistical parsing to extract information from text. In 1st Meeting of the North American Chapter of the Association for Computational Linguistics. Hao Peng, Roy Schwartz, and Noah A. Smith. 2019. PaLM: A hybrid parser and language model. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3644– 3651, Hong Kong, China. Association for Computational Linguistics. Hoifung Poon and Pedro Domingos. 2009. Unsupervised semantic parsing. In Proceedings of the 2009 conference on empirical methods in natural language processing, pages 1–10. Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, and Samuel R Bowman. 2020. Intermediate-task transfer learning with pretrained models for natural language understanding: When and why does it work? arXiv preprint arXiv:2005.00628. Vasin Punyakanok, Dan Roth, and Wen-tau Yih. 2008. The importance of syntactic parsing and inference in semantic role labeling. Computational Linguistics, 34(2):257–287. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. Devendra Singh Sachan, Yuhao Zhang, Peng Qi, and William Hamilton. 2020. Do syntax trees help pretrained transformers extract information? arXiv preprint arXiv:2008.09084. Yikang Shen, Shawn Tan, Alessandro Sordoni, Siva Reddy, and Aaron Courville. 2021. Explicitly modeling syntax in language models with incremental parsing and a dynamic oracle. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1660–1672, Online. Association for Computational Linguistics. Anders Søgaard and Yoav Goldberg. 2016. Deep multitask learning with low level tasks supervised at lower layers. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 231–235, Berlin, Germany. Association for Computational Linguistics. Mitchell Stern, Daniel Fried, and Dan Klein. 2017. Effective inference for generative neural parsing. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1695–1700, Copenhagen, Denmark. Association for Computational Linguistics. Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-informed self-attention for semantic role labeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5027–5038. Charles Sutton and Andrew McCallum. 2005. Joint parsing and semantic role labeling. Technical report, MASSACHUSETTS UNIV AMHERST DEPT OF COMPUTER SCIENCE. Charles Sutton, Andrew McCallum, and Khashayar Rohanimanesh. 2007. Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data. Journal of Machine Learning Research, 8(3). Swabha Swayamdipta, Sam Thomson, Kenton Lee, Luke Zettlemoyer, Chris Dyer, and Noah A. Smith. 2018. Syntactic scaffolds for semantic structures. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3772–3782, Brussels, Belgium. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008. Curran Associates, Inc. Yaushian Wang, Hung-Yi Lee, and Yun-Nung Chen. 2019. Tree transformer: Integrating tree structures into self-attention. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 1061–1070, Hong Kong, China. Association for Computational Linguistics. Alex Warstadt, Yu Cao, Ioana Grosu, Wei Peng, Hagen Blix, Yining Nie, Anna Alsop, Shikha Bordia, Haokun Liu, Alicia Parrish, et al. 2019. Investigating bert’s knowledge of language: Five analysis methods with npis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 2870–2880. Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2020. BLiMP: The benchmark of linguistic minimal pairs for English. Transactions of the Association for Computational Linguistics, 8:377–392. Ethan Wilcox, Peng Qian, Richard Futrell, Miguel Ballesteros, and Roger Levy. 2019. Structural supervision improves learning of non-local grammatical dependencies. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3302–3312, Minneapolis, Minnesota. Association for Computational Linguistics. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Yuan Zhang and David Weiss. 2016. Stackpropagation: Improved representation learning for syntax. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1557–1566, Berlin, Germany. Association for Computational Linguistics. Zhuosheng Zhang, Yuwei Wu, Junru Zhou, Sufeng Duan, Hai Zhao, and Rui Wang. 2020. Sg-net: Syntax guided transformer for language representation. IEEE Transactions on Pattern Analysis and Machine Intelligence.
|
2021
|
289
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 340–350 August 1–6, 2021. ©2021 Association for Computational Linguistics 340 Aspect-Category-Opinion-Sentiment Quadruple Extraction with Implicit Aspects and Opinions Hongjie Cai∗ Rui Xia∗† Jianfei Yu School of Computer Science and Engineering, Nanjing University of Science and Technology, China {hjcai, rxia, jfyu}@njust.edu.cn Abstract Product reviews contain a large number of implicit aspects and opinions. However, most of the existing studies in aspect-based sentiment analysis ignored this problem. In this work, we introduce a new task, named AspectCategory-Opinion-Sentiment (ACOS) Quadruple Extraction, with the goal to extract all aspect-category-opinion-sentiment quadruples in a review sentence and provide full support for aspect-based sentiment analysis with implicit aspects and opinions. We further construct two new datasets Restaurant-ACOS and Laptop-ACOS for this new task. The former is an extension of the SemEval Restaurant dataset; the latter is a brand new Laptop dataset with much larger size than the SemEval Laptop dataset. Both contain the annotations of not only aspect-category-opinionsentiment quadruples but also implicit aspects and opinions. We finally benchmark the task with four baseline systems. Experiments demonstrate the feasibility of the new task and its advantage in extracting and describing implicit aspects and implicit opinions in ABSA. The two datasets and source code of four systems are publicly released at https: //github.com/NUSTM/ACOS. 1 Introduction As a fine-grained sentiment analysis task, aspectbased sentiment analysis (ABSA) has received continuous attention. Its core task is to extract the opinion target described by an entity and its aspect (collectively referred to as aspect) from product reviews, and identify the sentiment toward the aspect (Liu, 2012). The standard aspect-based sentiment analysis task includes two basic subtasks: aspect extraction and aspect-based sentiment classification. By integrating the two subtasks, one can ∗Equal contribution. † Corresponding author. Review Sentence Looks nice, and the surface is smooth, but certain apps take seconds to respond. Aspect-Category-Opinion-Sentiment Quadruple Extraction surface-Design-smooth-Positive NULL-Design-nice-Positive apps-Software-NULL-Negative Figure 1: An example of the Aspect-Category-OpinionSentiment Quadruple Extraction task. Restaurant Laptop Explicit Aspect & Explicit Opinion 63.34% 56.06% Implicit Aspect & Explicit Opinion 19.47% 17.54% Explicit Aspect & Implicit Opinion 12.38% 27.55% Implicit Aspect & Implicit Opinion 14.83% 8.24% Table 1: The percentage of review sentences with explicit and implicit aspect/opinion. identify an aspect-sentiment pair (g, s), where g is an aspect term, and s is the sentiment polarity toward the aspect. (Hu and Liu, 2004; Qiu et al., 2011) pointed out that the correlation between the aspect term and the opinion term is helpful for better ABSA. The following studies in this direction includes aspect-opinion co-extraction (Wang et al., 2016a, 2017; Yu et al., 2018; Li et al., 2018; Dai and Song, 2019), aspect-opinion pair extraction (Chen et al., 2020a; Zhao et al., 2020), and aspect-opinion-sentiment triple extraction (Peng et al., 2020; Xu et al., 2020; Wu et al., 2020; Mao et al., 2021), etc. However, most of the existing studies only considered the extraction of explicit aspects and opinions, while ignored the implicit ones. In fact, product reviews contain a large amount of implicit aspects and opinions. Table 1 summarizes the percentage of implicit aspects and opinions in the 341 SemEval Restaurant dataset and our new Laptop dataset. It can be seen that nearly 44% of the review sentences contain implicit aspects or implicit opinions in the Laptop domain, and the percentage of sentences containing both implicit aspects and implicit opinions also exceeds 8%. Similar percentages can be observed in the Restaurant domain. Although some studies have attempted to solve the implicit aspect problem (Liu et al., 2005; Poria et al., 2014; Chen and Chen, 2016; Wan et al., 2020) or the implicit opinion problem (Lazhar and Guiyassa, 2016) from respective perspectives, there is still a lack of a unified framework that fully discusses and solves the implicit aspect and implicit opinion problems. In this work, we introduce a new task named Aspect-Category-Opinion-Sentiment (ACOS) Quadruple Extraction, with the goal to extract all aspect-category-opinion-sentiment quadruples in a review sentence, and provide full support for aspect-level sentiment analysis with implicit aspects and opinions. As shown in Figure 1, in the review sentence “Looks nice and the surface is smooth, but certain apps take seconds to respond”, surface is an aspect, Design is its category, smooth is the opinion toward this aspect, and Positive is the corresponding sentiment. The four elements are combined into an explicit quadruple surfaceDesign-smooth-Positive. In addition to that, there are two other quadruples that need to be extracted: Null-Design-nice-Positive which contains an implicit aspect, and apps-Software-Null-Negative which contains an implicit opinion. The new ACOS Quadruple Extraction task has the following two challenges: • In term of dataset, so far there was no available dataset that is fully annotated with aspectcategory-opinion-sentiment quadruples including all implicit aspects and opinions; • In terms of modeling complexity, the task includes two extraction problems (aspect extraction, opinion extraction) and two classification problems (category classification, sentiment classification). It is challenging to effectively model the four subtasks together to construct quadruples containing implicit aspects and implicit opinions. To address these issues, we further construct two new datasets, Restaurant-ACOS and LaptopACOS, for the new task. The former is an extension of the existing SemEval Restaurant dataset, based on which we add the annotation of implicit aspects, implicit opinions, and the quadruples. The latter is a brand new one collected from the Amazon Laptop domain. It has twice size of the SemEval Loptop dataset, and is annotated with quadruples containing all explicit/implicit aspects and opinions. We finally benchmark the task by establishing four baseline systems, Double-PropagationACOS, JET-ACOS, TAS-BERT-ACOS and Extract-Classify-ACOS, by adapting the representative approaches in aspect-opinion pair extraction, aspect-category-opinion triple extraction or aspect-opinion-sentiment triple extraction to ACOS Quadruple Extraction. The experiments on the two ACOS datasets demonstrate the feasibility of the new ACOS Quadruple Extraction task and its effectiveness in extracting and describing implicit aspects and implicit opinions. The contributions of this work can be summarized as follows: • We introduce a new task named AspectCategory-Opinion-Sentiment Quadruple Extraction, to address the implicit aspects/opinions issues in ABSA; • We construct two new datasets for the task, with ACOS quadruple annotations including implicit aspects/opinions; • We benchmark the task with four baseline systems. The experiments demonstrate the new task’s advantage in addressing the implicit aspect/opinion issues. 2 Task We first define the four elements of the ACOS Quadruple Extraction task based on (Liu, 2012). (Peng et al., 2020; Mao et al., 2021) provided good summaries of recent tasks and terminology in ABSA. For simplicity, in this paper we use aspect, category, opinion and sentiment to denote aspect term, aspect category, opinion term and sentiment polarity, respectively. They are defined as follows: • Aspect denotes an entity and its aspect indicating the opinion target, which is normally a word or phrase in the text; • Category represents a unique predefined category for the aspect in a particular domain; • Opinion refers the subjective statement on an aspect, which is normally a subjective word or phrase in the text; 342 • Sentiment is the predefined semantic orientation (e.g., Positive, Negative, or Neutral) toward the aspect. Aspect-Category-Opinion-Sentiment (ACOS) Quadruple Extraction is then defined as a task to extract a set of aspect-category-opinion-sentiment quadruples described in a review sentence containing n words r=[w1, . . . , wn]: SACOS = {. . . , ai-cj-ok-sl, . . .}, (1) where ai-cj-ok-sl denotes an aspect-categoryopinion-sentiment quadruple, ai is the extracted aspect, cj ∈C is its category, ok is the extracted opinion, and sl ∈{Positive, Neutral, Negative} is its corresponding sentiment.1 Note that a review sentence usually contains multiple aspects and opinions. The ACOS Quadruple Extraction task does not only identify four elements, but also combine them into a set of valid quadruples, meanwhile considering implicit aspects/opinions. As the implicit aspect/opinion is not explicitly expressed as a word or phrase, in case of implicit aspect we set a as NULL and use category c to describe the opinion target, and in case of implicit opinion we set o as NULL and use sentiment s to describe the semantic orientation. 3 Datasets We construct two new datasets, Restaurant-ACOS and Laptop-ACOS, for the ACOS Quadruple Extraction task. 3.1 Source The Restaurant-ACOS dataset is constructed based on the SemEval 2016 Restaurant dataset (Pontiki et al., 2016) and its expansion datasets (Fan et al., 2019; Xu et al., 2020). Laptop-ACOS is a brand new Laptop dataset collected from the Amazon platform at the years of 2017 and 2018 (covering ten types of laptops under six brands such as ASUS, acer, Samsung, Lenovo, MBP, MSI and so on). It contains 4,076 review sentences, much larger than the SemEval Laptop datasets. 1Similarly, the previous representative tasks in ABSA can also be denoted by the combination of the above elements, e.g., aspect-sentiment (AS) pair extraction (Mitchell et al., 2013; Zhang et al., 2015), aspect-opinion (AO) pair extraction (Chen et al., 2020a; Zhao et al., 2020), aspect-opinion-sentiment (AOS) triple extraction (Peng et al., 2020; Xu et al., 2020; Wu et al., 2020; Mao et al., 2021; Chen et al., 2021), aspectcategory-sentiment (ACS) triple extraction (Wan et al., 2020), etc. Restaurant-ACOS Laptop-ACOS #Categories 13 121 #Sentences 2286 4076 #Quadruples EA & EO 2429 (66.40%) 3269 (56.77%) IA & EO 530 (14.49%) 910 (15.80%) EA & IO 350 (9.57%) 1237 (21.48%) IA & IO 349 (9.54%) 342 (5.94%) All 3658 5758 #Quadruples #Sentences 1.60 1.42 Table 2: Statistics of our two ACOS Quadruple datasets. EA, EO, IA and IO denote explicit aspect, explicit opinion, implicit aspect, and implicit opinion, respectively. #Categories represents the number of aspect categories which are consistent with that in (Pontiki et al., 2016). 3.2 Annotation The SemEval 2016 Restaurant dataset (Pontiki et al., 2016) was annotated with explicit and implicit aspects, categories, and sentiment. (Fan et al., 2019; Xu et al., 2020) further added the opinion annotations. We integrate their annotations to construct aspect-category-opinion-sentiment quadruples and further annotate the implicit opinions. For Laptop-ACOS, we annotate the four elements and their corresponding quadruples all by ourselves. We employ the aspect categories defined in the SemEval 2016 Laptop dataset. Two PhD students familiar with aspect-based sentiment analysis are selected as annotators for independent annotation with the annotation tool introduced by (Yang et al., 2017a). The strict quadruple matching F1 score between two annotators is 75.86%, which indicates a substantial agreement between two annotators (Kim and Klinger, 2018). In case of disagreement, a third expert will be asked to make the final decision. 3.3 Statistics and Analysis The basic statistics of the two datasets are reported in Table 2. The Restaurant-ACOS dataset contains 2286 sentences with 3658 quadruples, and the Laptop-ACOS dataset contains 4076 sentences with 5758 quadruples. As we have mentioned, a large percentage of the quadruples contain implicit aspects or implicit opinions. By comparing two datasets, it can be observed that Laptop-ACOS has higher percentage of implicit opinions than Restaurant-ACOS. In Table 3, we further compare our two ACOS datasets with the existing representative datasets 343 Sentence Aspect Category Opinion Sentiment AS AO AOS ACS ACOS Pair Pair Triple Triple Quadruple Restaurant-2014 (Pontiki et al., 2014) 3841 4827 4738 4534 4827 Laptop-2014 (Pontiki et al., 2014) 1910 3012 3012 3012 Restaurant-2016 (Pontiki et al., 2016) 2295 3122 3001 3122 3182 3364 Laptop-2016 (Pontiki et al., 2016) 2612 3705 3705 Restaurant-2014-AO (Fan et al., 2019) 2125 3503 3610 4092 Restaurant-2016-AO (Fan et al., 2019) 1407 1968 2146 2294 Restaurant-2014-AOS (Xu et al., 2020) 2068 3399 3443 3399 3399 3908 3908 Restaurant-2016-AOS (Xu et al., 2020) 1393 1946 2101 1946 1946 2247 2247 Restaurant-ACOS (ours) 2286 3110 2967 3335 3110 3155 3571 3575 3335 3658 Laptop-ACOS (ours) 4076 4958 4992 5378 4958 5035 5726 5731 5227 5758 Table 3: The comparison between the sizes of our two ACOS Quadruple datasets and existing representative ABSA datasets. AS, AO, AOS, and ACS denote Aspect-Sentiment, Aspect-Opinion, Aspect-Opinion-Sentiment, and Aspect-Category-Sentiment, respectively. in ABSA. Restaurant 2014/2016 and Laptop 2014/2016 denote the SemEval 2014/2016 Restaurant and Laptop datasets, respectively. Restaurant 2014/2016 contains the annotations of aspect, category and sentiment. It should be noted the category definitions in two datasets are different. Laptop 2014 contains only the annotations of aspect and sentiment, while Laptop 2016 contains only the annotations of category and sentiment. Restaurant-2014-AO and Restaurant-2016-AO are two aspect-opinion pair datasets proposed by (Fan et al., 2019), based on Restaurant 2014 and 2016, respectively. They removed the sentences with implicit aspects and added the opinion annotations. (Xu et al., 2020) further added sentiment which was originally included in Resturant 2014/2016 to Restaurant-2014/2016-AO, and obtained two aspect-opinion-sentiment triple datasets: Restaurant-2014-AOS and Restaurant-2016-AOS. For Restaurant-ACOS, we integrate the above annotations to construct ACOS quadruples. But it should be noted that we keep the sentences with implicit aspects in Restaurant-2016, and further annotate the implicit opinions. As a result, the size (including sentences, AO pairs and AOS triples) of Restaurant-ACOS is about 1.6 times that of Restaurant-2016-AO and Restaurant-2016-AOS. The new Laptop-ACOS has 4076 review sentences. The numbers of annotations for aspect, category, opinion and sentiment are 4958, 4992, 5378 and 4958, respectively. By combining these elements, we construct 5035 AS pairs, 5726 AO pairs, 5731 AOS triples, 5227 ACS triples and 5758 ACOS quadruples, nearly twice the size of Restaurant-ACOS.2 2It is worth noting that the Restaurant-ACOS and Laptop4 Methods We benchmark the ACOS Quadruple Extraction task with four baseline systems, namely, Double-Propagation-ACOS, JET-ACOS, TASBERT-ACOS and Extract-Classify-ACOS, by adapting the representative approaches in aspectopinion pair extraction, aspect-category-opinion triple extraction or aspect-opinion-sentiment triple extraction to ACOS Quadruple Extraction. 4.1 Double-Propagation-ACOS Since Double Propagation (DP) is one of the representative rule-based methods for aspect-opinionsentiment triple extraction (Qiu et al., 2011), we propose to adapt it to our ACOS quadruple extraction task by first extracting all the aspect-opinionsentiment triples, followed by assigning the aspect category for each extracted triple. We name the adapted approach as Double-Propagation-ACOS. Specifically, we first follow the DP algorithm to extract the aspect-opinion-sentiment triples, where we utilize the syntactic relations between aspects and opinions to iteratively extract them in each review, and rely on the sentiment lexicon to assign sentiments (i.e., Positive, Negative, and Neutral) to aspects and opinions in a bootstrapping manner. Second, to identify the aspect category of each extracted triple, we use the following strategy: if the aspect in the triple is in the training set, we take its most co-occurred aspect category as the final aspect category; otherwise, we adopt the aspect ACOS datasets are available for all subtasks in ABSA, including aspect-based sentiment classification, aspect-sentiment pair extraction, aspect-opinion pair extraction, aspect-opinionsentiment triple extraction, aspect-category-sentiment triple extraction, etc. 344 category of the nearest aspect in the input review as the final aspect category. Based on the two steps mentioned above, we can extract the ACOS quadruples in each review sentence. 4.2 JET-ACOS As one of the state-of-the-art approaches for aspectopinion-sentiment triple extraction, JET (Xu et al., 2020) introduced an end-to-end framework to this task, by combining the identification of aspects, their corresponding opinions, and their sentiment polarities with a position-aware tagging scheme3. Similar to Double-Propagation-ACOS, we adapt JET to our task by first extracting the triple with JET, followed by predicting the aspect category for each extracted triple. Specifically, we first obtain the candidate aspectopinion-sentiment triples based on JET, and then design a BERT-based model to get the aspect category of the extracted triples. Given the review sentence r, we first feed it to BERT to get the context-aware token representation H as follows: H =[h[CLS], hr, h[SEP]], (2) where hr = [h1, . . . , hn] is the output representation for r. Next, given an extracted triple a-o-s, we can obtain the representation of the aspect and the opinion as ua = avg(ha) and uo = avg(ho), where avg(ha) and avg(ho) are the average vectors of words in the aspect ha and the opinion ho, respectively. We then concatenate ua and uo, and feed it to a fully-connected layer with the Sigmoid function for each category c: yc = Sigmoid(W ⊤ c [ua; uo] + bc). (3) Given a-o-s and c, yc = 1 indicates a valid quadruple, and yc = 0 indicates an invalid quadruple. In the training stage, we adopt the standard binary cross-entropy loss for optimization. In the inference stage, we combine the extracted aspectopinion-sentiment triples from JET and our predicted aspect categories to get all the quadruples from each review sentence. 3JET contains two variants, i.e., JETt and JETo. JETt aims to identify the aspects, the offset of their corresponding opinions, and their sentiment polarity; whereas JETo aims to identify the opinions, the offset of their corresponding aspects, and their sentiment polarity. We employ JETo to extract the aspect-opinion-sentiment triple, as it has been shown to obtain better performance than JETt. 4.3 TAS-BERT-ACOS TAS-BERT (Wan et al., 2020) is one of the state-ofthe-art method for aspect-category-sentiment triple extraction, which integrates aspect category-based sentiment classification and aspect extraction in a unified framework by attaching the aspect category and the sentiment polarity to the review sentence and using it as the input of BERT. To adapt TASBERT to our ACOS extraction task, we propose to adopt the input transformation strategy in TASBERT to perform category-sentiment conditional aspect-opinion co-extraction, following by filtering out the invalid aspect-opinion pairs to form the final quadruples. Specifically, given a review sentence r, an aspect category c ∈C, and a sentiment s ∈S, the input is constructed as follows: x =[[CLS], r, [SEP], c, s, [SEP]], (4) We then feed x to BERT to get the context-aware token representation H: H =[h[CLS], hr, h[SEP], hcs, h[SEP]], (5) where hr = [h1, . . . , hn] is the output representation for r, hcs is the output representation for the concatenation of c and s, and h[CLS] is used for category-sentiment verification. We then perform aspect-opinion co-extraction over H by modeling it as a single sequence labeling task. Specifically, we employ a modified BeginInside-Outside (BIO) tagging scheme, which consists of five tags: {BA, IA, BO, IO, O}, indicating the beginning and inside of the aspect, the beginning and inside of the opinion, and others. We feed hr to a CRF layer to extract the aspects and opinions in r with respect to the input category c and sentiment s as follows: Y ao = [yao 1 , . . . , yao n ] = CRF(h1, . . . , hn); (6) Next, we perform Cartesian Product on the extracted aspects and opinions to obtain a set of candidate aspect-category-opinion-sentiment quadruples: SACOS = {a1-c1-o1-s1, ..., a|A|-c|C|-o|O|-s|S|}, (7) where |A| and |O| are the number of extracted aspects and opinions, |C| and |S| are the number of detected categories and sentiment. 345 … … … Explicit Aspect Opinion Co-Extraction [CLS] Looks nice, and the surface … to respond. [CLS] BERT ℎ2 … ℎ𝑛−1 Aspect-Opinion Pairing Category-Sentiment Classification Candidate Aspect-Opinion Pairs ℎCLS Implicit Aspect Prediction ℎCLS Implicit Opinion Prediction ℎ1 ℎ𝑛 Candidate Aspects 𝑎|A| 𝑎1 … 𝑜1 Candidate Opinions 𝑜|𝑂| … … … 𝑎1-𝑐𝑗1-𝑜2-𝑠𝑙1 𝑎2-𝑐𝑗2-𝑜1-𝑠𝑙2 … Valid Aspect-Category-Opinion-Sentiment Quadruples … 𝑎1-o1 𝑎1-o2 𝑎1-o|𝑂| 𝑎|𝐴|-o1 𝑎|𝐴|-o2 𝑎|𝐴|-o|𝑂| Figure 2: The Structure of Extract-Classify-ACOS. On the basis of SACOS, we average the vectors of tokens in the aspect and opinion, and then feed their concatenation [ua; uo] to a quadruple filter: yacos = Sigmoid(W ⊤[ua; uo] + b), (8) where yacos = 1 indicates a valid quadruple, and yacos = 0 indicates an invalid quadruple. 4.4 Extract-Classify-ACOS Finally, we propose Extract-Classify-ACOS by adapting one of the representative aspect-opinion co-extraction system (Wang et al., 2017) to our ACOS quadruple extraction task. Specifically, the first step performs aspect-opinion co-extraction, and the second step predicts category-sentiment given the extracted aspect-opinion pairs. As shown in Figure 2, we first insert two [CLS] tokens at the beginning and the end of the review sentence r, and then feed the transformed input to BERT to obtain the context-aware token representations H as follows: H =[h[CLS], hr, h[CLS]], (9) Similar to the method in TAS-BERT-ACOS, the explicit aspect-opinion co-extraction is based on a CRF layer with the modified BIO tagging scheme. Training Validation Testing Restaurant-ACOS 1531 170 585 Laptop-ACOS 2934 326 816 Table 4: The division of training, validation, and testing sets. We further apply two binary classification tasks on the [CLS] tokens to predict whether there is implicit aspect or implicit opinion. Thus, we can obtain the potential aspect set SA, opinion set SO, and perform Cartesian Product on SA and SO to obtain a set of candidate aspect-opinion pairs: SAO = {a1-o1, ..., a|A|-o|O|}. (10) Next, we model the category-sentiment classification as a multiple multi-class classification problem. Specifically, for each category c, we concatenate the average vectors of each aspect-opinion pair a-o, and feed them to a fully-connected layer with Softmax function as follows: saoc = Softmax(W ⊤ aoc[ua; uo] + baoc), (11) where saoc ∈{Positive, Negative, Neutral, Invalid} denotes its sentiment given current a-o and c, or indicates an invalid quadruple. 5 Experiments We evaluate the performance of four baselines systems on two ACOS quadruple datasets. 5.1 Experimental Settings and Evaluation Metrics In Extract-Classify-ACOS, we adopt BERTbase (Devlin et al., 2018) as the basic encoder, which consists of 12 stacked Transformer blocks. During training, we use the AdamW optimizer of BERT with weight decay fix. The maximum length of the review sentence is set to 128, covering all sentences in two datasets. We set the batch size and learning rates in aspect opinion co-extraction and categorysentiment classification as [32, 2e-5] and [16, 3e5], respectively. The dropout rate is set as 0.1. The batch size and learning rate in the category classification of JET-ACOS and the aspect-opinion pair filtering in TAS-BERT-ACOS are all set as [8, 5e-5], other settings of these two modules are the same as Extract-Classify-ACOS. We divide the original dataset into a training set, a validation set and a testing set according to Table 4. 346 Method Restaurant-ACOS Laptop-ACOS P R F1 P R F1 Double-Propagation-ACOS 0.3467 0.1508 0.2104 0.1304 0.0057 0.0800 JET-ACOS 0.5981 0.2894 0.3901 0.4452 0.1625 0.2381 TAS-BERT-ACOS 0.2629 0.4629 0.3353 0.4715 0.1922 0.2731 Extract-Classify-ACOS 0.3854 0.5296 0.4461 0.4556 0.2948 0.3580 Table 5: Main results of the Aspect-Category-Opinion-Sentiment Quadruple Extraction task. Method Restaurant-ACOS Laptop-ACOS EA & EO IA & EO EA & IO IA & IO EA & EO IA & EO EA & IO IA & IO Double-Propagation-ACOS 0.2602 N/A N/A N/A 0.0980 N/A N/A N/A JET-ACOS 0.5230 N/A N/A N/A 0.3570 N/A N/A N/A TAS-BERT-ACOS 0.3360 0.3184 0.1403 0.3976 0.2610 0.4154 0.1090 0.2115 Extract-Classify-ACOS 0.4496 0.3466 0.2386 0.3370 0.3539 0.3900 0.1682 0.1858 Table 6: F1 score on testing subsets with different aspect & opinion types. EA, EO, IA and IO denote explicit aspect, explicit opinion, implicit aspect and implicit opinion, respectively. N/A means the model can not deal with the corresponding type. In evaluation, a quadruple is viewed as correct if and only if the four elements as well as their combination are exactly the same as those in the gold quadruple. On this basis, we calculate the Precision and Recall, and use F1 score as the final evaluation metric for AOCS Quadruple Extraction. 5.2 Main Results Table 5 reports the ACOS quadruple extraction performance of four different systems on the two datasets. It can be seen that Double-PropagationACOS gets the lowest performance. It is reasonable that only using rules is somehow difficult to identify multiple implicit elements and their complex combinations in reviews. JET-ACOS and TAS-BERT-ACOS achieve comparable F1 performance: the former is better on Restaurant-ACOS dataset and the latter is better on Laptop-ACOS. Extract-Classify-ACOS achieves the best performance among four baseline systems. It outperforms JET-ACOS by 5.60 percentage points on Restaurant-ACOS and outperforms TAS-BERTACOS by 8.49 percentage points on Laptop-ACOS, respectively. The main advantage is that ExtractClassify-ACOS can achieve robustly higher recall score. In comparison, JET-ACOS has higher or comparable precision score but its recall is much lower. It is also worth noting that the F1 score of Extract-Classify-ACOS on both datasets are not high (0.4461 and 0.3580). It is reasonable because the evaluation metric is based on exact matching and the ACOS Quadruple Extraction is a more complicated task than the traditional ABSA tasks. 5.3 Effectiveness of Modeling of Implicit Aspects/Opinions As we have mentioned, a large percentage of review sentences contain implicit aspects/opinions. Therefore, efficient modeling of implicit aspects/opinions is of great importance. To investigate the ability of different systems in addressing the implicit aspects/opinion problem, in Table 6 we split the testing set into four subsets and observe the performance on different subsets: 1) EA & EO denotes the subset with explicit aspects and explicit opinions; 2) IA & EO denotes the subset with implicit aspects and explicit opinions; 3) EA & IO denotes the subset with explicit aspects and implicit opinions; 4) IA & IO denotes the subset with both implicit aspects and implicit opinions. Among four systems, Double-PropagationACOS and JET-ACOS can only address EA & EO, while TAS-BERT-ACOS and Extract-ClassifyACOS can support both implicit aspects and implicit opinions. They show comparable ability in modeling the implicit aspects/opinions. ExtractClassify-ACOS is better in case of IA & EO and EA & IO on Restaurant-ACOS, while TAS-BERTACOS is better in case of IA & EO and IA & IO on Laptop-ACOS. But Extract-Classify-ACOS performs significantly better in case of EA & EO on two datasets. We further compare the performance on differ347 Aspect & Opinion Type EA & EO IA & EO EA & IO IA & IO Review Sentence Keyboard is comfortable and screen is sharp. Nice, I ordered this just for simple web browsing and personal use. I noticed the battery went down to 67% for no reason. We waited for an hour to be seated. AS RACL (Chen and Qian, 2020) screen-Pos N/A N/A Pair Keyboard-Pos AO SDRN (Chen et al., 2020a) screen-sharp N/A N/A N/A Pair Keyboard-comfortable ACS TAS-BERT (Wan et al., 2020) screen-Design&Feature-Pos battery-Performance-Neg Triple Keyboard-Usability-Pos AOS JET (Xu et al., 2020) screen-sharp-Pos N/A N/A N/A Triple Keyboard-comfortable-Pos JET-ACOS screen-Performance-sharp-Pos N/A N/A N/A Keyboard-Usability-comfortable-Pos ACOS TAS-BERT-ACOS screen-Design&Feature-sharp-Pos battery-Performance-NULL-Neg NULL-Service-NULL-Neg Quadruple Keyboard-Usability-comfortable-Pos Extract-Classify-ACOS screen-Design&Feature-sharp-Pos NULL-General-Nice-Pos battery-Performance-NULL-Neg NULL-Service-NULL-Neg Keyboard-Usability-comfortable-Pos Table 7: The predictions of some representative approaches in five ABSA tasks on review sentences with different aspect & opinion types. EA, EO, IA and IO denote explicit aspect, explicit opinion, implicit aspect and implicit opinion, respectively. N/A stands for non-available; and denote correct and false predictions, respectively. ent subsets. The result shows that the worst performance is obtained on EA & IO rather than IA & IO. One possible reason is that the categories corresponding to IA & IO are relatively regular than EA & IO, and is easier to predict. 5.4 Case study In Table 7, we further conduct case study by comparing the predictions of some representative approaches on five ABSA tasks including Aspect-Sentiment (AS) Pair extraction, AspectOpinion (AO) Pair extraction, Aspect-CategorySentiment (ACS) Triple extraction, AspectOpinion-Sentiment (AOS) Triple extraction, and ACOS extraction. We choose four different sentences according to whether the aspect/opinion is explicit or implicit, and observe the predictions obtained by different approaches. It can be observed that: 1) RACL (Chen and Qian, 2020) accurately extracts the AS pairs in case of EA & EO, but it does not support implicit aspects and it fails to make predictions in case of EA & IO on our testing sentence; 2) SDRN (Chen et al., 2020a) is only capable of aspect-opinion pair extraction in case of EA & EO; 3) JET (Xu et al., 2020) can only extract aspectopinion-sentiment triples in case of EA & EO; 4) Although TAS-BERT (Wan et al., 2020) supports aspect-category-sentiment triple extraction for either implicit aspect or implicit opinion, it fails to give accurate predictions in case of IA & EO and IA & IO on our testing sentences; 5) As for the three ACOS baseline systems, JET-ACOS is only capable of ACOS quadruple extraction in case of EA & EO, and has a false prediction. TAS-BERT-ACOS and Extract-Classify-ACOS support ACOS quadruple extraction in case of both implicit aspects and implicit opinions. TAS-BERT-ACOS performs better than JET-ACOS but still fails in the case of IA & EO. Extract-Classify-ACOS performs generally the best and produces more accurate predictions in all cases. 6 Related Work Aspect-based sentiment analysis (ABSA) has drawn wide attention during the last decade. As a core task of ABSA, aspect-based sentiment classification (ABSC) which aims to detect the sentiment of a given aspect has been extensively studied in the literature (Jiang et al., 2011; Vo and Zhang, 2015; Tang et al., 2015; Wang et al., 2016b; Tang et al., 2016; Zhang et al., 2016; Yang et al., 2017b; Ma et al., 2017; Zhang et al., 2018; Wang et al., 2018, 2019; Xu et al., 2019; Tang et al., 2020; Chen et al., 2020b). In recent years, on the basis of traditional ABSC, a series of expansion tasks have appeared in this field. We divide these work into the following four categories: Aspect-Sentiment Pair Extraction. It also can be viewed as joint aspect extraction and ABSC. (Mitchell et al., 2013) first explored the opendomain aspect-sentiment extraction task by designing a variety of conditional random fieldbased models based on traditional discrete features. With the recent trend of deep learning, researchers have proposed various neural pipeline approaches (Zhang et al., 2015; Hu et al., 2019) or joint learning approaches for this task (Li et al., 2019; Luo et al., 2019; He et al., 2019; Chen and Qian, 2020). 348 Aspect-Opinion Pair Extraction. (Hu and Liu, 2004) first addressed the task in a pipeline manner. (Chen et al., 2020a) proposed to extract aspectopinion pairs with a double-channel recurrent network while taking the correlation between aspects and opinions into consideration. (Zhao et al., 2020) designed a span-based multi-task learning framework to extract aspect-opinion pairs jointly. The work on aspect-opinion co-extraction (Wang et al., 2016a, 2017; Yu et al., 2018) can be viewed as the first stage of aspect-opinion pair extraction. Aspect-Opinion-Sentiment Triple Extraction. Considering the relation between aspect and opinion, (Hu and Liu, 2004) designed a feature-based opinion summary system, which identifies explicit aspect, opinion and sentiment, and integrates them into review opinion summaries. (Qiu et al., 2011) further proposed a Double Propagation method to utilize the syntactic relations between aspects and opinions to iteratively extract the aspect-opinionsentiment triples. More recently, (Peng et al., 2020) proposed a two-stage framework to first extract aspect-sentiment pairs and opinions separately, followed by matching them to obtain aspect-opinionsentiment triples. (Xu et al., 2020) further proposed an end-to-end position-aware tagging scheme to model the relations among aspect, opinion and sentiment. (Wu et al., 2020) proposed a Grid Tagging Scheme to address this problem. (Mao et al., 2021; Chen et al., 2021) transformed the triple extraction task into multi-turn machine reading comprehension task and achieved state-of-the-art performances. Aspect-Category-Sentiment Triple Extraction. Previous two categories only focus on explicit aspect-based sentiment analysis, while ignoring the implicit aspects. To address this issue, (Liu et al., 2005) designed rule-based method to find the corresponding implicit aspects through the opinion existing in the review sentence. With the recent advances of pre-trained models, (Wan et al., 2020) proposed a BERT-based architecture to address this task in an end-to-end fashion. Since the problem of implicit aspect and implicit opinion has not been systematically addressed in previous studies, in this work we introduce a new task for Aspect-Category-Opinion-Sentiment (ACOS) Quadruple Extraction with implicit aspects and opinions, construct two ACOS Quadruple datasets, and benchmark the task with four baseline systems. 7 Conclusions and Future Work In this paper, we introduce a new task, AspectCategory-Opinion-Sentiment (ACOS) Quadruple Extraction, aiming to systematically address the implicit aspect/opinion problem. We construct two new datasets for this task, with ACOS annotations including implicit aspects and implicit opinions. We finally benchmark the task with four baseline systems. Experiments demonstrate the advantages of the new task in aspect-based sentiment analysis with implicit aspects/opinions. The focus of this paper is the introduction of the new task and datasets. The proposed four baseline systems are relatively simple and leave much room for further improvements. We welcome future work proposing stronger models on this task. We also welcome the usage of our datasets on the other ABSA tasks. Acknowledgments This work was supported by the Natural Science Foundation of China (No. 62076133 and 62006117), and the Natural Science Foundation of Jiangsu Province for Young Scholars (No. BK20200463) and Distinguished Young Scholars (No. BK20200018). References Huan-Yuan Chen and Hsin-Hsi Chen. 2016. Implicit polarity and implicit aspect recognition in opinion mining. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL), pages 20–25. Shaowei Chen, Jie Liu, Yu Wang, Wenzheng Zhang, and Ziming Chi. 2020a. Synchronous doublechannel recurrent network for aspect-opinion pair extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 6515–6524. Shaowei Chen, Yu Wang, Jie Liu, and Yuelin Wang. 2021. Bidirectional machine reading comprehension for aspect sentiment triplet extraction. In Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI), pages 12666–12674. Xiao Chen, Changlong Sun, Jingjing Wang, Shoushan Li, Luo Si, Min Zhang, and Guodong Zhou. 2020b. Aspect sentiment classification with document-level sentiment preference modeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3667–3677. 349 Zhuang Chen and Tieyun Qian. 2020. Relation-aware collaborative learning for unified aspect-based sentiment analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3685–3694. Hongliang Dai and Yangqiu Song. 2019. Neural aspect and opinion term extraction with mined rules as weak supervision. arXiv preprint arXiv:1907.03750. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Zhifang Fan, Zhen Wu, Xinyu Dai, Shujian Huang, and Jiajun Chen. 2019. Target-oriented opinion words extraction with target-fused neural sequence labeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2509–2518. Ruidan He, Wee Sun Lee, Hwee Tou Ng, and Daniel Dahlmeier. 2019. An interactive multi-task learning network for end-to-end aspect-based sentiment analysis. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 504–515. Minghao Hu, Yuxing Peng, Zhen Huang, Dongsheng Li, and Yiwei Lv. 2019. Open-domain targeted sentiment analysis via span-based extraction and classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 537–546. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168–177. Long Jiang, Mo Yu, Ming Zhou, Xiaohua Liu, and Tiejun Zhao. 2011. Target-dependent twitter sentiment classification. In Proceedings of the 49th annual Meeting of the association for computational linguistics (ACL), pages 151–160. Evgeny Kim and Roman Klinger. 2018. Who feels what and why? annotation of a literature corpus with semantic roles of emotions. In Proceedings of the 27th International Conference on Computational Linguistics (COLING), pages 1345–1359. Farek Lazhar and Yamina Tlili Guiyassa. 2016. Mining explicit and implicit opinions from reviews. Int. J. Data Min. Model. Manag., 8:75–92. Xin Li, Lidong Bing, Piji Li, and Wai Lam. 2019. A unified model for opinion target extraction and target sentiment prediction. In Proceedings of the 33rd AAAI Conference on Artificial Intelligence (AAAI), pages 6714–6721. Xin Li, Lidong Bing, Piji Li, Wai Lam, and Zhimou Yang. 2018. Aspect term extraction with history attention and selective transformation. arXiv preprint arXiv:1805.00760. Bing Liu. 2012. Sentiment analysis and opinion mining. Synthesis lectures on human language technologies, pages 1–167. Bing Liu, Minqing Hu, and Junsheng Cheng. 2005. Opinion observer: analyzing and comparing opinions on the web. In Proceedings of the 14th international conference on World Wide Web (WWW), pages 342–351. Huaishao Luo, Tianrui Li, Bing Liu, and Junbo Zhang. 2019. Doer: Dual cross-shared rnn for aspect term-polarity co-extraction. arXiv preprint arXiv:1906.01794. Dehong Ma, Sujian Li, Xiaodong Zhang, and Houfeng Wang. 2017. Interactive attention networks for aspect-level sentiment classification. arXiv preprint arXiv:1709.00893. Yue Mao, Yi Shen, Chao Yu, and Longjun Cai. 2021. A joint training dual-mrc framework for aspect based sentiment analysis. arXiv preprint arXiv:2101.00816. Margaret Mitchell, Jacqui Aguilar, Theresa Wilson, and Benjamin Van Durme. 2013. Open domain targeted sentiment. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1643–1654. Haiyun Peng, Lu Xu, Lidong Bing, Fei Huang, Wei Lu, and Luo Si. 2020. Knowing what, how and why: A near complete solution for aspect-based sentiment analysis. In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI), pages 8600– 8607. Maria Pontiki, Dimitrios Galanis, Haris Papageorgiou, Ion Androutsopoulos, Suresh Manandhar, Mohammad Al-Smadi, Mahmoud Al-Ayyoub, Yanyan Zhao, Bing Qin, Orph´ee De Clercq, et al. 2016. Semeval-2016 task 5: Aspect based sentiment analysis. In International workshop on semantic evaluation, pages 19–30. Maria Pontiki, Dimitris Galanis, John Pavlopoulos, Harris Papageorgiou, Ion Androutsopoulos, and Suresh Manandhar. 2014. SemEval-2014 task 4: Aspect based sentiment analysis. In Proceedings of the 8th International Workshop on Semantic Evaluation (SemEval 2014), pages 27–35, Dublin, Ireland. Association for Computational Linguistics. Soujanya Poria, Erik Cambria, Lun-Wei Ku, Chen Gui, and Alexander Gelbukh. 2014. A rule-based approach to aspect extraction from product reviews. In Proceedings of the second workshop on natural language processing for social media (SocialNLP), pages 28–37. 350 Guang Qiu, Bing Liu, Jiajun Bu, and Chun Chen. 2011. Opinion word expansion and target extraction through double propagation. Computational linguistics, 37(1):9–27. Duyu Tang, Bing Qin, Xiaocheng Feng, and Ting Liu. 2015. Effective lstms for targetdependent sentiment classification. arXiv preprint arXiv:1512.01100. Duyu Tang, Bing Qin, and Ting Liu. 2016. Aspect level sentiment classification with deep memory network. arXiv preprint arXiv:1605.08900. Hao Tang, Donghong Ji, Chenliang Li, and Qiji Zhou. 2020. Dependency graph enhanced dualtransformer structure for aspect-based sentiment classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 6578–6588. Duy-Tin Vo and Yue Zhang. 2015. Target-dependent twitter sentiment classification with rich automatic features. In Proceedings of the 24th International Joint Conference on Artificial Intelligence (IJCAI), pages 1347–1353. Hai Wan, Yufei Yang, Jianfeng Du, Yanan Liu, Kunxun Qi, and Jeff Z Pan. 2020. Target-aspect-sentiment joint detection for aspect-based sentiment analysis. In Proceedings of the 34th AAAI Conference on Artificial Intelligence (AAAI), pages 9122–9129. Jingjing Wang, Changlong Sun, Shoushan Li, Xiaozhong Liu, Luo Si, Min Zhang, and Guodong Zhou. 2019. Aspect sentiment classification towards question-answering with reinforced bidirectional attention network. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3548–3557. Shuai Wang, Sahisnu Mazumder, Bing Liu, Mianwei Zhou, and Yi Chang. 2018. Target-sensitive memory networks for aspect sentiment classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (ACL), pages 957–967. Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, and Xiaokui Xiao. 2016a. Recursive neural conditional random fields for aspect-based sentiment analysis. arXiv preprint arXiv:1603.06679. Wenya Wang, Sinno Jialin Pan, Daniel Dahlmeier, and Xiaokui Xiao. 2017. Coupled multi-layer attentions for co-extraction of aspect and opinion terms. In Proceedings of the 31st AAAI Conference on Artificial Intelligence (AAAI), pages 3316–3322. Yequan Wang, Minlie Huang, Xiaoyan Zhu, and Li Zhao. 2016b. Attention-based lstm for aspectlevel sentiment classification. In Proceedings of the 2016 conference on empirical methods in natural language processing (EMNLP), pages 606–615. Zhen Wu, Chengcan Ying, Fei Zhao, Zhifang Fan, Xinyu Dai, and Rui Xia. 2020. Grid tagging scheme for end-to-end fine-grained opinion extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 2576–2585. Hu Xu, Bing Liu, Lei Shu, and S Yu Philip. 2019. Bert post-training for review reading comprehension and aspect-based sentiment analysis. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), pages 2324–2335. Lu Xu, Hao Li, Wei Lu, and Lidong Bing. 2020. Position-aware tagging for aspect sentiment triplet extraction. arXiv preprint arXiv:2010.02609. Jie Yang, Yue Zhang, Linwei Li, and Xingxuan Li. 2017a. Yedda: A lightweight collaborative text span annotation tool. arXiv preprint arXiv:1711.03759. Min Yang, Wenting Tu, Jingxuan Wang, Fei Xu, and Xiaojun Chen. 2017b. Attention based lstm for target dependent sentiment classification. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI), pages 5013–5014. Jianfei Yu, Jing Jiang, and Rui Xia. 2018. Global inference for aspect and opinion terms co-extraction based on multi-task neural networks. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(1):168–177. Lei Zhang, Shuai Wang, and Bing Liu. 2018. Deep learning for sentiment analysis: A survey. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 8(4):e1253. Meishan Zhang, Yue Zhang, and Duy-Tin Vo. 2015. Neural networks for open domain targeted sentiment. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 612–621. Meishan Zhang, Yue Zhang, and Duy-Tin Vo. 2016. Gated neural networks for targeted sentiment analysis. In Proceedings of the 30th AAAI Conference on Artificial Intelligence (AAAI), pages 3087–3093. He Zhao, Longtao Huang, Rong Zhang, Quan Lu, et al. 2020. Spanmlt: A span-based multi-task learning framework for pair-wise aspect and opinion terms extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 3239–3248.
|
2021
|
29
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3746–3757 August 1–6, 2021. ©2021 Association for Computational Linguistics 3746 Surprisal Estimators for Human Reading Times Need Character Models Byung-Doh Oh Christian Clark William Schuler Department of Linguistics The Ohio State University {oh.531, clark.3664, schuler.77}@osu.edu Abstract While the use of character models has been popular in NLP applications, it has not been explored much in the context of psycholinguistic modeling. This paper presents a character model that can be applied to a structural parser-based processing model to calculate word generation probabilities. Experimental results show that surprisal estimates from a structural processing model using this character model deliver substantially better fits to self-paced reading, eye-tracking, and fMRI data than those from large-scale language models trained on much more data. This may suggest that the proposed processing model provides a more humanlike account of sentence processing, which assumes a larger role of morphology, phonotactics, and orthographic complexity than was previously thought. 1 Introduction and Related Work Expectation-based theories of sentence processing (Hale, 2001; Levy, 2008) posit that processing difficulty is determined by predictability in context. In support of this position, predictability quantified through surprisal has been shown to correlate with behavioral measures of word processing difficulty (Goodkind and Bicknell, 2018; Hale, 2001; Levy, 2008; Shain, 2019; Smith and Levy, 2013). However, surprisal itself makes no representational assumptions about sentence processing, leaving open the question of how best to estimate its underlying probability model. In natural language processing (NLP) applications, the use of character models has been popular for several years (Al-Rfou et al., 2019; Kim et al., 2016; Lee et al., 2017). Character models have been shown not only to alleviate problems with out-of-vocabulary words but also to embody morphological information available at the subword level. For this reason, they have been extensively used to model morphological processes (Elsner et al., 2019; Kann and Schütze, 2016) or incorporate morphological information into models of syntactic acquisition (Jin et al., 2019). Nonetheless, the use of character models has been slow to catch on in psycholinguistic surprisal estimation, which has recently focused on evaluating largescale language models that make predictions at the word level (e.g. Futrell et al. 2019; Goodkind and Bicknell 2018; Hale et al. 2018; Hao et al. 2020). This raises the question of whether incorporating character-level information into an incremental processing model will result in surprisal estimates that better characterize predictability in context. To answer this question, this paper presents a character model that can be used to estimate word generation probabilities in a structural parser-based processing model.1 The proposed model defines a process of generating a word from an underlying lemma and a morphological rule, which allows the processing model to capture the predictability of a given word form in a fine-grained manner. Regression analyses on self-paced reading, eye-tracking, and fMRI data demonstrate that surprisal estimates calculated from this character-based structural processing model contribute to substantially better fits compared to those calculated from large-scale language models, despite the fact that these other models are trained on much more data and show lower perplexities on test data. This finding deviates from the monotonic relationship between test perplexity and predictive power observed in previous studies (Goodkind and Bicknell, 2018; Wilcox et al., 2020). Furthermore, it suggests that the character-based structural processing model may provide a more humanlike account of processing difficulty and may suggest a larger role of morphology, phonotactics, and orthographic complexity than was previously 1Code for model and experiments is available at https: //github.com/byungdoh/acl21_semproc. 3747 thought. 2 Background The experiments presented in this paper use surprisal predictors (Shannon, 1948) calculated by an incremental processing model based on a leftcorner parser (Johnson-Laird, 1983; van Schijndel et al., 2013). This incremental processing model provides a probabilistic account of sentence processing by making a single lexical attachment decision and a single grammatical attachment decision for each input word. Surprisal. Surprisal can be defined as the negative log ratio of prefix probabilities of word sequences w1..t at consecutive time steps t −1 and t: S(wt) def= −log P(w1..t) P(w1..t−1) (1) These prefix probabilities can be calculated by marginalizing over the hidden states qt of the forward probabilities of an incremental processing model: P(w1..t) = X qt P(w1..t qt) (2) These forward probabilities are in turn defined recursively using a transition model: P(w1..t qt) def= X qt−1 P(wt qt | qt−1) · P(w1..t−1 qt−1) (3) Left-corner parsing. The transition model presented in this paper is based on a probabilistic leftcorner parser (Johnson-Laird, 1983; van Schijndel et al., 2013). Left-corner parsers have been used to model human sentence processing because they define a fixed number of decisions at every time step and also require only a bounded amount of working memory, in keeping with experimental observations of human memory limits (Miller and Isard, 1963). The transition model maintains a distribution over possible working memory store states qt at every time step t, each of which consists of a bounded number D of nested derivation fragments ad t /bd t . Each derivation fragment spans a part of a derivation tree from some apex node ad t lacking a base node bd t yet to come. Previous work has shown that large annotated corpora such as the Penn Treebank (Marcus et al., 1993) do not require more than D = 4 of such fragments (Schuler et al., 2010). At each time step, a left-corner parsing model generates a new word wt and a new store state qt in two phases (see Figure 1). First, it makes a lexical decision ℓt regarding whether to use the word to complete the most recent derivation fragment (match), or to use the word to create a new preterminal node aℓt (no-match). Subsequently, the model makes a grammatical decision gt regarding whether to use a predicted grammar rule to combine the node constructed in the lexical phase aℓt with the next most recent derivation fragment (match), or to use the grammar rule to convert this node into a new derivation fragment agt/bgt (no-match):2 P(wt qt | qt−1) = X ℓt,gt P(ℓt | qt−1) · P(wt | qt−1 ℓt) · P(gt | qt−1 ℓt wt) · P(qt | qt−1 ℓt wt gt) (4) Thus, the parser creates a hierarchically organized sequence of derivation fragments and joins these fragments up whenever expectations are satisfied. In order to update the store state based on the lexical and grammatical decisions, derivation fragments above the most recent nonterminal node are carried forward, and derivation fragments below it are set to null (⊥): P(qt | . . .) def= D Y d′=1 Jad′ t , bd′ t = ad′ t−1, bd′ t−1K if d′ < d Jad′ t , bd′ t = agt, bgtK if d′ = d Jad′ t , bd′ t = ⊥, ⊥K if d′ > d (5) where the indicator function JϕK = 1 if ϕ is true and 0 otherwise, and d = argmaxd′{ad′ t−1,⊥} + 1 − mℓt −mgt. Together, these probabilistic decisions generate the n unary branches and n −1 binary branches of a parse tree in Chomsky normal form for an n-word sentence. 3 Model 3.1 Processing Model The processing model extends the above left-corner parser to maintain lemmatized predicate information by augmenting each preterminal, apex, and base node to consist not only of a syntactic category label cpt, cad t , or cbd t , but also of a binary predicate context vector hpt, had t , or hbd t ∈{0, 1}K+V·K, where K is the size of the set of predicate contexts and V is the maximum valence of any syntactic 2Johnson-Laird (1983) refers to lexical and grammatical decisions as ‘shift’ and ‘predict’ respectively. 3748 a) lexical decision ℓt b) grammatical decision gt ad t−1 bd t−1 wt ⇒ aℓt mℓt = 1 ad t−1 bd t−1 wt ⇒ ad t−1 bd t−1 aℓt mℓt = 0 ad−mℓt t−1 bd−mℓt t−1 aℓt ⇒ agt bgt mgt = 1 ad−mℓt t−1 bd−mℓt t−1 aℓt ⇒ ad−mℓt t−1 bd−mℓt t−1 agt bgt mgt = 0 Figure 1: Left-corner parser operations: a) lexical match (mℓt=1) and no-match (mℓt=0) operations, creating new apex aℓt, and b) grammatical match (mgt=1) and no-match (mgt=0) operations, creating new apex agt and base bgt. category.3 Each 0 or 1 element of this vector represents a unique predicate context, which consists of a ⟨predicate, role⟩pair that specifies the content constraints of a node in a predicate-argument structure. These predicate contexts are obtained by reannotating the training corpus using a generalized categorial grammar of English (Nguyen et al., 2012),4 which is sensitive to syntactic valence and non-local dependencies. Lexical decisions. Each lexical decision of the parser includes a match decision mℓt and decisions about a syntactic category cℓt and a predicate context vector hℓt that together specify a preterminal node pℓt. The probability of generating the match decision and the predicate context vector depends on the base node bd t−1 of the previous derivation fragment (i.e. its syntactic category and predicate context vector). The first term of Equation 4 can therefore be decomposed into the following: P(ℓt | qt−1) = SOFTMAX mℓthℓt ( FFθL[δd⊤, [δ⊤ cbd t−1 , h⊤ bd t−1] EL] ) · P(cℓt | qt−1 mℓt hℓt) (6) where FF is a feedforward neural network, and δi is a Kronecker delta vector consisting of a one at element i and zeros elsewhere. Depth d = argmaxd′{ad′ t−1,⊥} is the number of non-null derivation fragments at the previous time step, and EL is a matrix of jointly trained dense embeddings for each syntactic category and predicate context. The syntactic category and predicate context vector 3The valence of a category is the number of unsatisfied syntactic arguments it has. Separate vectors for syntactic arguments are needed in order to correctly model cases such as passives where syntactic arguments do not align with predicate arguments. 4The predicates in this annotation scheme come from words that have been lemmatized by a set of rules that have been manually written and corrected in order to account for common irregular inflections. together define a complete preterminal node pℓt for use in the word generation model: pℓt def= cbd t−1, hbd t−1+ hℓt if mℓt = 1 cℓt, hℓt if mℓt = 0 (7) and a new apex node aℓt for use in the grammatical decision model: aℓt def= ad t−1 if mℓt = 1 pℓt if mℓt = 0 (8) Grammatical decisions. Each grammatical decision includes a match decision mgt and decisions about a pair of syntactic category labels cgt and c′ gt, as well as a predicate context composition operator ogt, which governs how the newly generated predicate context vector hℓt is propagated through its new derivation fragment agt/bgt. The probability of generating the match decision and the composition operators depends on the base node bd−mℓt t−1 of the previous derivation fragment and the apex node aℓt from the current lexical decision (i.e. their syntactic categories and predicate context vectors). The third term of Equation 4 can accordingly be decomposed into the following: P(gt | qt−1 ℓt wt) = SOFTMAX mgtogt ( FFθG[δd⊤, [δ⊤ c b d−mℓt t−1 , h⊤ b d−mℓt t−1 , δ⊤ caℓt, h⊤ aℓt] EG] ) · P(cgt | qt−1 ℓt wt mgt ogt) · P(c′ gt | qt−1 ℓt wt mgt ogt cgt) (9) where EG is a matrix of jointly trained dense embeddings for each syntactic category and predicate context. The composition operators are associated with sparse composition matrices Aogt which can be used to compose predicate context vectors associated with the apex node agt: agt def= ad−mℓt t−1 if mgt = 1 cgt, Aogthaℓt if mgt = 0 (10) 3749 and sparse composition matrices Bogt which can be used to compose predicate context vectors associated with the base node bgt: bgt def= c′ gt, Bogt[hb d−mℓt t−1 ⊤, haℓt ⊤]⊤if mgt=1 c′ gt, Bogt[0⊤, haℓt ⊤]⊤ if mgt=0 (11) 3.2 Character-based Word Model The baseline version of the word model P(wt | qt−1 ℓt) uses relative frequency estimation with backoff probabilities for out-of-vocabulary words trained using hapax legomena. A character-based test version of this model instead applies a morphological rule rt to a lemma xt to generate an inflected form wt. The set of rules model affixation through string substitution and are inverses of lemmatization rules that are used to derive predicates in the generalized categorial grammar annotation (Nguyen et al., 2012). For example, the rule %ay→%aid can apply to the word say to derive its past tense form said. There are around 600 such rules that account for inflection in Sections 02 to 21 of the Wall Street Journal corpus of the Penn Treebank (Marcus et al., 1993), which includes an identity rule for words in bare form and a ‘no semantics’ rule for generating certain function words. For an observed input word wt, the model first generates a list of ⟨xt, rt⟩pairs that deterministically generate wt. This allows the model to capture morphological regularity and estimate how expected a word form is given its predicted syntactic category and predicate context, which have been generated as part of the preceding lexical decision. In addition, this lets the model hypothesize the underlying morphological structure of out-of-vocabulary words and assign probabilities to them. The second term of Equation 4 can thus be decomposed into the following: P(wt | qt−1 ℓt) = X xt,rt P(xt | qt−1 ℓt) · P(rt | qt−1 ℓt xt) · P(wt | qt−1 ℓt xt rt) (12) The probability of generating the lemma sequence depends on the syntactic category cpℓt and predicate context hℓt resulting from the preceding lexical decision ℓt: P(xt | qt−1 ℓt) = Y i SOFTMAX xt,i ( WX xt,i + bX ) (13) where xt,1, xt,2, ..., xt,I is the character sequence of lemma xt, with xt,1 = ⟨s⟩and xt,I = ⟨e⟩as special start and end characters. WX and bX are respectively a weight matrix and bias vector of a softmax classifier. A recurrent neural network (RNN) calculates a hidden state xt,i for each character from an input vector at that time step and the hidden state after the previous character xt,i−1: xt,i = RNNθX( [δ⊤ cpℓt , h⊤ ℓt, δ⊤ xt,i] EX, x⊤ t,i−1 ) (14) where EX is a matrix of jointly trained dense embeddings for each syntactic category, predicate context, and character. Subsequently, the probability of applying a particular morphological rule to the generated lemma depends on the syntactic category cpℓt and predicate context hℓt from the preceding lexical decision as well as the character sequence of the lemma: P(rt | qt−1 ℓt xt) = SOFTMAX rt ( WR rt,I + bR ) (15) where WR and bR are respectively a weight matrix and bias vector of a softmax classifier. rt,I is the last hidden state of an RNN that takes as input the syntactic category, predicate context, and character sequence of the lemma xt,2, xt,3, ..., xt,I−1 without the special start and end characters: rt,i = RNNθR( [δ⊤ cpℓt , h⊤ ℓt, δ⊤ xt,i] ER, r⊤ t,i−1 ) (16) where ER is a matrix of jointly trained dense embeddings for each syntactic category, predicate context, and character. Finally, as the model calculates probabilities only for ⟨xt, rt⟩pairs that deterministically generate wt, the word probability conditioned on these variables P(wt | qt−1 ℓt xt rt) is deterministic. 4 Experiment 1: Effect of Character Model In order to assess the influence of the characterbased word generation model over the baseline word generation model on the predictive quality of surprisal estimates, linear mixed-effects models containing common baseline predictors and one or more surprisal predictors were fitted to self-paced reading times. Subsequently, a series of likelihood ratio tests were conducted in order to evaluate the relative contribution of each surprisal predictor to regression model fit. 3750 4.1 Response Data The first experiment described in this paper used the Natural Stories Corpus (Futrell et al., 2018), which contains self-paced reading times from 181 subjects that read 10 naturalistic stories consisting of 10,245 tokens. The data were filtered to exclude observations corresponding to sentenceinitial and sentence-final words, observations from subjects who answered fewer than four comprehension questions correctly, and observations with durations shorter than 100 ms or longer than 3000 ms. This resulted in a total of 768,584 observations, which were subsequently partitioned into an exploratory set of 383,906 observations and a held-out set of 384,678 observations. The partitioning allows model selection (e.g. making decisions about predictors and random effects structure) to be conducted on the exploratory set and a single hypothesis test to be conducted on the held-out set, thus eliminating the need for multiple trials correction. All observations were log-transformed prior to model fitting. 4.2 Predictors The baseline predictors commonly included in all regression models are word length measured in characters and index of word position within each sentence.5 In addition to the baseline predictors, surprisal predictors were calculated from two variants of the processing model in which word generation probabilities P(wt | qt−1 ℓt) are calculated using relative frequency estimation (FreqWSurp) and using the character-based model described in Section 3.2 (CharWSurp). Both variants of the processing model were trained on a generalized categorial grammar (Nguyen et al., 2012) reannotation of Sections 02 to 21 of the Wall Street Journal (WSJ) corpus of the Penn Treebank (Marcus et al., 1993). Beam search decoding with a beam size of 5,000 was used to estimate prefix probabilities and surprisal predictors for both variants. To account for the time the brain takes to process and respond to linguistic input, it is standard practice in psycholinguistic modeling to include ‘spillover’ variants of predictors from preceding words (Rayner et al., 1983; Vasishth, 2006). However, as including multiple spillover variants of predictors leads to identifiability issues in mixed5Although unigram surprisal or 5-gram surprisal is also commonly included as a baseline predictor, it was not included in this experiment due to convergence issues. Model comparison χ2 p-value Full vs. No CharWSurp 204.48 0.0001∗∗∗ Full vs. No FreqWSurp 0.024 0.8779 Table 1: Likelihood ratio test evaluating the contribution of CharWSurp and FreqWSurp in predicting selfpaced reading times from the Natural Stories Corpus. effects modeling (Shain and Schuler, 2019), CharWSurp and FreqWSurp were both spilled over by one position. All predictors were centered and scaled prior to model fitting, and all regression models included by-subject random slopes for all fixed effects as well as random intercepts for each word and subject-sentence interaction, following the convention of keeping the random effects structure maximal in psycholinguistic modeling (Barr et al., 2013). 4.3 Likelihood Ratio Testing A total of three linear mixed-effects models were fitted to reading times in the held-out set using lme4 (Bates et al., 2015); the full model included the fixed effects of both CharWSurp and FreqWSurp, and the two ablated models included the fixed effect of either CharWSurp or FreqWSurp. This resulted in two pairs of nested models whose fit could be compared through a likelihood ratio test (LRT). The first LRT tested the contribution of CharWSurp by comparing the fit of the full regression model to that of the regression model without the fixed effect of CharWSurp. Similarly, the second LRT tested the contribution of FreqWSurp by comparing the fit of the full regression model to that of the regression model without its fixed effect. 4.4 Results The results in Table 1 show that the contribution of CharWSurp in predicting reading times is statistically significant over and above that of FreqWSurp (p < 0.0001), while the converse is not significant (p = 0.8779). This demonstrates that incorporating a character-based word generation model to the structural processing model better captures predictability in context, subsuming the effects of the processing model without it. 5 Experiment 2: Comparison to Other Models To further examine the impact of the characterbased word generation model, CharWSurp and Fre3751 qWSurp were evaluated against surprisal predictors calculated from a number of other large-scale pretrained language models and smaller parser-based models. To compare the predictive power of surprisal estimates from different language models on equal footing, we calculated the increase in loglikelihood (∆LL) to a baseline regression model as a result of including a surprisal predictor, following recent work (Goodkind and Bicknell, 2018; Hao et al., 2020). 5.1 Surprisal Estimates from Other Models A total of three pretrained language models were used to calculate surprisal estimates at each word.6 • GLSTMSurp (Gulordava et al., 2018): A twolayer LSTM model trained on ∼80M tokens of the English Wikipedia. • JLSTMSurp (Jozefowicz et al., 2016): A twolayer LSTM model with CNN character inputs trained on ∼800M tokens of the 1B Word Benchmark (Chelba et al., 2014). • GPT2Surp (Radford et al., 2019): GPT-2 XL, a 48-layer decoder-only transformer model trained on the WebText dataset (∼8M web documents). In addition, three incremental parsing models were used to calculate surprisal estimates: • RNNGSurp (Hale et al., 2018; Dyer et al., 2016): An LSTM-based model with explicit phrase structure, trained on Sections 02 to 21 of the WSJ corpus. • vSLCSurp (van Schijndel et al., 2013): A leftcorner parser based on a PCFG with subcategorized syntactic categories (Petrov et al., 2006), trained on a generalized categorial grammar reannotation of Sections 02 to 21 of the WSJ corpus. • JLCSurp (Jin and Schuler, 2020): A neural leftcorner parser based on stack LSTMs (Dyer et al., 2015), trained on Sections 02 to 21 of the WSJ corpus. 5.2 Procedures The set of self-paced reading times from the Natural Stories Corpus after applying the same data exclusion criteria as Experiment 1 provided the response variable for the regression models. In addition to the full dataset, regression models were 6Please refer to the appendix for surprisal calculation, outof-vocabulary handling, and re-initialization procedures. also fitted to a ‘no out-of-vocabulary (No-OOV)’ version of the dataset, in which observations corresponding to out-of-vocabulary words for the LSTM language model with the smallest vocabulary (i.e. Gulordava et al., 2018) were also excluded. This exclusion criterion was included in order to avoid putting the LSTM language models that may have unreliable surprisal estimates for out-of-vocabulary words at an unfair disadvantage. This resulted in a total of 744,607 observations in the No-OOV dataset, which were subsequently partitioned into an exploratory set of 371,937 observations and a held-out set of 372,670 observations. All models were fitted to the held-out set, and all observations were log-transformed prior to model fitting. The predictors included in the baseline linear mixed-effects model were word length, word position in sentence, and unigram surprisal. Unigram surprisal was calculated using the KenLM toolkit (Heafield et al., 2013) with parameters trained on the Gigaword 4 corpus (Parker et al., 2009). In order to calculate the increase in log-likelihood (∆LL) attributable to each surprisal predictor, a ‘full’ linear-mixed effects model, which includes one surprisal predictor on top of the baseline model, was fitted for each surprisal predictor. As with Experiment 1, the surprisal predictors were spilled over by one position. All predictors were centered and scaled prior to model fitting, and all regression models included by-subject random slopes for all fixed effects and random intercepts for each word and subject-sentence interaction. Additionally, in order to examine whether any of the models fail to generalize across domains, their perplexity on the entire Natural Stories Corpus was also calculated. 5.3 Results The results show that surprisal from the characterbased structural model (CharWSurp) made the biggest contribution to model fit compared to surprisal from other models on both full and No-OOV sets of self-paced reading times (Figure 2; the difference between the model with CharWSurp and other models is significant with p < 0.001 by a paired permutation test using by-item errors). The exclusion of OOV words did not make a notable difference in the overall trend of ∆LL across models. This finding, despite the fact that the pretrained language models were trained on much 3752 (a) Baseline LL: -20445.4 (b) Baseline LL: -17485.2 Figure 2: Perplexity measures from each model, and improvements in regression model log-likelihood from including each surprisal estimate on Natural Stories self-paced reading data. larger datasets and also show lower perplexities on test data,7 suggests that this model may provide a more humanlike account of processing difficulty. In other words, accurately predicting the next word alone does not fully explain humanlike processing costs that manifest in self-paced reading times. The analysis of residuals grouped by the lowest base category of the previous time step (cbd t−1) from manual annotations (Shain et al., 2018) shows that the improvement of CharWSurp over GPT2Surp was broad-based across categories (see Figure 3). 6 Experiment 3: Eye-tracking Data In order to examine whether these results generalize to other latency-based measures, linear-mixed effects models were fitted on the Dundee eyetracking corpus (Kennedy et al., 2003) to test the contribution of each surprisal predictor, following similar procedures to Experiment 2. 6.1 Procedures The set of go-past durations from the Dundee Corpus (Kennedy et al., 2003) provided the response 7Perplexity of the parsing models is higher partly because they optimize for a joint distribution over words and trees. Figure 3: Residual error from the regression model with GPT2Surp and change in error from the regression model with CharWSurp. Circle widths show the frequency of each syntactic category in the Natural Stories self-paced reading data. variable for the regression models. The Dundee Corpus contains gaze durations from 10 subjects that read 20 newspaper editorials consisting of 51,502 tokens. The data were filtered to exclude unfixated words, words following saccades longer than four words, and words at starts and ends of sentences, screens, documents, and lines. This resulted in the full set with a total of 195,296 observations, which were subsequently partitioned into an exploratory set of 97,391 observations and a held-out set of 97,905 observations. As with Experiment 2, regression models were also fitted to a No OOV version of the dataset, in which observations corresponding to out-of-vocabulary words for the Gulordava et al. (2018) model were also excluded. This resulted in a subset with a total of 184,894 observations (exploratory set of 92,272 observations, held-out set of 92,622 observations). All models were fitted to the held-out set, and all observations were log-transformed prior to model fitting. The predictors included in the baseline linear mixed-effects models were word length, word position, and saccade length. In order to calculate the increase in log-likelihood from including each surprisal predictor, a full model including one sur3753 (a) Baseline LL: -65100.6 (b) Baseline LL: -60807.5 Figure 4: Perplexity measures from each model, and improvements in regression model log-likelihood from including each surprisal estimate on Dundee eyetracking data. prisal predictor on top of the baseline model was fitted for each surprisal predictor. All surprisal predictors were spilled over by one position, and all predictors were centered and scaled prior to model fitting. All regression models included by-subject random slopes for all fixed effects and random intercepts for each word and sentence. 6.2 Results The results in Figure 4 show that as with Experiment 2, surprisal from the character-based structural model (CharWSurp) made the biggest contribution to model fit on both full and No-OOV sets of go-past durations (the difference between model with CharWSurp and other models is significant with p < 0.001 by a paired permutation test using by-item errors). In contrast to Natural Stories, surprisal from the two left-corner parsing models (i.e. vSLCSurp and JLCSurp) did not contribute to as much model fit compared to other models. The exclusion of OOV words again did not make a notable difference in the general trend across different models, although it led to an increase in ∆LL for GLSTMSurp and RNNGSurp. Residuals grouped by the lowest base category from the previous time Figure 5: Residual error from the regression model with GPT2Surp and change in error from the regression model with CharWSurp. Circle widths show the frequency of each syntactic category in the Dundee eyetracking data. step show that, similarly to Natural Stories, the improvement of CharWSurp over GPT2Surp was broad-based across different categories (see Figure 5). These results provide further support for the observation that language models that are trained to predict the next word accurately do not fully explain processing cost in the form of latency-based measures. 7 Experiment 4: fMRI Data Finally, to examine whether a similar tendency is observed in brain responses, we analyzed the time series of blood oxygenation level-dependent (BOLD) signals in the language network, which were identified using functional magnetic resonance imaging (fMRI). To this end, the novel statistical framework of continuous-time deconvolutional regression (CDR; Shain and Schuler, 2019) was employed. As CDR allows the data-driven estimation of continuous impulse response functions from variably spaced linguistic input, it is more appropriate for modeling fMRI responses, which are typically measured in fixed time intervals. Similarly to the previous experiments, the increase in CDR model log-likelihood as a result of including a 3754 surprisal predictor on top of a baseline CDR model was calculated for evaluation. 7.1 Procedures This experiment used the same fMRI data used by Shain et al. (2019), which were collected from 78 subjects that listened to a recorded version of the Natural Stories Corpus. The functional regions of interest (fROI) corresponding to the domainspecific language network were identified for each subject based on the results of a localizer task that they conducted. This resulted in a total of 202,295 observations, which were subsequently partitioned into an exploratory set of 100,325 observations and a held-out set of 101,970 observations by assigning alternate 60-second intervals of BOLD series to different partitions for each participant. All models were fitted to the BOLD signals in the held-out set. The predictors included in the baseline CDR model were the index of current fMRI sample within the current scan, unigram surprisal, and the deconvolutional intercept which captures the influence of stimulus timing. Following Shain et al. (2019), the CDR models assumed the twoparameter HRF based on the double-gamma canonical HRF (Lindquist et al., 2009). Furthermore, the two parameters of the HRF were tied across predictors, modeling the assumption that the shape of the blood oxygenation response to neural activity is identical in a given region. However, to allow the HRFs to have differing amplitudes, a coefficient that rescales the HRF was estimated for each predictor. The models also included a by-fROI random effect for the amplitude coefficient and a by-subject random intercept. To calculate the increase in log-likelihood from including each predictor, a full CDR model including the fixed effects of one surprisal predictor was also fitted for each surprisal predictor. All surprisal predictors were included without spillover,8 and all predictors were centered prior to model fitting. 7.2 Results The results in Figure 6 show that surprisal from GPT-2 (GPT2Surp) made the biggest contribution to model fit in comparison to surprisal from other models (difference between model with GPT2Surp and other models significant with p < 0.001 by a paired permutation test using by-item errors). Most 8As CDR estimates continuous HRFs from variably spaced linguistic input, consideration of spillover variants of surprisal predictors was not necessary. (a) Baseline LL: -269825.1 Figure 6: Perplexity measures from each model, and improvements in regression model log-likelihood from including each surprisal estimate on Natural Stories fMRI data. notably, in contrast to self-paced reading times and eye-gaze durations, CharWSurp did not contribute as much to model fit on fMRI data, with a ∆LL lower than those of the LSTM language models. This differential contribution of CharWSurp across datasets suggests that latency-based measures and blood oxygenation levels may capture different aspects of online processing difficulty. 8 Conclusion This paper presents a character model that can be used to estimate word generation probabilities in a structural parser-based processing model. Experiments demonstrate that surprisal estimates calculated from this processing model generally contribute to substantially better fits to human response data than those calculated from large-scale pretrained language models or other incremental parsers. These results add a new nuance to the relationship between perplexity and predictive power reported in previous work (Goodkind and Bicknell, 2018; Wilcox et al., 2020). In addition, they suggest that structural parser-based processing models may provide a more humanlike account of sentence processing, and may suggest a larger role of morphology, phonotactics, and orthographic complexity than was previously thought. Acknowledgments The authors would like to thank the anonymous reviewers for their helpful comments. This work was supported by the National Science Foundation grant #1816891. All views expressed are those of the authors and do not necessarily reflect the views of the National Science Foundation. 3755 Ethical Considerations Experiments presented in this work used datasets from previously published research (Futrell et al., 2018; Kennedy et al., 2003; Marcus et al., 1993; Shain et al., 2019), in which the procedures for data collection and validation are outlined. References Rami Al-Rfou, Do Kook Choe, Noah Constant, Mandy Guo, and Llion Jones. 2019. Character-level language modeling with deeper self-attention. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence, pages 3159–3166. Dale J. Barr, Roger Levy, Christoph Scheepers, and Harry J. Tily. 2013. Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68:255–278. Douglas Bates, Martin Mächler, Ben Bolker, and Steve Walker. 2015. Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1):1– 48. Ciprian Chelba, Tomas Mikolov, Mike Schuster, Qi Ge, Thorsten Brants, and Phillipp Koehn. 2014. One billion word benchmark for measuring progress in statistical language modeling. In Proceedings of INTERSPEECH, pages 2635–2639. Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A. Smith. 2015. Transitionbased dependency parsing with stack long shortterm memory. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 334–343. Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A. Smith. 2016. Recurrent neural network grammars. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 199–209. Micha Elsner, Andrea D. Sims, Alexander Erdmann, Antonio Hernandez, Evan Jaffe, Lifeng Jin, Martha Booker Johnson, Shuan Karim, David L. King, Luana Lamberti Nunes, Byung-Doh Oh, Nathan Rasmussen, Cory Shain, Stephanie Antetomaso, Kendra V. Dickinson, Noah Diewald, Michelle McKenzie, and Symon Stevens-Guille. 2019. Modeling morphological learning, typology, and change: What can the neural sequence-tosequence framework contribute? Journal of Language Modelling, 7(1):53–98. Richard Futrell, Edward Gibson, Harry J. Tily, Idan Blank, Anastasia Vishnevetsky, Steven Piantadosi, and Evelina Fedorenko. 2018. The Natural Stories Corpus. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, pages 76–82. Richard Futrell, Ethan Wilcox, Takashi Morita, Peng Qian, Miguel Ballesteros, and Roger Levy. 2019. Neural language models as psycholinguistic subjects: Representations of syntactic state. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 32– 42. Adam Goodkind and Klinton Bicknell. 2018. Predictive power of word surprisal for reading times is a linear function of language model quality. In Proceedings of the 8th Workshop on Cognitive Modeling and Computational Linguistics, pages 10–18. Kristina Gulordava, Piotr Bojanowski, Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1195–1205. John Hale. 2001. A probabilistic Earley parser as a psycholinguistic model. In Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics on Language Technologies, pages 1–8. John Hale, Chris Dyer, Adhiguna Kuncoro, and Jonathan Brennan. 2018. Finding syntax in human encephalography with beam search. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2727–2736. Yiding Hao, Simon Mendelsohn, Rachel Sterneck, Randi Martinez, and Robert Frank. 2020. Probabilistic predictions of people perusing: Evaluating metrics of language model performance for psycholinguistic modeling. In Proceedings of the 10th Workshop on Cognitive Modeling and Computational Linguistics, pages 75–86. Kenneth Heafield, Ivan Pouzyrevsky, Jonathan H. Clark, and Philipp Koehn. 2013. Scalable modified Kneser-Ney language model estimation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 690–696. Lifeng Jin, Finale Doshi-Velez, Timothy Miller, Lane Schwartz, and William Schuler. 2019. Unsupervised learning of PCFGs with normalizing flow. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2442– 2452. Lifeng Jin and William Schuler. 2020. Memorybounded neural incremental parsing for psycholinguistic prediction. In Proceedings of the 16th International Conference on Parsing Technologies and the IWPT 2020 Shared Task on Parsing into Enhanced Universal Dependencies, pages 48–61. 3756 Philip N. Johnson-Laird. 1983. Mental models: Towards a cognitive science of language, inference, and consciousness. Harvard University Press, Cambridge, MA. Rafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, and Yonghui Wu. 2016. Exploring the limits of language modeling. arXiv. Katharina Kann and Hinrich Schütze. 2016. MED: The LMU system for the SIGMORPHON 2016 shared task on morphological reinflection. In Proceedings of the 14th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 62–70. Alan Kennedy, Robin Hill, and Joël Pynte. 2003. The Dundee Corpus. In Proceedings of the 12th European conference on eye movement. Yoon Kim, Yacine Jernite, David Sontag, and Alexander M. Rush. 2016. Character-aware neural language models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pages 2741– 2749. Jason Lee, Kyunghyun Cho, and Thomas Hofmann. 2017. Fully character-level neural machine translation without explicit segmentation. Transactions of the Association for Computational Linguistics, 5:365–378. Roger Levy. 2008. Expectation-based syntactic comprehension. Cognition, 106(3):1126–1177. Martin A. Lindquist, Ji Meng Loh, Lauren Y. Atlas, and Tor D. Wager. 2009. Modeling the hemodynamic response function in fMRI: Efficiency, bias and mismodeling. NeuroImage, 45(1, Supplement 1):S187– S198. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. George A. Miller and Stephen Isard. 1963. Some perceptual consequences of linguistic rules. Journal of Verbal Learning and Verbal Behavior, 2(3):217– 228. Luan Nguyen, Marten van Schijndel, and William Schuler. 2012. Accurate unbounded dependency recovery using generalized categorial grammars. In Proceedings of the 24th International Conference on Computational Linguistics, pages 2125–2140. Robert Parker, David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2009. English Gigaword LDC2009T13. Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 2006. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 433–440. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Technical Report. Keith Rayner, Marcia Carlson, and Lyn Frazier. 1983. The interaction of syntax and semantics during sentence processing: Eye movements in the analysis of semantically biased sentences. Journal of verbal learning and verbal behavior, 22(3):358–374. Marten van Schijndel, Andy Exley, and William Schuler. 2013. A model of language processing as hierarchic sequential prediction. Topics in Cognitive Science, 5(3):522–540. William Schuler, Samir AbdelRahman, Tim Miller, and Lane Schwartz. 2010. Broad-coverage incremental parsing using human-like memory constraints. Computational Linguistics, 36(1):1–30. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1715–1725. Cory Shain. 2019. A large-scale study of the effects of word frequency and predictability in naturalistic reading. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4086–4094. Cory Shain, Idan Asher Blank, Marten van Schijndel, William Schuler, and Evelina Fedorenko. 2019. fMRI reveals language-specific predictive coding during naturalistic sentence comprehension. Neuropsychologia, 138. Cory Shain, Marten van Schijndel, and William Schuler. 2018. Deep syntactic annotations for broad-coverage psycholinguistic modeling. In Workshop on Linguistic and Neuro-Cognitive Resources (LREC 2018). Cory Shain and William Schuler. 2019. ContinuousTime Deconvolutional Regression for Psycholinguistic Modeling. PsyArXiv. Claude Elwood Shannon. 1948. A mathematical theory of communication. Bell System Technical Journal, 27:379–423. Nathaniel J. Smith and Roger Levy. 2013. The effect of word predictability on reading time is logarithmic. Cognition, 128:302–319. Mitchell Stern, Daniel Fried, and Dan Klein. 2017. Effective inference for generative neural parsing. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1695–1700. 3757 Shravan Vasishth. 2006. On the proper treatment of spillover in real-time reading studies: Consequences for psycholinguistic theories. In Proceedings of the International Conference on Linguistic Evidence, pages 96–100. Ethan Gotlieb Wilcox, Jon Gauthier, Jennifer Hu, Peng Qian, and Roger P. Levy. 2020. On the predictive power of neural language models for human realtime comprehension behavior. In Proceedings of the 42nd Annual Meeting of the Cognitive Science Society, pages 1707–1713. A Procedures for Surprisal Calculation • GLSTMSurp, JLSTMSurp: These models directly estimate P(wt | w1..t−1), which can be used to calculate S(wt) = −log P(wt | w1..t−1). • GPT2Surp: Since GPT-2 relies on byte-pair encoding (Sennrich et al., 2016), negative log probabilities of word pieces corresponding to wt were added together to calculate S(wt) = −log P(wt | w1..t−1). • RNNGSurp: Since the generative RNNG model defines a joint distribution over words and trees, we marginalize over trees to calculate P(wt | w1..t−1). To keep this tractable, a wordsynchronous beam search (Stern et al., 2017) was used with beam size 100, fast-track beam size 5, and word beam size 10. • vSLCSurp, JLCSurp: Beam search decoding with a beam size of 5,000 and 2,000 respectively was used to estimate prefix probabilities and surprisal predictors. B Procedures for Out-of-vocabulary Handling • GLSTMSurp, JLSTMSurp, JLCSurp: Out-ofvocabulary (OOV) words in the test corpus were replaced with a corresponding “UNK” symbol prior to surprisal estimation. • GPT2Surp: Special OOV handling was not necessary because GPT-2 uses byte-pair encoding (Sennrich et al., 2016). • RNNGSurp, vSLCSurp: Mapping rules from the Berkeley parser9 were used to replace OOV words with a set of unknown word classes (e.g. “UNK-LC-ing”). 9https://github.com/slavpetrov/ berkeleyparser C Procedures for Hidden State Re-initialization • GLSTMSurp, JLSTMSurp, GPT2Surp: The hidden states of these models were re-initialized at the end of every article before making predictions on the next article. • RNNGSurp, vSLCSurp, JLCSurp: Since these models predict parsing operations while making word predictions, their hidden states were reinitialized after each sentence.
|
2021
|
290
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3758–3769 August 1–6, 2021. ©2021 Association for Computational Linguistics 3758 CogAlign: Learning to Align Textual Neural Representations to Cognitive Language Processing Signals Yuqi Ren and Deyi Xiong ∗ College of Intelligence and Computing, Tianjin University, Tianjin, China {ryq20, dyxiong}@tju.edu.cn Abstract Most previous studies integrate cognitive language processing signals (e.g., eye-tracking or EEG data) into neural models of natural language processing (NLP) just by directly concatenating word embeddings with cognitive features, ignoring the gap between the two modalities (i.e., textual vs. cognitive) and noise in cognitive features. In this paper, we propose a CogAlign approach to these issues, which learns to align textual neural representations to cognitive features. In CogAlign, we use a shared encoder equipped with a modality discriminator to alternatively encode textual and cognitive inputs to capture their differences and commonalities. Additionally, a text-aware attention mechanism is proposed to detect task-related information and to avoid using noise in cognitive features. Experimental results on three NLP tasks, namely named entity recognition, sentiment analysis and relation extraction, show that CogAlign achieves significant improvements with multiple cognitive features over state-of-the-art models on public datasets. Moreover, our model is able to transfer cognitive information to other datasets that do not have any cognitive processing signals. The source code for CogAlign is available at https://github. com/tjunlp-lab/CogAlign.git. 1 Introduction Cognitive neuroscience, from a perspective of language processing, studies the biological and cognitive processes and aspects that underlie the mental language processing procedures in human brains while natural language processing (NLP) teaches machines to read, analyze, translate and generate human language sequences (Muttenthaler et al., 2020). The commonality of language processing shared by these two areas forms the base of ∗Corresponding author cognitively-inspired NLP, which uses cognitive language processing signals generated by human brains to enhance or probe neural models in solving a variety of NLP tasks, such as sentiment analysis (Mishra et al., 2017; Barrett et al., 2018), named entity recognition (NER) (Hollenstein and Zhang, 2019), dependency parsing (Strzyz et al., 2019), relation extraction (Hollenstein et al., 2019a), etc. In spite of the success of cognitively-inspired NLP in some tasks, there are some issues in the use of cognitive features in NLP. First, for the integration of cognitive processing signals into neural models of NLP tasks, most previous studies have just directly concatenated word embeddings with cognitive features from eye-tracking or EEG, ignoring the huge differences between these two types of representations. Word embeddings are usually learned as static or contextualized representations of words in large-scale spoken or written texts generated by humans. In contrast, cognitive language processing signals are collected by specific medical equipments, which record the activity of human brains during the cognitive process of language processing. These cognitive processing signals are usually assumed to represent psycholinguistic information (Mathias et al., 2020) or cognitive load (Antonenko et al., 2010). Intuitively, information in these two types of features (i.e., word embeddings and cognitive features) is not directly comparable to each other. As a result, directly concatenating them could be not optimal for neural models to solve NLP tasks. The second issue with the incorporation of cognitive processing signals into neural models of NLP is that not all information in cognitive processing signals is useful for NLP. The recorded signals contain information covering a wide variety of cognitive processes, particularly for EEG (Williams et al., 2019; Eugster et al., 2014). For different tasks, we may need to detect elements in the recorded signals, 3759 Figure 1: Neural Architecture of the proposed CogAlign. For inference, only the components in the red dashed box are used. which are closely related to specific NLP tasks, and neglect features that are noisy to the tasks. In order to address the two issues, we propose CogAlign, a multi-task neural network that learns to align neural representations of texts to cognitive processing signals, for several NLP tasks. As shown in Figure 1, instead of simply concatenating cognitive features with word embeddings, we use two private encoders to separately encode cognitive processing signals and word embeddings. The two encoders will learn task-specific representations for cognitive and textual inputs in two disentangled spaces. To align the representations of neural network with cognitive processing signals, we further introduce an additional encoder that is shared by both data sources. We alternatively feed cognitive and textual inputs into the shared encoder and force it to minimize an adversarial loss of the discriminator stacked over the shared encoder. The discriminator is task-agnostic so that it can focus on learning both differences and deep commonalities between neural representations of cognitive and textual features in the shared encoder. We want the shared encoder to be able to transfer knowledge of cognitive language processing signals to other datasets even if cognitive processing signals are not available for those datasets. Therefore, CogAlign does not require cognitive processing signals as inputs during inference. Partially inspired by the attentive pooling network (Santos et al., 2016), we propose a text-aware attention mechanism to further align textual inputs and cognitive processing signals at the word level. The attention network learns a compatibility matrix of textual inputs to cognitive processing signals. The learned text-aware representations of cognitive processing signals also help the model to detect task-related information and to avoid using other noisy information contained in cognitive processing signals. In a nutshell, our contributions are listed as follows: • We present CogAlign that learns to align neural representations of natural language to cognitive processing signals at both word and sentence level. Our analyses show that it can learn task-related specific cognitive processing signals. • We propose a text-aware attention mechanism that extracts useful cognitive information via a compatibility matrix. • With the adversarially trained shared encoder, CogAlign is capable of transferring cognitive knowledge into other datasets for the same task, where no recorded cognitive processing signals are available. • We conduct experiments on incorporating eyetracking and EEG signals into 3 different NLP tasks: NER, sentiment analysis and relation extraction, which show CogAlign achieves new state-of-the-art results and significant improvements over strong baselines. 3760 2 Related Work Eye-tracking for NLP. Eye-tracking data have proved to be associated with language comprehension activity in human brains by numerous research in neuroscience (Rayner, 1998; Henderson and Ferreira, 1993). In cognitively motivated NLP, several studies have investigated the impact of eye-tracking data on NLP tasks. In early works, these signals have been used in machine learning approaches to NLP tasks, such as part-of-speech tagging (Barrett et al., 2016), multiword expression extraction (Rohanian et al., 2017), syntactic category prediction (Barrett and Søgaard, 2015). In neural models, eyetracking data are combined with word embeddings to improve various NLP tasks, such as sentiment analysis (Mishra et al., 2017) and NER (Hollenstein and Zhang, 2019). Eye-tracking data have also been used to enhance or constrain neural attention in (Barrett et al., 2018; Sood et al., 2020b,a; Takmaz et al., 2020). EEG for NLP. Electroencephalography (EEG) measures potentials fluctuations caused by the activity of neurons in cerebral cortex. The exploration of EEG data in NLP tasks is relatively limited. Chen et al. (2012) improve the performance of automatic speech recognition (ASR) by using EEG signals to classify the speaker’s mental state. Hollenstein et al. (2019a) incorporate EEG signals into NLP tasks, including NER, relation extraction and sentiment analysis. Additionally, Muttenthaler et al. (2020) leverage EEG features to regularize attention on relation extraction. Adversarial Learning. The concept of adversarial training originates from the Generative Adversarial Nets (GAN) (Goodfellow et al., 2014) in computer vision. Since then, it has been also applied in NLP (Denton et al., 2015; Ganin et al., 2016). Recently, a great variety of studies attempt to introduce adversarial training into multi-task learning in NLP tasks, such as Chinese NER (Cao et al., 2018), crowdsourcing learning (Yang et al., 2018), cross-lingual transfer learning (Chen et al., 2018; Kim et al., 2017), just name a few. Different from these studies, we use adversarial learning to deeply align cognitive modality to textual modality at the sentence level. 3 CogAlign CogAlign is a general framework for incorporating cognitive processing signals into various NLP tasks. The target task can be specified at the predictor layer with corresponding task-specific neural network. CogAlign focuses on aligning cognitive processing signals to textual features at the word and encoder level. The text-aware attention aims at learning task-related useful cognitive information (thus filtering out noises) while the shared encoder and discriminator collectively learns to align representations of cognitive processing signals to those of textual inputs in a unified semantic space. The matched neural representations can be transferred to another datasets of the target task even though cognitive processing signals is not present. The neural architecture of CogAlign is visualized in Figure 1. We will elaborate the components of model in the following subsections. 3.1 Input Layer The inputs to our model include textual word embeddings and cognitive processing signals. Word Embeddings. For a given word xi from the dataset of a target NLP task (e.g., NER), we obtain the vector representation hword i by looking up a pre-trained embedding matrix. The obtained word embeddings are fixed during training. For NER, previous studies have shown that character-level features can improve the performance of sequence labeling (Lin et al., 2018). We therefore apply a character-level CNN framework (Chiu and Nichols, 2016; Ma and Hovy, 2016) to capture the characterlevel embedding. The word representation of word xi in NER task is the concatenation of word embedding and character-level embedding. Cognitive Processing Signals. For cognitive inputs, we can obtain word-level eye-tracking and EEG via data preprocessing (see details in Section 5.1). Thus, for each word xi, we employ two cognitive processing signals heye i and heeg i . The cognitive input hcog i can be either a single type of signal or a concatenation of different cognitive processing signals. 3.2 Text-Aware Attention As not all information contained in cognitive processing signals is useful for the target NLP task, we propose a text-aware attention mechanism to assign text sensitive weights to cognitive processing signals. The main process of attention mechanism consists of learning a compatibility matrix between word embeddings Hword ∈Rdw×N and cognitive representations Hcog ∈Rdc×N from the input 3761 layer and preforming cognitive-wise max-pooling operation over the matrix. The compatibility matrix G ∈Rdw×dc can be computed as follows: G = tanh(HwordUHcogT ) (1) where dw and dc are the dimension of word embeddings and cognitive representations, respectively, N is the length of the input, and U ∈RN×N is a trainable parameter matrix. We then obtain a vector gcog ∈Rdc, which is computed as the importance score for each element in the cognitive processing signals with regard to the word embeddings, by row-wise max-pooling over G. Finally, we compute attention weights and the text-aware representation of cognitive processing signals Hcog′ as follows: αcog = softmax(gcog) (2) Hcog′ = αcogHcog (3) 3.3 Encoder Layer We adopt Bi-LSTMs to encode both cognitive and textual inputs following previous works (Hollenstein and Zhang, 2019; Hollenstein et al., 2019a). In this work, we employ two private Bi-LSTMs and one shared Bi-LSTM as shown in Figure 1, where private Bi-LSTMs are used to encode cognitive and textual inputs respectively and the shared Bi-LSTM is used for learning shared semantics of both types of inputs. We concatenate the outputs of private Bi-LSTMs and shared Bi-LSTM as input to the task-specific predictors of subsequent NLP tasks. The hidden states of the shared Bi-LSTM are also fed into the discriminator. 3.4 Modality Discriminator We alternatively feed cognitive and textual inputs into the shared Bi-LSTM encoder. Our goal is that the shared encoder is able to map the representations of the two different sources of inputs into the same semantic space so as to learn the deep commonalities of two modalities (cognitive and textual). For this, we use a self-supervised discriminator to provide supervision for training the shared encoder. Particularly, the discriminator is acted as a classifier to categorize the alternatively fed inputs into either the textual or cognitive input. For the hidden state of modality k, we use a self-attention mechanism to first reduce the dimension of the output of the shared Bi-LSTM Hs k ∈Rdh×N: α = softmax(vT tanh(WsHs k + bs)) (4) hs k = N X i=1 αiHs ki (5) where Ws ∈Rdh×dh, bs ∈Rdh, v ∈Rdh are trainable parameters in the model, hs k is the output of self-attention mechanism. Then we predict the category of the input by softmax function: D(hs k) = softmax(Wdhs k + bd) (6) where D(hs k) is the probability that the shared encoder is encoding an input with modality k. 3.5 Predictor Layer Given a sample X, the final cognitively augmented representation after the encoder layer can be formulated as H ′ = [Hp; Hs] ∈R2dh×N. Hp and Hs are the result of private Bi-LSTM and shared Bi-LSTM, respectively. For sequence labeling tasks like NER, we employ the conditional random field (CRF) (Lafferty et al., 2001) as the predictor as Bi-LSTM-CRF is widely used in many sequence labeling tasks (Ma and Hovy, 2016; Luo et al., 2018) due to the excellent performance and also in cognitively inspired NLP (Hollenstein and Zhang, 2019; Hollenstein et al., 2019a). Firstly, we project the feature representation H ′ onto another space of which dimension is equal to the number of NER tags as follows: oi = Wnh ′ i + bn (7) We then compute the score of a predicted tag sequence y for the given sample X: score(X, y) = N X i=1 (oi,yi + Tyi−1,yi) (8) where T is a transition score matrix which defines the transition probability of two successive labels. Sentiment analysis and relation extraction can be regarded as multi-class classification tasks, with 3 and 11 classes, respectively. For these two tasks, we use a self attention mechanism to reduce the dimension of H ′ and obtain the probability of a predicted class via the softmax function. 3762 4 Training and Inference 4.1 Adversarial Learning In order to learn the deep interaction between cognitive and textual modalities in the same semantic space, we want the shared Bi-LSTM encoder to output representations that can fool the discriminator. Therefore we adopt the adversarial learning strategy. Particularly, the shared encoder acts as the generator that tries to align the textual and cognitive modalities as close as possible so as to mislead the discriminator. The shared encoder and discriminator works in an adversarial way. Additionally, to further increase the difficulty for the discriminator to distinguish modalities, we add a gradient reversal layer (GRL) (Ganin and Lempitsky, 2015) in between the encoder layer and predictor layer. The gradient reversal layer does nothing in the forward pass but reverses the gradients and passes them to the preceding layer during the backward pass. That is, gradients with respect to the adversarial loss ∂LAdv ∂θ are replaced with −∂LAdv ∂θ after going through GRL. 4.2 Training Objective CogAlign is established on a multi-task learning framework, where the final training objective is composed of the adversarial loss LAdv and the loss of the target task LTask. For NER, we exploit the negative log-likelihood objective as the loss function. Given T training examples (Xi; yi)1, LTask is defined as follows: LTask = − T X i=1 logp(yi|Xi) (9) where y denotes the ground-truth tag sequence. The probability of y is computed by the softmax function: p(y|X) = escore(X,y) P ey∈Y escore(X,ey) (10) For sentiment analysis and relation extraction tasks, the task objective is similar to that of NER. The only difference is that the label of the task is changed from a tag sequence to a single class. The adversarial loss LAdv is defined as: LAdv = min θs (max θd K X k=1 Tk X i=1 logD(S(Xi k))) (11) 1X can be either textual or cognitive input as we alternatively feed word embeddings and cognitive processing signals into CogAlign. where θs and θd denote the parameters of the shared Bi-LSTM encoders S and modality discriminator D, respectively, Xi k is the representation of sentence i in a modality k. The joint loss of CogAlign is therefore defined as: L = LTask + LAdv (12) 4.3 Inference After training, the shared encoder learns a unified semantic space for representations of both cognitive and textual modality. We believe that the shared space embeds knowledge from cognitive processing signals. For inference, we therefore only use the textual part and the shared encoder (components in the red dashed box in Figure 1). The private encoder outputs textual-modality-only representations while the shared encoder generates cognitive-augmented representations. The two representations are concatenated to feed into the predictor layer of the target task. This indicates that we do not need cognitive processing signals for the inference of the target task. It also means that we can pretrain CogAlign with cognitive processing signals and then transfer it to other datasets where cognitive processing signals are not available for the same target task. 5 Experiments We conducted experiments on three NLP tasks, namely NER, sentiment analysis and relation extraction with two types of cognitive processing signals (eye-tracking and EEG) to validate the effectiveness of the proposed CogAlign. 5.1 Dataset and Cognitive Processing Signals We chose a dataset2 with multiple cognitive processing signals: Zurich Cognitive Language Processing Corpus (ZuCo) (Hollenstein et al., 2018). This corpus contains simultaneous eye-tracking and EEG signals collected when 12 native English speakers are reading 1,100 English sentences. Word-level signals can be divided by the duration of each word. The dataset includes two reading paradigms: normal reading and task-specific reading where subjects exercise some specific task. In this work, we only used the data of normal reading, since this paradigm accords with human natural reading. The materials for normal reading paradigm 2The data is available here: https://osf.io/q3zws/ 3763 EARLY first fixation duration (FFD) the duration of word w that is first fixated first pass duration (FPD) the sum of the fixations before eyes leave the word w LATE number of fixations (NFIX) the number of times word w that is fixated fixation probability (FP) the probability that word w is fixated mean fixation duration (MFD) the average fixation durations for word w total fixation duration (TFD) the total duration of word w that is fixated n re-fixations (NR) the number of times word w that is fixated after the first fixation re-read probability (RRP) the probability of word w that is fixated more than once CONTEXT total regression-from duration (TRD) the total duration of regressions from word w w-2 fixation probability (w-2 FP) the fixation probability of the word w-2 w-1 fixation probability (w-1 FP) the fixation probability of the word w-1 w+1 fixation probability (w+1 FP) the fixation probability of the word w+1 w+2 fixation probability (w+2 FP) the fixation probability of the word w+2 w-2 fixation duration (w-2 FD) the fixation duration of the word w-2 w-1 fixation duration (w-1 FD) the fixation duration of the word w-1 w+1 fixation duration (w+1 FD) the fixation duration of the word w+1 w+2 fixation duration (w+2 FD) the fixation duration of the word w+2 Table 1: Eye-tracking features used in the NER task. consist of two datasets: 400 movie reviews from Stanford Sentiment Treebank (Socher et al., 2013) with manually annotated sentiment labels, including 123 neutral, 137 negative and 140 positive sentences; 300 paragraphs about famous people from Wikipedia relation extraction corpus (Culotta et al., 2006) labeled with 11 relationship types, such as award, education. We also tested our model on NER task. For NER, the selected 700 sentences in the above two tasks are annotated with three types of entities: PERSON, ORGANIZATION, and LOCATION. All annotated datasets3 are publicly available. The cognitive processing signals and textual features used for each task in this work are the same as (Hollenstein et al., 2019a). Eye-tracking Features. Eye-tracking signals record human gaze behavior while reading. The eye-tracking data of ZuCo are collected by an infrared video-based eye tracker EyeLink 1000 Plus with a sampling rate of 500 Hz. For NER, we used 17 eye-tracking features that cover all stages of gaze behaviors and the effect of context. According to the reading process, these features are divided into three groups: EARLY, the gaze behavior when a word is fixated for the first time; LATE, the gaze behavior over a word that is fixated many times; CONTEXT, the eye-tracking features over neighboring words of the current word. The 17 eyetracking features used in the NER task are shown in the Table 1. In the other two tasks, we employed 5 gaze behaviors, including the first fixation duration (FFD), the number of fixations (NFIX), the total fixation duration (TFD), the first pass duration 3https://github.com/DS3Lab/zuco-nlp/ (FPD), the gaze duration (GD) that is the duration of the first time eyes move to the current word until eyes leave the word. EEG Features. EEG signals record the brain’s electrical activity in the cerebral cortex by placing electrodes on the scalp of the subject. In the datasets we used, EEG signals are recorded by a 128-channel EEG Geodesic Hydrocel system (Electrical Geodesics, Eugene, Oregon) at a sampling rate of 500 Hz with a bandpass of 0.1 to 100 Hz. The original EEG signals recorded are of 128 dimensions. Among them, 23 EEG signals are removed during preprocessing since they are not related to the cognitive processing (Hollenstein et al., 2018). After preprocessing, we obtained 105 EEG signals. The left EEG signals are divided into 8 frequency bands by the frequency of brain’s electrical signals: theta1 (t1, 4-6 Hz), theta2 (t2, 6.5-8 Hz), alpha1 (a1, 8.5-10 Hz), alpha2 (a2, 10.5-13 Hz), beta1 (b1, 13.5-18 Hz), beta2 (b2, 18.5-30 Hz), gamma1 (g1, 30.5-40 Hz) and gamma2 (g2, 40-49.5 Hz). The frequency bands reflects the different functions of brain cognitive processing. For NER, we used 8 EEG features that are obtained by averaging the 105 EEG signals at each frequency band. For the other two tasks, EEG features were obtained by averaging the 105 signals over all frequency bands. All used EEG features are obtained by averaging over all subjects and normalization. 5.2 Settings We evaluated three NLP tasks in terms of precision, recall and F1 in our experiments. Word embeddings of all NLP tasks were initialized with the publicly available pretrained GloVe (Pennington 3764 Signals Model NER Sentiment Analysis Relation Extraction P (%) R (%) F1 (%) P (%) R (%) F1 (%) P (%) R (%) F1 (%) Base∗ 89.34 78.60 83.48 59.47 59.42 58.27 79.52 75.67 75.25 eye (Hollenstein et al., 2019a) 86.2 84.3 85.1 65.1 61.9 62.0 61.4 61.7 61.5 Base 90.56 81.05 85.43∗ 64.26 61.96 61.19∗ 82.01 78.23 77.95∗ Base+TA 90.75 81.77 85.93∗ 64.63 62.71 61.41∗ 83.26 76.47 78.04∗ CogAlign 90.76 82.52 86.41∗ 62.86 64.10 62.30∗ 78.33 82.06 78.56∗ EEG (Hollenstein et al., 2019a) 86.7 81.5 83.9 68.3 64.8 65.1 60.5 60.2 60.3 Base 89.82 80.55 84.76∗ 64.09 60.29 59.79∗ 82.79 77.16 77.61∗ Base+TA 89.54 82.22 85.62∗ 62.20 62.19 60.91∗ 80.83 78.46 77.81∗ CogAlign 89.87 83.08 86.21∗ 63.11 65.38 62.81∗ 77.94 82.60 78.66∗ eye +EEG (Hollenstein et al., 2019a) 85.1 83.2 84.0 66.3 59.3 60.8 59.8 60.0 59.8 Base 89.70 81.11 85.11∗ 62.86 61.49 60.84∗ 79.00 76.52 77.72∗ Base+TA 90.75 82.94 86.31∗ 65.22 63.88 63.23∗ 82.24 77.53 78.12∗ CogAlign 91.28 83.02 86.79∗ 65.11 65.94 65.40∗ 78.66 82.07 78.93∗ Table 2: Results of CogAlign and other methods on the three NLP tasks augmented with eye-tracking features (eye), EEG features (EEG), and both (eye+EEG). ‘Base∗’ denotes that the model does not use any cognitive processing signals. ‘Base’ is a neural model that consist of a textual private encoder and textual predictor, and combines cognitive processing signals with word embeddings via direct concatenation, similar to previous works. ‘Base+TA’ is a neural model where direct concatenation in the base model is replaced by the text-aware attention mechanism. Significance is indicated with the asterisks: * = p<0.01. et al., 2014) vectors of 300 dimensions. For NER, we used 30-dimensional randomly initialized character embeddings. We set the dimension of hidden states of LSTM to 50 for both the private Bi-LSTM and shared Bi-LSTM. We performed 10-fold cross validation for NER and sentiment analysis and 5fold cross validation for relation extraction. 5.3 Baselines We compared our model with previous state-ofthe-art methods on ZuCo dataset. The method by Hollenstein et al. (2019a) incorporates cognitive processing signals into their model via direct concatenation mentioned before. 5.4 Results Results of CogAlign on the three NLP tasks are shown in Table 2. From the table, we observe that: • By just simply concatenating word embeddings with cognitive processing signals, the Base model is better than the model without using any cognitive processing signals, indicating that cognitive processing signals (either eye-tracking or EEG signals) can improve all three NLP tasks. Notably, the improvements gained by eye-tracking features are larger than those obtained by EEG signals while the combination of both does not improve over only using one of them. We conjecture that this may be due to the low signal-to-noise ratio of EEG signals, which further decreases when two signals are combined together. • Compared with the Base model, the Base+TA achieves better results on all NLP tasks. The text-aware attention gains an absolute improvement of 0.88, 2.04, 0.17 F1 on NER, sentiment analysis, and relation extraction, respectively. With Base+TA, the best results for most tasks are obtained by the combination of eye-tracking and EEG signals. This suggests that the proposed text-aware attention may have alleviated the noise problem of cognitive processing signals. • The proposed CogAlign achieves the highest F1 over all three tasks, with improvements of 0.48, 2.17 and 0.87 F1 over Base+TA on NER, sentiment analysis and relation extraction, respectively, which demonstrates the effectiveness of our proposed model. In addition, CogAlign with both cognitive processing signals obtains new state-of-the-art performance in all NLP tasks. This suggests that CogAlign is able to effectively augment neural models with cognitive processing signals. 5.5 Ablation Study To take a deep look into the improvements contributed by each part of our model, we perform ablation study on all three NLP tasks with two cognitive processing signals. The ablation test includes: (1) w/o text-aware attention, removing text-aware attention mechanism; (2) w/o cognitive loss, discarding the loss of the cognitive predictor whose inputs are cognitive processing signals; (3) w/o modality discriminator, removing the discriminator to train parameters with the task loss. Table 3 reports the ablation study results. 3765 Model NER Sentiment Analysis Relation Extraction P (%) R (%) F1 (%) P (%) R (%) F1 (%) P (%) R (%) F1 (%) CogAlign (eye+EEG) 91.28 83.02 86.79∗ 65.11 65.94 65.40∗ 78.66 82.07 78.93∗ - text-aware attention 90.51 82.45 86.19∗ 64.75 65.30 63.90∗ 77.67 83.14 78.68∗ - cognitive loss 90.20 81.11 85.45∗ 64.48 65.42 63.77∗ 77.79 81.24 77.75∗ - modality discriminator 89.63 83.66 86.09∗ 64.11 66.24 63.28∗ 78.61 80.71 78.46∗ Table 3: Ablation study on the three NLP tasks. Significance is indicated with the asterisks: * = p<0.01. (a) without adv (b) with adv Figure 2: The visualization of hidden states from the shared Bi-LSTM layer. ‘adv’ denotes the adversarial learning. Red dots are the hidden representations of cognitive processing signals while blue dots hidden representations of textual inputs. Both are at the word level via t-SNE (Van der Maaten and Hinton, 2008). The absence of the text-aware attention, cognitive loss and modality discriminator results in a significant drop in performance. This demonstrates that these components all contribute to the effective incorporation of cognitive processing signals into neural models of the three target tasks. CogAlign outperforms both (2) w/o cognitive loss and (3) w/o modality discriminator by a great margin, indicating that the cognitive features can significantly enhance neural models. Furthermore, we visualize the distribution of hidden states learned by the shared Bi-LSTM to give a more intuitive demonstration of the effect of adversarial learning. In Figure 2, clearly, the modality discriminator with adversarial learning forces the shared Bi-LSTM encoder to align textual inputs to cognitive processing signals in the same space. 6 Analysis 6.1 Text-aware Attention Analysis In addition to denoising the cognitive processing signals, the text-aware attention mechanism also obtains the task-specific features. To have a clear view of the role that the text-aware attention mechanism plays in CogAlign, we randomly choose samples and visualize the average attention weights over each signal in Figure 3. For eye-tracking, signals reflecting the late syn(a) eye-tracking (b) EEG Figure 3: The visualization of attention weights over cognitive processing signals by the text-aware attention in the three NLP tasks. Darker colors represent higher attention weights. tactic processing, such as ‘NFIX’ (number of fixation), ‘TFD’ (total fixation duration), play an important role in the three tasks. These results are consistent with findings in cognitive neuroscience. In cognitive neuroscience, researchers have shown that readers tend to gaze at nouns repeatedly (Furtner et al., 2009) (related to the eye-tracking signal NFIX, the number of fixations) and there is a dependency relationship between regression features and sentence syntactic structures (Lopopolo et al., 2019). In other NLP tasks that infused eye-tracking features, the late gaze features have also proved to be more important than early gaze features, such as multiword expression extraction (Rohanian et al., 2017). Moreover, from the additional eye-tracking used in NER, we can find that the cognitive features from the neighboring words are helpful to identify entity, such as ‘w-2 FP’ (w-2 fixation probability), ‘w+1 FP’ (w+1 fixation probability). Since a single EEG signal has no practical meaning, we only visualize the attention weights over EEG signals used in the NER task. Obviously, attentions to ‘t1’ (theta1) and ‘a2’ (alpha2) are stronger than other signals, suggesting that low frequency electric activities in the brain are obvious when we recognize an entity. 3766 Model Wikigold SST P (%) R (%) F1 (%) P (%) R (%) F1 (%) baseline 80.70 70.67 75.19 56.67 57.58 56.40 baseline (two encoders) 80.16 73.39 75.73 56.76 58.05 56.89 CogAlign (eye) 80.39 72.59 76.17 58.05 59.69 57.27 CogAlign (EEG) 80.54 71.91 75.93 57.25 58.34 57.10 CogAlign (eye+EEG) 81.71 74.17 77.76 58.60 58.33 58.32 Table 4: Results of CogAlign in transfer learning to other datasets without cognitive processing signals. ‘baseline’ is a model trained and tested with one encoder for textual inputs. ‘baseline (+ZuCo text)’ is the baseline trained with both Zuco textual data and target dataset (i.e., Wikigold or SST). ‘baseline (two encoders)’ is the same as CogAlign (the inference version), where cognitive processing signals are replaced by textual inputs. 6.2 Transfer Learning Analysis The cognitively-inspired NLP is limited by the collection of cognitive processing signals. Thus, we further investigate whether our model can transfer cognitive features to other datasets without cognitive processing signals for the same task. We enable transfer learning in CogAlign with a method similar to the alternating training approach (Luong et al., 2016) that optimizes each task for a fixed number of mini-batches before shifting to the next task. In our case, we alternately feed instances from the ZuCo dataset and those from other datasets built for the same target task but without cognitive processing signals into CogAlign. Since CogAlign is a multi-task learning framework, model parameters can be updated either by data with cognitive processing signals or by data without such signals, where task-specific loss is used in both situations. Please notice that only textual inputs are fed into trained CogAlign for inference. To evaluate the capacity of CogAlign in transferring cognitive features, we select benchmark datasets for NER and sentiment analysis: Wikigold (Balasuriya et al., 2009) and Stanford Sentiment Treebank (Socher et al., 2013). Since no other datasets use the same set of relation types as that in ZuCo dataset, we do not test the relation extraction task for transfer learning. To ensure that the same textual data are used for comparison, we add a new baseline model (baseline (+Zuco text)) that is trained on the combination of textual data in ZuCo and benchmark dataset. Additionally, as CogAlign uses two encoders for inference (i.e., the textual encoder and shared encoder), for a fair comparison, we setup another baseline (baseline (two encoders)) that also uses two encoders fed with the same textual inputs. The experimental setup is the same as mentioned before. Results are shown in the Table 4. We can observe that CogAlign consistently outperforms the two baselines. It indicates that CogAlign is able to effectively transfer cognitive knowledge (either eye-tracking or EEG) from ZuCo to other datasets. Results show that the best performance is achieved by transferring both eye-tracking and EEG signals at the same time. 7 Conclusions In this paper, we have presented CogAlign, a framework that can effectively fuse cognitive processing signals into neural models of various NLP tasks by learning to align the textual and cognitive modality at both word and sentence level. Experiments demonstrate that CogAlign achieves new state-ofthe-art results on three NLP tasks on the Zuco dataset. Analyses suggest that the text-aware attention in CogAlign can learn task-related cognitive processing signals by attention weights while the modality discriminator with adversarial learning forces CogAlign to learn cognitive and textual representations in the unified space. Further experiments exhibit that CogAlign is able to transfer cognitive information from Zuco to other datasets without cognitive processing signals. Acknowledgments The present research was partially supported by the National Key Research and Development Program of China (Grant No. 2019QY1802) and Natural Science Foundation of Tianjin (Grant No. 19JCZDJC31400). We would like to thank the anonymous reviewers for their insightful comments. References Pavlo Antonenko, Fred Paas, Roland Grabner, and Tamara Van Gog. 2010. Using electroencephalography to measure cognitive load. Educational Psychology Review, 22(4):425–438. 3767 Dominic Balasuriya, Nicky Ringland, Joel Nothman, Tara Murphy, and James R. Curran. 2009. Named entity recognition in wikipedia. In Proceedings of the 1st 2009 Workshop on The People’s Web Meets NLP: Collaboratively Constructed Semantic Resources@IJCNLP 2009, Suntec, Singapore, August 7, 2009, pages 10–18. Association for Computational Linguistics. Maria Barrett, Joachim Bingel, Nora Hollenstein, Marek Rei, and Anders Søgaard. 2018. Sequence classification with human attention. In Proceedings of the 22nd Conference on Computational Natural Language Learning, pages 302–312. Maria Barrett, Joachim Bingel, Frank Keller, and Anders Søgaard. 2016. Weakly supervised part-ofspeech tagging using eye-tracking data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 579–584. Maria Barrett and Anders Søgaard. 2015. Reading behavior predicts syntactic categories. In Proceedings of the 19th Conference on Computational Natural Language Learning, CoNLL 2015, Beijing, China, July 30-31, 2015, pages 345–349. ACL. Pengfei Cao, Yubo Chen, Kang Liu, Jun Zhao, and Shengping Liu. 2018. Adversarial transfer learning for chinese named entity recognition with selfattention mechanism. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 182–192. Xilun Chen, Yu Sun, Ben Athiwaratkun, Claire Cardie, and Kilian Weinberger. 2018. Adversarial deep averaging networks for cross-lingual sentiment classification. Transactions of the Association for Computational Linguistics, 6:557–570. Yun-Nung Chen, Kai-min Chang, and Jack Mostow. 2012. Towards using EEG to improve ASR accuracy. In Human Language Technologies: Conference of the North American Chapter of the Association of Computational Linguistics, Proceedings, June 3-8, 2012, Montr´eal, Canada, pages 382–385. The Association for Computational Linguistics. Jason PC Chiu and Eric Nichols. 2016. Named entity recognition with bidirectional LSTM-CNNs. Transactions of the Association for Computational Linguistics, 4:357–370. Aron Culotta, Andrew McCallum, and Jonathan Betz. 2006. Integrating probabilistic extraction models and data mining to discover relations and patterns in text. In Proceedings of the Human Language Technology Conference of the NAACL, Main Conference, pages 296–303. Emily L Denton, Soumith Chintala, Rob Fergus, et al. 2015. Deep generative image models using a Laplacian Pyramid of Adversarial Networks. Advances in neural information processing systems, 28:1486– 1494. Manuel J. A. Eugster, Tuukka Ruotsalo, Michiel M. A. Spap´e, Ilkka Kosunen, Oswald Barral, Niklas Ravaja, Giulio Jacucci, and Samuel Kaski. 2014. Predicting term-relevance from brain signals. In The 37th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’14, Gold Coast , QLD, Australia - July 06 - 11, 2014, pages 425–434. ACM. Marco R. Furtner, John F. Rauthmann, and Pierre Sachse. 2009. Nomen est omen: Investigating the dominance of nouns in word comprehension with eye movement analyses. Advances in Cognitive Psychology, 5. Yaroslav Ganin and Victor S. Lempitsky. 2015. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 1180–1189. JMLR.org. Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Franc¸ois Laviolette, Mario Marchand, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17(1):2096–2030. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. Advances in neural information processing systems, 27:2672–2680. John M Henderson and Fernanda Ferreira. 1993. Eye movement control during reading: Fixation measures reflect foveal but not parafoveal processing difficulty. Canadian Journal of Experimental Psychology/Revue canadienne de psychologie exp´erimentale, 47(2):201. Nora Hollenstein, Maria Barrett, Marius Troendle, Francesco Bigiolli, Nicolas Langer, and Ce Zhang. 2019a. Advancing NLP with cognitive language processing signals. arXiv preprint arXiv:1904.02682. Nora Hollenstein, Jonathan Rotsztejn, Marius Troendle, Andreas Pedroni, Ce Zhang, and Nicolas Langer. 2018. Zuco, a simultaneous EEG and eye-tracking resource for natural sentence reading. Scientific data, 5(1):1–13. Nora Hollenstein and Ce Zhang. 2019. Entity recognition at first sight: Improving NER with eye movement information. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1–10. Association for Computational Linguistics. 3768 Joo-Kyung Kim, Young-Bum Kim, Ruhi Sarikaya, and Eric Fosler-Lussier. 2017. Cross-lingual transfer learning for pos tagging without cross-lingual resources. In Proceedings of the 2017 conference on empirical methods in natural language processing, pages 2832–2838. John D. Lafferty, Andrew McCallum, and Fernando C. N. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning (ICML 2001), Williams College, Williamstown, MA, USA, June 28 - July 1, 2001, pages 282–289. Morgan Kaufmann. Ying Lin, Shengqi Yang, Veselin Stoyanov, and Heng Ji. 2018. A multi-lingual multi-task architecture for low-resource sequence labeling. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 799–809. Association for Computational Linguistics. Alessandro Lopopolo, Stefan L. Frank, Antal Van Den Bosch, and Roel Willems. 2019. Dependency parsing with your eyes: Dependency structure predicts eye regressions during reading. In Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics. Ling Luo, Zhihao Yang, Pei Yang, Yin Zhang, Lei Wang, Hongfei Lin, and Jian Wang. 2018. An attention-based BiLSTM-CRF approach to document-level chemical named entity recognition. Bioinform., 34(8):1381–1388. Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2016. Multitask sequence to sequence learning. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings. Xuezhe Ma and Eduard H. Hovy. 2016. End-to-end sequence labeling via bi-directional LSTM-CNNsCRF. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics. Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research, 9(11). Sandeep Mathias, Diptesh Kanojia, Abhijit Mishra, and Pushpak Bhattacharya. 2020. A survey on using gaze behaviour for natural language processing. In Twenty-Ninth International Joint Conference on Artificial Intelligence and Seventeenth Pacific Rim International Conference on Artificial Intelligence IJCAI-PRICAI-20. Abhijit Mishra, Diptesh Kanojia, Seema Nagar, Kuntal Dey, and Pushpak Bhattacharyya. 2017. Leveraging cognitive features for sentiment analysis. arXiv preprint arXiv:1701.05581. Lukas Muttenthaler, Nora Hollenstein, and Maria Barrett. 2020. Human brain activity for machine attention. CoRR, abs/2006.05113. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1532–1543. ACL. Keith Rayner. 1998. Eye movements in reading and information processing: 20 years of research. Psychological bulletin, 124(3):372. Omid Rohanian, Shiva Taslimipoor, Victoria Yaneva, and Le An Ha. 2017. Using gaze data to predict multiword expressions. In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, Varna, Bulgaria, September 2 - 8, 2017, pages 601–609. INCOMA Ltd. Cicero dos Santos, Ming Tan, Bing Xiang, and Bowen Zhou. 2016. Attentive pooling networks. arXiv preprint arXiv:1602.03609. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642. Ekta Sood, Simon Tannert, Diego Frassinelli, Andreas Bulling, and Ngoc Thang Vu. 2020a. Interpreting attention models with human visual attention in machine reading comprehension. arXiv preprint arXiv:2010.06396. Ekta Sood, Simon Tannert, Philipp M¨uller, and Andreas Bulling. 2020b. Improving natural language processing tasks with human gaze-guided neural attention. arXiv preprint arXiv:2010.07891. Michalina Strzyz, David Vilares, and Carlos G´omezRodr´ıguez. 2019. Towards making a dependency parser see. arXiv preprint arXiv:1909.01053. Ece Takmaz, Sandro Pezzelle, Lisa Beinborn, and Raquel Fern´andez. 2020. Generating image descriptions via sequential cross-modal alignment guided by human gaze. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 4664–4677. Association for Computational Linguistics. 3769 Chad C. Williams, Mitchel Kappen, Cameron D. Hassall, Bruce Wright, and Olave E. Krigolson. 2019. Thinking theta and alpha: Mechanisms of intuitive and analytical reasoning. NeuroImage, 189:574– 580. YaoSheng Yang, Meishan Zhang, Wenliang Chen, Wei Zhang, Haofen Wang, and Min Zhang. 2018. Adversarial learning for chinese NER from crowd annotations. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 1627–1635. AAAI Press.
|
2021
|
291
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3770–3785 August 1–6, 2021. ©2021 Association for Computational Linguistics 3770 Self-Attention Networks Can Process Bounded Hierarchical Languages Shunyu Yao† Binghui Peng‡ Christos Papadimitriou‡ Karthik Narasimhan† †Princeton University ‡Columbia University {shunyuy, karthikn}@princeton.edu {bp2601, christos}@columbia.edu Abstract Despite their impressive performance in NLP, self-attention networks were recently proved to be limited for processing formal languages with hierarchical structure, such as Dyckk, the language consisting of well-nested parentheses of k types. This suggested that natural language can be approximated well with models that are too weak for formal languages, or that the role of hierarchy and recursion in natural language might be limited. We qualify this implication by proving that self-attention networks can process Dyckk,D, the subset of Dyckk with depth bounded by D, which arguably better captures the bounded hierarchical structure of natural language. Specifically, we construct a hard-attention network with D + 1 layers and O(log k) memory size (per token per layer) that recognizes Dyckk,D, and a soft-attention network with two layers and O(log k) memory size that generates Dyckk,D. Experiments show that self-attention networks trained on Dyckk,D generalize to longer inputs with near-perfect accuracy, and also verify the theoretical memory advantage of self-attention networks over recurrent networks.1 1 Introduction Transformers (Vaswani et al., 2017) are now the undisputed champions across several benchmark leaderboards in NLP. The major innovation of this architecture, self-attention, processes input tokens in a distributed way, enabling efficient parallel computation as well as long-range dependency modelling. The empirical success of self-attention in NLP has led to a growing interest in studying its properties, with an eye towards a better understanding of the nature and characteristics of natural language (Tran et al., 2018; Papadimitriou and Jurafsky, 2020). 1Code is available at https://github.com/ princeton-nlp/dyck-transformer. In particular, it was recently shown that selfattention networks cannot process various kinds of formal languages (Hahn, 2020; Bhattamishra et al., 2020a), among which particularly notable is Dyckk, the language of well-balanced brackets of k types. By the Chomsky-Schützenberger Theorem (Chomsky and Schützenberger, 1959), any context-free language can be obtained from a Dyckk language through intersections with regular languages and homomorphisms. In other words, this simple language contains the essence of all context-free languages, i.e. hierarchical structure, center embedding, and recursion – features which have been long claimed to be at the foundation of human language syntax (Chomsky, 1956). Consider for example the long-range and nested dependencies in English subject-verb agreement: (Laws (the lawmaker) [writes] [and revises]) [pass]. . . . The sentence structure is captured by Dyck2 string (()[][])[]. Given the state-of-the-art performance of Transformers in parsing natural language (Zhang et al., 2020; He and Choi, 2019), the Dyckk blind spot seems very suggestive. If the world’s best NLP models cannot deal with this simple language — generated by a grammar with k + 2 rules and recognized by a single-state pushdown automaton — does this not mean that the role of hierarchy and recursion in natural language must be limited? This question has of course, been extensively debated by linguists on the basis of both theoretical and psycholinguistic evidence (Hauser et al., 2002; Frank et al., 2012; Nelson et al., 2017; Brennan and Hale, 2019; Frank and Christiansen, 2018). So, what can self-attention networks tell us about natural language and recursion? Here we provide a new twist to this question by considering Dyckk,D, the subset of Dyckk with nesting depth at most D, and show that Transformers can process 3771 Input ( [ ] { [ ] ( ) } ) Layer 1 ( [ ] { [ ] ( ) } ) Layer 2 ( [ ] { [ ] ( ) } ) Layer 3 ( [ ] { [ ] ( ) } ) 햣헒햼헄3,3 Input [ ] [ ( ( [ ] ) ) Layer 1 1 0 1 2 3 4 3 2 1 Layer 2 1 0 1 2 3 4 3 2 1 햣헒햼헄2,4 (a) (b) next token prediction: ( [ ] Figure 1: Illustrations of our self-attention network constructions to recognize and generate Dyckk,D. In construction (a), at each layer, the innermost brackets attend to their matching brackets and “cancel” each other, yielding “shallower” spans for successive layers to process. In construction (b), the first layer computes the depth of each token by attending to all previous tokens, while the second layer uses depth information to find the most recent unclosed open bractket in the history. it. Dyckk,D models bounded (or finite) recursion, thus captures the hierarchical structure of human language much more realistically. For example, center-embedding depth of natural language sentences is known to rarely exceed three (Karlsson, 2007; Jin et al., 2018), and while pragmatics, discourse, and narrative can result in deeper recursion in language (Levinson, 2014), there is arguably a relatively small limit to the depth as well. In particular, we prove that self-attention networks can both recognize and generate Dyckk,D, with two conceptually simple yet different constructions (Figure 1). The first network requires D + 1 layers and a memory size of O(log k) (per layer per token) to recognize Dyckk,D, using a distributed mechanism of parenthesis matching. The second network has two layers and memory size O(log k). It works by attending to all previous tokens to count the depth for each token in the first layer, and then uses this depth information to attend to the most recent unclosed open bracket in the second layer. Our constructions help reconcile the result in Hahn (2020) with the success of Transformers in handling natural languages. Our proof requires certain assumptions about the positional encodings, an issue that is often considered in empirical papers (Ke et al., 2021; Shaw et al., 2018; Wang et al., 2020; Shiv and Quirk, 2019) but not in the more theoretical literature. First, positional encodings must have log n bits when the input length is n, as otherwise different positions would share the same representation. More importantly, positional encodings should support easy position comparisons, since token order is vital in formal language processing. Our experiments show that two standard practices, namely learnable or fixed sine/cosine positional encodings, cannot generalize well on Dyckk,D beyond the training input lengths. In contrast, using a single fixed scalar monotonic positional encoding such as pos/n achieves near-perfect accuracy even on inputs significantly longer than the training ones. Our findings provide a novel perspective on the function of positional encodings, and implies that different applications of self-attention networks (in this case, natural vs. formal language) may require different model choices. Our theoretical results also bring about interesting comparisons to recurrent networks (e.g. RNNs, LSTMs) in terms of the resource need to process hierarchical structure. While recurrent networks with finite precision need at least Ω(D log k) memory to process Dyckk,D (Hewitt et al., 2020), our second construction requires only O(log k) memory but a O(log n) precision. In experiments where precision is not an issue for practical input lengths (< 104), we confirm that a Transformer requires less memory than a LSTM to reach high test accuracies. This may help explain why Transformers outperform RNNs/LSTMs in syntactical tasks in NLP, and shed light into fundamental differences between recurrent and non-recurrent sequence processing. 2 Related work Our work primarily relates to the ongoing effort of characterizing theoretical abilities (Pérez et al., 2019; Bhattamishra et al., 2020b; Yun et al., 2020) 3772 and limitations of self-attention networks, particularly through formal hierarchical structures like Dyckk. Hahn (2020) proves that (even with positional encodings) hard-attention Transformers cannot model Dyckk, and soft-attention Transformers with bounded Lipschitz continuity cannot model Dyckk with perfect cross entropy. Bhattamishra et al. (2020a) prove a soft-attention network with positional masking (but no positional encodings) can solve Dyck1 but not Dyck2. Despite the expressivity issues theoretically posed by the above work, empirical findings have shown Transformers can learn Dyckk from finite samples and outperform LSTM (Ebrahimi et al., 2020). Our work addresses the theory-practice discrepancy by using positional encodings and modeling Dyckk,D. A parallel line of work with much lengthier tradition (Elman, 1990; Das et al., 1992; Steijvers and Grünwald, 1996) investigates the abilities and limitations of recurrent networks to process hierarchical structures. In particular, RNNs or LSTMs are proved capable of solving context-free languages like Dyckk given infinite precision (Korsky and Berwick, 2019) or external memory (Suzgun et al., 2019; Merrill et al., 2020). However, Merrill et al. (2020) also prove RNNs/LSTMs cannot process Dyckk without such assumptions, which aligns with experimental findings that recurrent networks perform or generalize poorly on Dyckk (Bernardy, 2018; Sennhauser and Berwick, 2018; Yu et al., 2019). Hewitt et al. (2020) propose to consider Dyckk,D as it better captures natural language, and show finite-precision RNNs can solve Dyckk,D with Θ(D log k) memory. For the broader NLP community, our results also contribute to settling whether self-attention networks are restricted to model hierarchical structures due to non-recurrence, a concern (Tran et al., 2018) often turned into proposals to equip Transformers with recurrence (Dehghani et al., 2019; Shen et al., 2018; Chen et al., 2018; Hao et al., 2019). On one hand, Transformers are shown to encode syntactic (Lin et al., 2019; Tenney et al., 2019; Manning et al., 2020) and word order (Yang et al., 2019) information, and dominate syntactical tasks in NLP such as constituency (Zhang et al., 2020) and dependency (He and Choi, 2019) parsing. On the other hand, on several linguistically-motivated tasks like English subject-verb agreement (Tran et al., 2018), recurrent models are reported to outperform Transformers. Our results help address the issue by confirming that distributed and recurrent sequence processing can both model hierarchical structure, albeit with different mechanisms and tradeoffs. 3 Preliminaries 3.1 Dyck Languages Consider the vocabulary of k types of open and close brackets Σ = ∪i∈[k]{⟨i, ⟩i}, and define Dyckk ⊂γΣ∗ω (γ, ω being special start and end tokens) to be the formal language of well-nested brackets of k types. It is generated starting from γXω through the following context-free grammar: X →ϵ | ⟨i X ⟩i X (i ∈[k]) (1) where ϵ denotes the empty string. Intuitively, Dyckk can be recognized by sequential scanning with a stack (i.e., a pushdown automaton). Open brackets are pushed into the stack, while a close bracket causes the stack to pop, and the popped open bracket is compared with the current close bracket (they should be of the same type). The depth of a string w1:n at position i is the stack size after scanning w1:i, that is, the number of open brackets left in the stack: d(w1:i) = count(w1:i, ⟨) −count(w1:i, ⟩) (2) Finally, we define Dyckk,D to be the subset of Dyckk strings with depth bounded by D: Dyckk,D = w1:n ∈Dyckk max i∈[n] d(w1:i) ≤D That is, a string in Dyckk,D only requires a stack with bounded size D to process. 3.2 Self-attention Networks We consider the encoder part of the original Transformer (Vaswani et al., 2017), which has multiple layers of two blocks each: (i) a self-attention block and (ii) a feed-forward network (FFN). For an input string w1:n ∈Σ∗, each input token wi is converted into a token embedding via fe : Σ →Rdmodel, then added with a position encoding pi ∈Rdmodel. Let xi,ℓ∈Rdmodel be the i-th representation of the ℓ-th layer (i ∈[n], ℓ∈[L]). Then xi,0 = fe(wi) + pi (3) ai,ℓ= Attℓ(Qℓ(xi), Kℓ(x), Vℓ(x)) (4) xi,ℓ+1 = Fℓ(ai,ℓ) (5) 3773 Attention In each head of a self-attention block, the input vectors x1:n undergo linear transforms Q, K, V yielding query, key, and value vectors. They are taken as input to a self-attention module, whose t-th output, Att(Qxi, Kx, V x), is a vector ai = P j∈[T] αjV xj, where α1:n = softmax(⟨Qxi, Kx1⟩, · · · , ⟨Qxi, Kxn⟩). The final attention output is the concatenation of multihead attention outputs. We also consider variants of the basic model along these directions: (i) Hard attention, as opposed to soft attention described above, where hardmax is used in place for softmax (i.e. Att(Qxi, Kx, V x) = V xj′ where j′ = arg maxj⟨Qxi, Kxj⟩). Though impractical for NLP, it has been used to model formal languages (Hahn, 2020). (ii) Positional masking, where α1:i (past) or αi:n (future) is masked for position i. Future-positional masking is usually used to train auto-regressive models like GPT-2 (Radford et al., 2019). Feed-forward network A feed-forward network F transforms each self-attention output vector ai →F(ai) individually. It is usually implemented as a multi-layer perceptron (MLP) with ReLU activations. Residual connections (He et al., 2016) and layer normalization (Ba et al., 2016) are two optional components to aid learning. Positional encodings Vaswani et al. (2017) proposes two kinds of positional encoding: (i) Fourier features (Rahimi and Recht, 2007), i.e. sine/cosine values of different frequencies; (ii) learnable features for each position. In this work we propose to use a single scalar i/n to encode position i ∈[n], and show that it helps process formal languages like Dyckk,D, both theoretically and empirically. Precision and memory size We define precision to be the number of binary bits used to represent each scalar, and memory size per layer (dmodel) to be the number of scalars used to represent each token at each layer. The memory size (L · dmodel) is the total memory used for each token. 3.3 Language Generation and Recognition For a Transformer with L layers and input w1:i, we can use a decoder (MLP + softmax) on the final token output xi,L to predict wi+1. This defines a language model fθ(wi+1|wi) where θ denotes Transformer and decoder parameters. We follow previous work (Hewitt et al., 2020) to define how a language model can generate a formal language: Definition 3.1 (Language generation). Language model fθ over Σ⋆generates a language L ⊆Σ⋆if there exists ϵ > 0 such that L = {w1:n ∈Σ⋆| ∀i ∈ [n], fθ(wi|w1:i−1) ≥ϵ}. We also consider language recognition by a language classifier gθ(w1:i), where a decoder on xi,L instead predicts a binary label. Definition 3.2 (Language recognition). Language classifier gθ over Σ⋆recognizes a language L ⊆ Σ⋆if L = {w1:n ∈Σ⋆|gθ(w1:n) = 1}. 4 Theoretical Results In this section we state our theoretical results along with some remarks. Proof sketches are provided in the next section, and details in Appendix A,B,C. Theorem 4.1 (Hard-attention, Dyckk,D recognition). For all k, D ∈N+, there exists a (D + 1)layer hard-attention network that can recognize Dyckk,D. It uses both future and past positional masking heads, positional encoding of the form i/n for position i, O(log k) memory size per layer, and O(log n) precision, where n is the input length. Theorem 4.2 (Soft-attention, Dyckk,D generation). For all k, D ∈N+, there exists a 2-layer softattention network that can generate Dyckk,D. It uses future positional masking, positional encoding of form i/n for position i, O(log k) memory size per layer, and O(log n) precision, where n is the input length. The feed-forward networks use residual connection and layer normalization. Theorem 4.3 (Precision lower bound). For all k ∈N+, no hard-attention network with o(log n) precision can recognize Dyckk,2 where n is the input length. Required precision Both constructions require a precision increasing with input length, as indicated by Theorem 4.3. The proof of the lower bound is inspired by the proof in Hahn (2020), but several technical improvements are necessary; see Appendix C. Intuitively, a vector with a fixed dimension and o(log n) precision cannot even represent n positions uniquely. The required precision is not unreasonable, since log n is a small overhead to the n tokens the system has to store. Comparison to recurrent processing Hewitt et al. (2020) constructs a 1-layer RNN to generate Dyckk,D with Θ(D log k) memory, and proves it is optimal for any recurrent network. Thus Theorem 4.2 establishes a memory advantage of selfattention networks over recurrent ones. However, 3774 this is based on two tradeoffs: (i) Precision. Hewitt et al. (2020) assumes O(1) precision while we require O(log n). (ii) Runtime. Runtime of recurrent and self-attention networks usually scale linearly and quadratically in n, respectively. Comparison between two constructions Theorem 4.2 requires fewer layers (2 vs. D) and memory size (O(log k) vs. O(D log k)) than Theorem 4.1, thanks to the use of soft-attention, residual connection and layer normalization. Though the two constructions are more suited to the tasks of recognition and generation respectively (Section 5), each of them can also be modified for the other task. Connection to Dyckk In Hahn (2020) it is shown that no hard-attention network can recognize Dyckk even for k = 1. Theorem 4.1 establishes that this impossibility can be circumvented by bounding the depth of the Dyck language. Hahn (2020) also points out soft-attention networks can be limited due to bounded Lipschitz continuity. In fact, our Theorem 4.2 construction can also work on Dyckk with some additional assumptions (e.g. feed n also in input embeddings), and we circumvent the impossibility by using laying normalization, which may have an O(n) Lipschitz constant. More details are in Appendix B.4. 5 Constructions 5.1 (D + 1)-layer Hard-Attention Network Our insight underlying the construction in Theorem 4.1 is that, by recursively removing matched brackets from innermost positions to outside, each token only needs to attend to nearest unmatched brackets to find its matching bracket or detect error within D layers. Specifically, at each layer ℓ≤D, each token will be in one of three states (Figure 2 (c)): (i) Matched, (ii) Error, (iii) Unmatched, and we leverage hard-attention to implement a dynamic state updating process to recognize Dyckk,D. Representation For an input w1:n ∈γΣ∗ω, the representation at position i of layer ℓhas five parts xi,ℓ= [ti, oi, pi, mi,ℓ, ei,ℓ]: (i) a bracket type embedding ti ∈R⌈log k⌉that denotes which bracket type (1 · · · k) the token is (or if the token is start/end token); (ii) a bracket openness bit oi ∈{0, 1}, where 1 denotes open brackets (or start token) and 0 denotes close one (or end token); (iii) a positional encoding scalar pi = i/n; (iv) a match bit mi,ℓ∈{0, 1}, where 1 denotes matched and 0 unmatched; (v) an error bit ei,ℓ∈{0, 1}, where 1 denotes error and 0 no error. Token identity parts ti, oi, pi are maintained unchanged throughout layers. The match and error bits are initialized as ei,0 = mi,0 = 0. The first D layers have identical self-attention blocks and feed-forward networks, detailed below. Attention Consider the ℓ-th self-attention layer (ℓ∈[D]), and denote xi = xi,ℓ−1, mi = mi,ℓ−1, ai = ai,ℓ, yi = xi,ℓfor short. We have 3 attention heads: (i) an identity head Attid, where each token only attends to itself with attention output aid i = xi; (ii) a left head Attleft with future positional masking; (iii) a right head Attright with past positional masking. The query, key, and value vectors for Attleft are defined as Qxi = 1 ∈R, Kxi = pi −mi ∈R, V xi = xi ∈Rdmodel, so that aleft i = xj1, j1 = arg max j<i (j/n −mj) is the representation of the nearest unmatched token to i on its left side. Similarly aright i = xj2, j2 = arg max j>i (1 −j/n −mj) is the representation of the nearest unmatched token to i on its right side. The attention output for position i is the concatenation of these three outputs: ai = [aid i , aleft i , aright i ] = [xi, xj1, xj2]. Feed-forward network (FFN) Following the notation above, the feed-forward network F : ai → yi serves to update each position’s state using information from xj1, xj2. The high level logic (Figure 2 (c)) is that, if wi is an open bracket, its potential matching half should be wj = wj2 (j2 > i), otherwise it should be wj = wj1 (j1 < i). If wi and wj are one open and one close, they either match (same type) or cause error (different types). If wi and wj are both open or both close, no state update is done for position i. Besides, token identity parts ti, oi, pi are copied from aid i to pass on. The idea can be translated into a language of logical operations (∧, ∨, ¬) plus a SAME(t, t′) operation, which returns 1 if vectors t = t′ and 0 otherwise: yi = [ti, oi, pi, m′ i, e′ i] m′ i = mi ∨(oi ∧¬oj2 ∧s1) ∨(¬oi ∧oj1 ∧s2) e′ i = ei ∨(oi ∧¬oj2 ∧¬s1) ∨(¬oi ∧oj1 ∧¬s2) s1 = SAME(ti, tj1) s2 = SAME(ti, tj2) 3775 ( [ ] ( ] ) Layer 1 ( [ ] ( ] ) Layer 2 ( [ ] ( ] ) Layer 3 output: 0/1 ( [ ] ( ] ) FFN ( [ ] ( ] ) ( [ ] ( ] ) ( [ ] ( ] ) ( [ ] ( ] ) ( [ ] ( ] ) ( [ ] ( ] ) 햠헍헍헅햾햿헍 햠헍헍헋헂헀헁헍 햠헍헍헂햽 ( [ ] ( ] ) matched error unmatched unmatched w1:n x1:n ( [ ] ( ] ) γ γ γ ω ω ω ω γ γ ω γ ω ω ω γ γ γ γ γ γ ω ω ω ω (a) (b) (c) y1:n Figure 2: Our construction for Theorem 4.1. (a) The network has multiple identical layers to match brackets and detect errors. (b) Each layer consists of three hard-attention heads so that a token attends to itself and the nearest unmatched tokens on both sides, and uses representations from these positions to update its state. (c) Each position can be in three states: matched, error, or unmatched. As we show in Appendix A, a multi-layer perception with ReLU activations can simulate all operations (∧, ∨, ¬, SAME), thus the existence of our desired FFN. Final layer At the (D + 1)-th layer, the self attention is designed as Qxi = 1 ∈R, Kxi = ei+1−mi ∈R, V xi = (ei, mi) ∈R2. If all brackets are matched without error ((ei, mi) = (0, 1)), all keys would be 0, and the attention output of the last token an would be (0, 1). If any bracket finds error (ei = 1) or is not matched (mi = 0), the key would be at least 1 and an would not be (0, 1). An FNN that emulates (a, b) 7→¬a ∧b will deliver yn as the recognition answer. 5.2 Two-layer Soft-Attention Network Our Theorem 4.2 construction takes advantage of soft attention, residual connection, and layer normalization to calculate each token depth and translate it into a vector form at the first layer. Using the depth information, at the second layer each wi can attend to the stack-top open bracket at the position, in order to decide if open brackets or which type of close brackets can be generated as the next token (Figure 3). Representation The representation at position i, layer ℓhas four parts xi,ℓ= [ti, oi, pi, di,ℓ], with bracket type embedding ti, bracket openness bit oi, position encoding pi already specified in Section 5.1. The last part di,ℓ∈R2 is used to store depth information for position i, and initialized as di,0 = (0, 0). First Layer – Depth Counting The first selfattention layer has two heads, where an Attid head is still used to inherit ti, oi, pi, and a future positional masking head2 Attd aims to count depth with Qxi = Kxi = 1 and V xi = 2oi −1, resulting in uniform attention scores and attention output ad i = P j≤i 1 i · (2oj −1) = d(w1:i)/i. However, our goal is to enable matching based on depth di = d(w1:i), and the attention output di/i isn’t readily usable for such a purpose: the denominator i is undesirable, and even a scalar di cannot easily attend to the same value using dotproduct attention. Thus in the first feed-forward network, we leverage residual connection and layer normalization to transform di/i 7→di = (cos(θ(di)), sin(θ(di))) (6) where θ(d) = arctan d D+2−d has an unique 2Here we assume wi+1:n is masked for position i, just for convenience of description. 3776 ҁ [ ] ( ) ( [ ] ( ) ( [ ] ( ) MLP 0 1 ( 2 [ 1 ] 2 ( 1 ) ( [ ] ( ) 0 1 2 1 2 1 MLP Attention layer 1 MLP layer 1 Attention layer 2 MLP layer 2 depth w1:i 0 # 1 ( 2 [ 1 ] 2 ( 1 ) γ γ γ γ γ prediction: ( [ ) wi+1 x1:i,0 Figure 3: Our construction for Theorem 4.2. The first self-attention layer calculates token depths, while the second layer uses them so that each token attends to the closest unmatched open bracket ign the history, which is useful for next token prediction. value for every d ∈{0, · · · , D + 1}, so that di · dj ( = 1 di = dj < 1 − 1 10D2 di ̸= dj (7) The representation by the end of first layer is xi,1 = [ti, oi, pi, di]. The full detail for the first FFN is in Appendix B.1. Second layer – Depth Matching The second self-attention layer has a depth matching hardattention head Attmatch, with query, key, value vectors as Qxi = [20D2 · di, 1, 2] ∈R4, Kxi = [di, pi, oi] ∈R4, V xi = xi, so that attention score ⟨Qxi, Kxj⟩= 20D2di · dj + j/n + 2oj ( = 20D2 + 2 + j/n di = dj, oj = 1 ≤20D2 + 1 otherwise would achieve its maximum when wj (j ≤i) is the open bracket (or start token) closest to wi with dj = di. The attention output is ai = [aid i , amatch i ] = [xi, xj] where j = max{j ≤i|di = dj ∧oj = 1}. With such a [xi, xj], the second-layer FFN can readily predict what wi+1 could be. It could be any open bracket when di < D (i.e. cos(θ(di)) > cos(θ(D))), and it could be a close bracket with type as tj (or end token if wj is start token). The detailed construction for such a FFN is in Appendix B.2. On Dyckk Generation In fact, this theoretical construction can also generate Dyckk, as intuitively the O(log n) precision assumption allows counting depth up to O(n). But it involves extra conditions like feeding n into network input, and may not be effectively learned in practice. Please refer to details in Appendix B.4. Connection to Empirical Findings Our theoretical construction explains the observation in Ebrahimi et al. (2020): the second layer of a twolayer Transformer trained on Dyckk often produces virtually hard attention, where tokens attend to the stack-top open bracket (or start token). It also explains why such a pattern is found less systematically as input depth increases, as (6) is hard to learn and generalize to unbounded depth in practice. 6 Experiments Our constructions show the existence of selfattention networks that are capable of recognizing and generating Dyckk,D. Now we bridge theoretical insights into experiments, and study whether such networks can be learned from finite samples and generalize to longer input. The answer is affirmative when the right positional encodings and memory size are chosen according to our theory. We first present results on Dyck8,10 (Section 6.1) as an example Dyckk,D language to investigate the effect of different positional encoding schemes, number of layers, and hidden size on the Transformer performance, and to compare with the LSTM performance. We then extend the Transformer vs. LSTM comparison on more Dyckk,D languages (k ∈{2, 8, 32, 128}, D ∈ {3, 5, 10, 15}) in Section 6.2. Finally, we apply 3777 1 2 3 4 5 10 # Layers 0.6 0.7 0.8 0.9 1.0 Close Accuracy (a) Transformers (Dyck-(8, 10) Test) Positional Encoding cos learn pos/N 20 40 60 80 100 Memory Dim. 0.8 0.9 1.0 Close Accuracy (b) Transformer v. LSTM (Dyck-(8, 10) Validation) Model Transformer (pos/N) LSTM 20 40 60 80 100 Memory Dim. 0.8 0.9 1.0 Close Accuracy (c) Transformer v. LSTM (Dyck-(8, 10) Test) Model Transformer (pos/N) LSTM Figure 4: Results on Dyck8,10 validation set (same input lengths as training) and test set (longer inputs). (a) compares Transformers of different layers (L ∈{1, 2, 3, 4, 5, 10}) and with different positional encodings (COS, LEARN,POS/N) on the test set. (b) and (c) compare a 2-layer Transformer (POS/N) with a 1-layer LSTM over varying memory sizes on the validation and test sets respectively. the novel scalar positional encoding to natural language modeling with some preliminary findings (Section 6.3). 6.1 Evaluation on Dyck8,10 Setup For Dyck8,10, we generate training and validation sets with input length n ≤700, and test set with length 700 < n ≤1400. We train randomly initialized Transformers using the Huggingface library (Wolf et al., 2019), with one future positional masking head, L ∈{1, 2, 3, 4, 5, 10} layers, and a default memory size dmodel = 30. We search for learning rates in {0.01, 0.001}, run each model with 3 trials, and report the average accuracy of generating close brackets, the major challenge of Dyckk,D. More setup details are in Appendix D.1. Positional Encodings We compare 3 types of positional encodings: (i) Fourier features (COS); (ii) learnable features (LEARN); (iii) a scalar i/6000 for position i (POS/N). Note that (i, ii) are original proposals in Vaswani et al. (2017), where positional encoding vectors are added to the token embeddings, while our proposal (iii) encodes the position as a fixed scalar separated from token embeddings. On the validation set of Dyck8,10 (see Appendix D.2), all three models achieve near-perfect accuracy with L ≥2 layers. On the test set (Figure 4(a)) however, only POS/N maintains nearperfect accuracy, even with L = 10 layers. Meanwhile, LEARN and COS fail to generalize, because encodings for position 700 < i ≤1400 are not learned (for LEARN) or experienced (for COS) during training. The result validates our theoretical construction, and points to the need for separate and systemic positional encodings for processing long and order-sensitive sequences like Dyckk,D. Memory Size and Comparison with LSTM We compare a two-layer Transformer (POS/N) with a one-layer LSTM3 (Hochreiter and Schmidhuber, 1997) using varying per-layer memory sizes dmodel ∈{10, 20, · · · , 100}. As Figure 4 (b) shows, the Transformer consistently outperforms the LSTM on the validation set. On the test set (Figure 4 (c)), the Transformer and the LSTM first achieve a > 90% accuracy using dmodel = 20 and 40 respectively, and an accuracy of > 95% with dmodel = 30 and 50, respectively. These findings agree with our theoretical characterization that selfattention networks have a memory advantage over recurrent ones. 6.2 Evaluation on More Dyckk,D Languages Setup In order to generalize some of the above results, we generate a wide range of Dyckk,D languages with different vocabulary sizes (k ∈ {2, 8, 32, 128}) and recursion bounds (D ∈ {3, 5, 10, 15}). We continue to compare the onelayer LSTM versus the two-layer Transformer (POS/N). For each model on each language, we perform a hyperparameter search for learning rate in {0.01, 0.001} and memory size dmodel ∈ {10, 30, 50}, and report results from the best setting based on two trials for each setting. 3LSTMs only need one layer to process Dyckk,D (Hewitt et al., 2020), while Transformers at least need two in our constructions. We also experimented with two-layer LSTMs but did not find improved performance. 3778 3 5 10 15 D 0.75 0.80 0.85 0.90 0.95 1.00 Close Accuracy (a) Dyck-(k, D) Validation Model Transformer LSTM k 2 8 32 128 3 5 10 15 D 0.75 0.80 0.85 0.90 0.95 1.00 Close Accuracy (b) Dyck-(k, D) Test Figure 5: Results on more Dyckk,D languages. 0 50 100 150 Epoch 2 4 6 8 10 Loss RoBERTa (WikiText-103) Positional Encoding learn pos/N Split Train Validation Figure 6: Results on WikiText-103. Results The validation and test accuracy of the models are reported in Figure 5, and more finegrained results for each dmodel ∈{10, 30, 50} are in Appendix D.2. The Transformer attains a > 99.9% validation accuracy and a > 94% test accuracy across all languages, strengthening the main claim that self-attention networks can learn Dyckk,D languages and generalize to longer input. On the other hand, the validation and test accuracy of the LSTM model are less than 80% when the vocabulary size and recursion depth are large, i.e. (k, D) ∈ {(32, 15), (128, 10), (128, 15)}4, which reconfirms Transformers’ memory advantage under limited memory (dmodel ≤50). 6.3 Evaluation on WikiText-103 In Section 6.1, we show a Transformer with the scalar positional encoding scheme (POS/N) can learn Dyckk,D and generalize to longer input, while traditional positional encoding schemes ((COS), (LEARN)) lead to degraded test performance. To investigate whether such a novel scheme is also useful in NLP tasks, we train two RoBERTa5 models (POS/N, LEARN) from scratch on the WikiText103 dataset (Merity et al., 2017) for 150 epochs. Figure 6 shows the masked language modeling loss on both training and validation sets. By the end of the training, POS/N has a slightly larger validation loss (1.55) than LEARN (1.31). But throughout the optimization, POS/N shows a gradual decrease of loss while LEARN has a sudden drop of loss around 20-30 epochs. We believe it will be interest4Note that Hewitt et al. (2020) only reports D ∈{3, 5}. 5We also tried language modeling with GPT-2 models, and POS/N has slightly larger train/validation losses than LEARN throughout the training. Interestingly, using no positional encoding leads to the same loss curves as LEARN, as positional masking leaks positional information. ing for future work to explore how POS/N performs on different downstream tasks, and why POS/N seems slightly worse than LEARN (at least on this MLM task), though theoretically it provides the complete positional information for Transformers. These topics will contribute to a deeper understanding of positional encodings and how Transformers leverage positional information to succeed on different tasks. 7 Discussion In this paper, we theoretically and experimentally demonstrate that self-attention networks can process bounded hierarchical languages Dyckk,D, even with a memory advantage over recurrent networks, despite performing distributed processing of sequences without explicit recursive elements. Our results may explain their widespread success at modeling long pieces of text with hierarchical structures and long-range, nested dependencies, including coreference, discourse and narratives. We hope these insights can enhance knowledge about the nature of recurrence and parallelism in sequence processing, and lead to better NLP models. Acknowledgement We thank Xi Chen, members of the Princeton NLP Group, and anonymous reviewers for suggestions and comments. Ethical Consideration Our work is mainly theoretical with no foreseeable ethical issues. 3779 References Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. 2016. Layer normalization. arXiv preprint arXiv:1607.06450. Jean-Phillipe Bernardy. 2018. Can recurrent neural networks learn nested recursion? In Linguistic Issues in Language Technology, Volume 16, 2018. CSLI Publications. Satwik Bhattamishra, Kabir Ahuja, and Navin Goyal. 2020a. On the ability of self-attention networks to recognize counter languages. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7096–7116. Satwik Bhattamishra, Arkil Patel, and Navin Goyal. 2020b. On the computational power of transformers and its implications in sequence modeling. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 455–475, Online. Association for Computational Linguistics. Jonathan R Brennan and John T Hale. 2019. Hierarchical structure guides rapid linguistic predictions during naturalistic listening. PloS one, 14(1):e0207741. Mia Xu Chen, Orhan Firat, Ankur Bapna, Melvin Johnson, Wolfgang Macherey, George Foster, Llion Jones, Mike Schuster, Noam Shazeer, Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Zhifeng Chen, Yonghui Wu, and Macduff Hughes. 2018. The best of both worlds: Combining recent advances in neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 76–86, Melbourne, Australia. Association for Computational Linguistics. Noam Chomsky. 1956. Three models for the description of language. IRE Transactions on information theory, 2(3):113–124. Noam Chomsky and Marcel P Schützenberger. 1959. The algebraic theory of context-free languages. In Studies in Logic and the Foundations of Mathematics, volume 26, pages 118–161. Elsevier. Sreerupa Das, C Lee Giles, and Guo-Zheng Sun. 1992. Learning context-free grammars: Capabilities and limitations of a recurrent neural network with an external stack memory. In Proceedings of The Fourteenth Annual Conference of Cognitive Science Society. Indiana University, page 14. Citeseer. Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Lukasz Kaiser. 2019. Universal transformers. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Javid Ebrahimi, Dhruv Gelda, and Wei Zhang. 2020. How can self-attention networks recognize Dyck-n languages? In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4301– 4306, Online. Association for Computational Linguistics. Jeffrey L Elman. 1990. Finding structure in time. Cognitive science, 14(2):179–211. Stefan L Frank, Rens Bod, and Morten H Christiansen. 2012. How hierarchical is language use? Proceedings of the Royal Society B: Biological Sciences, 279(1747):4522–4531. Stefan L Frank and Morten H Christiansen. 2018. Hierarchical and sequential processing of language: A response to: Ding, melloni, tian, and poeppel (2017). rule-based and word-level statistics-based processing of language: insights from neuroscience. language, cognition and neuroscience. Language, Cognition and Neuroscience, 33(9):1213–1218. Michael Hahn. 2020. Theoretical limitations of selfattention in neural sequence models. Transactions of the Association for Computational Linguistics, 8:156–171. Jie Hao, Xing Wang, Baosong Yang, Longyue Wang, Jinfeng Zhang, and Zhaopeng Tu. 2019. Modeling recurrence for transformer. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1198–1207, Minneapolis, Minnesota. Association for Computational Linguistics. Marc D Hauser, Noam Chomsky, and W Tecumseh Fitch. 2002. The faculty of language: what is it, who has it, and how did it evolve? science, 298(5598):1569–1579. Han He and Jinho D Choi. 2019. Establishing strong baselines for the new decade: Sequence tagging, syntactic and semantic parsing with bert. arXiv preprint arXiv:1908.04943. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770–778. IEEE Computer Society. John Hewitt, Michael Hahn, Surya Ganguli, Percy Liang, and Christopher D. Manning. 2020. RNNs can generate bounded hierarchical languages with optimal memory. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1978–2010, Online. Association for Computational Linguistics. Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. 3780 Lifeng Jin, Finale Doshi-Velez, Timothy Miller, William Schuler, and Lane Schwartz. 2018. Unsupervised grammar induction with depth-bounded PCFG. Transactions of the Association for Computational Linguistics, 6:211–224. Fred Karlsson. 2007. Constraints on multiple centerembedding of clauses. Journal of Linguistics, pages 365–392. Guolin Ke, Di He, and Tie-Yan Liu. 2021. Rethinking the positional encoding in language pre-training. In International Conference on Learning Representations, (ICLR 2021). Samuel A Korsky and Robert C Berwick. 2019. On the computational power of rnns. arXiv preprint arXiv:1906.06349. Stephen C Levinson. 2014. Pragmatics as the origin of recursion. In Language and recursion, pages 3–13. Springer. Yongjie Lin, Yi Chern Tan, and Robert Frank. 2019. Open sesame: Getting inside BERT’s linguistic knowledge. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 241–253, Florence, Italy. Association for Computational Linguistics. Christopher D Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proceedings of the National Academy of Sciences, 117(48):30046–30054. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer sentinel mixture models. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. William Merrill, Gail Weiss, Yoav Goldberg, Roy Schwartz, Noah A. Smith, and Eran Yahav. 2020. A formal hierarchy of RNN architectures. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 443–459, Online. Association for Computational Linguistics. Matthew J Nelson, Imen El Karoui, Kristof Giber, Xiaofang Yang, Laurent Cohen, Hilda Koopman, Sydney S Cash, Lionel Naccache, John T Hale, Christophe Pallier, et al. 2017. Neurophysiological dynamics of phrase-structure building during sentence processing. Proceedings of the National Academy of Sciences, 114(18):E3669–E3678. Isabel Papadimitriou and Dan Jurafsky. 2020. Learning Music Helps You Read: Using transfer to study linguistic structure in language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6829–6839, Online. Association for Computational Linguistics. Jorge Pérez, Javier Marinkovic, and Pablo Barceló. 2019. On the turing completeness of modern neural network architectures. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. Ali Rahimi and Benjamin Recht. 2007. Random features for large-scale kernel machines. In Advances in Neural Information Processing Systems 20, Proceedings of the Twenty-First Annual Conference on Neural Information Processing Systems, Vancouver, British Columbia, Canada, December 3-6, 2007, pages 1177–1184. Curran Associates, Inc. Luzi Sennhauser and Robert Berwick. 2018. Evaluating the ability of LSTMs to learn context-free grammars. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 115–124, Brussels, Belgium. Association for Computational Linguistics. Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. Self-attention with relative position representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 464–468, New Orleans, Louisiana. Association for Computational Linguistics. Tao Shen, Tianyi Zhou, Guodong Long, Jing Jiang, Shirui Pan, and Chengqi Zhang. 2018. Disan: Directional self-attention network for rnn/cnn-free language understanding. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 5446–5455. AAAI Press. Vighnesh Leonardo Shiv and Chris Quirk. 2019. Novel positional encodings to enable tree-based transformers. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 12058–12068. Mark Steijvers and Peter Grünwald. 1996. A recurrent network that performs a context-sensitive prediction task. In Proceedings of the 18th annual conference of the cognitive science society, pages 335–339. Mirac Suzgun, Sebastian Gehrmann, Yonatan Belinkov, and Stuart M Shieber. 2019. Memory-augmented recurrent neural networks can learn generalized dyck languages. arXiv preprint arXiv:1911.03329. 3781 Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593– 4601, Florence, Italy. Association for Computational Linguistics. Ke Tran, Arianna Bisazza, and Christof Monz. 2018. The importance of being recurrent for modeling hierarchical structure. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4731–4736, Brussels, Belgium. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 49, 2017, Long Beach, CA, USA, pages 5998–6008. Benyou Wang, Donghao Zhao, Christina Lioma, Qiuchi Li, Peng Zhang, and Jakob Grue Simonsen. 2020. Encoding word order in complex embeddings. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2019. Huggingface’s transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771. Baosong Yang, Longyue Wang, Derek F. Wong, Lidia S. Chao, and Zhaopeng Tu. 2019. Assessing the ability of self-attention networks to learn word order. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3635–3644, Florence, Italy. Association for Computational Linguistics. Xiang Yu, Ngoc Thang Vu, and Jonas Kuhn. 2019. Learning the Dyck language with attention-based Seq2Seq models. In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 138–146, Florence, Italy. Association for Computational Linguistics. Chulhee Yun, Srinadh Bhojanapalli, Ankit Singh Rawat, Sashank J. Reddi, and Sanjiv Kumar. 2020. Are transformers universal approximators of sequence-to-sequence functions? In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Yu Zhang, Houquan Zhou, and Zhenghua Li. 2020. Fast and accurate neural CRF constituency parsing. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 4046–4053. ijcai.org. 3782 A Construction Details of Section 5.1 We provide missing details on the construction of (D + 1)-layer Transformer with hard attention. In particular, we prove that neural networks are capable of simulating logic gates: AND, OR, NOT, SAME and arithmic gates: GREATERTHAN and EQUAL gate. For input x, y ∈R, the GREATERTHAN satisfies that GREATERTHAN(x, y) = 1 if x ≥y + c and GREATERTHAN(x, y) = 0 when x < y; the EQUAL gate satisfies EQUAL(x, y) = 1 if x = y and EQUAL(x, y) = 0 when x < y−c or x > y+c. Here c is a constant independent of x, y. Lemma A.1. A constant layer neural network can simulate logic gates: AND, OR, NOT, SAME and arithmic gates: GREATERTHAN, EQUAL. Proof. Our construction is as follows. (1) AND gate. Given input x1, . . . , xm ∈{0, 1}, we compute z = max{x1 +· · ·+xm −m+1, 0}. We conclude that z = 1 iff x1 = · · · = xm = 1 and z = 0 otherwise. (2) NOT gate. Given input x ∈{0, 1}, it suffices to compute z = max{1 −x, 0}. (3) OR gate. Given input x1, . . . , xm ∈{0, 1}, we compute z = max{1 −max{1 −x1 −· · · − xm, 0}, 0}. It is easy to see that z = 1 iff one of xi = 1 (i ∈[m]) and z = 0 otherwise. (3) SAME gate. Given input x1, . . . , xm ∈{0, 1} and y1, . . . , ym ∈{0, 1}. The SAME gate is equivalent to z = ((x1 ∨y1) ∧(x1 ∨y1)) ∨· · · ∨ ((xm ∨ym) ∧(xm ∨ym)). We can construct it using logic gates: AND, OR, NOT . (4) GREATERTHAN gate. Given x, y ∈R, compute z1 = 1 c max{c −max{x −y, 0}, 0}, we have that z1 = 0 when x > y + c and z = 1 when x ≤y. Taking z = max{1 −z1, 0} completes the construction. (5) EQUAL gate. Given x, y ∈ R. Let z1 = GREATEREQUAL(x, y) and z2 = GREATEREQUAL(y, x). It suffices to take z = ¬z1 ∧¬z2. With some extra effort, one can extend the construction for recognition task to generation task and prove that a D-layer Transformer is capable of generating Dyckk,D. Corollary A.2. ∀k, D ∈N+, there exists a Dlayer hard-attention network that can generate Dyckk,D. It uses both a future-position masking head and a past-position masking head, a O(log k) memory size, and O(log n) precision for processing input length up to n. Soft attention Both Theorem 4.1 and Corollary A.2 can be adapted to soft attention, by setting the temperature parameter η in softmax operator to be sufficient large, say η = Ω(n log nD). Then one can use soft attention to simulate hard attention. In order to fit the precision, for the soft attention distribution p = [p1, · · · , pm], we round pi to the closest multiple of 1 Cn, where C is a large constant. B Construction Details of Section 5.2 We provide missing details of the construction in Section 5.2. B.1 First Layer FFN Recall the output of the first attention layer is ai,1 = [ti, oi, pi, di,1], where ti, oi, pi are the bracket type embedding, the bracket openness bit and the position encoding. di,1 ∈R2 contains the information di/i, where di = d(w1:i) equals the depth at position i. For ease of presentation, we assume it also contains an entry with 1/i, this can be derived with an extra attention head in the first layer or be inherited from an extra position encoding. Define θ(d) = arctan d D+2−d . We prove Lemma B.1. With residual connection and layer normalization, a two-layer MLP can perform the following transformation (di/i, 1/i) 7→di = (cos(θ(di)), sin(θ(di))) while keeping ti, oi, pi unchanged. Proof. Consider the following series of operations. ti, oi, pi, di i , 1 i , 0, 0 7→ 0, 0, 0, −di i , di −D −2 i , di i , D + 2 −di i 7→ 0, 0, 0, −1 2 sin(θ(di)), −1 2 cos(θ(di)), 1 2 sin(θ(di)), 1 2 cos(θ(di)) 7→ 0, 0, 0, 0, 0, 1 2 sin(θ(di)), 1 2 cos(θ(di)) 7→ ti, oi, pi, di i , 1 i , 1 2 sin(θ(di)), 1 2 cos(θ(di)) 7→(ti, oi, pi, cos(θ(di)), sin(θ(di)), 0, 0)) The first steps can be achieved with a linear transformation, the second step can be achieved by layer 3783 normalization and the third step follows from the ReLU activation gate, the fourth step comes from the residual connection and the last step can be obtained with an extra layer of MLP. We conclude the proof here. B.2 Second Layer FFN We can choose between k open brackets and the matched close bracket, with the exception on a few boundary cases: (1) The depth of the current bracket reaches the maximum; (2) The length of the sequence is about to reach the maximum. Let emi be the bracket type of the matched bracket at position i, we implement the last layer as follow. yi = [oi, zi, zi] oi = ¬(di1 = sin(θ(D))) ∧¬(di1 = sin(θ( eD))) eD = min{n −i, D + 1} zi = ¬(di1 = 0) ∧emi zi = 1 −zi. We elaborate on a few details here. (1) We can derive the term sin(θ( eD)) via the similar method in Lemma B.1. (2) Since | sin(θ(i))−sin(θ(j))| = Ω 1 D2 holds for any i ̸= j ∈{0, 1, · · · , D + 1}, we know that the input gap (i.e. the constant c in Lemma A.1) for of all three EQUAL gates is at least Ω 1 d2 . Thus we can apply Lemma A.1. (3) We can obtain n −i by either augmenting the position encoding with n and i, or normalizing (i/n, 1 −i/n) (see Lemma B.1). Output mechanism The final output is determined by on V yT+2, where V ∈R2k×2⌈log k⌉+1 satisfies Vi,1 = 0 and Vi,1: is the binary encoding of the i-th close bracket and its complement when i ∈{1, · · · , k}; Vi,1 = ⌈log k⌉and Vi,j = 0 when i ≤{k + 1, · · · , 2k} and j > 1. Let S ⊆[2k] denote the index of valid output, we conclude that (V yT+2)i = ⌈log k⌉for i ∈S and (V yT+2)i ≤⌈log k⌉−1 for i /∈S. B.3 Extension to Recognition task Our construction can be adapted to recognition task with some extra efforts. Corollary B.2. For all k, D ∈N+, there exists a 3-layer soft-attention network that can generate Dyckk,D. It uses future positional masking, positional encoding of form i/n for position i, O(log k) memory size per layer, and O(log n) precision where n is the input length. The feed-forward networks use residual connection and layer normalization. B.4 Extension to Dyckk We can extend the above construction to recognize language Dyckk. Our construction bypasses the lower bound in Hahn (2020) since the layer normalization operation is not constant Lipschitz (it can be O(n) in the proof). Theorem B.3 (Soft-attention, Dyckk generation). For all k ∈N+, there exists a 2-layer soft-attention network that can generate Dyckk. It uses future positional masking, O(log k) memory size per layer, and O(log n) precision where n is the input length. The feed-forward networks use residual connection and layer normalization. Due to space limits, we omit the detailed proof and only outline the major difference from the proof of Theorem 4.2. 1. We need position encoding i/n3 instead of i/n, and add an extra position encoding of n. 2. For the first FNN, we replace D with n. In particular, for Lemma B.1, we need an extra input of n/i, this can be derived with either an extra attention head or an extra position encoding. 3. For the second FNN, we make some adjustment to the input of the EQUAL gate, since the gap between two input could be very small, i.e., O(1/n2). Nevertheless, we can use the same trick of Lemma B.1 to amplify the gap between two input a, b to be of order Ω(1), the later one suffices to our purpose. C Theoretical limits for finite position encoding We prove that a Transformer with finite precision can not recognize Dyckk,D language. In fact, we show a stronger result: no transformer with o(log n) precision can recognize Dyckk,D language of length more than n. Theorem C.1 (Formal statement of Theorem 4.3). For any k ∈N, using hard attention, no transformer with o(log n) encoding precision can recognize Dyckk,2 language with input length n. Our proof is inspired by Hahn (2020) but with several different technique ingredient: (1) we allow arbitrary attention masking (both future and past 3784 position masking); (2) we allow arbitrary position encoding (3) our lower bounds holds for bounded depth language Dyckk,D; (4) we provide an quantitative bound for precision in terms of input length n. In general, our lower bound is incomparable with Hahn (2020), we prove a fine grained bound on the precision requirement for bounded depth language Dyckk,D, while the proof in Hahn (2020) applies only for language with Depth Ω(n) but allows arbitrary precision on position encoding. The high level intuition behind our proof is that the attention head can only catch o(n) input positions when we properly fix a small number of symbol in the input sequence. This limits the capability of a Transformer and makes it fail to recognize Dyckk,D language. We consider a L-layer transformer and assume 3H attention heads in total: H normal attention heads, H attention heads with future position masking, H attention heads with past position masking. To make our hardness result general, we allow residual connection for the attention layer, and we assume the FNN can be arbitrary function defining on the attention outcome. In the proof, we would gradually fix o(n) positions of the input sequence. We only perform the follow two kinds of assignment (1) we assign matching brackets to position i, i + 1 where i is odd; (2) we assign matching brackets (e.g., we assign ‘[’, ‘(’, ‘)’, ‘]’) to position i, i + 1, i + 2, i + 3 for odd i. A partial assignment to the input sequence is said to be well-aligned if it follows these two rules. Throughout the proof, we guarantee that for any i ∈[n], ℓ∈[L], the output of the ℓ-th layer xi,ℓdepends only the input symbol at position i. This is clearly satisfied for ℓ= 0, given the it is composed by position embedding and word embedding only. We gradually fix the input and conduction induction on ℓ. We use cℓto denote the number of positions we fixed before the ℓ-th layer, and we use sℓto denote the number of consecutive assigned blocks of the input sequence. It is clear that sℓ≤2cℓ. The following Lemma is key to our analysis. Due to space limits, we omit the detailed proof. Lemma C.2. For any ℓ∈{1, · · · , L}, given a well-aligned partially assigned input sequence, suppose the input of ℓ-th layer xi,ℓ−1 depends on the symbol at position i only. Then by fixing cℓH2(k + 1)O(ℓH)2O(ℓHp) additional positions of the input sequence, we guarantee that the output of ℓ-th layer xi,ℓalso depends solely on the symbol at position i. Proof of Theorem C.1. We apply Lemma C.2 and compute the number of positions cL+1 we need to restrict, in order to guarantee that the output of L-th layer xi,L+1 depends only on the input at position (i ∈[n]). Since cℓ+1 ≤cℓH2(k + 1)O(ℓH)2O(ℓHp) and c1 = O(1), we have cL+1 ≲HO(L)(k + 1)O(L2H)2O(L2Hp). By taking HO(L)(k + 1)O(L2H)2O(L2Hp) ≤0.01n. We know the partial assigned sequence is wellaligned, has depth at most two, and the number of assignment is only 0.01. Thus, we assert that that when p = o(log n), the output of Transformer is completely determined by the partial assignment and it do not detect whether there exists error in the unassigned positions and thus can not recognize Dyckk,2 language. We conclude the proof here. D Experiment Details D.1 Setup Data We follow Hewitt et al. (2020) to generate Dyckk,D by randomly sampling stack decisions (push, pop, or end) and maintaining length conditions (Table 1) for a O(D2) hitting time of different DFA states. The number of tokens for train, validation, and test set is 2 × 106, 2 × 105, 106 respectively. D 3 5 10 15 Train/val lengths 1:84 1:180 1:700 1:1620 Test lengths 85:168 181:360 701:1400 1621:3240 Table 1: Input lengths for Dyckk,D with different D. Models We use the LSTM model implemented in Hewitt et al. (2020). For Transformer models, we turn off all drop outs as we find them to hurt performance greatly. We also use only 1 head as we find more heads to hurt performance. We use Adam optimizer with initial learning rate being 0.01 or 0.001, and choose the better learning rate in terms of validation accuracy for each experiment. We train for at most 100 epochs but allow early stopping if the validation loss converges. 3785 Metric We follow Hewitt et al. (2020) and use the accuracy of correct close bracket predictions: p(⟩j|⟩) = p(⟩j) P i p(⟩i) Let pl be the empirical probability that the model confidently predicts a close bracket (defined as p(⟩j|⟩) > .8), conditioned on it being separated from its open bracket by l tokens. Unlike Hewitt et al. (2020) where meanlpl is reported, we report Elpl for two reasons: (i) when l is large pl might be only defined by one trail, thus meanlpl amplifies the randomness; (ii) the findings remain similar with either metrics. D.2 More Results In Figure 7, we show the validation performance for Transformers of different positional encoding schemes. They all reach near-perfect accuracy when having at least 2 layers. In Figure 8, we break down the results in Section 6.2 when dmodel ∈{10, 30, 50}. We also add results for a five-layer Transformer, which performs similarly as the two-layer Transformer. This shows (i) a two-layer Transformer, as suggested by our theory, is enough to process Dyckk,D, and (ii) Transformers with more layers can also learn to process Dyckk,D without overfitting or degraded performance. 1 2 3 4 5 10 # Layers 0.6 0.7 0.8 0.9 1.0 Close Accuracy Transformers (Dyck-(8, 10) Validation) Positional Encoding cos learn pos/N Figure 7: Validation results on Dyck8,10. 0.0 0.2 0.4 0.6 0.8 1.0 dev_close_acc k = 2 | D = 3 Model LSTM (1 layer) Transformer (2 layers) Transformers (5 layers) k = 2 | D = 5 k = 2 | D = 10 k = 2 | D = 15 0.0 0.2 0.4 0.6 0.8 1.0 dev_close_acc k = 8 | D = 3 k = 8 | D = 5 k = 8 | D = 10 k = 8 | D = 15 0.0 0.2 0.4 0.6 0.8 1.0 dev_close_acc k = 32 | D = 3 k = 32 | D = 5 k = 32 | D = 10 k = 32 | D = 15 10 30 50 hidden_dim 0.0 0.2 0.4 0.6 0.8 1.0 dev_close_acc k = 128 | D = 3 10 30 50 hidden_dim k = 128 | D = 5 10 30 50 hidden_dim k = 128 | D = 10 10 30 50 hidden_dim k = 128 | D = 15 Transformers v. LSTM (Dyck-(k, D) Validation) 0.0 0.2 0.4 0.6 0.8 1.0 test_close_acc k = 2 | D = 3 k = 2 | D = 5 k = 2 | D = 10 k = 2 | D = 15 0.0 0.2 0.4 0.6 0.8 1.0 test_close_acc k = 8 | D = 3 k = 8 | D = 5 k = 8 | D = 10 k = 8 | D = 15 0.0 0.2 0.4 0.6 0.8 1.0 test_close_acc k = 32 | D = 3 k = 32 | D = 5 k = 32 | D = 10 k = 32 | D = 15 10 30 50 hidden_dim 0.0 0.2 0.4 0.6 0.8 1.0 test_close_acc k = 128 | D = 3 10 30 50 hidden_dim k = 128 | D = 5 10 30 50 hidden_dim k = 128 | D = 10 10 30 50 hidden_dim k = 128 | D = 15 Transformers v. LSTM (Dyck-(k, D) Test) Figure 8: Validation and test results on Dyckk,D (k ∈ {2, 8, 32, 128} and D ∈{3, 5, 10, 15}). Enlarge for details.
|
2021
|
292
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3786–3800 August 1–6, 2021. ©2021 Association for Computational Linguistics 3786 TextSETTR: Few-Shot Text Style Extraction and Tunable Targeted Restyling Parker Rileya∗, Noah Constantb, Mandy Guob, Girish Kumarc∗, David Uthusb, Zarana Parekhb aUniversity of Rochester bGoogle Research cStanford University Abstract We present a novel approach to the problem of text style transfer. Unlike previous approaches requiring style-labeled training data, our method makes use of readily-available unlabeled text by relying on the implicit connection in style between adjacent sentences, and uses labeled data only at inference time. We adapt T5 (Raffel et al., 2020), a strong pretrained text-to-text model, to extract a style vector from text and use it to condition the decoder to perform style transfer. As our labelfree training results in a style vector space encoding many facets of style, we recast transfers as “targeted restyling” vector operations that adjust specific attributes of the input while preserving others. We demonstrate that training on unlabeled Amazon reviews data results in a model that is competitive on sentiment transfer, even compared to models trained fully on labeled data. Furthermore, applying our novel method to a diverse corpus of unlabeled web text results in a single model capable of transferring along multiple dimensions of style (dialect, emotiveness, formality, politeness, sentiment) despite no additional training and using only a handful of exemplars at inference time. 1 Introduction There has been a recent surge of interest in text style transfer, with the aim of training models able to modify specific attributes of input text (e.g., sentiment or formality) while preserving the remaining content. For example, a sentiment transfer model might transform the input “best book ever!” into “worst book ever!”, while a formality transfer model might change the same input into “This is the best book I have ever read.” In these contexts, we define “style” as the attributes intended to be changed, ∗Work done while at Google Research. Please direct correspondence to [email protected], [email protected] and [email protected]. while “content” consists of the attributes intended to be preserved.1 Work in this area falls into three categories. Supervised approaches like Jhamtani et al. (2017) transfer between pre-selected styles, and rely on parallel training data to learn the desired input/output correspondence. This method is limited by the availability of parallel corpora. So-called “unsupervised” approaches like Li et al. (2018) and Lample et al. (2019) remove the need for parallel data, but still require that all training examples have style labels, and are limited to transfer between a pre-specified set of styles. Few-shot approaches like that of Xu et al. (2020) remove the need for any training labels, instead using a small number of labeled examples during inference. While the most challenging, this offers the potential for transferring between arbitrary styles at inference time and has significant value, as curated datasets are not available for many style attributes. In this work, we explore the hypothesis that large pretrained text-to-text models like T5 (Raffel et al., 2020) already contain a strong representation of textual style, which can be extracted and used to condition the decoder of a style transfer model through a relatively lightweight fine-tuning procedure. To isolate style information in the absence of labels, we rely on the observation that style is a “slow-moving” feature, which tends to be consistent over large spans of text. Specifically, given two adjacent sentences from an unlabeled corpus, we train our model to extract a “style vector” from the first and use that vector to perform denoising and other reconstruction tasks on the second. This technique extends the approach of Lample et al. (2019) to the few-shot setting, and is loosely reminiscent of the work of Akama et al. (2018), who found 1Krishna et al. (2020) use a different definition of style, under which certain transfers such as sentiment would instead be examples of attribute transfer. 3787 large context windows useful for encoding style information in word embeddings. Our approach also allows us to reformulate the style transfer operation as a directional operation in style vector space using the difference between target and source style vectors; we call this “targeted restyling”. When combined with a novel “tunable inference” technique for controlling token add/delete rates, this gives our final model: Text Style Extraction and Tunable Targeted Restyling (TextSETTR). Our main contributions are to: (1) present a new, flexible approach to few-shot style transfer, (2) use sentence adjacency as a means for inducing text style representations, (3) reframe style transfer as “targeted restyling” directional operations in style space, (4) introduce “tunable inference” for finergrained control of transfers, (5) show the effectiveness of “noisy” back-translation training, and (6) illustrate few-shot generalization to a range of style attributes including dialect, emotiveness, formality, politeness, and sentiment. 2 Method Figure 1 illustrates our proposed TextSETTR architecture. At a high level, our approach follows Lample et al. (2019), who train a denoising autoencoder conditioned on a fixed-width style vector. The key difference in our case is that the true style is unknown at training time. To overcome this, we jointly train a “style extractor” component to induce a useful style representation (that can aid in reconstruction) from text in the nearby context. We describe this in more detail below. 2.1 Model Architecture We conduct our experiments using a modified version of the Text-to-Text Transfer Transformer (T5) (Raffel et al., 2020). Like T5, our model includes a transformer-based encoder and decoder. As in T5 pretraining, the input to the encoder is a corrupted version of the target, resulting in a reconstruction task. Our goal is to design a type of corruption that results in this training task resembling style transfer, despite the lack of labeled training data. Our core addition to T5 is the style extractor. This component’s architecture is based on that of the encoder, and its input is an uncorrupted sentence in the same style as the target; relying on our assumption that style is a slow-moving feature, we use the sentence preceding the target (the “context”) for this. This encourages extracting a style representation that is useful for repairing the corrupted input. We note that this can result in a representation that encodes slow-moving attributes in general, which may include some features that do not fit an intuitive definition of textual style (such as topic). The only architectural difference between the encoder and style extractor is that we mean-pool the style extractor’s hidden state sequence into a single fixed-width “style vector”; in our experiments, the dimensionality of this vector and the encoder hidden states is 1024. To incorporate the style vector into the rest of the model, we simply add it to each of the final encoder hidden states. We initialize the weights of our model with those of a pretrained T5 model. We initialize both the style extractor and encoder from the pretrained encoder, but the weights are not tied during training. 2.2 Corruption Strategies We experiment with combinations of three different reconstruction tasks, each contributing a loss term. All three share the same overall structure, where a sentence si in the dataset is corrupted by some function f to produce ˜si = f(si). The crossentropy loss is calculated using the uncorrupted sentence si as the target, the corrupted sentence ˜si as the input, and the uncorrupted preceding sentence si−1 as the context. The three choices of f are Noise (N), Back-Translation (BT), and Noisy Back-Translation (NBT), described below. Noise (N) This function corrupts the input by (i) dropping, (ii) replacing, and/or (iii) shuffling tokens, in that order. For each example we sample a separate noise probability p for each sub-type of noise from a uniform distribution in the range 20–60%; doing so should widen the model’s range of possible style transfers at test time. For drop noise, we drop each token in si with probability p. For replace noise, let sik be the kth token within si. For each si, a random other example sj is chosen, and then each token sik is replaced by sjk with probability p. If sj has fewer than k tokens, then the replacement does not occur. For shuffle noise, each token in si is chosen with probability p, and then all chosen tokens are randomly shuffled to the position of another chosen token, leaving non-chosen tokens in place. The use of drop and shuffle noise results in a loss term similar to the denoising loss used by Lample et al. (2019). Their motivation for this loss was 3788 Style Target λ × (A − B) + Inp Tuning Ranges Add Delete 40-70% 25-35% It doesn’t work Input Encoder Decoder Output It works great Ex Ex Style A Exemplars Ex Ex Style B Exemplars Ex A great product. Context I really love it Input Corruption cat really it Encoder Style Extractor Decoder Target I really love it Training Inference Tuning Ranges Add Delete 10-50% 10-50% Figure 1: TextSETTR architecture for few-shot style transfer. The Encoder, Decoder and Style Extractor (Ex) are transformer stacks initialized from pretrained T5. During training, the model reconstructs a corrupted input, conditioned on a fixed-width “style vector” extracted from the preceding sentence. At inference time, a new style vector is formed via “targeted restyling”: adding a directional delta to the extracted style of the input text. Stochastic tuning ranges provide extra conditioning for the decoder, and enable fine-grained control of inference. to encourage language modeling. As we fine-tune an already-strong T5 language model in our experiments, our motivation is rather to introduce a conditional element to the language model, in the form of the extracted style vector input. Back-Translation (BT) This corruption function, used by Lample et al. (2019), runs the current version of the model in inference mode to transfer si into a different style, giving the corrupted ˜si. In prior work using labels, specifying a different target style was straightforward. In our case, because we do not have access to labels, we simply sample a random sentence sj to use as the context. To increase diversity of the generated examples, we decode with sampling instead of greedy decoding. Because ˜si is produced by a strong language model, BT should result in examples where both the input and output are coherent sentences, matching our inference setting. By contrast, Noise corruption does not resemble test-time inputs. Noisy Back-Translation (NBT) This novel corruption function is a composition of the previous two. Noise is first applied to si as described above, and the result is used as the input (with randomlysampled sj as the context) to the model in inference mode to produce ˜si via sampling, as in BT. Once the model has learned to undo random noise, NBT should produce training examples where some of the tokens are preserved from si while others were generated by the model itself under the influence of the “incorrect” context sj. This is similar to BT, but we hypothesize that it may be better suited to style transfer. BT was originally used for machine translation (Sennrich et al., 2016), a setting where most or all input tokens need to change. In contrast, style transfer within a single language usually requires only changing a subset of tokens; the training examples resulting from NBT should have this property. We believe that this will encourage the model to identify which tokens in the input do not match the target style indicated by si−1 and change them, which is exactly what we want a style transfer model to do. Final Loss The final loss term used for training is the sum of the above loss terms, each calculated from the same input si. However, not every model we experiment with includes all three losses. 2.3 Inference Procedure Tunable Add/Delete Rates In preliminary experiments, we observed a recurring problem that the model would often change either far too little (failing to achieve the target style), or far too much (failing to preserve the input content). To address this problem, we introduce a “tunable inference” mechanism to constrain how much content should be added and deleted at inference time. For every input/output pair during training, we calculate the proportions of tokens that were added and deleted. The “add rate” is the proportion of output tokens absent from the input, and the “delete rate” is the proportion of input tokens absent from the output.2 We provide these rates to the decoder as ranges covering but not necessarily centered 2This calculation ignores word order. As one example, if a token appears three times in the input and five times in the output, two of the five occurrences are counted as “added”. 3789 on the true rates.3 This approach provides more flexibility at inference time, so we can enforce tight or loose constraints on each rate. Targeted Restyling While previous work on style transfer has largely assumed a fixed set of discrete styles, we expect our model’s learned style representations to capture a rich summary of the sentence covering many attributes without specifying them beforehand. For example, a given style vector might encode that a sentence is informal, humorous, in British English, and so on. In this framework, transferring a single attribute (e.g., informal →formal) is not as simple as just providing a vanilla “formal” style target, as this would ignore all the other attributes that defined the original input. Rather, we must operate in style space to construct a new target style that is simultaneously formal, humorous, British, and so on. Concretely, at inference time, we assume access to a small set of “exemplar” sentences (between 1 and 100) for both the source value (e.g., informal) and target value (e.g., formal) of the attribute being modified. We infer style vectors for each exemplar using the style extractor, and take the mean of each class, giving vectors vsrc and vtrg. Assuming the exemplar pools are relatively diverse, this averaging should “wash out” most untargeted attributes. To transfer an input sentence x, we apply a targeted restyling in the appropriate direction. After extracting the original style from the input itself, vx, we compute the target output style by moving in the direction of the delta between the source and target attributes values, as in (1), producing the style vector used for decoding. In practice, we find that the delta scale λ is an important hyperparameter to tune. Generally values in the range [1.0, 10.0] work well, with the best values depending on the attribute and the exemplars in question. vx + λ × vtrg −vsrc (1) 3 Experiments on Sentiment Transfer To evaluate our approach and better understand the effects of our various design choices, we test on few-shot sentiment transfer, using the Amazon reviews dataset of Li et al. (2018). However, as their training split doesn’t indicate which sentences 3Specifically, we sample each range width uniformly from [0,1], and uniformly sample the “alignment” of the true rate within the range. The final ranges are clipped to [0,1], and a vector containing the upper and lower bound of each range is prepended to the encoder hidden state sequence. were adjacent in the original reviews, we make use of a different source of raw review text. Training Procedure Our unlabeled training data comes from the 233.1M Amazon reviews provided by Ni et al. (2019). Ignoring the star ratings completely, we extract adjacent lines from multi-line reviews to use as the context and input for our training procedure, giving 23.6M examples. We also preprocess all text to match the format of the Li et al. (2018) data, as detailed in Appendix A.4. Initializing our model from pretrained T5 (t5.1.1.large), we fine-tune on these examples, optimizing the joint reconstruction loss from Section 2. Our default TextSETTR configuration is selected based on preliminary experiments (on development data) varying the set of reconstruction tasks and inference procedures. The model uses an equally weighted combination of the Noise (N) and Noisy Back-Translation (NBT) tasks. For both tasks, we use drop and replace noise, but no shuffle noise. We fine-tune for 10k steps, with a batch size of 65,536 tokens, and a fixed learning rate of 1e-3. Evaluation Procedure Following prior work, we use automatic metrics to assess attribute control (sentiment) and content preservation on the data from Li et al. (2018). To estimate the sentiment of the output, we fine-tune a BERT-Large classifier (Devlin et al., 2019) on the train split, scoring 87.8% accuracy on the dev split. For content preservation, we follow Sudhakar et al. (2019) and Xu et al. (2020) and calculate self-BLEU between the output and input, using SacreBLEU (Post, 2018).4,5 Following Xu et al. (2018), we report “G-score” (the geometric mean of accuracy and content) as a summary of overall model quality. To perform transfers, we follow the procedure from Section 2.3. For our default setup, we sample 100 positive and 100 negative exemplars from the Li et al. (2018) train split. Unless otherwise specified, we use greedy decoding, a delta scale of λ=8, and add/delete tuning ranges of 20–40%. Core Results Figure 2 shows our core results. Our default TextSETTR configuration (N+NBT training, tuning ranges 20–40%) achieves 73.7% classifier-judged accuracy at swapping sentiment, while still staying somewhat close to the original 4Version string: BLEU+case.mixed+numrefs.1+ smooth.exp+tok.13a+version.1.4.13 5Some prior work reports instead BLEU scores between outputs and human-generated transfers from Li et al. (2018); we found this to be highly correlated with self-BLEU but report it in Appendix A.3 for completeness. 3790 Model Acc. Content G Few-Shot TextSETTR (10–30%) 54.0 55.8 54.9 TextSETTR (20–40%) 73.7 34.7 50.6 N 23.4 84.4 44.4 NBT 70.0 27.8 44.1 N + BT 13.3 98.7 36.2 −replace noise 66.1 42.1 52.8 +shuffle noise 70.3 34.1 49.0 manual exemplars 52.4 44.2 48.1 1000 exemplars 74.5 37.2 52.6 −tunable inference 71.5 39.4 53.1 overwrite style 25.3 55.8 37.6 small train set 74.5 33.4 49.9 CP-G 51.1 35.5 42.6 CP-B 36.3 39.8 38.0 Labeled CrossAligned 68.2 2.9 14.1 Delete&Retrieve 49.4 56.9 53.0 B-GST 60.2 54.2 57.1 0 20 40 60 80 100 Content Preservation (Self-BLEU) 0 20 40 60 80 100 Sentiment Transfer Accuracy 0 10% 0 20% 10 30% 20 40% 30 50% 40 60% 50 70% N N(50k) NBT N+BT N+BT (50k) tunable +shuffle replace manual overwrite 1000-exemplars small-train CP-G CP-B CrossAligned Delete&Retrieve B-GST TextSETTR TextSETTR ablations Other label-free models Models trained with labels Figure 2: Automatic evaluation metrics comparing our TextSETTR model, ablations, and previous work. Upand-right is better. We train for 10k steps and use add/delete:20–40% unless otherwise specified. We recalculate metrics for previous approaches, using our BERT classifier for accuracy, ensuring direct comparability. Model Accuracy Content G TextSETTR (10–30%) 72.7 60.2 66.2 TextSETTR (20–40%) 83.6 39.4 57.4 Lample et al. 2019 82.6 54.8 67.3 Table 1: Comparison with Lample et al. (2019) on the setting that includes pos→pos and neg→neg transfers. input text (self-BLEU 34.7). Due to our tunable inference technique, we can also trade off accuracy for content preservation by adjusting the add/delete rates, as seen in the points along the green line. Notably, TextSETTR outperforms the few-shot CP-G and CP-B models of Xu et al. (2020). More remarkably, TextSETTR outperforms several approaches that rely on training labels: CrossAligned (Shen et al., 2017) and Delete&Retrieve (Li et al., 2018). However there is still a small gap between our fewshot approach and the best labeled model, B-GST (Sudhakar et al., 2019). In Table 1, we compare with Lample et al. (2019) on the evaluation setting including pos→pos and neg→neg transfers. This setting doesn’t match our inference procedure, which assumes that the input and output styles differ. Nevertheless, TextSETTR comes close to the performance of Lample et al. (2019), despite not benefiting from training labels. As automatic metrics can diverge from human judgment (Sudhakar et al., 2019), we also conduct human evaluations of the three strongest models from Figure 2. We sample 200 examples per transfer direction from the Li et al. (2018) test set, and ask three annotators to evaluate each input/output Model Sentiment Preservation Fluency TextSETTR (10–30%) 2.0 3.5 2.9 TextSETTR (20–40%) 2.5 2.6 4.0 Delete&Retrieve 2.5 3.1 3.3 B-GST 2.2 2.9 3.6 Table 2: Human evaluation metrics. pair on three metrics: sentiment transfer (how well the model changed the sentiment), content preservation, and fluency, on scales of 1–5. The results in Table 2 confirm that TextSETTR achieves similar quality to models that benefit from training labels. Further details are presented in Appendix A.5. 3.1 Ablations Modifying Inference Procedure To better understand the value of our proposed “targeted restyling” mechanism, we consider an alternative inference procedure where we ignore the style of the input and simply use the average target exemplar style vtrg as the style vector. We expect that since our learned style space covers multiple attributes, this will result in setting the target attribute (e.g. sentiment) while simultaneously overwriting all other style attributes (e.g. formality) using the average style of the target exemplars. This is borne out in our “overwrite style” ablation, which performs significantly worse than our baseline: accuracy drops from 54.0% to 25.3% with no gain in self-BLEU. To assess the value of tunable add/delete rates, we also train a model (−tunable) without this feature. While the automatic metrics are slightly above the TextSETTR line, we observe several advan3791 tages to the tunable model. For one, we observe it significantly reduces the variance in self-BLEU across different inputs. For example, focusing on the case of overly high self-BLEU, we find that without tunable inference, 14.6% of dev eval outputs are identical to their inputs, whereas with tunable inference, this goes to 0.9%. Additionally, through qualitative analysis in Section 4, we find that tunable inference allows more flexibility for controlling different types of transfer. Adjusting Data Sizes While our unlabeled training data set consists of 23.6M examples, our model only sees 5.1M of these over its 10k steps of training. Yet this is still nearly 10× more data than the 0.6M examples in the Li et al. (2018) training set used by previous approaches. For a more direct comparison, we experiment with a “small train set”, sampling 0.6M examples from our training set. Remarkably, the results in Figure 2 are nearly identical to our baseline, supporting our hypothesis that a fairly lightweight adaptation is sufficient to allow T5 to extract and transfer textual style. To test the limits of our model’s generalization, we reduce the set of exemplars to four manually selected examples of each class. In this setting, we also find reducing delta scale to λ=4 is beneficial. The results, shown as “manual exemplars” in Figure 2, are still competitive, indicating that our approach generalizes well to this very-few-shot inference setting. In the other direction, we find that increasing the number of sampled exemplars from 100 to 1000 only provides small additional gains. Modifying Training Task Lample et al. (2019) showed promising results by combining noise (N) with back-translation (BT). However we find this combination unstable.6 When training for 10k steps, our N and N+BT models nearly always copy their input. Training for 50k steps recovers reasonable performance, but the metrics still fall below the TextSETTR line, using our novel NBT task. We also experiment with using NBT in isolation, but this again underperforms our baseline. We expect that the denoising task helps to ensure the NBT inputs (themselves the outputs of denoising) consist of realistic well-formed text. Finally, while Lample 6For all experiments in the paper, we use 0.0 for the add/delete rates during the forward pass of back-translation. However we later found that using random add/delete rates in back-translation can improve performance in the N+BT setting. On sentiment transfer, this improved our N+BT ablation to self-BLEU 42.4, accuracy 71.4%, G-score 55.0. et al. (2019) use drop and shuffle noise, we find that only drop and replace are valuable. 3.2 Embedding Visualization To demonstrate that our learned style extractor encodes multiple aspects of textual style, we compute style vectors for 12,000 lines of text from three review categories (Fashion, Software, Pantry) from the Ni et al. (2019) Amazon data. Within each category, we sample 2,000 positives (4 or 5 star) and 2,000 negatives (1 or 2 star), filtering examples where our BERT classifier disagrees with the label. Figure 3 (bottom) plots a 2D UMAP dimensionality reduction (McInnes et al., 2018) of the vectors, and shows clear separations among sentiments and product categories. The top row runs UMAP with the same settings, but over style vectors from our model before training, where the style extractor is initialized from pretrained T5. The contrast is a clear indication that our training procedure is helping to learn a representation space where sentiment and topic values are well separated. To confirm that the observed separation isn’t an artifact of dimensionality reduction, we compute the average distance between style vectors (a) within a class, and (b) across classes. We measure “separation” as the relative increase in mean distance between these two conditions. For product category, we find TextSETTR training improves separation from 1.7% to 8.1%. For sentiment, TextSETTR training improves separation from 0.9% to 4.7%. 4 One Model for All Styles An advantage of few-shot style transfer is that, in theory, a single model can perform transfer along any “dimension” of style given only a few exemplars, without the need for additional training. In this section, we investigate the degree to which our approach achieves this goal in practice. For this purpose, we train a single general-purpose TextSETTR model, with the same configuration as our model from Section 3, except fine-tuned for 200k steps on English Common Crawl data (the same “C4” data that T5 pretrained on) instead of Amazon reviews. Qualitative Evaluation Given that our architecture limits the style representation to 1024 dimensions, one may ask how the unsupervised model will make use of this capacity, and which style attributes will be encoded in the learned space. Encouragingly, we find that our model trained on un3792 Before TextSETTR training (pretrained T5 initialization) After TextSETTR training Figure 3: 2D UMAP embeddings of the style vectors extracted by our TextSETTR model before and after training, for text inputs from Amazon reviews covering three product categories and two sentiment labels. Within each row, the same embeddings are labeled with product category (left) and sentiment (right). We sub-sample to 3,000 points after dimensionality reduction. Note, we don’t expect perfect separation, as inputs may be underspecified for category (“I love this product”) or for sentiment (“I bought this last month”). We also don’t expect to see crisp linear separation within each attribute since we aim for the learned embedding space to encode many style attributes simultaneously. Reserved ⇒Emotive Emotive ⇒Reserved I liked the movie. ⇒I cannot even describe how amazing this movie was!! I loved every minute of the movie! ⇒I liked the movie. I was impressed with the results. ⇒I was absolutely blown away with the results!! I was shocked by the amazing results! ⇒I was surprised by the results. American ⇒British British ⇒American The elevator in my apartment isn’t working. ⇒The lift in my flat isn’t working. The lift in my flat isn’t working. ⇒The elevator in my apartment isn’t working. The senators will return to Washington next week. ⇒The MPs will return to Westminster next week. MPs will return to Westminster next week. ⇒Representatives will return to Washington next week. Polite ⇒Rude Rude ⇒Polite Are you positive you’ve understood my point? ⇒you’ve never understood my point! What the hell is wrong with your attitude? ⇒Perhaps the question is more about your attitude. Could you ask before using my phone? ⇒I ask you to stop using my phone! I could care less, go find somebody else to do this crap. ⇒I could be wrong, but I would try to find somebody else to do this. Formal ⇒Informal Informal ⇒Formal I hereby commit to never purchase anything from this institution in the future. ⇒i gonna never buy anything from this place again. best book ever!! ⇒The book is highly recommended. I couldn’t figure out what the author was trying to say. ⇒i dont know what ur trying to say. couldnt figure out what author tryna say ⇒The reader couldn’t figure out what the author was trying to say. Positive ⇒Negative Negative ⇒Positive I was pretty impressed with the results. ⇒I was pretty disappointed with the results. I was pretty disappointed with the results. ⇒I was pretty impressed with the results. I will definitely buy this brand again. ⇒I will definitely not buy this brand again. I definitely won’t buy this brand again. ⇒I definitely won’t hesitate to buy this brand again. Table 3: Examples of transferring along five different axes of style. The same model is used across all examples, with no additional training. Words deleted from the input are red, and words added in the output are blue. Within each category, a fixed tiny set of exemplars is chosen, and fixed delta scale and tuning rates are used. The exemplars and settings are provided in Appendix A.2. 3793 labeled Common Crawl data is capable of transferring along many independent axes of style. Table 3 shows selected successful examples of our Common Crawl model transferring emotiveness, dialect, politeness, formality and sentiment. The same model is used in each case, with no additional training. At inference time, a tiny set of exemplars (1–5 examples of each class) is the only labeled data used to compute the style vector delta; these exemplars are presented in Appendix A.2. Across each type of transfer, we see evidence of generalization beyond the specifics of the chosen exemplars. In making text more emotive, the model uses amazing and blown away, despite these terms not occurring in the exemplars. In making text more polite, the model inserts novel hedges like perhaps and I could be wrong. In transferring between American and British styles, the model generalizes to unseen vocabulary items (elevator ↔lift) and draws sound analogies (senators ↔MPs). We do note though that the latter case illustrates that the model is willing to change the semantic content of the input in cases where it would otherwise be outof-place in the target style. Future work includes investigating ways to control this in settings where such behavior is not desired. Quantitative Evaluation To assess the quality of our general-purpose TextSETTR model, we benchmark the same model on three distinct transfer tasks in Table 4.7 The sentiment transfer task follows the evaluation procedure from Section 3. While our generic model underperforms our model trained on Amazon reviews, it still outperforms other few-shot methods. For author transfer, we use the Shakespeare-to-modern task of Jhamtani et al. (2017). Here, TextSETTR outperforms the previous best model of He et al. (2020) that leveraged 36,790 labeled examples during training. For personality transfer, we use the task of Li et al. (2020), which requires transferring between three personalities: angry, happy, malicious. We compare8 TextSETTR, which sees no labels in training and only 100 of each class in inference, with CARA (Li et al., 2020), which trained on 2,604 labels. 7For each task, we set our tuning ranges to 20–40% and compute target styles using 100 exemplars of each class taken from the train set. We use λ values of sentiment:8, author:16, personality:8. To measure accuracy, we fine-tune BERT-Large classifiers over the training data, reaching validation accuracies of sentiment:87.8%, author:89.7%, personality:81.9%. 8Note, as Li et al. (2020) use a different classifier to assess accuracy, those numbers may not be directly comparable. Task Model Acc. Content G Sentiment CP-G 51.1 35.5 42.6 CP-B 36.3 39.8 38.0 TextSETTR 44.9 54.4 49.4 CrossAligned 68.2 2.9 14.1 Delete&Retrieve 49.4 56.9 53.0 B-GST 60.2 54.2 57.1 Author UNMT 68.5 7.8 23.1 BT+NLL 59.3 12.4 27.1 He et al. 2020 68.5 12.5 29.2 TextSETTR 81.7 13.8 33.5 Personality CARA 91.6 21.6 44.5 CARAAB 66.2 29.7 44.3 Ctrl-Gen 67.6 22.9 39.3 ARAE− 88.0 20.3 42.3 TextSETTR 49.3 46.0 47.6 Table 4: Automated metrics comparing our generalpurpose TextSETTR model with recent work on three transfer tasks. To enable direct comparison, “content” refers to reference-BLEU for author transfer, and selfBLEU elsewhere. Apart from CP-G/CP-B, all competitors are trained for only one type of transfer using labeled data. Personality transfer results are from Li et al. (2020), while all others are recalculated from scratch. 4.1 Dialect-Sensitive Completion In addition to performing style and attribute transfer, we find that our system can also be used as a style-aware language model capable of completing prompts in a specified style. Examples of completions in American and British English are given in Table 5. In each case, the input is of the form “My favorite X: ”. Despite the fact that TextSETTR is not trained specifically for completions, we can use the add/delete rates to encourage the model to insert a few additional tokens, while leaving the original prompt largely unchanged.9 The completions demonstrate knowledge of stereotypical American and British culture. It is remarkable that the model is able to generalize to “deeper” cultural differences such as music and drink preferences, given only the shallow vocabulary differences (e.g., neighbor vs. neighbour) presented in the limited set of exemplars in Table 9. It is also worth highlighting that, thanks to our directional transfer procedure, these completions are not merely “typical American” or “typical British” such as we would expect from a conditional language model trained on each sub-domain of text. Rather, since our inference procedure pushes the style away from one domain and towards the other, the resulting completions are distinctive representations of each dialect. As one example, we expect 9We note that in transferring American to British, the model prefers to change the prompt from favorite to favourite. 3794 American ⇒British British ⇒American My favourite food: fish and chips. My favorite food: quinoa. My favourite hot drink: a mug of tea. My favorite hot drink: Starbucks Coffee. My favourite dessert: a scone! My favorite dessert: a brownie. My favourite city: Cardiff. My favorite city: San Diego. My favourite band: The Beatles. My favorite band: The Black Keys. My favourite sports league: the English Premier League. My favorite sports league: the NFL. My favourite newspaper: The Daily Telegraph. My favorite newspaper: The Washington Post. My favourite museum: the British Museum. My favorite museum: The National Air and Space Museum. Table 5: Examples of dialect-sensitive completion (λ=8, add:40–70%, delete:0%). In each case, the input text consists of an unfinished phrase, for example: “My favorite food: ”. The three exemplars used for each dialect are the same as those used for the transfers in Table 3, as listed in Table 9. “quinoa” would not only be a common American favorite, but also an uncommon British favorite. Additional examples of using our model for tasks other than pure style transfer are presented in Appendix A.1. 5 Related Work As mentioned at the outset, recent work on text style transfer falls into three classes: supervised, “unsupervised”, and few-shot. Supervised style transfer has seen limited research due to the difficulty of obtaining parallel data. Examples include Jhamtani et al. (2017) and Carlson et al. (2018). Unsupervised Approaches The bulk of research has focused on “unsupervised” approaches, which rely on labeled but non-parallel data. Typically, labels are assumed to be available for both source and target styles (Shen et al. 2017, Li et al. 2018, Niu et al. 2018, and many others). Zhao et al. (2018) explore the case where only the target style is labeled. The use of labels at training time can aid modeling, but limits the applicability of these methods, as labeled datasets are not readily available for many attributes of interest. Our work differs from the above by removing the need for training labels, and offering a single model that can target an unrestricted set of style attributes. Despite these differences, our work shares some similarities with past work. For example, our encoder-decoder architecture and corruption methods are similar to Lample et al. (2019), and we leverage a strong pretrained language model, as in Sudhakar et al. (2019) and Wu et al. (2019). Few-Shot Approaches A few-shot approach has recently been explored by Xu et al. (2020). The authors train a variational auto-encoder on unlabeled text, where a “manipulable” portion of the latent representation is constrained to fall on a k-dimensional simplex. To perform transfer, they identify empirically the basis vector that most strongly corresponds to the target attribute, and manipulate its magnitude. Compared to our approach, a key difference is that the number of latent factors must be chosen ahead of time, which limits the number of attributes that may be controlled. Additionally, there is no guarantee that a single basis of the learned simplex will correspond to a target attribute such as dialect or politeness. Controlled Generation A separate strand of research explores “controlled generation” methods for supplementing generative language models to allow control of specific attributes of the output text. As with style transfer, this can be achieved either through labeled training examples, as in CTRL (Keskar et al., 2019) and PPLM (Dathathri et al., 2020), or a few-shot approach, as in CoCon (Chan et al., 2020). These models differ from style transfer models in that they aim to generate plausible continuations following a prompt, as opposed to transferring attributes of a fully-formed input while preserving as much content as possible. It is not clear if controlled generation models could be used to perform style transfer, and they have not to our knowledge been evaluated in this context. 6 Conclusion We have presented a unique approach to few-shot text style transfer that is competitive with systems trained with labels (an easier setting), while allowing control of how much of the input is changed. We demonstrate that this approach can produce a single system capable of transferring many different styles while requiring only a handful of exemplars at inference time. Acknowledgments We thank Llion Jones, Rami Al-Rfou, and Daniel Gildea for helpful discussion and comments on an earlier draft. 3795 References Reina Akama, Kento Watanabe, Sho Yokoi, Sosuke Kobayashi, and Kentaro Inui. 2018. Unsupervised learning of style-sensitive word vectors. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 572–578, Melbourne, Australia. Association for Computational Linguistics. Keith Carlson, Allen Riddell, and Daniel Rockmore. 2018. Evaluating prose style transfer with the bible. Royal Society Open Science, 5(10):171920. Alvin Chan, Yew-Soon Ong, Bill Pung, Aston Zhang, and Jie Fu. 2020. CoCon: A self-supervised approach for controlled text generation. CoRR, abs/2006.03535. Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representations. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Junxian He, Xinyi Wang, Graham Neubig, and Taylor Berg-Kirkpatrick. 2020. A probabilistic formulation of unsupervised text style transfer. In International Conference on Learning Representations. Harsh Jhamtani, Varun Gangal, Eduard Hovy, and Eric Nyberg. 2017. Shakespearizing modern language using copy-enriched sequence to sequence models. In Proceedings of the Workshop on Stylistic Variation, pages 10–19, Copenhagen, Denmark. Association for Computational Linguistics. Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard Socher. 2019. CTRL: A conditional transformer language model for controllable generation. CoRR, abs/1909.05858. Kalpesh Krishna, John Wieting, and Mohit Iyyer. 2020. Reformulating unsupervised style transfer as paraphrase generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 737–762, Online. Association for Computational Linguistics. Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer, Marc’Aurelio Ranzato, and YLan Boureau. 2019. Multiple-attribute text rewriting. In International Conference on Learning Representations. Juncen Li, Robin Jia, He He, and Percy Liang. 2018. Delete, retrieve, generate: a simple approach to sentiment and style transfer. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1865–1874, New Orleans, Louisiana. Association for Computational Linguistics. Yuan Li, Chunyuan Li, Yizhe Zhang, Xiujun Li, Guoqing Zheng, Lawrence Carin, and Jianfeng Gao. 2020. Complementary auxiliary classifiers for labelconditional text generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8303–8310. Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. UMAP: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861. Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 188–197, Hong Kong, China. Association for Computational Linguistics. Xing Niu, Sudha Rao, and Marine Carpuat. 2018. Multi-task neural models for translating between styles within and across languages. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1008–1021, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186– 191, Belgium, Brussels. Association for Computational Linguistics. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-totext transformer. Journal of Machine Learning Research, 21(140):1–67. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86–96, Berlin, Germany. Association for Computational Linguistics. Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, 3796 and R. Garnett, editors, Advances in Neural Information Processing Systems 30, pages 6830–6841. Curran Associates, Inc. Akhilesh Sudhakar, Bhargav Upadhyay, and Arjun Maheswaran. 2019. “Transforming” delete, retrieve, generate approach for controlled text style transfer. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3269– 3279, Hong Kong, China. Association for Computational Linguistics. Xing Wu, Tao Zhang, Liangjun Zang, Jizhong Han, and Songlin Hu. 2019. Mask and infill: Applying masked language model for sentiment transfer. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI19, pages 5271–5277. International Joint Conferences on Artificial Intelligence Organization. Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xuancheng Ren, Houfeng Wang, and Wenjie Li. 2018. Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 979–988, Melbourne, Australia. Association for Computational Linguistics. Peng Xu, Jackie Chi Kit Cheung, and Yanshuai Cao. 2020. On variational learning of controllable representations for text without supervision. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 10534–10543. PMLR. Yanpeng Zhao, Wei Bi, Deng Cai, Xiaojiang Liu, Kewei Tu, and Shuming Shi. 2018. Language style transfer from sentences with arbitrary unknown styles. CoRR, abs/1808.04071. 3797 A Appendix A.1 Beyond Style Transfer In this section, we provide additional examples illustrating the abilities of our TextSETTR model trained on Common Crawl data, beyond typical style transfer. Examples of shortening are given in Table 6, with inputs taken from the first five sentences of the Wikipedia article “Artificial neural network”. As shortening may require minor rephrases, we set our tuning ranges to add:0–5%, delete:40–90%. Since our intention is to leave the style unchanged (apart from length), we extract the target style directly from the input text, with no delta added. The model is largely successful at identifying and removing “superfluous” content, and finding ways of rephrasing to shorten while preserving meaning. Examples of random augmentations are given in Table 7. In each case, we transfer the input sentence “What’ll the weather be tomorrow?” to a slightly different style. Specifically, for each transfer, we extract this sentence’s style vector and apply a small amount of noise, with each component of the noise vector sampled from a Gaussian N(0, 0.08). Note that apart from the noise in the style vector, the transfer process is deterministic, as we use greedy decoding. The cells of Table 7 apply different tuning ranges, conditioning the model to change a little or a lot. Within each cell, we repeatedly sample the noised style, and present the first five unique outputs. The results indicate that many random changes in style are largely meaning preserving, especially when a small change is requested. With larger add/delete rates, the outputs are still closely related in meaning, despite low lexical overlap. A.2 Settings used for Qualitative Analysis For each of the transfer types (e.g., formal ↔informal) in Table 3, we specify the intended target styles through a tiny set of exemplars. These exemplars are provided in Tables 8–12. Additionally, for each transfer type, we select a delta scale λ and add/delete rates. These settings are selected through initial experiments, and are held fixed across all examples of transfer shown. A.3 Human Reference BLEU Li et al. (2018) provide human reference transfers for their Amazon test data, and report BLEU scores of model outputs against these targets. In principle, we believe this metric is less informative than selfBLEU, as style transfer is a relatively open-ended task, and successful transfers may differ significantly from the single human reference. However, for completeness, we report “reference BLEU” of our models and those of prior work in Figure 4. We observe BLEU and self-BLEU are highly correlated, and the “Accuracy vs. BLEU” plot conveys the same relationships we saw in Figure 2. As before, all BLEU scores are calculated using SacreBLEU (Post, 2018). A.4 Amazon Reviews Preprocessing We use the code in Figure 5 to process raw Amazon reviews from the Ni et al. (2019) dataset and extract pairs of adjacent lines, preprocessed to have a similar format to Li et al. (2018) dataset. We split reviews on newlines, and clip lines to 100 characters, always ending with a period. This gives results similar to Li et al. (2018), where one line may contain multiple sentences, and may consists of a “half-sentence” ending with “e.g.” or a similar non-sentence-final period. Additionally, we apply various tokenization and normalization operations to roughly match the observed Li et al. (2018) text. A.5 Human Evaluation Setup For the human evaluations of our models, we employed 3 in-house annotators. The annotators were paid hourly wages that are competitive for their locale and have standard rights as contractors. They spoke native English. For the evaluation task, the annotators were shown both the original and transformed pieces of text. They were then asked to evaluate for three metrics: fluency, meaning preservation, and sentiment change. For fluency, they were asked, “For the new text, how do you rate the fluency, i.e., the quality and readability of the text, with 1 being not fluent at all and 5 being very fluent.” For meaning preservation, they were asked, “Comparing the new text against the original text, and ignoring the change of style, how well does the new text preserve as much of the original meaning, with 1 being all meaning is lost and 5 being preserving as much as possible given the sentiment change?” And for sentiment change, they were asked, “Comparing the new text against the original text, how well did the sentiment of the new text become more positive, with 1 being not more positive and 5 being a lot more positive?” 3798 Artificial neural networks (ANN) or connectionist systems are computing systems that are inspired by, but not identical to, biological neural networks that constitute animal brains. ⇒Artificial neural networks (ANNs) are computing systems that are inspired by the biological neural networks that constitute animal brains. Such systems “learn” to perform tasks by considering examples, generally without being programmed with task-specific rules. ⇒Such systems learn to perform tasks by considering examples, generally without explicit rules. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as “cat” or “no cat” and using the results to identify cats in other images. ⇒For example, image recognition systems might learn to identify images that contain cats by analyzing images that have been manually classified as “cat” or “no cat”. They do this without any prior knowledge of cats, for example, that they have fur, tails, whiskers and cat-like faces. ⇒They do not know that cats have fur, tails, whiskers and cat-like faces. Instead, they automatically generate identifying characteristics from the examples that they process. ⇒Instead, they automatically generate identifying characteristics. Table 6: Examples of shortening (add:0–5%, delete:40-90%), using the first five sentences from the Wikipedia article “Artificial neural network”. For each sentence, the target style is extracted directly from the input text, and no delta is added. Add/Delete: 10–30% Add/Delete: 30–50% What’ll the weather be like? What’s the weather like? What’ll the weather be like tomorrow? What will the weather be like tomorrow? What’s the weather like tomorrow? Will the weather be better tomorrow? What’ll the weather be tomorrow? What’s the weather forecast for tomorrow? What’s the weather supposed to be tomorrow? How will the weather be tomorrow? Add/Delete: 50–70% Add/Delete: 70–90% Will the weather be perfect tomorrow? How do you know what the weather will be like? What’s the weather for tomorrow? Is it supposed to be cold tomorrow? What’s the weather like on the course? What will the weather be like in the South? Hopefully the weather will be better tomorrow. I’m not a fan of the weather. What’s the weather like for the next day? What is the temperature and what is the humidity. Table 7: Random augmentations of input text “What’ll the weather be tomorrow?”, using random style vector deltas with components sampled from N(0, 0.08). Reserved Exemplars Emotive Exemplars 1. That is a very pretty painting. 2. I’m excited to see the show. 3. I’m surprised they rescheduled the meeting. 4. This specimen is an example of the baroque style. 5. After the performance, we ate a meal. 1. OMG, that’s such a beautiful painting! 2. I’m sooo excited to see the show, it’s going to be stellar!! 3. I absolutely can not believe that they rescheduled the meeting! 4. This wonderful specimen is a truly spectacular example of the baroque style. 5. After the superb performance, we ate a delicious meal. Table 8: Emotiveness transfer exemplars. Transfer settings: λ=9, add/delete rates: 0–100%. American Exemplars British Exemplars 1. It cost ten bucks. 2. My neighbor apologized. 3. I’m heading out to the bar with some friends. 1. It cost ten quid. 2. My neighbour apologised. 3. I’m heading out to the pub with some mates. Table 9: Dialect transfer exemplars. Transfer settings: λ=8, add/delete rates: 10–30%. 3799 Polite Exemplars Rude Exemplars 1. No thank you, I’d prefer not to. 2. This game could have been better designed. 3. Do you know why they might have delayed the launch? 4. Sorry, I wasn’t certain if you were joking. 1. Hell no, you can’t make me do that. 2. This game is such a piece of garbage! 3. Why in god’s name would they delay the damn launch? 4. Are you frigging kidding me? Table 10: Politeness transfer exemplars. Transfer settings: λ=5, add/delete rates: 20–50%. Formal Exemplars Informal Exemplars 1. This was a remarkably thought-provoking read. 2. It is certainly amongst my favorites. 3. We humbly request your presence at our gala on the 12th. 1. reading this rly makes u think 2. Its def one of my favs 3. come swing by our bbq next week if ya can make it Table 11: Formality transfer exemplars. Transfer settings: λ=4, add/delete rates: 40–80%. Positive Exemplars Negative Exemplars 1. Five stars, I love it. 1. Zero stars, I hate it. Table 12: Sentiment transfer exemplars. Transfer settings: λ=3, add/delete rates: 0–100%. Model BLEU Self-BLEU CrossAligned 2.0 2.9 Delete&Retrieve 29.7 56.9 B-GST 29.0 54.2 CP-G 17.0 35.5 CP-B 19.4 39.8 TextSETTR (0–20%) 39.0 73.3 TextSETTR (10–30%) 30.7 55.8 TextSETTR (20–40%) 20.0 34.7 TextSETTR (30–50%) 10.6 18.4 TextSETTR (40–60%) 5.5 9.1 TextSETTR (50–70%) 2.2 3.6 0 5 10 15 20 25 30 35 40 Reference BLEU 0 20 40 60 80 100 Sentiment Transfer Accuracy 0 20% 10 30% 20 40% 30 50% 40 60% 50 70% CP-G CP-B CrossAligned Delete&Retrieve B-GST TextSETTR Other label-free models Models trained with labels Figure 4: BLEU scores between model outputs and human references provided by Li et al. (2018), along with self-BLEU for comparison. The first group of models in the table had access to labels at training time, while the second group did not. TextSETTR (X–Y%) refers to our model with add/delete rate ranges set to X–Y%. 3800 import re from html.parser import HTMLParser html_parser = HTMLParser() def preprocess(line): """Simulate Li et al. preprocessing of one review line.""" # Lowercase. line = line.lower() # Replace apostrophes, parens and quotes with spaces. line = re.sub("[’()\"]", " ", line) # Replace dollar values ==> $ line = re.sub("\$[\d.]*", "$", line) # Replace percent values ==> % line = re.sub("[\d.]*%", "%", line) # Replace single digits ==> num_num line = re.sub(" \d[ ,]", " num_num ", line) # Replace multi-digits and codes ==> num_extend line = re.sub(" \d[ˆ ]*", " num_extend", line) # Remove remaining numbers, including decimals. line = re.sub("\d[\d.]*", "", line) # Add spaces around certain punctuation marks. line = re.sub("([.,?!:])", r" \1 ", line) # Remove double spaces after periods before words. return re.sub(r"\. ([a-z])", r". \1", line) def acceptable_line(line): """Check if text looks like an acceptable line from Li et al.""" if not line or len(line) < 30 or len(line) >= 100: return False # Avoid lines with any char absent from Li et al. train. if re.search(’[ˆ !$%+,.:;>?@\ˆ_‘a-z{|}]’, line): return False return True def clip_to_last_period(line): return line[:len(line) - line[::-1].index(’.’)] def adjacent_lines(review): """Extract a list of adjacent line pairs from review text.""" review = html_parser.unescape(review) review = review.replace(’\\"’, ’"’) # Simulate Li et al. splitting and filtering. if ’\n’ not in review: return lines = review.split(’\n’) lines = [preprocess(clip_to_last_period(l[:100])) for l in lines if l and "." in l[:100]] lines = [preprocess(l) for l in lines] lines = [l for l in lines if acceptable_line(l)] if len(lines) < 2: return return list(zip(lines[:-1], lines[1:])) Figure 5: Python code to extract adjacent lines of text from raw Amazon reviews, producing outputs in a similar style to Li et al. (2018).
|
2021
|
293
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3801–3815 August 1–6, 2021. ©2021 Association for Computational Linguistics 3801 H-Transformer-1D: Fast One-Dimensional Hierarchical Attention for Sequences Zhenhai Zhu Google Research [email protected] Radu Soricut Google Research [email protected] Abstract We describe an efficient hierarchical method to compute attention in the Transformer architecture. The proposed attention mechanism exploits a matrix structure similar to the Hierarchical Matrix (H-Matrix) developed by the numerical analysis community, and has linear run time and memory complexity. We perform extensive experiments to show that the inductive bias embodied by our hierarchical attention is effective in capturing the hierarchical structure in the sequences typical for natural language and vision tasks. Our method is superior to alternative sub-quadratic proposals by over +6 points on average on the Long Range Arena benchmark. It also sets a new SOTA test perplexity on One-Billion Word dataset with 5x fewer model parameters than that of the previous-best Transformer-based models. 1 Introduction Linearly combining information using contentbased weights, a method generically known as attention, is a key building block in many deep neural networks such as recurrent neural networks (RNN) (Luong et al., 2015), convolutional neural networks (CNN) (Bello et al., 2019) and graph convolutional networks (GCN) (Velickovic et al., 2018). One particular type of such attention, called multi-head scaled dot-product attention, is one of the main components of the Transformer architecture proposed by Vaswani et al. (2017), which has been shown to push the state-of-theart (SOTA) performance for various understanding and generation tasks. These include standard natural language processing (NLP) tasks such as machine translation, document classification, entailment, summarization and question answering (Zaheer et al., 2020; Dai et al., 2019; Baevski and Auli, 2019), as well as music generation (Huang et al., 2018), image generation (Parmar et al., 2018; Chen et al., 2020) and genomics (Zaheer et al., 2020; Choromanski et al., 2020). The Transformer is also the backbone architecture for models such as BERT (Devlin et al., 2019) (and its numerous relatives) and GPT3 (Brown et al., 2020), which have delivered impressive performance across many NLP tasks. However, the standard attention mechanism of the Transformer has a run time and memory usage that scales quadratically with sequence length. Therefore, this quadratic complexity has become a critical bottleneck in processing long sequences (over 1,000 tokens), and has since motivated many new attention algorithms, see (Tay et al., 2020d) for a survey of such work. In this paper, we draw inspiration from two branches in numerical analysis: Hierarchical Matrix (H-Matrix) (Hackbusch, 1999, 2000) and Multigrid method (Briggs et al., 2000). We propose a hierarchical attention that has linear complexity in run time and memory, and only utilizes dense linear algebra operations optimized for GPUs or TPUs. We hypothesize that the inductive bias embodied by the proposed hierarchical structure for the attention matrix is effective in capturing the hierarchical structure in the sequences typically seen in many natural language processing and computer vision tasks. The main benchmark we use in this paper is the Long Range Arena (LRA) benchmark (Tay et al., 2020c), which has been specifically designed to evaluate and compare various sub-quadratic attention algorithms. Our new hierarchical attention mechanism achieves best average performance to-date on the LRA benchmark by more than 6 points over the previous-best BigBird algorithm (Zaheer et al., 2020), while pushing SOTA performance higher in 4 of the 5 successful tasks. Furthermore, using this new atten3802 tion, a Transformer-based language model trained on the One-Billion Word dataset (Chelba et al., 2014) sets a new SOTA performance record by reducing the test perplexity by 1.55 points comparing to the previous-best Transformer-XL (Dai et al., 2019) with 5x more parameters. Overall, these empirical results both validate the soundness of our approximation method for computing attention weights, as well as the the appropriateness of the inductive bias present in the proposed hierarchical attention. 2 Related Works It is well established in the NLP literature that the embeddings of nearby tokens tend to be more similar than the distant ones (Manning and Sch¨utze, 1999). This leads to the intuition that token similarity and hence the attention should decrease with the sequence distance between a query token and a key token1. This motivates the sliding-window local attention (Parmar et al., 2018; Ramachandran et al., 2019; Qiu et al., 2019) which amounts to truncating off-diagonal entries in the attention matrix beyond a user-specified sequence distance. A second approach is to keep O(1) number of nonzeros per row in the attention matrix. The nonzero entry selection is either content-based (Kitaev et al., 2020; Roy et al., 2020; Tay et al., 2020b; Zhou et al., 2020), hand-crafted (Beltagy et al., 2020; Brown et al., 2020; Child et al., 2019; Ho et al., 2019) or simply random (Zaheer et al., 2020). It is also well known in the NLP literature that long-range contextual information is necessary for many NLP tasks (Khandelwal et al., 2018; Liu and Lapata, 2019). So a set of global tokens are also considered. This adds O(1) number of dense rows and columns to the attention matrix (Zaheer et al., 2020; Ainslie et al., 2020; Beltagy et al., 2020). A third approach is to approximate the attention matrix with a low-rank factored form (Choromanski et al., 2020; Wang et al., 2020; Tay et al., 2020a). The first two approaches are based on the premise that one needs to explicitly zero out entries in the attention matrix in order to reduce the quadratic complexity. Decades of research by the scientific computing and numerical analysis community has resulted in more sophisticated algorithms to sparsify matrices. A 1Eq. (11) and (12) offer a simple illustration of this intuition. small set of samples of these algorithms and their engineering applications include Fast Multipole Method (Greengard and Rokhlin, 1987; Greengard, 1994; Nabors et al., 1994; Shi et al., 1998), Pre-corrected FFT (Phillips and White, 1997; Zhu et al., 2005), Hierarchical Singular Value Decomposition (SVD) (Kapur and Long, 1997) and Hierarchical Matrix (H-Matrix) (Hackbusch, 1999, 2000; Zhu and White, 2005). These are generally called Multilevel Methods (Brandt and Lubrecht, 1990). The hierarchical attention proposed in this paper is inspired by these Multilevel Methods in general and the H-Matrix in particular. The hierarchical matrix structure allows a linear complexity in both constructing and applying the attention matrix. 3 Definition and Notation Given matrices Q, K and V , with rows representing sequences of token embedding or feature vectors for query, key and value respectively, the output weighted by the scaled dot-product attention in the Transformer (Vaswani et al., 2017) is defined as Z = softmax(QKT √ d )V (1) where Z, Q, K, V ∈RL×d, L is the length of the sequences, and d is the embedding or feature size. In a more compact matrix form, Eq. (1) can be written as Z = D−1AV (2) where A = eS (3) Si,j = QiKT j √ d (4) D = diag{A · 1L} (5) 1L = [1, 1, ..., 1]T . (6) Here, A, S ∈RL×L, 1L ∈RL is a vector with all ones, and Si,j represents the unnormalized cosine similarity between query embedding Qi (the i-th row in Q) and key embedding Kj (the j-th row in K). For the sake of clarity, we focus on the singlehead attention in the exposition of the proposed algorithm. Extension to the multi-head case is straightforward since each attention head is computed independently (Vaswani et al., 2017). 3803 Computing the similarity matrix S in Eq. (4) and the attention matrix A in Eq. (3) takes O(L2d) time and O(L2) memory. Similarly, computing AV in Eq. (2) takes O(L2d) time, and computing A · 1L in Eq. (5) takes O(L2) time. The O(L2d) and O(L2) complexities are the bottlenecks for applying the attention mechanism over very long sequences. 4 Introduction on H-Matrix and Multigrid Method 4.1 H-Matrix The singular-value decomposition of the attention matrix A in Eq. (3) is A = UΣV T (7) where Σ = diag{σ1, σ2, ..., σL} and σi is the i-th singular value. The numerical rank of matrix A is r if PL i=r+1 σi < ϵ for a given tolerance ϵ (Trefethen and Bau, 1997). The standard rank-r approximation to matrix A is A ≈ˆU ˆΣ ˆV T = ˆU ˜V T (8) where ˆΣ = diag{σ1, σ2, ..., σr}, ˆU, ˆV ∈RL×r have the first r columns of U and V , and ˜V = ˆV ˆΣ. This is the low-rank approximation used in (Choromanski et al., 2020; Wang et al., 2020; Tay et al., 2020a). This approximation compresses L2 entries in A to 2rL entries in ˆU and ˜V T . So the compression rate is L 2r. The H-Matrix generalizes this low-rank approximation by using matrix block hierarchy. Consider a two-level H-Matrix with 4 × 4 and 2 × 2 block partition at level-0 and level-1, respectively. Matrix A is partitioned as A = A(0) 11 A(0) 12 A(0) 21 A(0) 22 A(1) 12 A(1) 21 A(0) 33 A(0) 34 A(0) 43 A(0) 44 . (9) The low-rank approximation in Eq. (8) is applied to the off-diagonal blocks at each level. For example, A(l) 12 ≈ˆU (l) 12 ˜V (l) 12 T (10) where l = 0, 1. To give a concrete example, suppose each entry in matrix A has the analytical form Ai,j = eSi,j (11) Si,j = 2e−(i−j)2 −1 (12) where i, j = 0, 1, 2, ..., 15 2. With the block hierarchy defined in Eq. (9), the size of the matrix block at level-1 and level-0 is 8 × 8 and 4 × 4, respectively. For tolerance ϵ = 10−3, one can verify that the numerical rank map of matrix A is 4 2 2 4 2 2 4 2 2 4 (13) where the number in each block is the numerical rank of the corresponding block in Eq. (9). Note that matrix A still has full numerical rank of 16 at a looser tolerance 10−1. So the standard lowrank approximation is ineffective in this case. But even this simple two-level H-matrix already offers a compression rate of 4 3 since storing an H-matrix with the rank map in Eq. (13) takes 192 entries 3. In addition, one can verify that no entry Ai,j in Eq. (11) is very small, since Si,j ∈[−1, 1] in Eq. (12). Therefore, truncating off-diagonal entries of matrix A, as proposed in (Parmar et al., 2018), would produce a poor approximation. In practice, the number of levels is adapted to the underlining governing equations that result in matrix A and it can easily be over 10 (Kapur and Long, 1997; Hackbusch, 2000; Zhu and White, 2005). In turn, this can substantially increase the compression rate. In general, the computation complexity of the H-Matrix is either O(L) or O(L log L), depending on the underlining physics (Hackbusch, 1999, 2000). 4.2 Elements of the Multigrid Method Multigrid Method is a multi-level nested iterative method for solving large-scale sparse matrices resulting from discretized partial-differential equations (PDEs) (Briggs et al., 2000; Trottenberg et al., 2000). At its core are two simple but powerfully complementary ideas: relaxation and correction. Our proposed hierarchical attention only uses the correction scheme as a building block since there is no sparse matrix to relax on. The correction scheme has two components: restriction or coarsening, and interpolation or pro2Matrix A in Eq.(11) is a symmetric Toeplitz matrix (Golub and Loan, 1996) and hence only has 16 unique entries. But we ignore this fact and treat A as a general matrix here. 3Each one of four diagonal blocks at level-0 takes 16 entries. Each one of four off-diagonal blocks at level-0 takes 16 entries. Each one of two off-diagonal blocks at level-1 takes 32 entries. 3804 longation. Consider a vector ¯vh of scalar values defined on a set of N grids with uniform interval h. The simplest coarsening is to take the average of the scalar values on each pair of grids, i.e., ¯v2h j = 1 2(¯vh 2j + ¯vh 2j+1) (14) where j = 0, 1, 2, ...N/2 −1. The superscript in Eq. (14) indicates that the grid interval at these two levels is h and 2h, respectively. The simplest interpolation is to duplicate the value on each coarse grid to values on a pair of fine grids, i.e., ¯vh 2j = ¯v2h j , ¯vh 2j+1 = ¯v2h j (15) where j = 0, 1, 2, ...N/2 −1. 5 Intuition for Hierarchical Attention The hierarchical low-rank structure like Eq. (13) turns out to be pervasive in many if not all physics phenomena. Much of the theoretical analysis by (Greengard and Rokhlin, 1987; Hackbusch, 1999) is concerned with quantifying such aspects. The key insight into these Multilevel Methods can be summarized as follows: perform no approximation for near interactions, and apply progressively lower-precision approximation for progressively longer distance interactions. The simple case shown in Eq. (9)-(13) is a good example. To satisfy the tolerance of 10−3, we need full rank (no approximation) for the diagonal blocks (near interactions), higher precision approximation (rank-2 vs full-rank of 4) for the 4 × 4 off-diagonal blocks at level-0 (mid-distance) and lower precision approximation (rank-2 vs full-rank of 8) for the 8×8 off-diagonal blocks at level-1 (long-distance). In this section, we present some intuition to answer two important questions: 1) Does the hierarchical low-rank structure hold for the attention matrix A in Eq. (3)? 2) What is the algorithm to efficiently compute the hierarchical low-rank structure? We only give an informal exposition of the hierarchical attention. The formal mathematical derivation is deferred to the Appendix. 5.1 Hierarchical Structure As Inductive Bias The error analysis in (Greengard and Rokhlin, 1987; Hackbusch, 1999) offers little direct insight since the attention matrix A in Eq. (3) is data dependent by definition and hence its analytical form like Eq. (11) and (12) is generally unknown. So gathering empirical evidences seems the only viable path to answer the first question listed above. The ablation studies by (Khandelwal et al., 2018) examine the effect of context words on a language model. Within the context range of about 200 tokens, word order is only relevant within the 20 most recent tokens or about a sentence. In the long-range context, order has almost no effect on performance, suggesting that the model maintains a high-level, rough semantic representation of faraway words. The observation is succinctly summarized by the title of the paper ”sharp nearby, fuzzy far away”. Remarkably, this is in spirit very close to the key insight into the Multilevel Methods. A few recent attention-related studies have explored this direction with some success, such as word-level and sentence-level attentions in (Miculicich et al., 2018; Abreu et al., 2019), and sentence-level and paragraph-level attentions in (Liu and Lapata, 2019). Even though the proposed hierarchical attention in these studies only has two levels, as opposed to ten or more levels typically used by the Multilevel Methods, the reported positive results are quite suggestive. We therefore hypothesize that the same hierarchical low-rank structure as shown in Eq (13) might also hold for the attention matrix in many NLP tasks. And we treat it as the inductive bias in the hierarchical attention mechanism proposed in this paper. As pointed out in (Goyal and Bengio, 2020), inductive biases encourage the learning algorithm to prioritise solutions with certain properties. Hence good benchmark performance delivered by a Transformer-based model with proposed hierarchical attention can be regarded as a positive evidence to support the hierarchical low-rank structure hypothesis. 5.2 Informal Exposition of Hierarchical Attention In the standard definition of attention in Eq. (3) and (4), there is no preference given to any keys based on the sequence distance between a query and keys. The observation in (Khandelwal et al., 2018) clearly suggests that a distance-dependent attention mechanism should be a better alternative. We will take three steps to informally explain the hierarchical attention mechanism. First, the attention matrix blocks for nearby, mid-distance and long-distance attention are separated in sec3805 tion 5.2.1. This is the first step toward the distance-dependent attention mentioned above. Second, a token hierarchy is established in section 5.2.2. Third, the hierarchical attention is constructed in section 5.2.3 5.2.1 Attention Partition Consider a 16-word sentence in Fig. 1. The sentence is partitioned at three segment granularity. This induces a three-level partition of the attention matrix A for the original sequence: A = A(2) + A(1) + A(0) (16) where A(2) = " 0 A(2) 12 A(2) 21 0 # (17) A(1) = A(1) 12 A(1) 21 A(1) 23 A(1) 32 A(1) 34 A(1) 43 (18) A(0) = A(0) 11 A(0) 12 A(0) 21 A(0) 22 A(0) 23 ... ... ... A(0) 87 A(0) 88 . (19) Note that the nonzero entries in A(0), A(1) and A(2) are the same as the corresponding entries of matrix A in Eq. (3). Matrix block size of A(0) ij , A(1) ij and A(2) ij is 2×2, 4×4 and 8×8, respectively. Following the key insight into Multilevel Methods, we perform no approximation to any level-0 matrix block A(0) ij and apply a low-rank approximation to off-diagonal matrix blocks in A(1) and A(2). If we set the numerical rank of all these blocks to 2, then we can assemble the three rank maps into a single rank map as 4 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 . (20) 4We omit some of implementation details to handle the overlapping entries between adjacent levels. this sentence is to illustrate how to set up token hierarchy level by level with aggregation a) Level-0: 16 tokens partitioned into 8 segments b) Level-1: 16 tokens partitioned in 4 segments c) Level-2: 16 tokens partitioned in 2 segments this sentence is to to set up token illustrate how hierarchy level by level with aggregation this sentence is to illustrate how to set up token hierarchy level by level with aggregation Figure 1: Token sequence partitions in three segment granularity. The hierarchical structure embodied by the predetermined rank map in Eq. (20) represents the inductive bias for the attention matrix A in Eq. (16). But this construction step is inefficient because we need to form the original attention matrix and then perform SVD to discover the low-rank approximation. 5.2.2 Token Hierarchy To illustrate the notion of token hierarchy, consider the same 16-word sentence in Fig. 2. A simple 3-level binary-tree hierarchy can be set up by following the simple coarsening defined in Eq. (14): 1) At level-0, each one of the 16 words is mapped to its word embedding; 2) At level-1, each token (parent node) corresponds to a pair of adjacent words at level-0 (child nodes), which are shown inside each box. The embedding of each parent token is simply the average of its child token embeddings; 3) At level-2, each token (parent node) corresponds to one pair of adjacent tokens at level-1 (child nodes) or 4 adjacent words at level-0 (grand child nodes), which are shown inside each box. The embedding of each parent token is simply the average of its child token embeddings. In general, the height of the binary tree is O(log2(L) and the total number of tree nodes is O(2L), where L is the sequence length. We only need word embeddings for the leaf nodes since the embeddings of all other tree nodes can be recursively computed. The formal definition and notations of the recursion for query and key are detailed in section 6.1. 5.2.3 Informal Construction of Hierarchical Attention It is clear from Fig. 2 that the embeddings of higher level tokens represent a coarser level representation of a larger chunk of the text. The tokens at different levels can be understood as multi-scale snapshots of the original token sequence at level-0. 3806 this sentence is to illustrate how to set up token hierarchy level by level with aggregation this sentence is to illustrate how to set up token hierarchy level by level with aggregation this sentence is to illustrate how ho set up token hierarchy level by level with aggregation a) Level-0: 16 tokens partitioned into 8 segments b) Level-1: 8 tokens partitioned into 4 segments c) Level-2: 4 tokens partitioned into 2 segments Figure 2: A three-level token hierarchy. Dashed boxes represent segmentation and solid boxes represents tokens. Hence this token hierarchy naturally induces a set of multi-scale attention matrices. Let ˜A(i) be the attention matrix induced by the tokens at level-i. It is clear from Fig. 2 that the size of ˜A(0), ˜A(1) and ˜A(2) is 16×16, 8×8 and 4×4, respectively. This multi-scale viewpoint does not directly lead to a useful algorithm since matrix ˜A(0) contains all the information and there is little additional information from ˜A(1) and ˜A(2). A key step to arrive at the hierarchical attention is to apply the contextual sliding window at each hierarchy level. The tokens at each level are partitioned into segments of size 2 in Fig. 2. One way to implement the local attention is to allow each query token segment to attend only two adjacent key token segments, one to its left and another to its right. At level-0, each query token segment also attends to the collocated key token segment. The token segment partition and local attention lead to a tri-diagonal block sparse matrix structure for ˜A(0) and bi-diagonal block sparse matrix structure for ˜A(1) and ˜A(2). Their sparsity patterns are ˜A(0) ∝ 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 (21) ˜A(1) ∝ 2 2 2 2 2 2 (22) ˜A(2) ∝ 2 2 (23) where the 2 in the nonzero blocks indicates that these are dense blocks of size 2 × 2. It is clear that ˜A(0) is identical to A(0) in Eq. (19). The efficiency gain comes from ˜A(2) and ˜A(1). Each nonzero entry in ˜A(2) and ˜A(1) captures the aggregated or coarse attention between two disjoint chunk of four and two tokens, respectively. Progressively larger token chunks lead to progressively lower-precision approximation to the original attention blocks. This is precisely the intention of the rank map in Eq. (20). We can now see that ˜A(2) and ˜A(1) provide an efficient way to approximate A(2) in Eq. (17) and A(1) in Eq. (18), respectively. 6 Key Components in Hierarchical Attention 6.1 Constructing Hierarchical Attention The simple example in Fig. 2 can be easily generalized. Eq. (14) is used to coarsen or merge rows in matrices Q, K and V in Eq. (1). For sequence length L = 2M+1, the coarsening establishes a binary tree of depth M for Q, K and V , respectively. Each tree node represents a matrix row and there are 2M+1−l nodes or rows at level-l. To facilitate the discussion, we define a few hierarchy related notations here. Let ˜Q(l), ˜K(l) and ˜V (l) be coarsened versions of Q, K and V at level-l in the binary tree. We note that l = 0 is a special case, which is defined as ˜Q(0) = Q, ˜K(0) = K, ˜V (0) = V. (24) Following Eq. (14), the recursion to coarsen Q, K and V is: ˜Q(l+1) j = 1 2( ˜Q(l) 2j + ˜Q(l) 2j+1) (25) ˜K(l+1) j = 1 2( ˜K(l) 2j + ˜K(l) 2j+1) (26) ˜V (l+1) j = ( ˜V (l) 2j + ˜V (l) 2j+1) (27) where l = 0, 1, ..., M −2 and j = 0, 1, 2, ..., 2M−l. It should be noted that the coarsening of V in Eq. (27) does not have the averaging factor 1 2. We defer more details on coarsening to Appendix Section A.1. Now we are ready to compute the nonzero entries in Eq. (21), (22) and (23) and construct hierarchical attention matrix ˜A(l). Substituting Eq. (25) and (26) into (4) and then into (3), we obtain ˜A(l) ij = e ˜S(l) ij = e ˜ Q(l) i ( ˜ K(l) j )T √ d (28) 3807 Again, we note that l = 0 is a special case because ˜A(0) ij = Aij. 6.2 Applying Hierarchical Attention The hierarchical matrix structure in Eq. (17), (18) and (19) naturally leads to a hierarchical approach to the matrix-matrix multiplication in Eq. (2) and the matrix-vector multiplication in Eq. (5). We use the matrix-matrix multiplication as an example since matrix-vector multiplication is just a special case of the matrix-matrix multiplication. In view of Eq. (17), (18) and (19), we write the matrix-matrix multiplication in Eq. (2) as Y = AV ≈A(0)V (0) + ˜A(1) ˜V (1) + ˜A(2) ˜V (2) = Y (0) + P (0) ˜Y (1) + P (1) ˜Y (2) (29) where ˜Y (l) = ˜A(l) ˜V (l), l = 1, 2 (30) We defer the detailed derivation of Eq. (29) to Appendix Section A.5 and A.6. 7 Algorithm And Computational Complexity To facilitate the description and the complexity analysis of the algorithm, we define a few more hierarchy-related notations. In addition to sequence length L, number of hierarchy levels M and embedding or feature size d in Eq. (1), the new notations include: 1) Nr : numerical rank of the off-diagonal blocks (for instance, 2 in Eq. (20)). This is also the diagonal block size at level-0; 2) N(l) b : number of blocks at level-l. Note that L and d are usually data-dependent hyper-parameters, while Nr is the only model hyper-parameter responsible for our method’s inductive bias. In turn, N(l) b and M are derived parameters, computed as: N(0) b = L Nr , N(l+1) b = N(l) b 2 (31) M = log2(N(0) b ). (32) It is easy to verify that M−1 X l=0 N(l) b = M−1 X l=0 N(0) b 2l ≈2N(0) b . (33) It is important to note that only the diagonal blocks at level-0 and the super-diagonal and subdiagonal blocks at level-l are needed in applying the hierarchical attention matrix. This is clearly shown in Eq. (21)- (23). This means that only N(l) b −1 super-diagonal and sub-diagonal blocks are computed at level-l. This is crucial to the overall linear complexity in run time and memory. We should also note that all matrix blocks in coarse attention matrix ˜A(l) have the same size Nr × Nr. This is due to the rank map in Eq. (20). This is crucial for efficiency reason since the single-instruction-multiple-data (SIMD) programming style supported by the dense linear algebra libraries for GPU and TPU encourages uniform tensor shapes. We summarize the main steps to construct and apply the hierarchical attention in Algorithm 1. Algorithm 1 H-Transformer-1D Input: Q(query), K(key), V (value) Output: Z Coarsen Q using Eq. (25) and coarsen K using Eq. (26) Compute diagonal blocks in ˜A(0) and superdiagonal and sub-diagonal blocks in ˜A(l) using Eq. (28) Coarsen V using Eq. (27) Compute Y = AV in Eq. (2) using Eq. (29) Compute D in Eq. (5) using Eq. (29) Compute Z = D−1Y The computational cost for Algorithm 1 has two parts: 1. Computing the hierarchical attention matrix: (a) diagonal blocks at level-0: dN2 r N(0) b (b) Super- and sub-diagonal blocks at levell: 4dN2 r (N(l) b −1) (c) total: 5dLNr = O(dL) 2. Computing matrix-matrix (MM) multiplication in Eq. (2) and matrix-vector (MV) multiplication in Eq. (5): (a) MM: 5dLNr (b) MV: 5LNr (c) total: 5(d + 1)LNr = O(dL) So the overall run time complexity of the hierarchical attention algorithm is O(dL). Likewise, the memory complexity can be shown to be O(dL) as well. We defer the detailed analysis to appendix Section A.5 and A.6. 3808 Model ListOps Text Retrieval Image Pathfinder Path-X Avg Chance 10.00 50.00 50.00 10.00 50.00 50.00 44.00 Transformer 36.37 64.27 57.46 42.44 71.40 FAIL 54.39 Local Attention 15.82 52.98 53.39 41.46 66.63 FAIL 46.06 Sparse Trans. 17.07 63.58 59.59 44.24 71.71 FAIL 51.24 Longformer 35.63 62.85 56.89 42.22 69.71 FAIL 53.46 Linformer 35.70 53.94 52.27 38.56 76.34 FAIL 51.36 Reformer 37.27 56.10 53.40 38.07 68.50 FAIL 50.67 Sinkhorn Trans. 33.67 61.20 53.83 41.23 67.45 FAIL 51.39 Synthesizer 36.99 61.68 54.67 41.61 69.45 FAIL 52.88 BigBird 36.05 64.02 59.29 40.83 74.87 FAIL 55.01 Linear Trans. 16.13 65.90 53.09 42.34 75.30 FAIL 50.55 Performer 18.01 65.40 53.82 42.77 77.05 FAIL 51.41 H-Transformer-1D 49.53 78.69 63.99 46.05 68.78 FAIL 61.41 Table 1: Experimental results on long-range arena benchmark. Best model is in boldface and second best is underlined. All models do not learn anything on Path-X task, contrary to the Pathfinder task and this is denoted by FAIL. Path-X is not counted toward the Average score as it has no impact on relative performance. 8 Experiments And Results We have implemented the proposed hierarchical attention using Jax, an open source library 5 for automatic gradient computation and linear algebra operations on GPUs and TPUs. All numerical operations in our algorithm use the Numpy native linear algebra functions supported by Jax. In all our experiments in this section, we use the standard Transformer architecture described in (Vaswani et al., 2017) as the backbone for our HTransformer-1D model. Unless specified otherwise, the model parameters are: number of layers is 6, number of heads is 8, word embedding size is 512 and the feed-forward module (FFN) size is 2048. We follow the API for the standard multihead scaled dot-product attention implementation 6 so that we can perform a simple drop-in replacement of the standard multihead attention with our hierarchical attention implementation. This allows for an easy and fair comparison. 8.1 Long-Range Arena The open-source Long-Range Arena (LRA) benchmark 7 has been proposed as a standard way to probe and quantify the capabilities of various xformer (long-range Transformer) architectures (Tay et al., 2020c). In our case, it also serves to highlight the effectiveness of the inductive bias 5https://github.com/google/jax 6https://github.com/google/flax/blob/master/flax/nn 7https://github.com/google-research/long-range-arena inspired by the H-Matrix method, as well as the capability of our hierarchical attention to handle long sequences. The LRA has several desirable qualities that made us focus on it as a primary evaluation benchmark: generality (restricted to encoder-only tasks to accommodate most proposals); simplicity (no pretraining, no data augmentation allowed); difficulty (large headroom with existing approaches); long-input focus (so that modeling improvements in this area are visible); diverse (6 tasks, covering math, language, image, and spatial modeling); and lightweight (so that modeling improvements are measurable independently of the ability to train and run high-capacity models). The tasks that comprise LRA are: ListOps (sequences of arithmetical expressions of lengths of up to 2K that tests the ability to reason hierarchically while handling long context); Text (byte/character-level text classification at document level, which both simulates longer input sequences – max length 4K – and increases the difficulty level); Retrieval (byte/character-level document retrieval, which simulates the ability to model document similarity as a score between two independently-encoded long input sequences – max length 4K + 4K = 8K); Image (image classification based on the CIFAR-10 dataset, where an NxN image is flattened to a sequence of length N2 pixels); Pathfinder (long-range spatial dependency task, with images consisting of two small 3809 Model perplexity parameters (Dai et al., 2019) 21.8 800M (Baevski and Auli, 2019) 23.02 1000M (Dai et al., 2019) 23.5 465M (Baevski and Auli, 2019) 23.91 465M (Shazeer et al., 2018) 24.0 4900M Transformer baseline 30.04 53M Transformer baseline 24.8 144M H-Transformer-1D Nr = 16 23.95 53M H-Transformer-1D Nr = 16 20.25 144M Table 2: Experimental results on one-billion word benchmark. We compare previous SOTA results obtained with models of size 465M-4900M parameters against the performance of the quadratic attention baseline and the HTransformer-1D models. circles and dash-line paths that either connect the two circles or not – image dimensions of 32x32 for a pixel sequence of length 1,024); Path-X (same as Pathfinder, but for image dimensions of 128x128 for a total pixel sequence of length 16,384). The default Transformer model parameters such as number of layers and number of heads etc are pre-determined by the benchmark configuration for each task. The results obtained by our H-Transformer-1D model on the LRA benchmark are given in Table 1. Overall, the H-Transformer-1D model achieves 61.41 average accuracy, a +6.4 points improvement over the previous-best average performance from BigBird (Zaheer et al., 2020). We want to highlight ListOps, Text and Retrieval because they all involve long sequences and H-Transformer-1D model improves SOTA performance by relatively large margins. These should be strong evidences to support our hypothesis in section 5.1 and validate the inductive bias due to the hierarchical attention. 8.2 Language Models Trained on One-Billion Words We have used Flax, an open-source library 8 to train neural networks, as the code base for the model training. Our H-Transformer-1D model uses the standard Transformer decoder implementation in Flax as the backbone. Only the attention is replaced with our hierarchical attention. We trained both the Transformer baseline and HTransformer-1D on the One-Billion Word benchmark (Chelba et al., 2014). We tried different Nr 8https://github.com/google/flax (numerical rank) in our H-Transformer-1D model. These represent different inductive bias. We found that H-Transformer-1D with Nr = 16 generated text with quality comparable to that of the baseline Transformer. For both Transformer baseline and H-Transformer-1D, we also tried two sets of model parameters: 1) embedding size is 512 and feed-forward module size is 2048 and hence the parameter count is 53M; 2) embedding size is 1024 and feed-forward module size is 4096 and hence the parameter count is 144M. The test perplexity results of these four models and various SOTA models are shown in table 2. H-Transformer-1D delivers the lowest perplexity to-date while using 5× smaller model capacity than that of the previous SOTA model Transformer-XL (Dai et al., 2019). This is another strong evidence to support our hypothesis in section 5.1 and validate the inductive bias due to the hierarchical attention. 9 Conclusions and Future Work We have proposed a new Transformer attention using the inductive bias inspired by the HMatrix. The new algorithm has linear complexity in run time and memory usage and is fully compatible with dense linear algebra libraries on GPU and TPU. The effectiveness of this new attention is demonstrated by the empirical evidences from long-range arena benchmark and One-Billion word language modeling. Future work include applying the new attention to music and genomics, developing proper inductive bias for cross-attention and extending the onedimensional hierarchical attention to 2D images. 3810 References Jader Abreu, Luis Fred, David Macˆedo, and C. Zanchettin. 2019. Hierarchical attentional hybrid neural networks for document classification. ArXiv, abs/1901.06610. Joshua Ainslie, S. Onta˜n´on, C. Alberti, V. Cvicek, Zachary Kenneth Fisher, Philip Pham, Anirudh Ravula, S. Sanghai, Qifan Wang, and L. Yang. 2020. Etc: Encoding long and structured inputs in transformers. In EMNLP. Alexei Baevski and M. Auli. 2019. Adaptive input representations for neural language modeling. ArXiv, abs/1809.10853. I. Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V. Le. 2019. Attention augmented convolutional networks. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3285–3294. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. ArXiv, abs/2004.05150. A. Brandt and A. A. Lubrecht. 1990. Multilevel matrix multiplication and fast solution of integral equations. 90:348–370. W.L. Briggs, V.E. Henson, and S.F. McCormick. 2000. A Multigrid Tutorial. SIAM. Tom B. Brown, Benjamin Pickman Mann, Nick Ryder, Melanie Subbiah, Jean Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, G. Kr¨uger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric J Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. ArXiv, abs/2005.14165. Ciprian Chelba, Tomas Mikolov, M. Schuster, Qi Ge, T. Brants, Phillipp Koehn, and T. Robinson. 2014. One billion word benchmark for measuring progress in statistical language modeling. ArXiv, abs/1312.3005. Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. 2020. Generative pretraining from pixels. Proceedings of the 37th International Conference on Machine Learning, PMLR 119. R. Child, Scott Gray, A. Radford, and Ilya Sutskever. 2019. Generating long sequences with sparse transformers. ArXiv, abs/1904.10509. Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Jared Davis, Tam´as Sarl´os, David Belanger, Lucy J. Colwell, and Adrian Weller. 2020. Masked language modeling for proteins via linearly scalable long-context transformers. ArXiv, abs/2006.03555. Zihang Dai, Z. Yang, Yiming Yang, J. Carbonell, Quoc V. Le, and R. Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. In ACL. J. Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT. G.H. Golub and C.F. Van Loan. 1996. Matrix Computation. The John Hopkins University Press, Baltimore. Anirudh Goyal and Yoshua Bengio. 2020. Inductive biases for deep learning of higher-level cognition. ArXiv, abs/2011.15091. L Greengard. 1994. Fast algorithms for classical physics. Science, 265:909–914. L Greengard and V Rokhlin. 1987. A fast algorithm for particle simulations. 73:325–348. W. Hackbusch. 1999. A sparse matrix arithmetic based on h-matrices. part I: Introduction to H-matrices. Computing, 62:89–108. W. Hackbusch. 2000. A sparse matrix arithmetic based on H-matrices. part II: Application to multidimensional problems. Computing, 64:21–47. Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans. 2019. Axial attention in multidimensional transformers. ArXiv, abs/1912.12180. Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M. Dai, Matthew D. Hoffman, Monica Dinculescu, and Douglas Eck. 2018. Music transformer. arXiv: Learning. S. Kapur and D.E. Long. 1997. IES3: A fast integral equation solver for efficient 3-dimensional extraction. International Conference on Computer AidedDesign, pages 448–455. Urvashi Khandelwal, He He, Peng Qi, and Dan Jurafsky. 2018. Sharp nearby, fuzzy far away: How neural language models use context. ArXiv, abs/1805.04623. Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. ArXiv, abs/2001.04451. Yang Liu and Mirella Lapata. 2019. Hierarchical transformers for multi-document summarization. In ACL. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. ArXiv, abs/1508.04025. 3811 Chris Manning and Hinrich Sch¨utze. 1999. Foundations of Statistical Natural Language Processing. MIT Press, Cambridge, MA. Lesly Miculicich, Dhananjay Ram, Nikolaos Pappas, and James Henderson. 2018. Document-level neural machine translation with hierarchical attention networks. In EMNLP. K. Nabors, T. Korsmeyer, and J. White. 1994. Multipole accelerated preconditioned iterative methods for three-dimensional potential integral equations of the first kind. SIAM J. Sci. and Stat. Comp. Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. 2018. Image transformer. ArXiv, abs/1802.05751. Joel R. Phillips and J. K. White. 1997. A precorrectedFFT method for electrostatic analysis of complicated 3D structures. IEEE Transactions on ComputerAided Design of Integrated Circuits and Systems, pages 1059–1072. Jiezhong Qiu, Hao Ma, Omer Levy, Scott Yih, Sinong Wang, and Jie Tang. 2019. Blockwise selfattention for long document understanding. ArXiv, abs/1911.02972. Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jonathon Shlens. 2019. Stand-alone self-attention in vision models. ArXiv, abs/1906.05909. Aurko Roy, M. Saffar, Ashish Vaswani, and David Grangier. 2020. Efficient content-based sparse attention with routing transformers. ArXiv, abs/2003.05997. Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, P. Hawkins, H. Lee, Mingsheng Hong, C. Young, Ryan Sepassi, and Blake A. Hechtman. 2018. Meshtensorflow: Deep learning for supercomputers. In NeurIPS. W. Shi, J. Liu, N. Kakani, and T. Yu. 1998. A fast hierarchical algorithm for 3-d capacitance extraction. ACM/IEEE Design Automation Conference. Yi Tay, Dara Bahri, Donald Metzler, D. Juan, Zhe Zhao, and Che Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. ArXiv, abs/2005.00743. Yi Tay, Dara Bahri, L. Yang, Donald Metzler, and D. Juan. 2020b. Sparse sinkhorn attention. In ICML. Yi Tay, M. Dehghani, Samira Abnar, Y. Shen, Dara Bahri, Philip Pham, J. Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2020c. Long range arena: A benchmark for efficient transformers. ArXiv, abs/2011.04006. Yi Tay, M. Dehghani, Dara Bahri, and Donald Metzler. 2020d. Efficient transformers: A survey. ArXiv, abs/2009.06732. L.N. Trefethen and D. Bau. 1997. Numerical linear algebra. SIAM, Philadelphia. Ulrich Trottenberg, Cornelius W. Oosterlee, and Anton Schuller. 2000. Multigrid. Academic Press. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762. Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li`o, and Yoshua Bengio. 2018. Graph attention networks. ArXiv, abs/1710.10903. Sinong Wang, Belinda Z. Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. ArXiv, abs/2006.04768. Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Onta˜n´on, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2020. Big bird: Transformers for longer sequences. Hao-Yi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2020. Informer: Beyond efficient transformer for long sequence time-series forecasting. ArXiv, abs/2012.07436. Zhenhai Zhu, Ben Song, and J. K. White. 2005. Algorithms in FastImp: A fast and wideband impedance extraction program for complicated 3D geometries. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. Zhenhai Zhu and J. K. White. 2005. Fastsies: a fast stochastic integral equation solver for modeling the rough surface effect. International Conference on Computer Aided-Design, pages 675–682. 3812 A Appendix A.1 Restriction or Coarsening Matrices For sequence length L = 2M, the coarsening establishes a binary tree of depth M for Q, K and V , respectively. The root of the binary tree at level(M −1) has two nodes which correspond to the two matrix rows coarsened from four matrix rows at level-(M −2). The piecewise constant restriction matrix at level-(M −2) is R(M−2) = 1 1 0 0 0 0 1 1 2×4 . (34) Likewise, the piecewise constant restriction matrix at level-(M −3) is R(M−3) = 1 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 1 1 4×8 = R(M−2) 0 0 R(M−2) . (35) In general, the restriction matrices follow the recursion R(l−1) = R(l) 0 0 R(l) (36) which starts from R(M−2) of size 2 × 4 and goes backward to R(0) of size L 2 × L. A.2 Interpolation Matrices Given Y (l) at level-l, the interpolated Y (l−1) at level-(l −1) can be written as Y (l−1) = P (l)Y (l) (37) where l = 1, 2, ..., M −1, sparse matrix P (l) has size L(l−1) × L(l), and L(l) = 2M−l is the node count at level-l of the binary tree. This recursion also follows the binary tree hierarchy. The four matrix rows at level-(M −2) are interpolated from the two matrix rows at level(M −1). Specifically, the piecewise constant interpolation matrix at level-(M −1) is P (M−1) = 1 0 1 0 0 1 0 1 4×2 . (38) Likewise, the piecewise constant interpolation matrix at level-(M −2) is P (M−2) = 1 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 0 1 0 0 0 0 1 0 0 0 1 8×4 = P (M−1) 0 0 P (M−1) . (39) In general, the interpolation matrices follow the recursion P (l−1) = P (l) 0 0 P (l) (40) which starts from P (M−1) of size 4 × 2 and goes backward to P (0) of size L× L 2 . In view of Eq. (34) and (38), it is obvious that P (M−1) = (R(M−2))T . (41) In view of the recursions in Eq. (36) and (40), it is easy to prove by induction that P (l) = (R(l−1))T . (42) A.3 Expansion Matrices For the purpose of factored low-rank approximation for the off-diagonal attention matrix blocks, we design a series of so-called expansion matrices. The first two expansion matrices in this series are T (M−1) = P (M−1) = 1 0 1 0 0 1 0 1 4×2 = 12 0 0 12 (43) and T (M−2) = P (M−2)P (M−1) = 1 0 1 0 1 0 1 0 0 1 0 1 0 1 0 1 8×2 = 14 0 0 14 (44) 3813 where 1N is a length-N vector of ones. The general form of matrix T (l) is defined as T (l) = ΠM−1 i=l P (i) (45) where l = 1, 2, ..., M −1. In view of Eq. (43), (45) and (40), it is easy to prove by induction that T (l) = 12M−l 0 0 12M−l (46) and it has size 2M−l+1 × 2. Further more, in view of Eq. (45) and (42), we have (T (l))T = Πl i=M−1R(i−1). (47) A.4 Low-Rank Factored Form Matrix T (l) plays a pivotal role in constructing the low-rank approximation to the off-diagonal attention matrix blocks. Let the ij-th block in the coarsened attention matrix at level-1 be ˜A(1) ij = a11 a12 a21 a22 (48) where aij is the entry resulted from the inner product between a row in ˜Q(1) and ˜K(1). The rank-2 approximation to the corresponding ij-th block in the original attention matrix A at level-1 can be written as A(1) ij ≈T (M−1) ˜A(1) ij (T (M−1))T (49) = 1 0 1 0 0 1 0 1 a11 a12 a21 a22 1 1 0 0 0 0 1 1 = a11 a11 a12 a12 a11 a11 a12 a12 a21 a21 a22 a22 a21 a21 a22 a22 . (50) It is clear that the resulting 4 × 4 matrix A(1) ij is essentially the piecewise constant interpolation of the 2 × 2 matrix ˜A(1) ij along row and column direction. And since both T (M−1) and ˜A(1) ij have full rank 2, A(1) ij necessarily has rank 2. One can also view aij as being similar to the average value at the ij-th cluster center in the K-mean method. The role of matrix T (M−1) is to expand from these 2×2 clusters to the 4×4 grid and hence the name expansion matrix. Since we maintain the same numerical rank 2 for all super- and sub-diagonal attention matrix blocks, the rank-2 approximation to the ij-th block in the original attention matrix A at level-l is A(l) ij ≈ T (M−l) ˜A(l) ij (T (M−l))T = ΠM−1 i=M−lP (i) ˜A(l) ij ΠM−l i=M−1R(i−1)(51) where the last equality is due to Eq. (45) and (47). We note that matrix T (l) has full column rank 2 by design and this can be easily shown from Eq. (46). We have used this fact to construct the rank-2 approximation in Eq. (51). A.5 Construct Hierarchical Attention Matrix To see how Eq. (51) can be used, consider a simple three-level partition of the attention matrix A for sequence length L = 16 A = " A(2) 11 A(2) 12 A(2) 21 A(2) 22 # (52) A(2) 11 = A(0) 11 A(0) 12 A(0) 21 A(0) 22 A(1) 12 A(1) 21 A(0) 33 A(0) 34 A(0) 43 A(0) 44 (53) A(2) 22 = A(0) 55 A(0) 56 A(0) 65 A(0) 66 A(1) 34 A(1) 43 A(0) 77 A(0) 78 A(0) 87 A(0) 88 (54) where the size of level-0, level-1 and level-2 matrix blocks is 2 × 2, 4 × 4 and 8 × 8, respectively. Note that the number of levels is M = log2(L/2) = 3. We use this simple three-level example to illustrate the key steps in both constructing and applying the hierarchical attention matrix. In view of Eq. (51), we have A ≈ " ˜A(2) 11 T (1) ˜A(2) 12 (T (1))T T (1) ˜A(2) 21 (T (1))T ˜A(2) 22 # (55) ˜A(2) 11 = A(0) 11 A(0) 12 A(0) 21 A(0) 22 T (2) ˜A(1) 12 (T (2))T T (2) ˜A(1) 21 (T (2))T A(0) 33 A(0) 34 A(0) 43 A(0) 44 (56) 3814 ˜A(2) 22 = A(0) 55 A(0) 56 A(0) 65 A(0) 66 T (2) ˜A(1) 34 (T (2))T T (2) ˜A(1) 43 (T (2))T A(0) 77 A(0) 78 A(0) 87 A(0) 88 . (57) We note that matrices T (l), l = 1, 2 are never explicitly formed and are only implicitly used, as shown in next section. So only the diagonal blocks at level-0 and super- and sub-diagonal blocks of the coarsened matrix ˜A at level-l need to be explicitly computed. By design, all these blocks have the same size 2 × 2 if we set the numerical rank to Nr = 2. The total number of superand sub-diagonal blocks in the binary tree hierarchy is upper bounded by twice the number of super- and sub-diagonal blocks at level-0, which is 2N(0) b . Hence the total number of entries is 5N(0) b N2 r = 5LNr = O(LNr). Each entry is equal to the inner product between ˜Q(l) i and ˜K(l) j and hence the run time cost per entry is O(d), where d is the embedding size. So the final total run time cost is O(Ld) and memory foot print is O(L). Here we leave out Nr since it is a constant model hyper parameter. A.6 Apply Hierarchical Attention Matrix Computing matrix-matrix product AV follows the hierarchical structure of matrix A in Eq. (55), (56) and (57). We first partition matrix V according to the three-level binary tree established by the coarsening process, i.e., V = V (0) 1 V (0) 2... V (0) 7 V (0) 8 = V (1) 1 V (1) 2 V (1) 3 V (1) 4 = " V (2) 1 V (2) 2 # . (58) Note that these are partitions of the same matrix V at 3 different levels. For sequence length L = 16, matrix V has size 16 × d, and the size of the partitioned blocks V (0) i , V (1) j and V (2) k are 2 × d, 4 × d and 8 × d, respectively. In the derivation to come, we may exchange partitions at different levels. For instance, in view of Eq. (58), we have V (2) 1 = " V (1) 1 V (1) 2 # . (59) So we may replace V (2) 1 with the right-hand side in Eq. (59). In view of Eq. (52) and (58), matrix-matrix product AV can be written as Y = AV = " A(2) 11 V (2) 1 A(2) 22 V (2) 2 # + " A(2) 12 V (2) 2 A(2) 21 V (2) 1 # = " A(2) 11 V (2) 1 A(2) 22 V (2) 2 # + Y (2). (60) In view of Eq. (55), we have Y (2) = " A(2) 12 V (2) 2 A(2) 21 V (2) 1 # ≈ " T (1) ˜A(2) 12 (T (1))T V (2) 2 T (1) ˜A(2) 21 (T (1))T V (2) 1 # = " P (1)P (2) ˜A(2) 12 R(1)R(0)V (2) 2 P (1)P (2) ˜A(2) 21 R(1)R(0)V (2) 1 # = P (0)P (1) " ˜A(2) 12 ˜V (2) 2 ˜A(2) 21 ˜V (2) 1 # = P (0)P (1) " ˜Y (2) 1 ˜Y (2) 2 # (61) where " ˜V (2) 1 ˜V (2) 2 # = " R(1)R(0)V (2) 1 R(1)R(0)V (2) 2 # . (62) The third equality in Eq. (61) is due to Eq. (45) and (47) where l = 1. The fourth equality in Eq. (61) is due to Eq. (40). In view of Eq. (56), we have A(2) 11 V (2) 1 ≈˜A(2) 11 V (2) 1 = A(0) 11 A(0) 12 A(0) 21 A(0) 22 T (2) ˜A(1) 12 (T (2))T T (2) ˜A(1) 21 (T (2))T A(0) 33 A(0) 34 A(0) 43 A(0) 44 V (2) 1 = Y (0) 1 Y (0) 2 Y (0) 3 Y (0) 4 + Y (1) 1 (63) 3815 where Y (1) 1 = " T (2) ˜A(1) 12 (T (2))T V (1) 2 T (2) ˜A(1) 21 (T (2))T V (1) 1 # = " P (2) ˜A(1) 12 R(1)V (1) 2 P (2) ˜A(1) 21 R(1)V (1) 1 # = P (1) " ˜A(1) 12 ˜V (1) 2 ˜A(1) 21 ˜V (1) 1 # = P (1) " ˜Y (1) 1 ˜Y (1) 2 # (64) and " ˜V (1) 1 ˜V (1) 2 # = " R(1)V (1) 1 R(1)V (1) 2 # . (65) The second equality in Eq. (64) is due to Eq. (45) and (47) where l = 2. The third equality in Eq. (64) is due to Eq. (40). In view of Eq.(57), we have A(2) 22 V (2) 2 ≈˜A(2) 22 V (2) 2 = A(0) 55 A(0) 56 A(0) 65 A(0) 66 T (1) ˜A(1) 34 (T (1))T T (1) ˜A(1) 43 (T (1))T A(0) 77 A(0) 78 A(0) 87 A(0) 88 V (2) 2 = Y (0) 5 Y (0) 6 Y (0) 7 Y (0) 8 + Y (1) 2 (66) where Y (1) 2 = " P (2) ˜A(1) 34 R(1)V (1) 4 P (2) ˜A(1) 43 R(1)V (1) 3 # = P (1) " ˜A(1) 34 ˜V (1) 4 ˜A(1) 43 ˜V (1) 3 # = P (1) " ˜Y (1) 3 ˜Y (1) 4 # (67) and " ˜V (1) 3 ˜V (1) 4 # = " R(1)V (1) 3 R(1)V (1) 4 # . (68) Substituting Eq. (61), (63) and (66) into (60), we obtain the final result for the matrix-matrix product Y = AV ≈Y (0) + P (0) ˜Y (1) + P (1) ˜Y (2) (69) where Y (0) = A(0) 11 V (0) 1 + A(0) 12 V (0) 2 A(0) 21 V (0) 1 + A(0) 22 V (0) 2 ... A(0) 87 V (0) 7 + A(0) 88 V (0) 8 (70) ˜Y (1) = ˜Y (1) 1 ˜Y (1) 2 ˜Y (1) 3 ˜Y (1) 4 = ˜A(1) 12 ˜V (1) 2 ˜A(1) 21 ˜V (1) 1 ˜A(1) 34 ˜V (1) 4 ˜A(1) 43 ˜V (1) 3 (71) ˜Y (2) = " ˜Y (2) 1 ˜Y (2) 2 # = " ˜A(2) 12 ˜V (2) 2 ˜A(2) 21 ˜V (2) 1 # (72) To summarize, matrix-matrix product computation includes the following steps: 1. Compute ˜V (1) in Eq. (65) and (68), and compute ˜V (2) in Eq. (62); 2. Compute Y (0) in Eq. (70), ˜Y (1) in Eq. (71) and ˜Y (2) in Eq. (72); 3. Interpolate and cumulative sum in Eq. (69); Note that all operations in step-2 are dense matrixmatrix product, well suited for dense linear algebra libraries optimized for GPU and TPU. The total number of super- and sub-diagonal blocks is upper bounded by twice the number of super- and sub-diagonal blocks at level-0, which is 2N(0) b . The run time of each dense matrix-matrix product is O(N2 r d). So the total run time is 5N(0) b N2 r d = 5LNrd = O(Ld). Here we leave out Nr since it is a constant model hyper-parameter. The coarsening in step-1 and interpolation in step-3 all use sparse matrices with fixed sparsity patterns. Hence matrices P (l) and R(l) are never explicitly formed and applying them can be easily done with standard library functions. Take Jax Numpy library as an example, coarsening can be done with sum() along row axis and interpolation can be done with repeat() along row axis. For this reason, step-1 and step-3 only have dense matrix operations as well. The formulation of the matrix-matrix product for the general level-M case is Y = AV = Y (0) + P (0)( ˜Y (1) + P (1)( ˜Y (2) + P (2)(· · · + P (M−2) ˜Y (M−1)) · · · )). (73) This formulation is a direct consequence of the nested attention matrix structure and can be derived similarly as Eq. (69).
|
2021
|
294
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3816–3830 August 1–6, 2021. ©2021 Association for Computational Linguistics 3816 Making Pre-trained Language Models Better Few-shot Learners Tianyu Gao†∗ Adam Fisch‡∗ Danqi Chen† †Princeton University ‡Massachusetts Institute of Technology {tianyug,danqic}@cs.princeton.edu [email protected] Abstract The recent GPT-3 model (Brown et al., 2020) achieves remarkable few-shot performance solely by leveraging a natural-language prompt and a few task demonstrations as input context. Inspired by their findings, we study few-shot learning in a more practical scenario, where we use smaller language models for which fine-tuning is computationally efficient. We present LM-BFF—better few-shot fine-tuning of language models1—a suite of simple and complementary techniques for finetuning language models on a small number of annotated examples. Our approach includes (1) prompt-based fine-tuning together with a novel pipeline for automating prompt generation; and (2) a refined strategy for dynamically and selectively incorporating demonstrations into each context. Finally, we present a systematic evaluation for analyzing few-shot performance on a range of NLP tasks, including classification and regression. Our experiments demonstrate that our methods combine to dramatically outperform standard fine-tuning procedures in this low resource setting, achieving up to 30% absolute improvement, and 11% on average across all tasks. Our approach makes minimal assumptions on task resources and domain expertise, and hence constitutes a strong task-agnostic method for few-shot learning.2 1 Introduction The GPT-3 model (Brown et al., 2020) has made waves in the NLP community by demonstrating astounding few-shot capabilities on myriad language understanding tasks. Given only a natural language prompt and a few demonstrations of the task, GPT-3 is able to make accurate predictions without updating any of the weights of its underlying lan*The first two authors contributed equally. 1Alternatively, language models’ best friends forever. 2Our implementation is publicly available at https:// github.com/princeton-nlp/LM-BFF. guage model. However, while remarkable, GPT-3 consists of 175B parameters, which makes it challenging to use in most real-wold applications. In this work, we study a more practical scenario in which we only assume access to a moderatelysized language model such as BERT (Devlin et al., 2019) or RoBERTa (Liu et al., 2019), and a small number of examples (i.e., a few-shot setting), which we can use to fine-tune the weights of the language model. This setting is appealing as (1) such models can be trained on typical research hardware; (2) few-shot settings are realistic, as it is generally both easy to acquire a few annotations (e.g., 32 examples) and efficient to train on them; and (3) updating parameters typically leads to better performance. Inspired by GPT-3’s findings, we propose several novel strategies for expanding its few-shot learning abilities to our setting, considering both classification and—for the first time—regression. First, we follow the route of prompt-based prediction, first developed by the GPT series (Radford et al., 2018, 2019; Brown et al., 2020) for zero-shot prediction and recently studied by PET (Schick and Sch¨utze, 2021a,b) for fine-tuning. Prompt-based prediction treats the downstream task as a (masked) language modeling problem, where the model directly generates a textual response (referred to as a label word) to a given prompt defined by a taskspecific template (see Figure 1(c)). Finding the right prompts, however, is an art—requiring both domain expertise and an understanding of the language model’s inner workings. Even if significant effort is invested, manual prompts are likely to be suboptimal. We address this issue by introducing automatic prompt generation, including a pruned brute-force search to identify the best working label words, and a novel decoding objective to automatically generate templates using the generative T5 model (Raffel et al., 2020)—all of which only require the few-shot training data. This allows us 3817 MLM head ··· no utterly ✔ ··· MLM head great (label:positive) terrible (label:negative) ✔ label:positive label:negative ✔ CLS head [CLS] No reason to watch . It was [MASK] . [SEP] A fun ride . It was great . [SEP] The drama discloses nothing . It was terrible . [SEP] [CLS] No reason to watch . [SEP] [CLS] it's a [MASK] movie in every regard , and [MASK] painful to watch . [SEP] MLM head ··· great terrible ✔ ··· (a) MLM pre-training (b) Fine-tuning (c) Prompt-based fine-tuning with demonstrations (our approach) Demonstration for label:positive Demonstration for label:negative Template Input Vocab Label space Label mapping Vocab Figure 1: An illustration of (a) masked language model (MLM) pre-training, (b) standard fine-tuning, and (c) our proposed LM-BFF using prompt-based fine-tuning with demonstrations. The underlined text is the task-specific template, and colored words are label words. to cheaply obtain effective prompts that match or outperform our manually chosen ones. Second, we adopt the idea of incorporating demonstrations as additional context. GPT-3’s naive “in-context learning” paradigm picks up to 32 randomly sampled examples, and concatenates them with the input. This method is not guaranteed to prioritize the most informative demonstrations, and mixing random examples from different classes together creates long contexts which can be hard to learn from. Additionally, the number of usable demonstrations is bounded by the model’s maximum input length. We develop a more refined strategy, where, for each input, we randomly sample a single example at a time from each class to create multiple, minimal demonstration sets. We also devise a novel sampling strategy that pairs inputs with similar examples, thereby providing the model with more discriminative comparisons. We present a systematic evaluation for analyzing few-shot performance on 8 single-sentence and 7 sentence-pair NLP tasks. We observe that given a small number of training examples, (1) promptbased fine-tuning largely outperforms standard finetuning; (2) our automatic prompt search method matches or outperforms manual prompts; and (3) incorporating demonstrations is effective for finetuning, and boosts few-shot performance. Together, these simple-yet-effective methods contribute towards a dramatic improvement across the tasks we evaluate on, and we obtain gains up to 30% absolute improvement (11% on average) compared to standard fine-tuning. For instance, we find that a RoBERTa-large model achieves around 90% accuracy on most binary sentence classification tasks, while only relying on 32 training examples. We refer to our approach as LM-BFF, better few-shot fine-tuning of language models: a strong, taskagnostic method for few-shot learning. 2 Related Work Language model prompting. The GPT series (Radford et al., 2018, 2019; Brown et al., 2020) fueled the development of prompt-based learning, and we follow many of its core concepts. We are also greatly inspired by the recent PET work (Schick and Sch¨utze, 2021a,b), although they mainly focus on a semi-supervised setting where a large set of unlabeled examples are provided. We only use a few annotated examples as supervision, and also explore automatically generated prompts and fine-tuning with demonstrations. Furthermore, we deviate from their evaluation by providing a more rigorous framework, as we will discuss in §3. Finally, there is a large body of work on prompting for mining knowledge from pre-trained models (Trinh and Le, 2018; Petroni et al., 2019; Davison et al., 2019; Talmor et al., 2020, inter alia). Different from these works, we focus on leveraging prompting for fine-tuning on downstream tasks. Automatic prompt search. Schick and Sch¨utze (2021a) and Schick et al. (2020) explore ways of identifying label words automatically, however, none of these results lead to better performance compared to hand-picked ones. In contrast, our method searches over both templates and label words, and is able to match or outperform our manual prompts. Several other attempts have been made in addition—yet these approaches either op3818 erate in limited domains, such as finding patterns to express specific relations (Jiang et al., 2020), or require a large number of examples for gradientguided search (Shin et al., 2020; Zhong et al., 2021). Our approach aims to develop general-purpose search methods that rely only on a few annotations. Fine-tuning of language models. A number of recent studies have focused on better methods for fine-tuning language models (Howard and Ruder, 2018; Dodge et al., 2020; Lee et al., 2020; Zhang et al., 2021). These works mainly focus on optimization and regularization techniques to stabilize fine-tuning. Here we use standard optimization techniques, and instead mainly focus our efforts on better prompt-based fine-tuning in a more extreme few-shot setting. We anticipate that results of these studies are largely complementary to ours. Few-shot learning. Broadly speaking, our setting is also connected to other few-shot learning paradigms in NLP, including (1) semi-supervised learning (Miyato et al., 2017; Xie et al., 2020; Chen et al., 2020), where a set of unlabeled examples are given; (2) meta-learning (Yu et al., 2018; Han et al., 2018; Bansal et al., 2020a,b; Bao et al., 2020), where a set of auxiliary tasks are given; and (3) intermediate training (Phang et al., 2018; Yin et al., 2020), where a related, intermediate task is given. We deviate from these settings by making minimal assumptions about available resources: we only assume a few annotated examples and a pre-trained language model. Our focus is on understanding how far we can push without any other advantages. 3 Problem Setup Task formulation. In this work, we assume access to a pre-trained language model L that we wish to fine-tune on a task D with a label space Y. For the task, we only assume K training examples per class3 for the task’s training set Dtrain, such that the total number of examples is Ktot = K × |Y|, and Dtrain = {(xi in, yi)}Ktot i=1. Our goal is then to develop task-agnostic learning strategies that generalize well to an unseen test set (xtest in , ytest) ∼Dtest. For model selection and hyper-parameter tuning, we assume a development set Ddev, of the same size as the few-shot training set, i.e., |Ddev| = |Dtrain|. This distinction is important: using a larger development set confers a significant advantage (see our 3For regression, we partition the data into two “classes” according to being above or below the median value. experiments in Appendix A), and subverts our initial goal of learning from limited data.4 For all of the following experiments (unless specified otherwise), we take L = RoBERTa-large and K = 16. Evaluation datasets. We conduct a systematic study across 8 single-sentence and 7 sentence-pair English tasks, including 8 tasks from the GLUE benchmark (Wang et al., 2019), SNLI (Bowman et al., 2015), and 6 other popular sentence classification tasks (SST-5, MR, CR, MPQA, Subj, TREC). All of the dataset details are provided in Appendix B. For single-sentence tasks, the goal is to make a prediction based on an input sentence xin = x1, such as whether a movie review is positive or not. For sentence-pair tasks, the goal is to take a pair of input sentences xin = (x1, x2) and predict the relationship between them. We also interchangeably refer to the inputs as <S1> or (<S1>, <S2>). Note that we mainly use SST-2 and SNLI for pilot experiments and model development, making it close to a true few-shot setting, at least for all the other datasets we evaluate on. Evaluation protocol. Systematically evaluating few-shot performance can be tricky. It is wellknown that fine-tuning on small datasets can suffer from instability (Dodge et al., 2020; Zhang et al., 2021), and results may change dramatically given a new split of data. To account for this, we measure average performance across 5 different randomly sampled Dtrain and Ddev splits. This issue has also been discussed in Schick and Sch¨utze (2021b)— they suggest using a fixed set of training examples. We argue that sampling multiple splits gives a more robust measure of performance, and a better estimate of the variance. We also observe that hyperparameters can make a significant difference, thus we sweep multiple hyper-parameters for each data sample, and take the best setting as measured on the Ddev of that sample (see Appendix C.1). 4 Prompt-based Fine-tuning Given a masked language model L, we first convert input xin to a token sequence ˜x, and the language model L then maps ˜x to a sequence of hidden vectors {hk ∈Rd}. During standard finetuning, we usually take ˜xsingle = [CLS]x1[SEP] or ˜xpair = [CLS]x1[SEP]x2[SEP]. For down4In contrast, Schick and Sch¨utze (2021a,b) do not use a development set, and adopt a set of hyper-parameters based on practical considerations. This is akin to “shooting in the dark” on a setting that we show can have unintuitive outcomes. 3819 Task Template Label words SST-2 <S1> It was [MASK] . positive: great, negative: terrible SST-5 <S1> It was [MASK] . v.positive: great, positive: good, neutral: okay, negative: bad, v.negative: terrible MR <S1> It was [MASK] . positive: great, negative: terrible CR <S1> It was [MASK] . positive: great, negative: terrible Subj <S1> This is [MASK] . subjective: subjective, objective: objective TREC [MASK] : <S1> abbreviation: Expression, entity: Entity, description: Description human: Human, location: Location, numeric: Number COLA <S1> This is [MASK] . grammatical: correct, not grammatical: incorrect MNLI <S1> ? [MASK] , <S2> entailment: Yes, netural: Maybe, contradiction: No SNLI <S1> ? [MASK] , <S2> entailment: Yes, netural: Maybe, contradiction: No QNLI <S1> ? [MASK] , <S2> entailment: Yes, not entailment: No RTE <S1> ? [MASK] , <S2> entailment: Yes, not entailment: No MRPC <S1> [MASK] , <S2> equivalent: Yes, not equivalent: No QQP <S1> [MASK] , <S2> equivalent: Yes, not equivalent: No STS-B <S1> [MASK] , <S2> yu: Yes, yl: No Table 1: Manual templates and label words that we used in our experiments. STS-B is a regression task (§4.2). stream classification tasks with a label space Y, we train a task-specific head, softmax(Woh[CLS]), by maximizing the log-probability of the correct label, where h[CLS] is the hidden vector of [CLS], and Wo ∈R|Y|×d is a set of randomly initialized parameters introduced at the start of fine-tuning. Similarly, for a regression task, we can introduce wo ∈Rd and optimize the mean squared error between wo·h[CLS] and the gold label. In either case, the number of new parameters can be substantial— for example, a simple binary classification task will introduce 2,048 new parameters for a RoBERTalarge model—making it challenging to learn from a small amount of annotated data (e.g., 32 examples). An alternative approach to solving this problem is prompt-based fine-tuning, in which L is directly tasked with “auto-completing” natural language prompts. For instance, we can formulate a binary sentiment classification task using a prompt with input x1 (e.g., “No reason to watch it .”) as: xprompt = [CLS] x1 It was [MASK] . [SEP] and let L decide whether it is more appropriate to fill in “great” (positive) or “terrible” (negative) for [MASK]. We now formalize this approach for classification and regression (§4.1 and §4.2), and discuss the importance of prompt selection (§4.3). 4.1 Classification Let M: Y →V be a mapping from the task label space to individual words5 in the vocabulary 5More generally, we can consider a one-to-many mapping M: Y →2|Y| in which we map labels to sets of words. However, we did not find significant gains in our experiments. V of L. Then for each xin, let the manipulation xprompt = T (xin) be a masked language modeling (MLM) input which contains one [MASK] token. In this way, we can treat our task as an MLM, and model the probability of predicting class y ∈Y as: p(y | xin) = p ([MASK] = M(y) | xprompt) = exp wM(y) · h[MASK] P y′∈Y exp wM(y′) · h[MASK] , (1) where h[MASK] is the hidden vector of [MASK] and wv denotes the pre-softmax vector corresponding to v ∈V. When supervised examples {(xin, y)} are available, L can be fine-tuned to minimize the cross-entropy loss. It is important to note that this approach re-uses the pre-trained weights wv and does not introduce any new parameters. It also reduces the gap between pre-training and fine-tuning, making it more effective in few-shot scenarios. 4.2 Regression We assume the same basic setup as in classification, but treat the label space Y as a bounded interval [vl, vu]. Inspired by Mettes et al. (2019), we model the problem as an interpolation between two opposing poles, {yl, yu}, with values vl and vu respectively. For instance, we can formulate our previous sentiment analysis task as a regression problem in the range [0, 1], where we slide between “terrible” (vl = 0) and “great” (vu = 1). In this way, we can express y as a mixture model: y = vl · p(yl | xin) + vu · p(yu | xin), (2) where p(yu | xin) is the probability of yu, and p(yl | xin) = 1 −p(yu | xin). Then we define 3820 Template Label words Accuracy SST-2 (positive/negative) mean (std) <S1> It was [MASK] . great/terrible 92.7 (0.9) <S1> It was [MASK] . good/bad 92.5 (1.0) <S1> It was [MASK] . cat/dog 91.5 (1.4) <S1> It was [MASK] . dog/cat 86.2 (5.4) <S1> It was [MASK] . terrible/great 83.2 (6.9) Fine-tuning 81.4 (3.8) SNLI (entailment/neutral/contradiction) mean (std) <S1> ? [MASK] , <S2> Yes/Maybe/No 77.2 (3.7) <S1> . [MASK] , <S2> Yes/Maybe/No 76.2 (3.3) <S1> ? [MASK] <S2> Yes/Maybe/No 74.9 (3.0) <S1> <S2> [MASK] Yes/Maybe/No 65.8 (2.4) <S2> ? [MASK] , <S1> Yes/Maybe/No 62.9 (4.1) <S1> ? [MASK] , <S2> Maybe/No/Yes 60.6 (4.8) Fine-tuning 48.4 (4.8) Table 2: The impact of templates and label words on prompt-based fine-tuning (K = 16). M: {yl, yu} →V, and model p(yu | xin) the same as Eq. (1). We fine-tune L to minimize the KL-divergence between the inferred p(yu | xin) and the observed mixture weight, (y−vl)/(vu−vl). 4.3 Manual prompts: the good and the bad The key challenge is to construct the template T and label words M(Y)—we refer to these two together as a prompt P. Previous works (Schick and Sch¨utze, 2021a,b) hand-craft both the templates and label words, which usually requires domain expertise and trial-and-error. Table 1 summarizes manual templates and label words chosen for each dataset in our experiments. These templates and label words were designed by intuition, and by considering formats used in previous literature. To better understand what constitutes a good template or label word, we conduct a pilot study on SST-2 and SNLI. Table 2 shows that different prompts can lead to substantial differences in final accuracy. Specifically, when a template is fixed, the better the label words match the “semantic classes”, the better the final accuracy is (great/terrible > good/bad > cat/dog). In extreme cases where we swap plausible label words (e.g., terrible/great), we achieve the worst overall performance.6 Furthermore, with the same set of label words, even a small change in the template can make a difference. For example, for SNLI, if we put [MASK] at the end, or swap sentence order, we observe a >10% drop. The above evidence clearly underlines the 6It is unclear, however, why RoBERTa thinks that “cat” is more positive than “dog”. The authors tend to disagree. importance of selecting good templates and label words. Searching for prompts, however, is hard, as the search space can be very large—especially for the template. Even worse, we only have a few examples to use to guide our search, which can easily overfit. We will address these issues next. 5 Automatic Prompt Generation We now explore principled ways of automating the search process for label words (§5.1) and templates (§5.2). Our goals are to reduce the human involvement required to design prompts, and to find more optimal settings than those that we manually choose. Here, we assume a classification task, but the process for regression is analogous. 5.1 Automatic selection of label words We first study how to construct a label word mapping M that maximizes accuracy on Ddev after fine-tuning, given a fixed template T . Naively searching all possible assignments, however, is (1) generally intractable, as the search space is exponential in the number of classes; and (2) prone to overfitting, as we will tend to uncover spurious correlations given only a few annotations. As a simple solution, for each class c ∈Y, we construct a pruned set Vc ⊂V of the top k vocabulary words based on their conditional likelihood using the initial L. That is, let Dc train ⊂Dtrain be the subset of all examples of class c. We take Vc as Top-k v∈V X xin∈Dc train log PL [MASK] = v | T (xin) , (3) where PL denotes the output probability distribution of L. To further narrow down the search space, we find the top n assignments over the pruned space that maximize zero-shot accuracy on Dtrain (both n and k are hyper-parameters, see Appendix C.2). Then we fine-tune all top n assignments, and rerank to find the best one using Ddev. This approach is similar to the automatic verbalizer search methods in Schick and Sch¨utze (2021a); Schick et al. (2020), except that we use a much simpler search process (brute-force) and also apply re-ranking— which we find to be quite helpful. 5.2 Automatic generation of templates Next, we study how to generate a diverse set of templates {T } automatically from a fixed set of label words M(Y). To address this challenging problem, we propose to use T5 (Raffel et al., 2020), 3821 Best template Generated templates Training examples for label:negative T5 … Training examples for label:positive … Decode <S1> A [MASK] one. <S1> This is [MASK]. … <S1> A [MASK] one. A fun ride. <X> great <Y> A pleasure to watch. <X> great <Y> No reason to watch. <X> terrible <Y> This junk. <X> terrible <Y> Fine-tune and evaluate positive: great, negative: terrible Label mapping Figure 2: Our approach for template generation. a large pre-trained text-to-text Transformer. T5 is pre-trained to fill in missing spans (replaced by T5 mask tokens, e.g., <X> or <Y>) in its input. For example, given the input “Thank you <X> me to your party <Y> week”, T5 is trained to generate “<X> for inviting <Y> last <Z>”, meaning that “for inviting” is the replacement for <X> and “last” is the replacement for <Y>. This is well suited for prompt generation: we can simply take input sentences from Dtrain and let the T5 model construct the template T , without having to specify a predefined number of tokens for it. Given an input example (xin, y) ∈Dtrain, we consider the following simple conversions, denoted as Tg(xin, y), for formulating the T5 model inputs:7 <S1> −→<X> M(y) <Y> <S1>, <S1> −→<S1> <X> M(y) <Y>, <S1>, <S2> −→<S1> <X> M(y) <Y> <S2>. As shown in Figure 2, we rely on the T5 model to fill in the placeholders. When decoding, our goal here is to find an output that can work well for all examples in Dtrain, i.e., the output template T that maximizes P (xin,y)∈Dtrain log PT5(T | Tg(xin, y)), where PT5 denotes the output probability distribution of T5. It can be decomposed according to: |T | X j=1 X (xin,y)∈Dtrain log PT5 tj | t1, ..., tj−1, Tg xin, y , (4) where (t1, . . . , t|T |) are the template tokens. We use beam search to decode multiple template candidates. Concretely, we use a wide beam width (e.g., 100) to cheaply obtain a large set of diverse templates. We then fine-tune each generated template on Dtrain and use Ddev to either pick the single template with the best performance (Table 3), or 7We consider putting the label word both before and after the input sentence for single-sentence tasks. However, we find that it is always better to put the label words in the middle (between the two sentences) for sentence-pair tasks. the top k templates to use as an ensemble (Table 4). Though it might appear to be expensive to fine-tune the model on each individual template, this is fast in practice due to the small size of Dtrain, and is also fully automated: making it easy to use, compared to manually tuning prompts for each dataset. 6 Fine-tuning with Demonstrations In this section, we study whether we can leverage demonstrations when fine-tuning medium-sized LMs, and find better ways to exploit them. 6.1 Training examples as demonstrations GPT-3’s naive approach to in-context learning simply involves concatenating the input with up to 32 examples randomly drawn from the training set. This approach is suboptimal as (1) the number of available demonstrations is bounded by the model’s maximum input length;8 and (2) mixing numerous random examples from different classes together creates extremely long contexts which can be hard to leverage, especially for a smaller model. To address these issues, we propose a simpler solution: at each training step, we randomly sample one9 example x(c) in , y(c) ∈Dtrain from each class, convert it into T x(c) in with [MASK] replaced by M(y(c))—we denote this as ˜T x(c) in , y(c) —and then concatenate them with xin (Figure 1(c)): T xin ⊕˜T x(1) in , y(1) ⊕· · · ⊕˜T x(|Y|) in , y(|Y|) . Here ⊕denotes concatenation of input sequences. During both training and inference we sample multiple demonstration sets for each xin. Note that both xin and demonstration examples are sampled from the same set Dtrain during training. At testing time, we still sample demonstration sets from Dtrain and ensemble predictions across all sets. 6.2 Sampling similar demonstrations We observe that controlling the construction of the demonstration examples {(x(c) in , y(c))} is crucial for good final performance. For example, if the set of contrastive demonstrations x(c) in are all dramatically different—from each other, or from the query xin—then it becomes challenging for the language model to decipher meaningful patterns. As a result, the model may simply ignore 8GPT-3 uses a context size of 2,048 while most smaller language models (e.g., RoBERTa) have a context size of 512. 9We also explored sampling multiple examples per class, but did not observe any improvements. 3822 SST-2 SST-5 MR CR MPQA Subj TREC CoLA (acc) (acc) (acc) (acc) (acc) (acc) (acc) (Matt.) Majority† 50.9 23.1 50.0 50.0 50.0 50.0 18.8 0.0 Prompt-based zero-shot‡ 83.6 35.0 80.8 79.5 67.6 51.4 32.0 2.0 “GPT-3” in-context learning 84.8 (1.3) 30.6 (0.9) 80.5 (1.7) 87.4 (0.8) 63.8 (2.1) 53.6 (1.0) 26.2 (2.4) -1.5 (2.4) Fine-tuning 81.4 (3.8) 43.9 (2.0) 76.9 (5.9) 75.8 (3.2) 72.0 (3.8) 90.8 (1.8) 88.8 (2.1) 33.9 (14.3) Prompt-based FT (man) 92.7 (0.9) 47.4 (2.5) 87.0 (1.2) 90.3 (1.0) 84.7 (2.2) 91.2 (1.1) 84.8 (5.1) 9.3 (7.3) + demonstrations 92.6 (0.5) 50.6 (1.4) 86.6 (2.2) 90.2 (1.2) 87.0 (1.1) 92.3 (0.8) 87.5 (3.2) 18.7 (8.8) Prompt-based FT (auto) 92.3 (1.0) 49.2 (1.6) 85.5 (2.8) 89.0 (1.4) 85.8 (1.9) 91.2 (1.1) 88.2 (2.0) 14.0 (14.1) + demonstrations 93.0 (0.6) 49.5 (1.7) 87.7 (1.4) 91.0 (0.9) 86.5 (2.6) 91.4 (1.8) 89.4 (1.7) 21.8 (15.9) Fine-tuning (full)† 95.0 58.7 90.8 89.4 87.8 97.0 97.4 62.6 MNLI MNLI-mm SNLI QNLI RTE MRPC QQP STS-B (acc) (acc) (acc) (acc) (acc) (F1) (F1) (Pear.) Majority† 32.7 33.0 33.8 49.5 52.7 81.2 0.0 Prompt-based zero-shot‡ 50.8 51.7 49.5 50.8 51.3 61.9 49.7 -3.2 “GPT-3” in-context learning 52.0 (0.7) 53.4 (0.6) 47.1 (0.6) 53.8 (0.4) 60.4 (1.4) 45.7 (6.0) 36.1 (5.2) 14.3 (2.8) Fine-tuning 45.8 (6.4) 47.8 (6.8) 48.4 (4.8) 60.2 (6.5) 54.4 (3.9) 76.6 (2.5) 60.7 (4.3) 53.5 (8.5) Prompt-based FT (man) 68.3 (2.3) 70.5 (1.9) 77.2 (3.7) 64.5 (4.2) 69.1 (3.6) 74.5 (5.3) 65.5 (5.3) 71.0 (7.0) + demonstrations 70.7 (1.3) 72.0 (1.2) 79.7 (1.5) 69.2 (1.9) 68.7 (2.3) 77.8 (2.0) 69.8 (1.8) 73.5 (5.1) Prompt-based FT (auto) 68.3 (2.5) 70.1 (2.6) 77.1 (2.1) 68.3 (7.4) 73.9 (2.2) 76.2 (2.3) 67.0 (3.0) 75.0 (3.3) + demonstrations 70.0 (3.6) 72.0 (3.1) 77.5 (3.5) 68.5 (5.4) 71.1 (5.3) 78.1 (3.4) 67.7 (5.8) 76.4 (6.2) Fine-tuning (full)† 89.8 89.5 92.6 93.3 80.9 91.4 81.7 91.9 Table 3: Our main results using RoBERTa-large. †: full training set is used (see dataset sizes in Table B.1); ‡: no training examples are used; otherwise we use K = 16 (per class) for few-shot experiments. We report mean (and standard deviation) performance over 5 different splits (§3). Majority: majority class; FT: fine-tuning; man: manual prompt (Table 1); auto: automatically searched templates (§5.2); “GPT-3” in-context learning: using the in-context learning proposed in Brown et al. (2020) with RoBERTa-large (no parameter updates). the context, or even get confused by the additional examples. To address this issue, we devise a simple strategy in which we only sample examples that are semantically close to xin. Specifically, we use a pre-trained SBERT (Reimers and Gurevych, 2019) model to obtain embeddings for all input sentences (for sentence-pair tasks, we use the concatenation of the two sentences). Here we just feed the raw sentences without the templates into SBERT. For each query xin and each label c ∈Y, we sort all training instances with the label x ∈Dc train by their similarity score to the query cos(e(xin), e(x)), and only sample from the top r = 50% instances for each class to use as demonstrations. 7 Experiments We present our main results, and address several research questions pertaining to our LM-BFF approach. Implementation details are in Appendix C. 7.1 Main results We use a RoBERTa-large model and set K = 16 in our experiments. A comparison of using RoBERTa vs BERT can be found in Appendix D. For automatic prompt search, in our main table we report automatic template search only (which consistently performs the best, see Table 5). To put our results in perspective, we compare to a number of baselines, namely (1) standard fine-tuning in our few-shot setting; (2) standard fine-tuning using the full training set; (3) simply taking the most frequent class (measured on the full training set); (4) prompt-based zero-shot prediction where we take our manual prompts and use L “out-of-thebox” without using any training examples; and (5) “GPT-3” in-context learning, where we use the same prompt-based zero-shot setting, but augment the context with randomly sampled 32 demonstrations (and still use RoBERTa-large, not GPT-3). Single-prompt results. Table 3 shows our main results using a single prompt, either from our manually designed ones (Table 1) , or the best generated ones. First, prompt-based zero-shot prediction achieves much better performance than the majority class, showing the pre-encoded knowledge in RoBERTa. Also, “GPT-3” in-context learning does not always improve over zero-shot prediction, likely because smaller language models are not expressive enough to use off-the-shelf like GPT-3. 3823 Prompt-based Fine-tuning MNLI RTE Our single manual P 68.3 (2.3) 69.1 (3.6) PPET 71.9 (1.5) 69.2 (4.0) Pours, |Pours| = |PPET| 70.4 (3.1) 73.0 (3.2) + demonstrations 74.0 (1.9) 71.9 (4.6) Pours, |Pours| = 20 72.7 (2.5) 73.1 (3.3) + demonstrations 75.4 (1.6) 72.3 (4.5) Table 4: Ensemble models using manual prompts from PET (Schick and Sch¨utze, 2021a,b) and our automatic templates. PET uses 4 prompts for MNLI and 5 for RTE. We also use an equal number of templates in |Pours| = |PPET| for a fair comparison. SST-2 SNLI TREC MRPC Manual 92.7 77.2 84.8 74.5 Auto T 92.3 77.1 88.2 76.2 Auto L 91.5 75.6 87.0 77.2 Auto T + L 92.1 77.0 89.2 74.0 Table 5: Comparison between manual prompts and different automatic prompt generation methods: autogenerated templates (Auto T), auto-generated label words (Auto L), and their combination (Auto T + L). Second, prompt-based fine-tuning can greatly outperform standard fine-tuning, both when using a manual prompt or a generated one. CoLA is one interesting exception, as the input may be a nongrammatical sentence which is out of the distribution of L. Generally, our automatically searched templates can achieve comparable or even higher results than manual ones, especially for tasks in which constructing strong manual templates is less intuitive (e.g., TREC, QNLI and MRPC). Finally, using demonstrations in context leads to consistent gains in a majority of tasks. In summary, our combined solution—fine-tuning with automatically searched templates and sampled demonstration sets—achieves a 30% gain on SNLI compared to standard fine-tuning, and 11% gain on average. Ensemble results. An advantage of automatic prompt search is that we can generate as many prompts as we want, train individual models, and create large ensembles. PET (Schick and Sch¨utze, 2021a,b) also ensembles multiple models trained with manual prompts.10 In Table 4, we make a direct comparison of our searched prompts and PET’s manual prompts on MNLI and RTE (two 10They then use unlabeled data and distillation to get a single model, which is outside of our scope. SST-2 (positive/negative) Auto T M(Y) = {great, terrible} #1. <S1> A [MASK] one . #2. <S1> A [MASK] piece . #3. <S1> All in all [MASK] . Auto L T (xin) = <S1> It was [MASK]. #1. irresistible/pathetic #2. wonderful/bad #3. delicious/bad SNLI (entailment/neutral/contradiction) Auto T M(Y) = {Yes, Maybe, No} #1. <S1> . [MASK] , no , <S2> #2. <S1> . [MASK] , in this case <S2> #3. <S1> . [MASK] this time <S2> Auto L T (xin) = <S1> ? [MASK] , <S2> #1. Alright/Watch/Except #2. Hi/Watch/Worse #3. Regardless/Fortunately/Unless Table 6: Examples of our automatically generated templates (Auto T) and label words (Auto L). datasets that we evaluate in common).11 As the results show, an ensemble with multiple templates always improves performance. An ensemble of the same number of automatic templates achieves comparable or better performance than the ensemble of PET’s manual prompts. Increasing the number of automatic templates brings further gains. 7.2 Analysis of generated prompts Table 5 gives the results of using manual vs automatic prompts. For automatic prompts, we compare template search (Auto T), label word search (Auto L), and a joint variant (Auto T + L) in which we start from manual label words, apply Auto T, and then Auto L. In most cases, Auto T achieves comparable or higher performance than manual ones, and is consistently the best variant. Auto L outperforms manual prompts on TREC and MRPC—but is considerably worse on SNLI. Auto T + L is often better than Auto L, but only sometimes better than Auto T. Table 6 shows examples from Auto T and Auto L (A full list in Appendix E). Auto T templates generally fit the context and label words well, but can contain biased peculiarities (e.g., “{Yes/No}, no” in SNLI). For Auto L words, things are mixed: while most look intuitively reasonable, there are also some mysterious abnormalities (e.g., “Hi” for the “entailment” class in SNLI). 11In the PET NLI templates, the hypothesis is put before the premise, which we actually found to be suboptimal. In our experiments, we swap the two and get better results. 3824 SST-2 SNLI TREC MRPC Prompt-based FT 92.7 77.2 84.8 74.5 Uniform sampling 92.3 78.8 85.6 70.9 + RoBERTa sel. 92.7 79.5 83.4 76.6 + SBERT sel. 92.6 79.7 87.5 77.8 Table 7: Impact of demonstration sampling strategies. Uniform sampling randomly samples demonstrations, while selective (sel.) sampling only takes top sentences measured by the sentence encoders (§6). 7.3 Analysis of demonstration sampling Table 7 compares the performance of demonstrations using uniform sampling to selective sampling by SBERT. We acknowledge that SBERT is trained on SNLI and MNLI datasets, thus we also tried a simple sentence encoder using mean pooling of hidden representations from RoBERTa-large. We find that in either case, using selective sampling outperforms uniform sampling, highlighting the importance of sampling similar examples for incorporating demonstrations in context. 7.4 Sample efficiency Figure 3 illustrates how standard fine-tuning and our LM-BFF compare as K increases. For a simple task such as SST-2 (also see MR, CR and MPQA in Table 3), despite using only 32 total examples, LMBFF has already nearly saturated its performance and is comparable to standard fine-tuning over the entire dataset. On the harder task of SNLI, LMBFF continues to improve as K increases while still maintaining a performance gap over standard finetuning, until the two converge around K = 256. 8 Discussion Reformulating NLP tasks as MLM has exciting implications for few-shot learning, but also has limitations. First, while LM-BFF greatly outperforms standard fine-tuning, Table 3 shows that, overall, the performance still substantially lags behind finetuning with thousands of examples, especially for harder tasks. Additionally, just like standard finetuning, our results also suffer from high variance. As described in §2, several recent studies have tried to counter instability in few-shot fine-tuning and we expect these methods to also help here. With respect to automatic prompt generation, despite its effectiveness, we still find it practically challenging to expand the search space, or generalize well based on only approximately 32 examples. 16 32 64 128 256 K 70 75 80 85 90 95 Accuracy (%) SST-2 Fine-tune LM-BFF 16 32 64 128 256 K 40 50 60 70 80 90 Accuracy (%) SNLI Fine-tune LM-BFF Figure 3: Standard fine-tuning vs our LM-BFF as a function of K (# instances per class). For lower K, our method consistently outperforms standard fine-tuning. This is partly due to our lingering reliance on some manual design—either manual templates (for label word search) or manual label words (for template search), which allows us to get our search off the ground, but does also bias it towards areas of the search space that we might have already imagined. Finally, it is important to clarify that LM-BFF favors certain tasks which (1) can be naturally posed as a “fill-in-the-blank” problem; (2) have relatively short input sequences; and (3) do not contain many output classes. Issues (2) and (3) might be ameliorated with longer-context language models (e.g., Beltagy et al., 2020). For tasks that are not straightforward to formulate in prompting, such as structured prediction, issue (1) is more fundamental. We leave it as an open question for future work. 9 Conclusion In this paper we presented LM-BFF, a set of simple but effective techniques for fine-tuning language models using only a few examples. Our approach proposes to (1) use prompt-based finetuning with automatically searched prompts; and (2) include selected task demonstrations (training examples) as part of the input context. We show that our method outperforms vanilla fine-tuning by up to 30% (and 11% on average). We concluded by discussing the limitations of our approach, and posed open questions for future study. Acknowledgements We thank the members of Princeton, MIT, Tsinghua NLP groups and the anonymous reviewers for their valuable feedback. TG is supported by a Graduate Fellowship at Princeton University and AF is supported by an NSF Graduate Research Fellowship. This research is also partly supported by a Google Research Scholar Award. 3825 References Trapit Bansal, Rishikesh Jha, and Andrew McCallum. 2020a. Learning to few-shot learn across diverse natural language classification tasks. In International Conference on Computational Linguistics (COLING). Trapit Bansal, Rishikesh Jha, Tsendsuren Munkhdalai, and Andrew McCallum. 2020b. Self-supervised meta-learning for few-shot natural language classification tasks. In Empirical Methods in Natural Language Processing (EMNLP). Yujia Bao, Menghua Wu, Shiyu Chang, and Regina Barzilay. 2020. Few-shot text classification with distributional signatures. In International Conference on Learning Representations (ICLR). Roy Bar Haim, Ido Dagan, Bill Dolan, Lisa Ferro, Danilo Giampiccolo, Bernardo Magnini, and Idan Szpektor. 2006. The second PASCAL recognising textual entailment challenge. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document Transformer. arXiv:2004.05150. Luisa Bentivogli, Peter Clark, Ido Dagan, and Danilo Giampiccolo. 2009. The fifth PASCAL recognizing textual entailment challenge. In TAC. Samuel Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Empirical Methods in Natural Language Processing (EMNLP). Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS). Daniel Cer, Mona Diab, Eneko Agirre, I˜nigo LopezGazpio, and Lucia Specia. 2017. SemEval-2017 task 1: Semantic textual similarity multilingual and crosslingual focused evaluation. In the 11th International Workshop on Semantic Evaluation (SemEval2017). Jiaao Chen, Zichao Yang, and Diyi Yang. 2020. MixText: Linguistically-informed interpolation of hidden space for semi-supervised text classification. In Association for Computational Linguistics (ACL). Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The PASCAL recognising textual entailment challenge. In the First International Conference on Machine Learning Challenges: Evaluating Predictive Uncertainty Visual Object Classification, and Recognizing Textual Entailment. Joe Davison, Joshua Feldman, and Alexander M Rush. 2019. Commonsense knowledge mining from pretrained models. In Empirical Methods in Natural Language Processing (EMNLP). Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional Transformers for language understanding. In North American Chapter of the Association for Computational Linguistics (NAACL). Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. 2020. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305. William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In the Third International Workshop on Paraphrasing (IWP2005). Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. 2007. The third PASCAL recognizing textual entailment challenge. In the ACLPASCAL Workshop on Textual Entailment and Paraphrasing. Xu Han, Hao Zhu, Pengfei Yu, Ziyun Wang, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2018. Fewrel: A large-scale supervised few-shot relation classification dataset with state-of-the-art evaluation. In Empirical Methods in Natural Language Processing (EMNLP). Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. In Association for Computational Linguistics (ACL). Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In ACM SIGKDD international conference on Knowledge discovery and data mining. Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association of Computational Linguistics (TACL). Cheolhyoung Lee, Kyunghyun Cho, and Wanmo Kang. 2020. Mixout: Effective regularization to finetune large-scale pretrained language models. In International Conference on Learning Representations (ICLR). Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692. Pascal Mettes, Elise van der Pol, and Cees Snoek. 2019. Hyperspherical prototype networks. In Advances in Neural Information Processing Systems (NeurIPS). Takeru Miyato, Andrew M Dai, and Ian Goodfellow. 2017. Adversarial training methods for semisupervised text classification. In International Conference on Learning Representations (ICLR). 3826 Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Association for Computational Linguistics (ACL). Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Association for Computational Linguistics (ACL). Fabio Petroni, Tim Rockt¨aschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Empirical Methods in Natural Language Processing (EMNLP). Jason Phang, Thibault F´evry, and Samuel R Bowman. 2018. Sentence encoders on STILTs: Supplementary training on intermediate labeled-data tasks. arXiv preprint arXiv:1811.01088. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Technical report, OpenAI. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Technical report, OpenAI. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text Transformer. The Journal of Machine Learning Research (JMLR), 21(140). Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Empirical Methods in Natural Language Processing (EMNLP). Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. In Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing (EMNLPIJCNLP). Timo Schick, Helmut Schmid, and Hinrich Sch¨utze. 2020. Automatically identifying words that can serve as labels for few-shot text classification. In International Conference on Computational Linguistics (COLING). Timo Schick and Hinrich Sch¨utze. 2021a. Exploiting cloze questions for few-shot text classification and natural language inference. In European Chapter of the Association for Computational Linguistics (EACL). Timo Schick and Hinrich Sch¨utze. 2021b. It’s not just size that matters: Small language models are also few-shot learners. In North American Chapter of the Association for Computational Linguistics (NAACL). Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Automatic prompt construction for masked language models. In Empirical Methods in Natural Language Processing (EMNLP). Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Empirical Methods in Natural Language Processing (EMNLP). Alon Talmor, Yanai Elazar, Yoav Goldberg, and Jonathan Berant. 2020. oLMpics-on what language model pre-training captures. Transactions of the Association of Computational Linguistics (TACL), 8. Trieu H Trinh and Quoc V Le. 2018. A simple method for commonsense reasoning. arXiv preprint arXiv:1806.02847. Ellen M Voorhees and Dawn M Tice. 2000. Building a question answering test collection. In the 23rd annual international ACM SIGIR conference on Research and development in information retrieval. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations (ICLR). Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. Transactions of the Association of Computational Linguistics (TACL), 7. Janyce Wiebe, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language resources and evaluation, 39(2-3). Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT). Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. 2020. Unsupervised data augmentation for consistency training. Advances in Neural Information Processing Systems (NeurIPS), 33. Wenpeng Yin, Nazneen Fatema Rajani, Dragomir Radev, Richard Socher, and Caiming Xiong. 2020. Universal natural language processing with limited annotations: Try few-shot textual entailment as a start. In Empirical Methods in Natural Language Processing (EMNLP). Mo Yu, Xiaoxiao Guo, Jinfeng Yi, Shiyu Chang, Saloni Potdar, Yu Cheng, Gerald Tesauro, Haoyu Wang, 3827 and Bowen Zhou. 2018. Diverse few-shot text classification with multiple metrics. In North American Chapter of the Association for Computational Linguistics (NAACL). Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q Weinberger, and Yoav Artzi. 2021. Revisiting fewsample BERT fine-tuning. In International Conference on Learning Representations (ICLR). Zexuan Zhong, Dan Friedman, and Danqi Chen. 2021. Factual probing is [MASK]: Learning vs. learning to recall. In North American Association for Computational Linguistics (NAACL). 3828 A Impact of Development Sets Table A.1 shows how the size of the development sets can affect the final performance of the model. For “No Ddev”, we take the same hyper-parameters from Schick and Sch¨utze (2021a,b): batch size = 16, learning rate = 1e-5 and training steps = 250. We also experiment with a variant that we sample a development set of 10 times larger than the training set. We can see that using larger development sets leads to better performance, and this is why we stick to |Dtrain| = |Ddev| in our few-shot setting. Fine-tuning SST-2 SNLI TREC MRPC No Ddev 79.5 49.2 83.9 77.8 |Ddev| = |Dtrain| 81.4 48.4 88.8 76.6 |Ddev| = 10|Dtrain| 83.5 52.0 89.4 79.6 Prompt-based FT SST-2 SNLI TREC MRPC No Ddev 92.1 75.3 84.8 70.2 |Ddev| = |Dtrain| 92.7 77.2 84.8 74.5 |Ddev| = 10|Dtrain| 93.0 79.7 89.3 80.9 Table A.1: Impact of different sizes of development sets. Standard deviations are omitted here to save space. For No |Ddev|, we use the same set of hyper-parameters as Schick and Sch¨utze (2021a,b). B Datasets For SNLI (Bowman et al., 2015) and datasets from GLUE (Wang et al., 2019), including SST2 (Socher et al., 2013), CoLA (Warstadt et al., 2019), MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016), RTE (Dagan et al., 2005; Bar Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), MRPC (Dolan and Brockett, 2005), QQP12 and STS-B (Cer et al., 2017), we follow Zhang et al. (2021) and use their original development sets for testing. For datasets which require a cross-validation evaluation—MR (Pang and Lee, 2005), CR (Hu and Liu, 2004), MPQA (Wiebe et al., 2005), Subj (Pang and Lee, 2004)—we simply randomly sample 2,000 examples as the testing set and leave them out from training. For SST5 (Socher et al., 2013) and TREC (Voorhees and Tice, 2000), we use their official test sets. We show dataset statistics in Table B.1. C Experimental Details C.1 Hyper-parameter selection For grid search, we take learning rates from {1e5, 2e-5, 5e-5} and batch sizes from {2, 4, 8}. These 12https://www.quora.com/q/quoradata/ numbers are picked by pilot experiments on the SST-2 and SNLI datasets. We also use early stopping to avoid overfitting. For each trial, we train the model for 1,000 steps, validate the performance every 100 steps, and take the best checkpoint. C.2 Prompt-based fine-tuning Table 1 shows all the manual templates and label words we use in experiment. For automatically template generation, we take the T5-3B13 model, which is the largest publicly available one that can fit on a single GPU. For automatically searching label words, we set k to 100 for all tasks except SST-5 and TREC. For SST-5 we set a smaller k = 30, as it is a 5-way classification task. For TREC, we observe that filtering Vc using conditional likelihood alone is still noisy, thus we set k = 1000, and then re-rank Vc by the nearest neighbors of the original manual label words and take the top 30 per class. We set n to 100 in all experiments. Due to the large number of trials in automatic search, we take a fixed set of hyper-parameters in this part: batch size of 8 and learning rate of 1e-5. Since the idea of prompt-based fine-tuning is to make the input and output distribution close to the pre-training, the implementation details are crucial. For templates, we put extra space before sentences if it is not at the beginning of the input. Also, we lowercase the first letter of the sentence if it is concatenated with a prefix (e.g., <S2> in Table 1). Also if one sentence is appended any punctuation (e.g., <S1> in Table 1), then the last character of the original sentence is discarded. Finally, we prepend a space for label words in M(Y). For example, we use “ great” instead of “great” in the RoBERTa vocabulary, where “ ” stands for space. C.3 Fine-tuning with demonstrations When using demonstrations, we sample 16 different sets of demonstrations for each input and average the predicted log probability for each class during inference. We find that further increasing the number of samples does not bring substantial improvement. Additional, we have tried different aggregation methods like taking the result with the maximum confidence and we did not find a meaningful improvement. For selective demonstrations, we take roberta-large-nli-stsb 13We take the T5 1.0 checkpoint, which is trained on both unsupervised and downstream task data. We compared it to T5 1.1 (without downstream task data) and did not find a significant difference in generated templates. 3829 Category Dataset |Y| L #Train #Test Type Labels (classification tasks) SST-2 2 19 6,920 872 sentiment positive, negative SST-5 5 18 8,544 2,210 sentiment v. pos., positive, neutral, negative, v. neg. MR 2 20 8,662 2,000 sentiment positive, negative singleCR 2 19 1,775 2,000 sentiment positive, negative sentence MPQA 2 3 8,606 2,000 opinion polarity positive, negative Subj 2 23 8,000 2,000 subjectivity subjective, objective TREC 6 10 5,452 500 question cls. abbr., entity, description, human, loc., num. CoLA 2 8 8,551 1,042 acceptability grammatical, not grammatical MNLI 3 22/11 392,702 9,815 NLI entailment, neutral, contradiction SNLI 3 14/8 549,367 9,842 NLI entailment, neutral, contradiction sentenceQNLI 2 11/30 104,743 5,463 NLI entailment, not entailment pair RTE 2 49/10 2,490 277 NLI entailment, not entailment MRPC 2 22/21 3,668 408 paraphrase equivalent, not equivalent QQP 2 12/12 363,846 40,431 paraphrase equivalent, not equivalent STS-B R 11/11 5,749 1,500 sent. similarity Table B.1: The datasets evaluated in this work. |Y|: # of classes for classification tasks (with one exception: STS-B is a real-valued regression task over the interval [0, 5]). L: average # of words in input sentence(s). Note that we only sample Dtrain and Ddev of K × |Y| examples from the original training set in our few-shot experiments (§3). BERT-large SST-2 SNLI TREC MRPC Fine-tuning 79.5 51.4 80.3 74.4 Prompt-based FT 85.6 59.2 79.0 66.8 + demo (1-seg) 87.5 50.4 77.2 68.5 + demo (2-seg) 86.1 61.3 77.9 73.2 + demo (n-seg) 86.4 58.6 79.6 71.0 RoBERTa-large SST-2 SNLI TREC MRPC Fine-tuning 81.4 48.4 88.8 76.6 Prompt-based FT 92.7 77.2 84.8 74.5 + demonstrations 92.6 79.7 87.5 77.8 Table D.1: A comparison of BERT-large vs RoBERTalarge. We use manual prompts in these experiments. mean-tokens14 from Reimers and Gurevych (2019) as our sentence embedding model. D Comparisons of BERT vs RoBERTa Table D.1 compares the results of BERT-large (uncased) and RoBERTa-large in our settings. Pretrained BERT provides two segment embeddings (A/B) for different parts of input. The common practice, when fine-tuning BERT, is that using only segment A for single-sentence tasks, and using segment A/B for the two sentences in sentence-pair tasks. In our case of incorporating demonstrations, however, we have more than two sentences. Thus we explore the following different strategies for segments: (1) using the A segment for all sentences 14https://github.com/UKPLab/ sentence-transformers (1-seg); (2) using the A segment for the original input and the B segment for the demonstrations (2-seg); (3) using different segment embeddings for each sentence (n-seg), e.g., for SNLI, we use different segments for each premise and hypothesis in both the original input and the demonstrations, which leads to a total number of 8 segment embeddings. This introduces new segment embeddings (randomly initialized and learned during fine-tuning) as the pre-trained BERT only has two. Table D.1 shows that prompt-based fine-tuning with demonstrations also works for BERT, and 2seg works the best when incorporating demonstrations. Still, we take RoBERTa-large as our main model, for RoBERTa performs much better than BERT and RoBERTa saves the trouble to tune the usage of segment embeddings. E Generated Prompts We demonstrate the top 3 automatically generated templates and label words for all tasks in Table E.1. In general, most automatic templates are reasonable and grammatically correct. For the label words, the generated results look intuitive for most single sentence tasks. For other tasks, the automatic ones can be counterintuitive in some cases. It is still unclear why the language model picks these words and sometimes they actually work well. We leave this for future study. 3830 Task Auto template Auto label words SST-2 (positive/negative) <S1> A [MASK] one . irresistible/pathetic <S1> A [MASK] piece . wonderful/bad <S1> All in all [MASK] . delicious/bad SST-5 (very positive/positive/neutral/negative/very negative) <S1> The movie is [MASK] . wonderful/remarkable/hilarious/better/awful <S1> The music is [MASK] . wonderful/perfect/hilarious/better/awful <S1> But it is [MASK] . unforgettable/extraordinary/good/better/terrible MR (positive/negative) It was [MASK] ! <S1> epic/terrible <S1> It’s [MASK] . epic/awful <S1> A [MASK] piece of work . exquisite/horrible CR (positive/negative) <S1> It’s [MASK] ! fantastic/horrible <S1> The quality is [MASK] . neat/pointless <S1> That is [MASK] . magnificent/unacceptable MPQA (positive/negative) <S1> is [MASK] . important/close <S1>, [MASK] ! needed/bad <S1>. [MASK] . unexpected/shocking Subj (subjective/objective) <S1> It’s all [MASK] . everywhere/tragic <S1> It’s [MASK] . everywhere/horrifying <S1> Is it [MASK] ? something/surreal TREC (abbreviation/entity/description/human/location/numeric) Q: [MASK] : <S1> Application/Advisor/Discussion/Culture/Assignment/Minute <S1> Why [MASK]? Production/AE/Context/Artist/Assignment/Minute <S1> Answer: [MASK] . Personality/Advisor/Conclusion/Hum/Assignment/Minute CoLA (grammatical/not grammatical) <S1> You are [MASK] . one/proof It is [MASK] . <S1> wrong/sad I am [MASK] . <S1> misleading/disappointing MNLI (entailment/neutral/contradiction) <S1> . [MASK] , you are right , <S2> Fine/Plus/Otherwise <S1> . [MASK] you’re right <S2> There/Plus/Otherwise <S1> . [MASK] ! <S2> Meaning/Plus/Otherwise SNLI (entailment/neutral/contradiction) <S1> . [MASK] , no , <S2> Alright/Watch/Except <S1> . [MASK] , in this case <S2> Hi/Watch/Worse <S1> . [MASK] this time <S2> Regardless/Fortunately/Unless QNLI (entailment/not entailment) <S1> ? [MASK] . Yes , <S2> Okay/Nonetheless <S1> ? [MASK] . It is known that <S2> Notably/Yet <S1> ? [MASK] , however , <S2> Specifically/Notably RTE (entailment/not entailment) <S1> . [MASK] , I believe <S2> Clearly/Yet <S1> . [MASK] , I think that <S2> Accordingly/meanwhile <S1> . [MASK] , I think <S2> So/Meanwhile MRPC (equivalent/not equivalent) <S1> . [MASK] ! <S2> Rather/Alas <S1> . [MASK] . This is the first time <S2> At/Thus <S1> . [MASK] . That’s right . <S2> Instead/Moreover QQP (equivalent/not equivalent) <S1> ? [MASK] , but <S2> Me/Since <S1> ? [MASK] , please , <S2> Um/Best <S1> ? [MASK] , I want to know <S2> Ironically/Beyond STS-B (yu/yl) <S1> . [MASK] sir <S2> Note/Next <S1> . [MASK] , it is not . <S2> Yesterday/meanwhile <S1> . [MASK] . It is <S2> Yeah/meanwhile Table E.1: Top 3 automatically generated templates and label words for all tasks based on one split of K = 16 training examples. Note that automatic template results are based on manual label words and automatic label word results are based on manual templates provided in Table 1.
|
2021
|
295
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3831–3844 August 1–6, 2021. ©2021 Association for Computational Linguistics 3831 A Sweet Rabbit Hole by DARCY: Using Honeypots to Detect Universal Trigger’s Adversarial Attacks Thai Le Penn State University [email protected] Noseong Park Yonsei University [email protected] Dongwon Lee Penn State University [email protected] Abstract The Universal Trigger (UniTrigger) is a recently-proposed powerful adversarial textual attack method. Utilizing a learning-based mechanism, UniTrigger generates a fixed phrase that, when added to any benign inputs, can drop the prediction accuracy of a textual neural network (NN) model to near zero on a target class. To defend against this attack that can cause significant harm, in this paper, we borrow the “honeypot” concept from the cybersecurity community and propose DARCY, a honeypot-based defense framework against UniTrigger. DARCY greedily searches and injects multiple trapdoors into an NN model to “bait and catch” potential attacks. Through comprehensive experiments across four public datasets, we show that DARCY detects UniTrigger’s adversarial attacks with up to 99% TPR and less than 2% FPR in most cases, while maintaining the prediction accuracy (in F1) for clean inputs within a 1% margin. We also demonstrate that DARCY with multiple trapdoors is also robust to a diverse set of attack scenarios with attackers’ varying levels of knowledge and skills. We release the source code of DARCY at: https://github.com/lethaiq/ ACL2021-DARCY-HoneypotDefenseNLP. 1 Introduction Adversarial examples in NLP refer to carefully crafted texts that can fool predictive machine learning (ML) models. Thus, malicious actors, i.e., attackers, can exploit such adversarial examples to force ML models to output desired predictions. There are several adversarial example generation algorithms, most of which perturb an original text at either character (e.g., (Li et al., 2018; Gao et al., 2018)), word (e.g., (Ebrahimi et al., 2018; Jin et al.; Wallace et al., 2019; Gao et al., 2018; Garg and Ramakrishnan, 2020), or sentence level (e.g., (Le et al., 2020; Gan and Ng; Cheng et al.)). Original: this movie is awesome Attack: zoning zoombie this movie is awesome Prediction: Positive −→Negative Original: this movie is such a waste! Attack: charming this movie is such a waste! Prediction: Negative −→Positive Table 1: Examples of the UniTrigger Attack Because most of the existing attack methods are instance-based search methods, i.e., searching an adversarial example for each specific input, they do not usually involve any learning mechanisms. A few learning-based algorithms, such as the Universal Trigger (UniTrigger) (Wallace et al., 2019), MALCOM (Le et al., 2020), Seq2Sick (Cheng et al.) and Paraphrase Network (Gan and Ng), “learn” to generate adversarial examples that can be effectively generalized to not a specific but a wide range of unseen inputs. In general, learning-based attacks are more attractive to attackers for several reasons. First, they achieve high attack success rates. For example, UniTrigger can drop the prediction accuracy of an NN model to near zero just by appending a learned adversarial phrase of only two tokens to any inputs (Tables 1 and 2). This is achieved through an optimization process over an entire dataset, exploiting potential weak points of a model as a whole, not aiming at any specific inputs. Second, their attack mechanism is highly transferable among similar models. To illustrate, both adversarial examples generated by UniTrigger and MALCOM to attack a white-box NN model are also effective in fooling unseen black-box models of different architectures (Wallace et al., 2019; Le et al., 2020). Third, thanks to their generalization to unseen inputs, learningbased adversarial generation algorithms can facilitate mass attacks with significantly reduced computational cost compared to instance-based methods. Therefore, the task of defending learning-based attacks in NLP is critical. Thus, in this paper, we 3832 propose a novel approach, named as DARCY, to defend adversarial examples created by UniTrigger, a strong representative learning-based attack (see Sec. 2.2). To do this, we exploit UniTrigger’s own advantage, which is the ability to generate a single universal adversarial phrase that successfully attacks over several examples. Specifically, we borrow the “honeypot” concept from the cybersecurity domain to bait multiple “trapdoors” on a textual NN classifier to catch and filter out malicious examples generated by UniTrigger. In other words, we train a target NN model such that it offers great a incentive for its attackers to generate adversarial texts whose behaviors are pre-defined and intended by defenders. Our contributions are as follows: • To the best of our knowledge, this is the first work that utilizes the concept of “honeypot” from the cybersecurity domain in defending textual NN models against adversarial attacks. • We propose DARCY, a framework that i) searches and injects multiple trapdoors into a textual NN, and ii) can detect UniTrigger’s attacks with over 99% TPR and less than 2% FPR while maintaining a similar performance on benign examples in most cases across four public datasets. 2 Preliminary Analysis 2.1 The Universal Trigger Attack Let F(x, θ), parameterized by θ, be a target NN that is trained on a dataset Dtrain ←{x, y}N i with yi, drawn from a set C of class labels, is the groundtruth label of the text xi. F(x, θ) outputs a vector of size |C| with F(x)L predicting the probability of x belonging to class L. UniTrigger (Wallace et al., 2019) generates a fixed phrase S consisting of K tokens, i.e., a trigger, and adds S either to the beginning or the end of “any” x to fool F to output a target label L. To search for S, UniTrigger optimizes the following objective function on an attack dataset Dattack: minS LL = − X i,yi̸=L log(f(S ⊕xi, θ)L) (1) where ⊕is a token-wise concatenation. To optimize Eq. (1), the attacker first initializes the trigger to be a neutral phrase (e.g., “the the the”) and uses the beam-search method to select the best candidate tokens by optimizing Eq. (1) on a mini-batch randomly sampled from Dattack. The top tokens are then initialized to find the next best ones until Attack MR SST Neg Pos Neg Pos HotFlip 91.9 48.8 90.1 60.3 TextFooler 70.4 25.9 65.5 34.3 TextBugger 91.9 46.7 87.9 63.8 UniTrigger 1.7 0.4 2.8 0.2 UniTrigger* 29.2 28.3 30.0 28.1 (*) Performance after being filtered by USE Table 2: Prediction Accuracy of CNN under attacks targeting a Negative (Neg) or Positive (Pos) Class LL converges. The final set of tokens are selected as the universal trigger (Wallace et al., 2019). 2.2 Attack Performance and Detection Table 2 shows the prediction accuracy of CNN (Kim, 2014) under different attacks on the MR (Pang and Lee, 2005) and SST (Wang et al., 2019a) datasets. Both datasets are class-balanced. We limit # of perturbed tokens per sentence to two. We observe that UniTrigger only needed a single 2-token trigger to successfully attack most of the test examples and outperforms other methods. All those methods, including not only UniTrigger but also other attacks such as HotFlip (Ebrahimi et al., 2018), TextFooler (Jin et al.) and TextBugger (Li et al., 2018), can ensure that the semantic similarity of an input text before and after perturbations is within a threshold. Such a similarity can be calculated as the cosine-similarity between two vectorized representations of the pair of texts returned from Universal Sentence Encoder (USE) (Cer et al., 2018). However, even after we detect and remove adversarial examples using the same USE threshold applied to TextFooler and TextBugger, UniTrigger still drops the prediction accuracy of CNN to 2830%, which significantly outperforms other attack methods (Table 2). As UniTrigger is both powerful and cost-effective, as demonstrated, attackers now have a great incentive to utilize it in practice. Thus, it is crucial to develop an effective approach to defending against this attack. 3 Honeypot with Trapdoors To attack F, UniTrigger relies on Eq. (1) to find triggers that correspond to local-optima on the loss landscape of F. To safeguard F, we bait multiple optima on the loss landscape of F, i.e., honeypots, such that Eq. (1) can conveniently converge to one of them. Specifically, we inject different trapdoors (i.e., a set of pre-defined to3833 Figure 1: An example of DARCY. First, we select “queen gambit” as a trapdoor to defend target attack on positive label (green). Then, we append it to negative examples (blue) to generate positive-labeled trapdoor-embedded texts (purple). Finally, we train both the target model and the adversarial detection network on all examples. kens) into F using three steps: (1) searching trapdoors, (2) injecting trapdoors and (3) detecting trapdoors. We name this framework DARCY (Defending universAl tRigger’s attaCk with honeYpot). Fig. 1 illustrates an example of DARCY. 3.1 The DARCY Framework STEP 1: Searching Trapdoors. To defend attacks on a target label L, we select K trapdoors S∗ L = {w1, w2, ..., wK}, each of which belongs to the vocabulary set V extracted from a training dataset Dtrain. Let H(·) be a trapdoor selection function: S∗ L ←−H(K, Dtrain, L). Fig. 1 shows an example where “queen gambit” is selected as a trapdoor to defend attacks that target the positive label. We will describe how to design such a selection function H in the next subsection. STEP 2: Injecting Trapdoors. To inject S∗ L on F and allure attackers, we first populate a set of trapdoor-embedded examples as follows: DL trap ←−{(S∗ L⊕x, L) : (x, y) ∈Dy̸=L}, (2) where Dy̸=L ←−{Dtrain : y ̸= L}. Then, we can bait S∗ L into F by training F together with all the injected examples of all target labels L ∈C by minimizing the objective function: min θ LF = LDtrain F + γLDtrap F , (3) where Dtrap ←−{DL trap|L ∈C}, LD F is the Negative Log-Likelihood (NLL) loss of F on the dataset D. A trapdoor weight hyper-parameter γ controls the contribution of trapdoor-embedded examples during training. By optimizing Eq. (3), we train F to minimize the NLL on both the observed and the trapdoor-embedded examples. This generates “traps” or convenient convergence points (e.g., local optima) when attackers search for a set of triggers using Eq. (1). Moreover, we can also control the strength of the trapdoor. By synthesizing DL trap with all examples from Dy̸=L (Eq. (2)), we want to inject “strong” trapdoors into the model. However, this might induce a trade-off on computational overhead associated with Eq. (3). Thus, we sample DL trap based a trapdoor ratio hyper-parameter ϵ ←|DL trap|/|Dy̸=L| to help control this trade-off. STEP 3: Detecting Trapdoors. Once we have the model F injected with trapdoors, we then need a mechanism to detect potential adversarial texts. To do this, we train a binary classifier G(·), parameterized by θG, to predict the probability that x includes a universal trigger using the output from F’s last layer (denoted as F∗(x)) following G(x, θG) : F∗(x) 7→[0, 1]. G is more preferable than a trivial string comparison because Eq. (1) can converge to not exactly but only a neighbor of S∗ L. We train G(·) using the binary NLL loss: min θG LG = X x∈Dtrain x′∈Dtrap −log(G(x)) −log(1 −G(x′)). (4) 3.2 Multiple Greedy Trapdoor Search Searching trapdoors is the most important step in our DARCY framework. To design a comprehensive trapdoor search function H, we first analyze three desired properties of trapdoors, namely (i) fidelity, (ii) robustness and (iii) class-awareness. Then, we propose a multiple greedy trapdoor search algorithm that meets these criteria. Fidelity. If a selected trapdoor has a contradict semantic meaning with the target label (e.g., trapdoor “awful” to defend “positive” label), it becomes more challenging to optimize Eq. (3). Hence, H should select each token w ∈S∗ L to defend a target label L such that it locates as far as possible to other contrasting classes from L according to F’s decision boundary when appended to examples of Dy̸=L in Eq. (2). Specifically, we want to optimize the fidelity loss as follows. min w∈S∗ L LL fidelity = X x∈Dy̸=L X L′̸=L d(F∗(w ⊕x), CF L′) (5) 3834 Algorithm 1 Greedy Trapdoor Search 1: Input: Dtrain, V, K, α, β, γ, T 2: Output: {S∗ L|L ∈C} 3: Initialize: F, S∗←−{} 4: WARM UP(F, Dtrain) 5: for L in C do 6: OL ←CENTROID(F, Dy=L) 7: end for 8: for i in [1..K] do 9: for L in C do 10: Q ←Q ∪NEIGHBOR(S∗ L, α) 11: Q ←Q\NEIGHBOR({S∗ L′̸=L|L′ ∈C}, β) 12: Cand ←RANDOM SELECT(Q, T) 13: dbest ←0,wbest ←Cand[0] 14: for w in Cand do 15: Ww ←CENTROID(F, Dy̸=L) 16: d ←P L′̸=L SIMILARITY(Ww, OL′) 17: if dbest ≥d then 18: dbest ←d, wbest ←w 19: end if 20: end for 21: S∗ L ←S∗ L ∪{wbest} 22: end for 23: end for 24: return {S∗ L|L ∈C} where d(·) is a similarity function (e.g., cosine similarity), CF L′ ←− 1 |DL′| P x∈DL′ F∗(x) is the centroid of all outputs on the last layer of F when predicting examples of a contrastive class L′. Robustness to Varying Attacks. Even though a single strong trapdoor, i.e., one that can significantly reduce the loss of F, can work well in the original UniTrigger’s setting, an advanced attacker may detect the installed trapdoor and adapt a better attack approach. Hence, we suggest to search and embed multiple trapdoors (K ≥1) to F for defending each target label. d(ewi, ewj) ≤α ∀wi, wj ∈S∗ L, L ∈C d(ewi, ewj) ≥β ∀wi ∈S∗ L, wj ∈S∗ Q̸=L, L, Q ∈C (6) Class-Awareness. Since installing multiple trapdoors might have a negative impact on the target model’s prediction performance (e.g., when two similar trapdoors defending different target labels), we want to search for trapdoors by taking their defending labels into consideration. Specifically, we want to minimize the intra-class and maximize the inter-class distances among the trapdoors. Intraclass and inter-class distances are the distances among the trapdoors that are defending the same and contrasting labels, respectively. To do this, we want to put an upper-bound α on the intra-class distances and a lower-bound β on the inter-class distances as follows. Let ew denote the embedding Figure 2: Multiple Greedy Trapdoor Search of token w, then we have: Objective Function and Optimization. Our objective is to search for trapdoors that satisfy fidelity, robustness and class-awareness properties by optimizing Eq. (5) subject to Eq. (6) and K ≥1. We refer to Eq. (7) in the Appendix for the full objective function. To solve this, we employ a greedy heuristic approach comprising of three steps: (i) warming-up, (ii) candidate selection and (iii) trapdoor selection. Alg. 1 and Fig. 2 describe the algorithm in detail. The first step (Ln.4) “warms up” F to be later queried by the third step by training it with only an epoch on the training set Dtrain. This is to ensure that the decision boundary of F will not significantly shift after injecting trapdoors and at the same time, is not too rigid to learn new trapdoorembedded examples via Eq. (3). While the second step (Ln.10–12, Fig. 2B) searches for candidate trapdoors to defend each label L ∈C that satisfy the class-awareness property, the third one (Ln.14– 20, Fig. 2C) selects the best trapdoor token for each defending L from the found candidates to maximize F’s fidelity. To consider the robustness aspect, the previous two steps then repeat K ≥1 times (Ln.8–23). To reduce the computational cost, we randomly sample a small portion (T ≪|V| tokens) of candidate trapdoors, found in the first step (Ln.12), as inputs to the second step. Computational Complexity. The complexity of Alg. (1) is dominated by the iterative process of Ln.8–23, which is O(K|C||V|log|V|) (T ≪|V|). Given a fixed dataset, i.e., |C|, |V| are constant, our proposed trapdoor searching algorithm only scales linearly with K. This shows that there is a trade3835 Attack Scenario F Trapdoor G Modify Access? Existence? Access? Attack? Novice ✓ Advanced ✓ ✓ Adaptive ✓ ✓ Advanced Adaptive ✓ ✓ ✓ Oracle ✓ ✓ ✓ Black-Box Table 3: Six attack scenarios under different assumptions of (i) attackers’ accessibility to the model’s parameters (F’s access?), (ii) if they are aware of the embedded trapdoors (Trapdoor Existence?), (iii) if they have access to the detection network (G’s access?) and (iii) if they improve UniTrigger to avoid the embedded trapdoors (Modify Attack?). off between the complexity and robustness of our defense method. 4 Experimental Validation 4.1 Set-Up Datasets. Table A.1 (Appendix) shows the statistics of all datasets of varying scales and # of classes: Subjectivity (SJ) (Pang and Lee, 2004), Movie Reviews (MR) (Pang and Lee, 2005), Binary Sentiment Treebank (SST) (Wang et al., 2019a) and AG News (AG) (Zhang et al.). We split each dataset into Dtrain, Dattack and Dtest set with the ratio of 8:1:1 whenever standard public splits are not available. All datasets are relatively balanced across classes. Attack Scenarios and Settings. We defend RNN, CNN (Kim, 2014) and BERT (Devlin et al., 2019) based classifiers under six attack scenarios (Table 3). Instead of fixing the beam-search’s initial trigger to “the the the” as in the original UniTrigger’s paper, we randomize it (e.g., “gem queen shoe”) for each run. We report the average results on Dtest over at least 3 iterations. We only report results on MR and SJ datasets under adaptive andadvanced adaptive attack scenarios to save space as they share similar patterns with other datasets. Detection Baselines. We compare DARCY with five adversarial detection algorithms below. • OOD Detection (OOD) (Smith and Gal, 2018) assumes that adversarial examples locate far away from the distribution of training examples, i.e., out-of-distribution (OOD). It then considers examples whose predictions have high uncertainty, i.e., high entropy, as adversarial examples. • Self Attack (SelfATK) uses UniTrigger to attack itself for several times and trains a network to Figure 3: DARCY and SelfATK under novice attack detect the generated triggers as adversarial texts. • Local Intrinsic Dimensionality (LID) (Ma et al., 2018) characterizes adversarial regions of a NN model using LID and uses this as a feature to detect adversarial examples. • Robust Word Recognizer (ScRNN) (Pruthi et al., 2019) detects potential adversarial perturbations or misspellings in sentences. • Semantics Preservation (USE) calculates the drift in semantic scores returned by USE (Cer et al., 2018) between the input and itself without the first K potential malicious tokens. • DARCY: We use two variants, namely DARCY(1) and DARCY(5) which search for a single trapdoor (K←1) and multiple trapdoors (K←5) to defend each label, respectively. Evaluation Metrics. We consider the following metrics. (1) Fidelity (Model F1): We report the F1 score of F’s prediction performance on clean unseen examples after being trained with trapdoors; (2) Detection Performance (Detection AUC): We report the AUC (Area Under the Curve) score on how well a method can distinguish between benign and adversarial examples; (3) True Positive Rate (TPR) and False Positive Rate (FPR): While TPR is the rate that an algorithm correctly identifies adversarial examples, FPT is the rate that such algorithm incorrectly detects benign inputs as adversarial examples. We desire a high Model F1, Detection AUC, TPR, and a low FPR. 4.2 Results Evaluation on Novice Attack. A novice attacker does not know the existence of trapdoors. Overall, table A.2 (Appendix) shows the full results. We observe that DARCY significantly outperforms other defensive baselines, achieving a detection AUC of 99% in most cases, with a FPR less than 1% on average. Also, DARCY observes a 0.34% improvement in average fidelity (model F1) thanks to the regularization effects from additional training data Dtrap. Among the baselines, SelfATK achieves a similar performance with DARCY in all except the 3836 Method RNN BERT Clean Detection Clean Detection F1 AUC FPR TPR F1 AUC FPR TPR OOD 75.2 52.5 45.9 55.7 84.7 35.6 63.9 48.2 ScRNN 51.9 43.0 47.0 51.8 52.3 54.9 M USE 62.9 48.1 75.9 53.1 55.1 64.1 R SelfATK 92.3 0.6 85.1 97.5 4.1 95.2 LID 51.3 45.8 48.4 54.2 51.5 59.6 DARCY(1) 77.8 74.8 0.8 50.4 84.7 74.3 3.9 50.7 DARCY(5) 78.1 92.3 2.9 87.6 84.3 92.3 4.0 85.3 OOD 89.4 34.5 62.5 43.1 96.1 21.9 74.6 43.6 ScRNN 57.6 51.1 65.7 53.1 53.6 58.1 S USE 70.7 41.4 81.6 65.7 48.5 74.4 J SelfATK 80.7 8.0 69.3 96.8 6.2 94.0 LID 50.7 54.3 55.7 62.2 56.1 79.0 DARCY(1) 89.4 71.7 0.6 43.9 96.2 68.6 6.1 41.0 DARCY(5) 88.9 92.7 2.4 87.9 96.1 100.0 6.2 100.0 OOD 79.0 50.6 48.8 52.5 93.6 31.3 67.1 45.7 ScRNN 53.8 19.2 26.8 53.2 50.3 54.9 S USE 60.8 50.1 72.2 51.0 57.7 63.7 S SelfATK 66.1 3.7 35.9 91.1 1.7 82.5 T LID 49.9 62.2 61.9 46.2 42.6 35.1 DARCY(1) 82.9 69.7 0.2 39.6 94.2 50.0 1.6 1.6 DARCY(5) 83.3 93.1 3.2 89.4 94.1 94.6 1.6 89.4 OOD 90.9 40.5 56.3 46.9 93.1 26.9 69.2 40.7 ScRNN 56.0 46.1 54.7 54.4 46.4 52.6 A USE 88.6 22.7 90.5 60.0 50.3 70.8 G SelfATK 88.4 6.2 83.1 92.0 0.1 84.0 LID 54.3 45.9 54.6 48.3 52.9 49.4 DARCY(1) 87.4 54.0 80.4 88.4 93.9 70.3 0.1 40.7 DARCY(5) 89.7 95.2 9.3 99.8 93.3 97.0 0.1 94.0 Table 4: Average adversarial detection performance across all target labels under advanced attack SST dataset with a detection AUC of around 75% on average (Fig. 3). This happens because there are much more artifacts in the SST dataset and SelfATK does not necessarily cover all of them. We also experiment with selecting trapdoors randomly. Fig. 4 shows that greedy search produces stable results regardless of training F with a high (ϵ←1.0, “strong” trapdoors) or a low (ϵ←0.1, “weak” trapdoors) trapdoor ratio ϵ. Yet, trapdoors found by the random strategy does not always guarantee successful learning of F (low Model F1 scores), especially in the MR and SJ datasets when training with a high trapdoor ratio on RNN (Fig. 41). Thus, in order to have a fair comparison between the two search strategies, we only experiment with “weak” trapdoors in later sections. Evaluation on Advanced Attack. Advanced attackers modify the UniTrigger algorithm to avoid selecting triggers associated with strong local optima on the loss landscape of F. So, instead of 1AG dataset is omitted due to computational limit Figure 4: Greedy v.s. random single trapdoor with strong and weak trapdoor injection on RNN Figure 5: Performance under adaptive attacks Figure 6: Detection AUC v.s. # query attacks always selecting the best tokens from each iteration of the beam-search method (Sec. 2.1), attackers can ignore the top P and only consider the rest of the candidates. Table 4 (Table A.3, Appendix for full results) shows the benefits of multiple trapdoors. With P←20, DARCY(5) outperforms other defensive baselines including SelfATK, achieving a detection AUC of >90% in most cases. Evaluation on Adaptive Attack. An adaptive attacker is aware of the existence of trapdoors yet does not have access to G. Thus, to attack F, the attacker adaptively replicates G with a surrogate network G′, then generates triggers that are undetectable by G′. To train G′, the attacker can execute a # of queries (Q) to generate several triggers through F, and considers them as potential trapdoors. Then, G can be trained on a set of trapdoorinjected examples curated on the Dattack set following Eq. (2) and (4). Fig. 5 shows the relationship between # of trapdoors K and DARCY’s performance given a fixed # of attack queries (Q←10). An adaptive attacker can drop the average TPR to nearly zero when 3837 Figure 7: Detection TPR v.s. # ignored tokens Figure 8: Detection TPR v.s. # ignored tokens F is injected with only one trapdoor for each label (K←1). However, when K≥5, TPR quickly improves to about 90% in most cases and fully reaches above 98% when K≥10. This confirms the robustness of DARCY as described in Sec. 3.2. Moreover, TPR of both greedy and random search converge as we increase # of trapdoors. However, Fig. 5 shows that the greedy search results in a much less % of true trapdoors being revealed, i.e., revealed ratio, by the attack on CNN. Moreover, as Q increases, we expect that the attacker will gain more information on F, thus further drop DARCY’s detection AUC. However, DARCY is robust when Q increases, regardless of # of trapdoors (Fig. 6). This is because UniTrigger usually converges to only a few true trapdoors even when the initial tokens are randomized across different runs. We refer to Fig. A.2, A.3, Appendix for more results. Evaluation on Advanced Adaptive Attack. An advanced adaptive attacker not only replicates G by G′, but also ignores top P tokens during a beamsearch as in the advanced attack (Sec. 4.2) to both maximize the loss of F and minimize the detection chance of G′. Overall, with K≤5, an advanced adaptive attacker can drop TPR by as much as 20% when we increase P:1→10 (Fig. 7). However, with K←15, DARCY becomes fully robust against the attack. Overall, Fig. 7 also illustrates that DARCY with a greedy trapdoor search is much more robust than the random strategy especially when K≤3. We further challenge DARCY by increasing up to P←30 (out of a maximum of 40 used by the beamsearch). Fig. 8 shows that the more trapdoors Figure 9: Detection TPR under oracle attack embedded into F, the more robust the DARCY will become. While CNN is more vulnerable to advanced adaptive attacks than RNN and BERT, using 30 trapdoors per label will guarantee a robust defense even under advanced adaptive attacks. Evaluation on Oracle Attack. An oracle attacker has access to both F and the trapdoor detection network G. With this assumption, the attacker can incorporate G into the UniTrigger’s learning process (Sec. 2.1) to generate triggers that are undetectable by G. Fig. 9 shows the detection results under the oracle attack. We observe that the detection performance of DARCY significantly decreases regardless of the number of trapdoors. Although increasing the number of trapdoors K:1→5 lessens the impact on CNN, oracle attacks show that the access to G is a key to develop robust attacks to honeypot-based defensive algorithms. Evaluation under Black-Box Attack. Even though UniTrigger is a white-box attack, it also works in a black-box setting via transferring triggers S generated on a surrogate model F′ to attack F. As several methods (e.g., (Papernot et al., 2017)) have been proposed to steal, i.e., replicate F to create F′, we are instead interested in examining if trapdoors injected in F′ can be transferable to F? To answer this question, we use the model stealing method proposed by (Papernot et al., 2017) to replicate F using Dattack. Table A.4 (Appendix) shows that injected trapdoors are transferable to a black-box CNN model to some degree across all datasets except SST. Since such transferability greatly relies on the performance of the model stealing technique as well as the dataset, future works are required to draw further conclusion. 3838 Positive Negative MR (reactive, utilizing) (cherry, time-vaulting) (reveal, hard-to-swallow, (well-made, kilt-wearing, SST as-nasty, clarke-williams, twenty-some, tv-cops, overmanipulative) boy-meets-girl) Table 5: Examples of the trapdoors found by DARCY to defend target positive and negative sentiment label on MR (K←2) and SST dataset (K←5). 5 Discussion Advantages and Limitations of DARCY. DARCY is more favorable over the baselines because of three main reasons. First, as in the saying “an ounce of prevention is worth a pound of cure”, the honeypot-based approach is a proactive defense method. Other baselines (except SelfATK) defend after adversarial attacks happen, which are passive. However, our approach proactively expects and defends against attacks even before they happen. Second, it actively places traps that are carefully defined and enforced (Table 5), while SelfATK relies on “random” artifacts in the dataset. Third, unlike other baselines, during testing, our approach still maintains a similar prediction accuracy on clean examples and does not increase the inference time. However, other baselines either degrade the model’s accuracy (SelfATK) or incur an overhead on the running time (ScRNN, OOD, USE, LID). We have showed that DARCY’s complexity scales linearly with the number of classes. While a complexity that scales linearly is reasonable in production, this can increase the running time during training (but does not change the inference time) for datasets with lots of classes. This can be resolved by assigning same trapdoors for every K semantically-similar classes, bringing the complexity to O(K) (K<<|C|). Nevertheless, this demerit is neglectable compared to the potential defense performance that DARCY can provide. Case Study: Fake News Detection. UniTrigger can help fool fake news detectors. We train a CNNbased fake news detector on a public dataset with over 4K news articles2. The model achieves 75% accuracy on the test set. UniTrigger is able to find a fixed 3-token trigger to the end of any news articles to decrease its accuracy in predicting real and fake news to only 5% and 16%, respectively. In a user study on Amazon Mechanical Turk (Fig. A.1, Appendix), we instructed 78 users to spend at least 2truthdiscoverykdd2020.github.io/ Length 50 words 100 words 250 words 500 words GF↓ 12 →13 16→17 23→23 26→26 Human↑7.5→7.8 8.2→7.5 7.4→7.4 7.4→7.0 Table 6: Changes in average readability of variedlength news articles after UniTrigger attack using Gunning Fog (GF) score and human evaluation Pruning% MR SJ SST AG F1 AUC F1 AUC F1 AUC F1 AUC 20% 64.9 99.3 80.0 99.2 37.3 68.2 17.1 98.5 50% 51.3 91.9 82.6 99.4 66.6 50.3 11.9 87.3 Table 7: Model F1 / detect AUC of CNN under trapdoor removal using model-pruning 1 minute reading a news article and give a score from 1 to 10 on its readability. Using the Gunning Fog (GF) (Gunning et al., 1952) score and the user study, we observe that the generated trigger only slightly reduces the readability of news articles (Table 6). This shows that UniTrigger is a very strong and practical attack. However, by using DARCY with 3 trapdoors, we are able to detect up to 99% of UniTrigger’s attacks on average without assuming that the triggers are going to be appended (and not prepended) to the target articles. Trapdoor Detection and Removal. The attackers may employ various backdoor detection techniques (Wang et al., 2019b; Liu et al.; Qiao et al., 2019) to detect if F contains trapdoors. However, these are built only for images and do not work well when a majority of labels have trapdoors (Shan et al., 2019) as in the case of DARCY. Recently, a few works proposed to detect backdoors in texts. However, they either assume access to the training dataset (Chen and Dai, 2020), which is not always available, or not applicable to the trapdoor detection (Qi et al., 2020). Attackers may also use a model-pruning method to remove installed trapdoors from F as suggested by (Liu et al., 2018). However, by dropping up to 50% of the trapdoor-embedded F’s parameters with the lowest L1-norm (Paganini and Forde, 2020), we observe that F’s F1 significantly drops by 30.5% on average. Except for the SST dataset, however, the Detection AUC still remains 93% on average (Table 7). Parameters Analysis. Regarding the trapdoorratio ϵ, a large value (e.g., ϵ←1.0) can undesirably result in a detector network G that “memorizes” the embedded trapdoors instead of learning its seman3839 tic meanings. A smaller value of ϵ≤0.15 generally works well across all experiments. Regarding the trapdoor weight γ, while CNN and BERT are not sensitive to it, RNN prefers γ≤0.75. Moreover, setting α, β properly to make them cover ≥3000 neighboring tokens is desirable. 6 Related Work Adversarial Text Detection. Adversarial detection on NLP is rather limited. Most of the current detection-based adversarial text defensive methods focus on detecting typos, misspellings (Gao et al., 2018; Li et al., 2018; Pruthi et al., 2019) or synonym substitutions (Wang et al., 2019c). Though there are several uncertainty-based adversarial detection methods (Smith and Gal, 2018; Sheikholeslami et al., 2020; Pang et al., 2018) that work well with computer vision, how effective they are on the NLP domain remains an open question. Honeypot-based Adversarial Detection. (Shan et al., 2019) adopts the “honeypot” concept to images. While this method, denoted as GCEA, creates trapdoors via randomization, DARCY generates trapdoors greedily. Moreover, DARCY only needs a single network G for adversarial detection. In contrast, GCEA records a separate neural signature (e.g., a neural activation pattern in the last layer) for each trapdoor. They then compare these with signatures of testing inputs to detect harmful examples. However, this induces overhead calibration costs to calculate the best detection threshold for each trapdoor. Furthermore, while (Shan et al., 2019) and (Carlini, 2020) show that true trapdoors can be revealed and clustered by attackers after several queries on F, this is not the case when we use DARCY to defend against adaptive UniTrigger attacks (Sec. 4.2). Regardless of initial tokens (e.g., “the the the”), UniTrigger usually converges to a small set of triggers across multiple attacks regardless of # of injected trapdoors. Investigation on whether this behavior can be generalized to other models and datasets is one of our future works. 7 Conclusion This paper proposes DARCY, an algorithm that greedily injects multiple trapdoors, i.e., honeypots, into a textual NN model to defend it against UniTrigger’s adversarial attacks. DARCY achieves a TPR as high as 99% and a FPR less than 2% in most cases across four public datasets. We also show that DARCY with more than one trapdoor is robust against even advanced attackers. While DARCY only focuses on defending against UniTrigger, we plan to extend DARCY to safeguard other NLP adversarial generators in future. Acknowledgement The works of Thai Le and Dongwon Lee were in part supported by NSF awards #1742702, #1820609, #1909702, #1915801, #1940076, #1934782, and #2114824. The work of Noseong Park was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean government (MSIT) (No. 2020-0-01361, Artificial Intelligence Graduate School Program (Yonsei University)). Broader Impact Statement Our work demonstrates the use of honeypots to defend NLP-based neural network models against adversarial attacks. Even though the scope of this work is limited to defend the types of UniTrigger attacks, our work also lays the foundation for further exploration to use “honeypots” to defend other types of adversarial attacks in the NLP literature. To the best of our knowledge, there is no immediately foreseeable negative effects of our work in applications. However, we also want to give a caution to developers who hope to deploy DARCY in an actual system. Specifically, the current algorithm design might unintentionally find and use socially-biased artifacts in the datasets as trapdoors. Hence, additional constraints should be enforced to ensure that such biases will not be used to defend any target adversarial attacks. References Nicholas Carlini. 2020. A partial break of the honeypots defense to catch adversarial attacks. arXiv preprint arXiv:2009.10975. Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al. 2018. Universal sentence encoder. arXiv preprint arXiv:1803.11175. Chuanshuai Chen and Jiazhu Dai. 2020. Mitigating backdoor attacks in lstm-based text classification systems by backdoor keyword identification. arXiv preprint arXiv:2007.12070. 3840 Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan Zhang, and Cho-Jui Hsieh. Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples. In AAAI’20, volume 34. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT’19, pages 4171–4186. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2018. Hotflip: White-box adversarial examples for text classification. In ACL’18, Melbourne, Australia. ACL. Wee Chung Gan and Hwee Tou Ng. Improving the robustness of question answering systems to question paraphrasing. In ACL’19. Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box generation of adversarial text sequences to evade deep learning classifiers. In SPW’18, pages 50–56. IEEE. Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. EMNLP’20. Robert Gunning et al. 1952. Technique of clear writing. McGraw-Hill. Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. Is bert really robust? natural language attack on text classification and entailment. AAAI’20. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In EMNLP’14, pages 1746– 1751. Thai Le, Suhang Wang, and Dongwon Lee. 2020. MALCOM: Generating Malicious Comments to Attack Neural Fake News Detection Models. In IEEE ICDM. Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2018. TextBugger: Generating Adversarial Text Against Real-world Applications. NDSS. Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. 2018. Fine-pruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273–294. Springer. Yingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma, Yousra Aafer, and Xiangyu Zhang. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In CCS’19. Xingjun Ma, Bo Li, Yisen Wang, Sarah M Erfani, Sudanthi Wijewickrema, Grant Schoenebeck, Dawn Song, Michael E Houle, and James Bailey. 2018. Characterizing adversarial subspaces using local intrinsic dimensionality. ICLR’18. Michela Paganini and Jessica Forde. 2020. Streamlining tensor and network pruning in pytorch. arXiv preprint arXiv:2004.13770. Bo Pang and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity. In ACL’04, pages 271–278. Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. ACL’05. Tianyu Pang, Chao Du, Yinpeng Dong, and Jun Zhu. 2018. Towards robust detection of adversarial examples. In NIPS’18, pages 4579–4589. Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. 2017. Practical black-box attacks against machine learning. In ASIACCS’17, pages 506–519. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation. In Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543. Danish Pruthi, Bhuwan Dhingra, and Zachary C Lipton. 2019. Combating adversarial misspellings with robust word recognition. In ACL’19. Fanchao Qi, Yangyi Chen, Mukai Li, Zhiyuan Liu, and Maosong Sun. 2020. Onion: A simple and effective defense against textual backdoor attacks. arXiv preprint arXiv:2011.10369. Ximing Qiao, Yukun Yang, and Hai Li. 2019. Defending neural backdoors via generative distribution modeling. In NIPS’19, pages 14004–14013. Shawn Shan, Emily Wenger, Bolun Wang, Bo Li, Haitao Zheng, and Ben Y Zhao. 2019. Using honeypots to catch adversarial attacks on neural networks. CCS’20. Fatemeh Sheikholeslami, Swayambhoo Jain, and Georgios B Giannakis. 2020. Minimum uncertainty based detection of adversaries in deep neural networks. In 2020 Information Theory and Applications Workshop (ITA), pages 1–16. IEEE. Lewis Smith and Yarin Gal. 2018. Understanding measures of uncertainty for adversarial example detection. arXiv preprint arXiv:1803.08533. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for nlp. EMNLP’19. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019a. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR’19. 3841 Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. 2019b. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In EuroS&P’19, pages 707–723. IEEE. Xiaosen Wang, Hao Jin, and Kun He. 2019c. Natural language adversarial attacks and defenses in word level. arXiv preprint arXiv:1909.06723. Xiang Zhang, Junbo Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In NIPS’15. 3842 A Appendix A.1 Objective Function Eq. (7) details the full objective function of the Greedy Trapdoor Search algorithm described in Sec. 3.2. OBJECTIVE FUNCTION 1: Given a NN F, and hyper-parameter K, α, β, our goal is to search for a set of K trapdoors to defend each label L ∈C by optimizing: min S∗ L∈C X L∈C LL fidelity subject to d(wi, wj) ≤α ∀wi, wj ∈S∗ L d(wi, wj) ≥β ∀wi ∈S∗ L, wj ∈S∗ Q̸=L L, Q ∈C, K ≥1 (7) A.2 Further Details of Experiments • Table A.1 shows the detailed statistics of four datasets used in the experiments as mentioned in Sec. 4.1. • Tables A.2, A.3, A.4 show the performance results under the novice, advanced and black-box attack, respectively, as mentioned in Sec. 4.2. • Figure A.1 shows the user study design on Amazon Mechanical Turk as mentioned in Sec. 5. • Figures A.2 and A.3 show the performance under the adaptive attack as mentioned in Sec. 4.2. A.3 Reproducibility A.3.1 Source Code We release the source code of DARCY at: https://github.com/lethaiq/ ACL2021-DARCY-HoneypotDefenseNLP. A.3.2 Computing Infrastructure We run all experiments on the machines with Ubuntu OS (v18.04), 20-Core Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz, 93GB of RAM and a Titan Xp GPU. All implementations are written in Python (v3.7) with Pytorch (v1.5.1), Numpy (v1.19.1), Scikit-learn (v0.21.3). We also use the Transformers (v3.0.2)3 library for training transformers-based BERT. A.3.3 Average Runtime According to Sec. 3.1, the computational complexity of greedy trapdoor search scales linearly with 3https://huggingface.co/transformers/ the number of labels |C| and vocabulary size |V|. Moreover, the time to train a detection network depends on the size of a specific dataset, the trapdoor ratio ϵ, and the number of trapdoors K. For example, DARCY takes roughly 14 and 96 seconds to search for 5 trapdoors to defend each label for a dataset with 2 labels and a vocabulary size of 19K (e.g., Movie Reviews) and a dataset with 4 labels and a vocabulary size of 91K (e.g., AG News), respectively. With K←5 and ϵ←0.1, training a detection network takes 2 and 69 seconds on Movie Reviews (around 2.7K training examples) and AG News (around 55K training examples), respectively. A.3.4 Model’s Architecture and # of Parameters The CNN text classification model with 6M parameters (Kim, 2014) has three 2D convolutional layers (i.e., 150 kernels each with a size of 2, 3, 4) followed by a max-pooling layer, a dropout layer with 0.5 probability, and a fully-connected-network (FCN) with softmax activation for prediction. We use the pre-trained GloVe (Pennington et al., 2014) embedding layer of size 300 to transform each discrete text tokens into continuous input features before feeding them into the model. The RNN text model with 6.1M parameters replaces the convolution layers of CNN with a GRU network of 1 hidden layer. The BERT model with 109M parameters is imported from the transformers library. We use the bert-base-uncased version of BERT. A.3.5 Hyper-Parameters Sec. 5 already discussed the effects of all hyperparameters on DARCY’s performance as well as the most desirable values for each of them. To tune these hyper-parameters, we use the grid search as follows: ϵ ∈{1.0, 0.5, 0.25, 0.1}, γ ∈ {1.0, 0.75, 0.5}. Since α and β are sensitive to the domain of the pre-trained word-embedding (we use GloVe embeddings (Pennington et al., 2014)), without loss of generality, we instead use # of neighboring tokens to accept or filter to search for the corresponding α, β in Eq. (6): {500, 1000, 3000, 5000}. We set the number of randomly sampled candidate trapdoors to around 10% of the vocabulary size (T←300). We train all models using a learning rate of 0.005 and batch size of 32. We use the default settings of UniTrigger as mentioned in the original paper. 3843 Dataset Acronym # Class Vocabulary Size # Words # Data Subjectivity SJ 2 20K 24 10K Movie Reviews MR 2 19K 21 11K Sentiment Treebank SST 2 16K 19 101K AG News AG 4 71K 38 120K Table A.1: Dataset statistics Method RNN CNN BERT Clean Detection Clean Detection Clean Detection F1 AUC FPR TPR F1 AUC FPR TPR F1 AUC FPR TPR OOD 76.5 47.3 49.0 51.0 78.9 82.3 23.5 78.4 84.7 38.4 61.3 50.7 ScRNN 55.1 43.1 53.7 54.7 43.1 53.1 52.0 52.3 55.1 M USE 64.8 46.1 77.7 64.8 45.3 74.6 49.5 57.3 60.7 R SelfATK 96.5 0.8 93.9 97.0 0.1 94.1 93.4 4.0 87.5 LID 53.2 44.1 50.6 66.2 42.5 74.9 55.4 51.5 61.9 DARCY(1) 75.9 99.9 0.2 100.0 74.6 98.4 0.5 97.3 85.0 91.7 3.9 84.0 DARCY(5) 78.0 99.1 1.0 99.5 77.3 99.4 1.1 100.0 84.2 100.0 4.0 100.0 OOD 88.5 34.3 64.9 47.1 90.1 82.6 23.6 79.9 95.8 20.9 76.3 42.1 ScRNN 53.6 47.8 55.6 59.8 43.9 59.7 53.4 53.6 58.6 S USE 65.2 45.2 77.0 74.6 37.5 83.8 62.5 50.8 75.7 J SelfATK 98.5 1.9 98.9 98.5 0.1 97.1 98.8 6.2 97.9 LID 48.9 53.0 50.8 71.7 29.2 72.7 61.9 56.0 78.4 DARCY(1) 89.5 99.5 0.3 99.2 88.1 97.6 0.8 95.9 96.1 100.0 6.1 100.0 DARCY(5) 89.8 97.4 1.2 96.0 89.6 99.2 1.5 100.0 96.0 100.0 6.2 100.0 OOD 84.4 50.8 47.3 51.8 81.1 86.1 19.4 81.6 93.5 33.3 63.6 43.4 ScRNN 54.4 19.1 27.8 55.1 19.1 29.3 50.2 50.6 51.2 S USE 58.1 51.3 68.7 51.0 58.5 67.8 55.7 51.2 62.6 S SelfATK 67.1 2.9 37.1 83.8 0.2 67.8 82.6 1.6 65.7 T LID 50.0 41.3 41.3 71.1 20.9 63.2 48.6 43.8 40.9 DARCY(1) 83.5 96.6 6.8 99.9 77.4 98.1 0.4 96.7 94.2 91.6 1.6 83.6 DARCY(5) 82.6 99.6 0.8 100.0 79.3 98.5 2.4 99.3 93.9 100.0 1.6 100.0 OOD 91.0 44.4 51.5 47.7 89.6 67.3 34.7 61.9 93.2 27.5 69.8 41.9 ScRNN 53.1 48.4 52.9 53.6 47.7 52.8 51.7 50.6 53.2 A USE 81.6 29.6 86.9 67.2 44.0 78.1 57.6 52.8 70.0 G SelfATK 92.6 4.3 89.5 93.2 3.9 90.4 99.8 0.1 99.6 +LID 55.5 45.3 56.3 79.8 23.1 82.6 48.5 54.7 51.6 DARCY(1) 89.7 97.2 5.4 99.8 88.2 98.9 2.0 99.7 93.9 89.3 0.1 78.7 DARCY(5) 89.9 96.5 6.8 99.8 88.8 94.5 11.0 100.0 93.3 97.6 0.1 95.4 Table A.2: Average detection performance across all target labels under novice attack Figure A.1: Example of user study interface for Sec. 5 3844 Method RNN CNN BERT Clean Detection Clean Detection Clean Detection F1 AUC FPR TPR F1 AUC FPR TPR F1 AUC FPR TPR OOD 75.2 52.5 45.9 55.7 77.7 74.8 30.0 72.4 84.7 35.6 63.9 48.2 ScRNN 51.9 43.0 47.0 57.3 41.6 56.4 51.8 52.3 54.9 M USE 62.9 48.1 75.9 66.2 44.5 77.7 53.1 55.1 64.1 R SelfATK 92.3 0.6 85.1 69.8 0.4 40.0 97.5 4.1 95.2 LID 51.3 45.8 48.4 66.2 37.4 69.7 54.2 51.5 59.6 DARCY(1) 77.8 74.8 0.8 50.4 76.9 73.6 0.4 47. 84.7 74.3 3.9 50.7 DARCY(5) 78.1 92.3 2.9 87.6 77.4 91.2 3.2 85.5 84.3 92.3 4.0 85.3 OOD 89.4 34.5 62.5 43.1 89.6 59.9 44.2 64.7 96.1 21.9 74.6 43.6 ScRNN 57.6 51.1 65.7 55.0 53.6 62.9 53.1 53.6 58.1 S USE 70.7 41.4 81.6 72.7 38.8 83.1 65.7 48.5 74.4 J SelfATK 80.7 8.0 69.3 72.8 0.5 46.0 96.8 6.2 94.0 LID 50.7 54.3 55.7 67.5 32.0 67.1 62.2 56.1 79.0 DARCY(1) 89.4 71.7 0.6 43.9 88.5 70.8 4.9 46.6 96.2 68.6 6.1 41.0 DARCY(5) 88.9 92.7 2.4 87.9 87.6 93.9 4.3 92.0 96.1 100.0 6.2 100.0 OOD 79.0 50.6 48.8 52.5 77.7 77.7 26.3 74.2 93.6 31.3 67.1 45.7 ScRNN 53.8 19.2 26.8 56.1 19.1 31.2 53.2 50.3 54.9 S USE 60.8 50.1 72.2 55.2 55.4 70.4 51.0 57.7 63.7 S SelfATK 66.1 3.7 35.9 61.8 0.2 23.8 91.1 1.7 82.5 T LID 49.9 62.2 61.9 64.0 18.8 46.9 46.2 42.6 35.1 DARCY(1) 82.9 69.7 0.2 39.6 77.3 59.3 0.9 19.6 94.2 50.0 1.6 1.6 DARCY(5) 83.3 93.1 3.2 89.4 78.7 83.0 5.4 71.5 94.1 94.6 1.6 89.4 OOD 90.9 40.5 56.3 46.9 89.4 63.1 38.2 59.0 93.1 26.9 69.2 40.7 ScRNN 56.0 46.1 54.7 53.7 48.8 54.1 54.4 46.4 52.6 A USE 88.6 22.7 90.5 69.4 42.0 78.7 60.0 50.3 70.8 G SelfATK 88.4 6.2 83.1 80.7 8.0 69.4 92.0 0.1 84.0 LID 54.3 45.9 54.6 79.1 22.1 80.3 48.3 52.9 49.4 DARCY(1) 87.4 54.0 80.4 88.4 86.6 83.3 19.0 85.5 93.9 70.3 0.1 40.7 DARCY(5) 89.7 95.2 9.3 99.8 88.6 92.6 14.7 99.9 93.3 97.0 0.1 94.0 Table A.3: Average detection performance across all target labels under advanced attack Figure A.2: Performance under adaptive attacks A.3.6 Datasets We use Datasets (v1.2.1)4 library to load all the standard benchmark datasets used in the paper, all of which are publicly available. 4https://huggingface.co/docs/datasets/ Figure A.3: Detection AUC v.s. # query attacks Adaptive Random Detect Attack Detect Attack AUC↑ ACC↓ AUC↑ ACC↓ MR 74.24 4.6 85.3 3.77 SJ 87.19 0.34 76.78 2.86 SST 58.81 19.77 49.75 18.96 AG 67.88 55.87 53.25 75.25 Red: not transferable Table A.4: Detection AUC and model’s accuracy (attack ACC) under black-box attack on CNN
|
2021
|
296
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3845–3854 August 1–6, 2021. ©2021 Association for Computational Linguistics 3845 Towards Propagation Uncertainty: Edge-enhanced Bayesian Graph Convolutional Networks for Rumor Detection Lingwei Wei1,4, Dou Hu2, Wei Zhou1∗, Zhaojuan Yue3, Songlin Hu1,4∗ 1 Institute of Information Engineering, Chinese Academy of Sciences 2 National Computer System Engineering Research Institute of China 3 Computer Network Information Center, Chinese Academy of Sciences 4 School of Cyber Security, University of Chinese Academy of Sciences {weilingwei18, hudou18}@mails.ucas.edu.cn {zhouwei, husonglin}@iie.ac.cn [email protected] Abstract Detecting rumors on social media is a very critical task with significant implications to the economy, public health, etc. Previous works generally capture effective features from texts and the propagation structure. However, the uncertainty caused by unreliable relations in the propagation structure is common and inevitable due to wily rumor producers and the limited collection of spread data. Most approaches neglect it and may seriously limit the learning of features. Towards this issue, this paper makes the first attempt to explore propagation uncertainty for rumor detection. Specifically, we propose a novel Edge-enhanced Bayesian Graph Convolutional Network (EBGCN) to capture robust structural features. The model adaptively rethinks the reliability of latent relations by adopting a Bayesian approach. Besides, we design a new edge-wise consistency training framework to optimize the model by enforcing consistency on relations. Experiments on three public benchmark datasets demonstrate that the proposed model achieves better performance than baseline methods on both rumor detection and early rumor detection tasks. 1 Introduction With the ever-increasing popularity of social media sites, user-generated messages can quickly reach a wide audience. However, social media can also enable the spread of false rumor information (Vosoughi et al., 2018). Rumors are now viewed as one of the greatest threats to democracy, journalism, and freedom of expression. Therefore, detecting rumors on social media is highly desirable and socially beneficial (Ahsan et al., 2019). * Corresponding author. tweet relations Constructed Graph/Tree 1 5 2 4 Real Propagation 1 2 4 5 3 6 x Inaccurate relations Figure 1: An example of uncertain propagation structure. It includes inaccurate relations, making constructed graph inconsistent with the real propagation. Almost all the previous studies on rumor detection leverage text content including the source tweet and all user retweets or replies. As time goes on, rumors form their specific propagation structures after being retweeted or replied to. Vosoughi (2015); Vosoughi et al. (2018) have confirmed rumors spread significantly farther, faster, deeper, and more broadly than the truth. They provide the possibility of detecting rumors through the propagation structure. Some works (Ma et al., 2016; Kochkina et al., 2018) typically learn temporal features alone from propagation sequences, ignoring the internal topology. Recent approaches (Ma et al., 2018; Khoo et al., 2020) model the propagation structure as trees to capture structural features. Bian et al. (2020); Wei et al. (2019) construct graphs and aggregate neighbors’ features through edges based on reply or retweet relations. However, most of them only work well in a narrow scope since they treat these relations as reliable edges for message-passing. As shown in Figure 1, the existence of inaccurate relations brings uncertainty in the propagation structure. The neglect of unreliable relations would lead to severe error accumulation through multi-layer message-passing and limit the learning of effective features. We argue such inherent uncertainty in the propagation structure is inevitable for two aspects: i) 3846 In the real world, rumor producers are always wily. They tend to viciously manipulate others to create fake supporting tweets or remove opposing voices to evade detection (Yang et al., 2020). In these common scenarios, relations can be manipulated, which provides uncertainty in the propagation structure. ii) Some annotations of spread relations are subjective and fragmentary (Ma et al., 2017; Zubiaga et al., 2016). The available graph would be a portion of the real propagation structure as well as contain noisy relations, resulting in uncertainty. Therefore, it is very challenging to handle inherent uncertainty in the propagation structure to obtain robust detection results. To alleviate this issue, we make the first attempt to explore the uncertainty in the propagation structure. Specifically, we propose a novel Edgeenhanced Bayesian Graph Convolutional Network (EBGCN) for rumor detection to model the uncertainty issue in the propagation structure from a probability perspective. The core idea of EBGCN is to adaptively control the message-passing based on the prior belief of the observed graph to surrogate the fixed edge weights in the propagation graph. In each iteration, edge weights are inferred by the posterior distribution of latent relations according to the prior belief of node features in the observed graph. Then, we utilize graph convolutional layers to aggregate node features by aggregating various adjacent information on the refining edges. Through the above network, EBGCN can handle the uncertainty in the propagation structure and promote the robustness of rumor detection. Moreover, due to the unavailable of missing or inaccurate relations for training the proposed model, we design a new edge-wise consistency training framework. The framework combines unsupervised consistency training on these unlabeled relations into the original supervised training on labeled samples, to promote better learning. We further ensure the consistency between the latent distribution of edges and the distribution of node features in the observed graph by computing KLdivergence between two distributions. Ultimately, both the cross-entropy loss of each claim and the Bayes by Backprop loss of latent relations will be optimized to train the proposed model. We conduct experiments on three real-world benchmark datasets (i.e., Twitter15, Twitter16, and PHEME). Extensive experimental results demonstrate the effectiveness of our model. EBGCN offers a superior uncertainty representation strategy and boosts the performance for rumor detection. The main contributions of this work are summarized as follows: • We propose novel Edge-enhanced Bayesian Graph Convolutional Networks (EBGCN) to handle the uncertainty in a probability manner. To the best of our knowledge, this is the first attempt to consider the inherent uncertainty in the propagation structure for rumor detection. • We design a new edge-wise consistency training framework to optimize the model with unlabeled latent relations. • Experiments on three real-world benchmark datasets demonstrate the effectiveness of our model on both rumor detection and early rumor detection tasks1. 2 Related Work 2.1 Rumor Detection Traditional methods on rumor detection adopted machine learning classifiers based on handcrafted features, such as sentiments (Castillo et al., 2011), bag of words (Enayet and El-Beltagy, 2017) and time patterns (Ma et al., 2015). Based on salient features of rumors spreading, Wu et al. (2015); Ma et al. (2017) modeled propagation trees and then used SVM with different kernels to detect rumors. Recent works have been devoted to deep learning methods. Ma et al. (2016) employed Recurrent Neural Networks (RNN) to sequentially process each timestep in the rumor propagation sequence. To improve it, many researchers captured more long-range dependency via attention mechanisms (Chen et al., 2018), convolutional neural networks (Yu et al., 2017; Chen et al., 2019), and Transformer (Khoo et al., 2020). However, most of them focused on learning temporal features alone, ignoring the internal topology structure. To capture topological-structural features, Ma et al. (2018) presented two recursive neural network (RvNN) based on bottom-up and top-down propagation trees. Yuan et al. (2019); Lu and Li (2020); Nguyen et al. (2020) formulated the propagation structure as graphs. Inspired by Graph Convolutional Network (GCN) (Kipf and Welling, 2017), Bian et al. (2020) first applied two GCNs 1The source code is available at https://github. com/weilingwei96/EBGCN. 3847 based on the propagation and dispersion graphs. Wei et al. (2019) jointly modeled the structural property by GCN and the temporal evolution by RNN. However, most of them treat the edge as the reliable topology connection for message-passing. Ignoring the uncertainty caused by unreliable relations could lead to lacking robustness and make it risky for rumor detection. Inspired by valuable research (Zhang et al., 2019a) that modeled uncertainty caused by finite available textual contents, this paper makes the first attempt to consider the uncertainty caused by unreliable relations in the propagation structure for rumor detection. 2.2 Graph Neural Networks Graph Neural Networks (GNNs) (Kipf and Welling, 2017; Schlichtkrull et al., 2018; Velickovic et al., 2018) have demonstrated remarkable performance in modeling structured data in a wide variety of fields, e.g., text classifcation (Yao et al., 2019), recommendation system (Wu et al., 2019) and emotion recognition (Ghosal et al., 2019). Although promising, they have limited capability to handle uncertainty in the graph structure. While the graphs employed in real-world applications are themselves derived from noisy data or modeling assumptions. To alleviate this issue, some valuable works (Luo et al., 2020; Zhang et al., 2019b) provide an approach for incorporating uncertain graph information by exploiting a Bayesian framework (Maddox et al., 2019). Inspired by them, this paper explores the uncertainty in the propagation structure from a probability perspective, to obtain more robust rumor detection results. 3 Problem Statement This paper develops EBGCN which processes text contents and propagation structure of each claim for rumor detection. In general, rumor detection commonly can be regarded as a multi-classification task, which aims to learn a classifier from training claims for predicting the label of a test claim. Formally, let C = {c1, c2, ..., cm} be the rumor detection dataset, where ci is the i-th claim and m is the number of claims. For each claim ci = {ri, xi 1, xi 2, ..., xi ni−1, Gi}, Gi indicates the propagation structure, ri is the source tweet, xi j refers to the j-th relevant retweet, and ni represents the number of tweets in the claim ci. Specifically, Gi is defined as a propagation graph Gi = ⟨Vi, Ei⟩ with the root node ri (Ma et al., 2018; Bian et al., 2020), where Vi = {ri, xi 1, xi 2, ..., xi ni−1} refers to the node set and Ei = {ei st|s, t = 0, ..., ni −1} represent a set of directed edges from a tweet to its corresponding retweets. Denote Ai ∈Rni×ni as an adjacency matrix where the initial value is αst = 1, if ei st ∈Ei 0, otherwise . Besides, each claim ci is annotated with a ground-truth label yi ∈Y, where Y represents finegrained classes. Our goal is to learn a classifier from the labeled claimed set, that is f : C →Y. 4 The Proposed Model In this section, we propose a novel edge-enhanced bayesian graph convolutional network (EBGCN) for rumor detection in Section 4.2. For better training, we design an edge-wise consistency training framework to optimize EBGCN in Section 4.3. 4.1 Overview The overall architecture of EBGCN is shown in Figure 2. Given the input sample including text contents and its propagation structure, we first formulate the propagation structure as directed graphs with two opposite directions, i.e., a top-down propagation graph and a bottom-up dispersion graph. Text contents are embedded by the text embedding layer. After that, we iteratively capture rich structural characteristics via two main components, node update module, and edge inference module. Then, we aggregate node embeddings to generate graph embedding and output the label of the claim. For training, we incorporate unsupervised consistency training on the Bayes by Backprop loss of unlabeled latent relations. Accordingly, we optimize the model by minimizing the weighted sum of the unsupervised loss and supervised loss. 4.2 Edge-enhanced Bayesian Graph Convolutional Networks 4.2.1 Graph Construction and Text Embedding The initial graph construction is similar to the previou work (Bian et al., 2020), i.e., build two distinct directed graphs for the propagation structure of each claim ci. The top-down propagation graph and bottom-up dispersion graph are denoted as GTD i and GBU i , respectively. Their corresponding initial adjacency matrices are ATD i = Ai and ABU i = A⊤ i . 3848 Unsupervised Consistency Loss Node features Input Sample Breaking: At least 10 dead, … @Samuel: The religion of peace strikes again. @Samuel: Hi, would you be willing to give… @... please call them terrorists … @Edward... @imranali27.Kill them Text Embedding Layer ࢟כ The groundtruth label ෝ࢟ The prediction label fc pooling pooling Retweet nodes Edges(top-down propagation) Edges(bottom-up dispersion) Source tweet node Node Embedding Graph Embedding Supervised Cross-entropy Loss Total Loss Edge Inference Adjust edge weights 3 5 6 1 2 4 0.21 0.61 0.80 0.42 0.37 0.91 0.12 0.18 0.62 0.46 0 8 3 5 6 1 2 4 GCL GCL l layers fc ݂ఏȉ ߤ௧ ߜ௧ ଶ Gaussian sampling The probability of latent relations ߤ௧ǡ ߜ௧ ଶ Edge Inference ݍ࣐ො࢘ࡴǡ ࡳ Graph(top-down) Graph(bottom-up) … ො࢘ࡴǡ ࡳ Figure 2: The architecture of the proposed rumor detection model EBGCN. Here, we leave out the superscript i in the following description for better presenting our method. The initial feature matrix of postings in the claim c can be extracted Top-5000 words in terms of TFIDF values, denoted as X = [x0, x1, ..., xn−1] ∈ Rn×d0, where x0 ∈Rd0 is the vector of the source tweet and d0 is the dimensionality of textual features. The initial feature matrices of nodes in propagation graph and dispersion graph are the same, i.e., XTD = XBU = X. 4.2.2 Node Update Graph convolutional networks (GCNs) (Kipf and Welling, 2017) are able to extract graph structure information and better characterize a node’s neighborhood. They define multiple Graph Conventional Layers (GCLs) to iteratively aggregate features of neighbors for each node and can be formulated as a simple differentiable message-passing framework. Motivated by GCNs, we employ the GCL to update node features in each graph. Formally, node features at the l-th layer H(l) = [h(l) 0 , h(l) 1 , ..., h(l) n−1] can be defined as, H(l) = σ(ˆA (l−1)H(l−1)W(l) + b(l)), (1) where ˆA (l−1) represents the normalization of adjacency matrix A(l−1) (Kipf and Welling, 2017). We initialize node representations by textual features, i.e., H(0) = X. 4.2.3 Edge Inference To alleviate the negative effects of unreliable relations, we rethink edge weights based on the currently observed graph by adopting a soft connection. Specifically, we adjust the weight between two nodes by computing a transformation fe(·; θt) based on node representations at the previous layer. Then, the adjacency matrix will be updated, i.e., g(l) t = fe ∥h(l−1) i −h(l−1) j ∥; θt , A(l) = T X t=1 σ(W(l) t g(l) t + b(l) t ) · A(l−1). (2) In practice, fe(·; θt) consists an convolutional layer and an activation function. T refers to the number of latent relation types. σ(·) refers to a sigmoid function. W(l) t and W(l) t are learnable parameters. We perform share parameters to the edge inference layer in two graphs GTD and GBU. After the stack of transformations in two layers, the model can effectively accumulate a normalized sum of features of the neighbors driven by latent relations, denoted as HTD and HBU. 4.2.4 Classification We regard the rumor detection task as a graph classification problem. To aggregate node representations in the graph, we employ aggregator to form the graph representations. Given the node representations in the propagation graph HTD and the node representations in the dispersion graph HBU, the graph representations can be computed as: CTD = meanpooling(HTD), CBU = meanpooling(HBU), (3) 3849 where meanpooling(·) refers to the mean-pooling aggregating function. Based on the concatenation of two distinct graph representations, label probabilities of all classes can be defined by a full connection layer and a softmax function, i.e., ˆy = softmax Wc[CTD; CBU] + bc , (4) where Wc and bc are learnable parameter matrices. 4.3 Edge-wise Consistency Training Framework For the supervised learning loss Lc, we compute the cross-entropy of the predictions and ground truth distributions C = {c1, c2, ..., cm}, i.e., Lc = − |Y| X i yilogˆyi, (5) where yi is a vector representing distribution of ground truth label for the i-th claim sample. For the unsupervised learning loss Le, we amortize the posterior distribution of the classification weight p(ϕ) as q(ϕ) to enable quick prediction at the test stage and learn parameters by minimizing the average expected loss over latent relations, i.e., ϕ∗= arg minϕ Le, where Le = E h DKL p(ˆr(l)|H(l−1), G)∥qϕ(ˆr(l)|H(l−1), G) i , ϕ∗= arg max ϕ E[log Z p(ˆr(l)|H(l−1), ϕ)qϕ(ϕ|H(l−1), G)dϕ], (6) where ˆr is the prediction distribution of latent relations. To ensure likelihood tractably, we model the prior distribution of each latent relation rt, t ∈ [1, T] independently. For each relation, we define a factorized Gaussian distribution for each latent relation qϕ(ϕ|H(l−1), G; Θ) with means µt and variances δ2 t set by the transformation layer, qϕ(ϕ|H(l−1), G; Θ)) = T Y t=1 qϕ(ϕt|{g(l) t }T t=1) = T Y t=1 N(µt, δ2 t ), µt = fµ({g(l) t }T t=1; θµ), δ2 t = fδ({g(l) t } T t=1; θδ), (7) where fµ(·; θµ) and fδ(·; θµ) refer to compute the mean and variance of input vectors, parameterized by θµ and θδ, respectively. Such that amounts to set the weight of each latent relation. Besides, we also consider the likelihood of latent relations when parameterizing the posterior distribution of prototype vectors. The likelihood of latent relations from the l-th layer based on node embeddings can be adaptively computed by, p(ˆr(l)|H(l−1), ϕ) = T Y t=1 p(ˆr(l) t |H(l−1), ϕt), p(ˆr(l) t |H(l−1), ϕt) = exp Wtg(l) t + bt PT t=1 exp Wtg(l) t + bt . (8) In this way, the weight of edges can be adaptively adjusted based on the observed graph, which can thus be used to effectively pass messages and learn more discriminative features for rumor detection. To sum up, in training, we optimize our model EBGCN by minimizing the cross-entropy loss of labeled claims Lc and Bayes by Backprop loss of unlabeled latent relations Le, i.e., Θ∗= arg min Θ γLc + (1 −γ)Le, (9) where γ is the trade-off coefficient. 5 Experimental Setup 5.1 Datasets We evaluate the model on three real-world benchmark datasets: Twitter15 (Ma et al., 2017), Twitter16 (Ma et al., 2017), and PHEME (Zubiaga et al., 2016). The statistics are shown in Table 1. Twitter15 and Twitter162 contain 1,490 and 818 claims, respectively. Each claim is labeled as Nonrumor (NR), False Rumor (F), True Rumor (T), or Unverified Rumor (U). Following (Ma et al., 2018; Bian et al., 2020), we randomly split the dataset into five parts and conduct 5-fold cross-validation to obtain robust results. PHEME dataset3 provides 2,402 claims covering nine events and contains three labels, False Rumor (F), True Rumor (T), and Unverified Rumor (U). Following the previous work (Wei et al., 2019), we conduct leave-oneevent-out cross-validation, i.e., in each fold, one event’s samples are used for testing, and all the rest are used for training. 5.2 Baselines For Twitter15 and Twitter16, we compare our proposed model with the following methods. DTC 2https://www.dropbox.com/s/ 7ewzdrbelpmrnxu/rumdetect2017.zip?dl=0 3https://figshare.com/articles/ dataset/PHEME_dataset_for_Rumour_ Detection_and_Veracity_Classification/ 6392078 3850 Dataset Twitter15 Twitter16 PHEME # of claims 1,490 818 2,402 # of false rumors 370 205 638 # of true rumors 374 205 1,067 # of unverified rumors 374 203 697 # of non-rumors 372 205 # of postings 331,612 204,820 105,354 Table 1: Statistics of the datasets. (Castillo et al., 2011) adopted a decision tree classifier based on information credibility. SVM-TS (Ma et al., 2015) leveraged time series to model the chronological variation of social context features via a linear SVM classifier. SVM-TK (Ma et al., 2017) applied an SVM classifier with a propagation tree kernel to model the propagation structure of rumors. GRU-RNN (Ma et al., 2016) employed RNNs to model the sequential structural features. RvNN (Ma et al., 2018) adopted two recursive neural models based on a bottom-up and a top-down propagation tree. StA-PLAN (Khoo et al., 2020) employed transformer networks to incorporate long-distance interactions among tweets with propagation tree structure. BiGCN (Bian et al., 2020) utilized bi-directional GCNs to model bottom-up propagation and top-down dispersion. For PHEME, we compare with several representative state-of-the-art baselines. NileTMRG (Enayet and El-Beltagy, 2017) used linear support vector classification based on bag of words. BranchLSTM (Kochkina et al., 2018) decomposed the propagation tree into multiple branches and adopted a shared LSTM to capture structural features. RvNN (Ma et al., 2018) consisted of two recursive neural networks to model propagation trees. Hierarchical GCN-RNN (Wei et al., 2019) modeled structural property based on GCN and RNN. BiGCN (Bian et al., 2020) consisted of propagation and dispersion GCNs to learn structural features from propagation graph. 5.3 Evaluation Metrics For Twitter15 and Twitter16, we follow (Ma et al., 2018; Bian et al., 2020; Khoo et al., 2020) and evaluate the accuracy (Acc.) over four categories and F1 score (F1) on each class. For PHEME, following (Enayet and El-Beltagy, 2017; Kochkina et al., 2018; Wei et al., 2019), we apply the accuracy (Acc.), macro-averaged F1 (mF1) as evaluation metrics. Also, we report the weighted-averaged F1 (wF1) because of the imbalanced class problem. 5.4 Parameter Settings Following comparison baselines, the dimension of hidden vectors in the GCL is set to 64. The number of latent relations T and the coefficient weight γ are set to [1, 5] and [0.0, 1.0], respectively. we train the model via backpropagation and a wildly used stochastic gradient descent named Adam (Kingma and Ba, 2015). The learning rate is set to {0.0002, 0.0005, 0.02} for Twitter15, Twitter16, and PHEME, respectively. The training process is iterated upon 200 epochs and early stopping (Yuan et al., 2007) is applied when the validation loss stops decreasing by 10 epochs. The optimal set of hyperparameters are determined by testing the performance on the fold-0 set of Twitter15 and Twitter16, and the class-balanced charlie hebdo event set of PHEME. Besides, on PHEME, following (Wei et al., 2019), we replace TF-IDF features with word embeddings by skip-gram with negative sampling (Mikolov et al., 2013) and set the dimension of textual features to 200. We implement this variant of BiGCN and EBGCN, denoted as BiGCN(SKP) and EBGCN(SKP), respectively. For results of baselines, we implement BiGCN according to their public project4 under the same environment. Other results of baselines are referenced from original papers (Khoo et al., 2020; Wei et al., 2019; Ma et al., 2018). 6 Results and Analysis 6.1 Performance Comparison with Baselines Table 2 shows results of rumor detection on Twitter15, Twitter16, and PHEME datasets. Our proposed model EBGCN obtains the best performance among baselines. Specifically, for Twitter15, EBGCN outperforms state-of-the-art models 2.4% accuracy and 3.6% F1 score of false rumor. For Twitter16, our model obtains 3.4% and 6.0% improvements on accuracy and F1 score of non-rumor, respectively. For PHEME, EBGCN significantly outperforms previous work by 40.2% accuracy, 34.7% mF1 , and 18.0% wF1. Deep learning-based (RvNN, StA-PLAN, BiGCN and EBGCN) outperform conventional methods using hand-crafted features (DTC, SVMTS), which reveals the superiority of learning high-level representations for detecting rumors. 4https://github.com/TianBian95/BiGCN 3851 Twitter15 Method Acc. NR F T U F1 F1 F1 F1 DTC 45.5 73.3 35.5 31.7 41.5 SVM-TS 54.4 79.6 47.2 40.4 48.3 GRU-RNN 64.1 68.4 63.4 68.8 57.1 SVM-TK 66.7 61.9 66.9 77.2 64.5 RvNN 72.3 68.2 75.8 82.1 65.4 StA-PLAN 85.2 84.0 84.6 88.4 83.7 BiGCN 87.1 86.0 86.7 91.4 85.4 EBGCN 89.2 86.9 89.7 93.4 86.7 Twitter16 Method Acc. NR F T U F1 F1 F1 F1 DTC 46.5 64.3 39.3 41.9 40.3 SVM-TS 54.4 79.6 47.2 40.4 48.3 GRU-RNN 63.6 61.7 71.5 57.7 52.7 SVM-TK 66.7 61.9 66.9 77.2 64.5 RvNN 72.3 68.2 75.8 82.1 65.4 StA-PLAN 85.2 84.0 84.6 88.4 83.7 BiGCN 88.5 82.9 89.9 93.2 88.2 EBGCN 91.5 87.9 90.6 94.7 91.0 PHEME Method Acc. mF1 wF1 NileTMRG 36.0 29.7 BranchLSTM 31.4 25.9 RvNN 34.1 26.4 Hierarchical GCN-RNN 35.6 31.7 BiGCN 49.2 46.7 63.2 BiGCN(SKP) 56.9 48.3 66.8 EBGCN 69.0 62.9 74.6 EBGCN(SKP) 71.5 57.5 79.1 Table 2: Results (%) of rumor detection. Moreover, compared with sequence-based models GRU-RNN, and StA-PLAN, EBGCN outperform them. It can attribute that they capture temporal features alone but ignore internal topology structures, which limit the learning of structural features. EBGCN can aggregate neighbor features in the graph to learn rich structural features. Furthermore, compared with state-of-the-art graph-based BiGCN, EBGCN also obtains better performance. We discuss the fact for two main reasons. First, BiGCN treats relations among tweet nodes as reliable edges, which may introduce inaccurate or irrelevant features. Thereby their performance lacks robustness. EBGCN considers the inherent uncertainty in the propagation structure. In the model, the unreliable relations can be refined (a) The effect of edge inference (b) The effect of unsupervised relation learning loss Figure 3: Results of model analysis on three datasets. in a probability manner, which boosts the bias of express uncertainty. Accordingly, the robustness of detection is enhanced. Second, the edge-wise consistency training framework ensures the consistency between uncertain edges and the current nodes, which is also beneficial to learn more effective structural features for rumor detection. Besides, EBGCN(SKP) and BiGCN(SKP) outperforms EBGCN and BiGCN that use TF-IDF features in terms of Acc. and wF1. It shows the superiority of word embedding to capture textual features. Our model consistently obtains better performance in different text embedding. It reveals the stability of EBGCN. 6.2 Model Analysis In this part, we further evaluate the effects of key components in the proposed model. The Effect of Edge Inference. The number of latent relation types T is a critical parameter in the edge inference module. Figure 3(a) shows the accuracy score against T. The best performance is obtained when T is 2, 3, and 4 on Twitter15, Twitter16, and PHEME, respectively. Besides, these best settings are different. An idea explanation is that complex relations among tweets are various in different periods and gradually tend to be more sophisticated in the real world with the development 3852 Figure 4: Performance of early rumor detection. of social media. The edge inference module can adaptively refine the reliability of these complex relations by the posterior distribution of latent relations. It enhances the bias of uncertain relations and promotes the robustness of rumor detection. The Effect of Unsupervised Relation Learning Loss. The trade-off parameter γ controls the effect of the proposed edge-wise consistency training framework. γ = 0.0 means this framework is omitted. The right in Figure 3 shows the accuracy score against γ. When this framework is removed, the model gains the worst performance. The optimal γ is 0.4, 0.3, and 0.3 on Twitter15, Twitter16, and PHEME, respectively. The results proves the effectiveness of this framework. Due to wily rumor producers and limited annotations of spread information, it is common and inevitable that datasets contains unreliable relations. This framework can ensure the consistency between edges and the corresponding node pairs to avoid the negative features. 6.3 Early Rumor Detection Rumor early detection is to detect a rumor at its early stage before it wide-spreads on social media so that one can take appropriate actions earlier. It is especially critical for a real-time rumor detection system. To evaluate the performance on rumor early detection, we follow (Ma et al., 2018) and control the detection deadline or tweet count since the source tweet was posted. The earlier the detection deadline or the less the tweet count, the less propagation information can be available. Figure 4 shows the performance of early rumor detection. First, all models climb as the detection deadline elapses or tweet count increases. Particularly, at each deadline or tweet count, our model EBGCN reaches a relatively high accuracy score than other comparable models. Second, compared with RvNN that captures temporal features alone and STM-TK based on handcrafted features, the superior performance of EBGCN and BiGCN that explored rich structural features reveals that structural features are more beneficial to the early detection of rumors. Third, EBGCN obtains better early detection results than BiGCN. It demonstrates that EBGCN can learn more conducive structural features to identify rumors by modeling uncertainty and enhance the robustness for early rumor detection. Overall, our model not only performs better longterm rumor detection but also boosts the performance of detecting rumors at an early stage. 6.4 The Case Study In this part, we perform the case study to show the existence of uncertainty in the propagation structure and explain why EBGCN performs well. We randomly sample a false rumor from PHEME, as depicted in Figure 5. The tweets are formulated as nodes and relations are modeled as edges in the graph, where node 1 refers to the source tweet and node 2-8 refer to the following retweets. As shown in the left of Figure 5, we observe that tweet 5 is irrelevant with tweet 1 although replying, which reveals the ubiquity of unreliable relations among tweets in the propagation structure and it is reasonable to consider the uncertainty caused by these unreliable relations. Right of Figure 5 indicates constructed graphs where the color shade indicates the value of edge weights. The darker the color, the greater the edge weight. The existing graph-based models always generate the representation of node 1 by aggregating the information of its all neighbors (node 2, 5, and 6) according to seemingly reliable edges. However, edge between node 1 and 5 would bring noise features and limit the learning of useful features for rumor detection. Our model EBGCN successfully weakens the negative effect of this edge by both the edge inference layer under the ingenious edge-wise consistency training framework. Accordingly, the 3853 ʹ ͳ 5 ͺ ͵ Ͷ Hi Henry would you be willing to give ITV News a phone interview for our Lunchtime bulletin in 2 hours? The religion of peace strikes again. if only people didn't hand out guns Explain. Tickets go on sale this week Kill them wherever you find them, and turn them out from where they have turned you out. Idiot strikes again with his stupid tweet. Breaking: At least 10 dead, 5 injured after to gunman open fire in offices of Charlie Hebdo, satirical mag that published Mohammed cartoons x Edge Inference Initial propagation structure Refined propagation structure ʹ ͳ 5 ͺ ͵ Ͷ ʹ ͳ 5 ͺ ͵ Ͷ ʹ ͳ 5 ͺ ͵ Ͷ ͳ ʹ ͵ Ͷ 5 ͺ 0.64 0.50 0.89 0.19 0.42 0.57 0.78 0.66 0.90 0.05 0.49 0.62 0.74 0.35 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 Figure 5: The case study. Left shows a false rumor sampled from PHEME. The gray-highlighted tweet is the irrelevant one towards this rumor propagation but included in. Right is the constructed directed graphs in topdown and bottom-up directions based on the propagation structure. Our model iteratively adjusts the weights of edges in each graph to strength the effect of reliable edges and weaken the effect of unreliable edges. model is capable of learning more conducive characteristics and enhances the robustness of results. 7 Conclusion In this paper, we have studied the uncertainty in the propagation structure from a probability perspective for rumor detection. Specifically, we propose Edge-enhanced Bayesian Graph Convolutional Networks (EBGCN) to handle uncertainty with a Bayesian method by adaptively adjusting weights of unreliable relations. Besides, we design an edge-wise consistency training framework incorporating unsupervised relation learning to enforce the consistency on latent relations. Extensive experiments on three commonly benchmark datasets have proved the effectiveness of modeling uncertainty in the propagation structure. EBGCN significantly outperforms baselines on both rumor detection and early rumor detection tasks. References Mohammad Ahsan, Madhu Kumari, and T. P. Sharma. 2019. Rumors detection, verification and controlling mechanisms in online social networks: A survey. Online Soc. Networks Media, 14. Tian Bian, Xi Xiao, Tingyang Xu, Peilin Zhao, Wenbing Huang, Yu Rong, and Junzhou Huang. 2020. Rumor detection on social media with bi-directional graph convolutional networks. In AAAI, pages 549– 556. AAAI Press. Carlos Castillo, Marcelo Mendoza, and Barbara Poblete. 2011. Information credibility on twitter. In WWW, pages 675–684. ACM. Tong Chen, Xue Li, Hongzhi Yin, and Jun Zhang. 2018. Call attention to rumors: Deep attention based recurrent neural networks for early rumor detection. In PAKDD (Workshops), volume 11154 of Lecture Notes in Computer Science, pages 40–52. Springer. Yixuan Chen, Jie Sui, Liang Hu, and Wei Gong. 2019. Attention-residual network with CNN for rumor detection. In CIKM, pages 1121–1130. ACM. Omar Enayet and Samhaa R. El-Beltagy. 2017. Niletmrg at semeval-2017 task 8: Determining rumour and veracity support for rumours on twitter. pages 470–474. Association for Computational Linguistics. Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander F. Gelbukh. 2019. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. In EMNLP/IJCNLP (1), pages 154–164. Association for Computational Linguistics. Ling Min Serena Khoo, Hai Leong Chieu, Zhong Qian, and Jing Jiang. 2020. Interpretable rumor detection in microblogs by attending to user interactions. In AAAI, pages 8783–8790. AAAI Press. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR (Poster). Thomas N. Kipf and Max Welling. 2017. Semisupervised classification with graph convolutional networks. In ICLR (Poster). OpenReview.net. Elena Kochkina, Maria Liakata, and Arkaitz Zubiaga. 2018. All-in-one: Multi-task learning for rumour verification. In COLING, pages 3402–3413. Association for Computational Linguistics. Yi-Ju Lu and Cheng-Te Li. 2020. GCAN: graph-aware co-attention networks for explainable fake news detection on social media. In ACL, pages 505–514. Association for Computational Linguistics. Yadan Luo, Zi Huang, Zheng Zhang, Ziwei Wang, Mahsa Baktashmotlagh, and Yang Yang. 2020. Learning from the past: Continual meta-learning with bayesian graph neural networks. In AAAI, pages 5021–5028. AAAI Press. 3854 Jing Ma, Wei Gao, Prasenjit Mitra, Sejeong Kwon, Bernard J. Jansen, Kam-Fai Wong, and Meeyoung Cha. 2016. Detecting rumors from microblogs with recurrent neural networks. In IJCAI, pages 3818– 3824. IJCAI/AAAI Press. Jing Ma, Wei Gao, Zhongyu Wei, Yueming Lu, and Kam-Fai Wong. 2015. Detect rumors using time series of social context information on microblogging websites. In CIKM, pages 1751–1754. ACM. Jing Ma, Wei Gao, and Kam-Fai Wong. 2017. Detect rumors in microblog posts using propagation structure via kernel learning. In ACL (1), pages 708–717. Association for Computational Linguistics. Jing Ma, Wei Gao, and Kam-Fai Wong. 2018. Rumor detection on twitter with tree-structured recursive neural networks. In ACL (1), pages 1980–1989. Association for Computational Linguistics. Wesley J. Maddox, Pavel Izmailov, Timur Garipov, Dmitry P. Vetrov, and Andrew Gordon Wilson. 2019. A simple baseline for bayesian uncertainty in deep learning. In NeurIPS, pages 13132–13143. Tom´as Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. In NIPS, pages 3111–3119. Van-Hoang Nguyen, Kazunari Sugiyama, Preslav Nakov, and Min-Yen Kan. 2020. FANG: leveraging social context for fake news detection using graph representation. In CIKM, pages 1165–1174. ACM. Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In ESWC, volume 10843 of Lecture Notes in Computer Science, pages 593–607. Springer. Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li`o, and Yoshua Bengio. 2018. Graph attention networks. In ICLR (Poster). OpenReview.net. Soroush Vosoughi. 2015. Automatic detection and verification of rumors on twitter. Soroush Vosoughi, Deb Roy, and Sinan Aral. 2018. The spread of true and false news online. Science, 359(6380):1146–1151. Penghui Wei, Nan Xu, and Wenji Mao. 2019. Modeling conversation structure and temporal dynamics for jointly predicting rumor stance and veracity. In EMNLP/IJCNLP (1), pages 4786–4797. Association for Computational Linguistics. Ke Wu, Song Yang, and Kenny Q. Zhu. 2015. False rumors detection on sina weibo by propagation structures. In ICDE, pages 651–662. Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan. 2019. Session-based recommendation with graph neural networks. In AAAI, pages 346–353. AAAI Press. Xiaoyu Yang, Yuefei Lyu, Tian Tian, Yifei Liu, Yudong Liu, and Xi Zhang. 2020. Rumor detection on social media with graph structured adversarial learning. In IJCAI, pages 1417–1423. ijcai.org. Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Graph convolutional networks for text classification. In AAAI, pages 7370–7377. AAAI Press. Feng Yu, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. 2017. A convolutional approach for misinformation identification. In IJCAI, pages 3901–3907. Chunyuan Yuan, Qianwen Ma, Wei Zhou, Jizhong Han, and Songlin Hu. 2019. Jointly embedding the local and global relations of heterogeneous graph for rumor detection. In ICDM, pages 796–805. IEEE. Yao Yuan, Lorenzo Rosasco, and Andrea Caponnetto. 2007. On early stopping in gradient descent learning. Constructive Approximation, 26(2):289 – 315. Qiang Zhang, Aldo Lipani, Shangsong Liang, and Emine Yilmaz. 2019a. Reply-aided detection of misinformation via bayesian deep learning. In WWW, pages 2333–2343. ACM. Yingxue Zhang, Soumyasundar Pal, Mark Coates, and Deniz ¨Ustebay. 2019b. Bayesian graph convolutional neural networks for semi-supervised classification. In AAAI, pages 5829–5836. AAAI Press. Arkaitz Zubiaga, Geraldine Wong Sak Hoi, Maria Liakata, Rob Procter, and Peter Tolmie. 2016. Analysing how people orient to and spread rumours in social media by looking at conversational threads. PLoS ONE, 11(3):e0150989.
|
2021
|
297
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3855–3864 August 1–6, 2021. ©2021 Association for Computational Linguistics 3855 Label-Specific Dual Graph Neural Network for Multi-Label Text Classification Qianwen Ma1,2, Chunyuan Yuan1,2, Wei Zhou1* and Songlin Hu1,2 1 Institute of Information Engineering, Chinese Academy of Sciences 2 School of Cyber Security, University of Chinese Academy of Sciences {maqianwen,yuanchunyuan,zhouwei,husonglin}@iie.ac.cn Abstract Multi-label text classification is one of the fundamental tasks in natural language processing. Previous studies have difficulties to distinguish similar labels well because they learn the same document representations for different labels, that is they do not explicitly extract label-specific semantic components from documents. Moreover, they do not fully explore the high-order interactions among these semantic components, which is very helpful to predict tail labels. In this paper, we propose a novel label-specific dual graph neural network (LDGN), which incorporates category information to learn label-specific components from documents, and employs dual Graph Convolution Network (GCN) to model complete and adaptive interactions among these components based on the statistical label cooccurrence and dynamic reconstruction graph in a joint way. Experimental results on three benchmark datasets demonstrate that LDGN significantly outperforms the state-of-the-art models, and also achieves better performance with respect to tail labels. 1 Introduction Automatically labeling multiple labels of documents is a fundamental and practical task in natural language processing. Recently, with the growth of data scale, multi-label text classification(MLTC) has attracted more attention, since it is often applied to many fields such as sentiment analysis (Liu and Chen, 2015; Li et al., 2016), emotion recognition (Wang et al., 2016; Jabreel and Moreno, 2019), web page tagging (Jain et al., 2016) and so on. However, the number of labels and documents and the complex relations of labels render it an unsolved and challenging task. Existing studies for multi-label text classification mainly focus on learning enhanced document *Corresponding Author representation (Liu et al., 2017) and modeling label dependency (Zhang et al., 2018; Yang et al., 2018; Tsai and Lee, 2019) to improve classification performance. Although they have explored the informative words in text content, or considered the label structure and label semantics to capture label correlations, these models cannot distinguish similar labels well (e.g., the categories Prices vs Consumer Prices in Reuters News). The main reason is that most of them neglect the semantic connections between labels and input documents and they learn the same document representations for different labels, which cannot issue the label similarity problem. More specifically, they do not explicitly consider the corresponding semantic parts of each label in the document. Recently, some studies (You et al., 2019; Xiao et al., 2019; Du et al., 2019) have used attention mechanism to explore the above semantic connections, and learn a label-specific document representation for classification. These methods have obtained promising results in MLTC, which shows the importance of exploring semantic connections. However, they did not further study the interactions between label-specific semantic components which can be guided by label correlations, and thus these models cannot work well on predicting tail labels which is also a challenging issue in MLTC. To handle these issues, a common way to explore the semantic interactions between labelspecific parts in document is to utilize the statistical correlations between categories to build a label co-occurrence graph for guiding interactions. Nevertheless, statistical correlations have three drawbacks. First, the co-occurrence patterns between label pairs obtained from training data are incomplete and noisy. Specifically, the label cooccurrences that appear in the test set but do not appear in the training set may be ignored, while 3856 some rare label co-occurrences in the statistical correlations may be noise. Second, the label cooccurrence graph is built in global, which may be biased for rare label correlations. And thus they are not flexible to every sample document. Third, statistical label correlations may form a long-tail distribution, i.e., some categories are very common while most categories have few of documents. This phenomenon may lead to models failing to predict low-frequency labels. Thus, our goal is to find a way to explore the complete and adaptive interactions among label-specific semantic components more accurately. In this paper, we investigate: (1) how to explicitly extract the semantic components related to the corresponding labels from each document; and (2) how to accurately capture the more complete and more adaptive interactions between label-specific semantic components according to label dependencies. To solve the first challenge, we exploit the attention mechanism to extract the labelspecific semantic components from the text content, which can alleviate the label similar problem. To capture the more accurate high-order interactions between these semantic components, we first employ one Graph Convolution Network (GCN) to learn component representations using the statistical label co-occurrence to guide the information propagation among nodes (components) in GCN. Then, we use the component representations to reconstruct the adjacency graph dynamically and re-learn the component representations with another GCN, and thus we can capture the latent interactions between these semantic components. Finally, we exploit final component representations to predict labels. We evaluate our model on three real-world datasets, and the results show that the proposed model LDGN outperforms all the comparison methods. Further studies demonstrate our ability to effectively alleviate the tail labels problem, and accurately capture the meaningful interactions between label-specific semantic components. The contributions of this paper are as follows: • We propose a novel label-specific dual graph neural network (LDGN), which incorporates category information to extract label-specific components from documents, and explores the interactions among these components. • To model the accurate and adaptive interactions, we jointly exploit global co-occurrence patterns and local dynamic relations. To make up the deficiency of co-occurrences, we employ the local reconstruction graph which is built by every document dynamically. • We conduct a series of experiments on three public datasets, and experimental results demonstrate that our model LDGN significantly outperforms the state-of-the-art models, and also achieves better performance with respect to tail labels. 2 Model As depicted in Figure 1, our model LDGN is composed of two major modules: 1) labelspecific document representation 2) dual graph neural network for semantic interaction learning. Specifically, label-specific document representation learning describes how to extract labelspecific semantic components from the mixture of label information in each document; and the dual graph neural network for semantic interaction learning illustrates how to accurately explore the complete interactions among these semantic components under the guidance of the prior knowledge of statistical label co-occurrence and the posterior information of dynamic reconstruction graph. Problem Formulation: Let D = {xi, yi}N be the set of documents, which consists of N document xi and its corresponding label yi ∈ {0, 1}|C|, where |C| denotes the total number of labels. Each document xi contains J words xi = wi1, wi2, . . . , wiJ. The target of multi-label text classification is to learn the mapping from input text sequence to the most relevant labels. 2.1 Label-specific Document Representation Given a document x with J words, we first embed each word wj in the text into a word vector ewj ∈Rd, where d is the dimensionality of word embedding vector. To capture contextual information from two directions of the word sequence, we first use a bidirectional LSTM to encode wordlevel semantic information in document representation. And we concatenate the forward and backward hidden states to obtain the final word sequence vector h ∈R|J|×D. After that, to explicitly extract the corresponding semantic component related to each label from documents, we use a label guided attention mechanism to learn label-specific text representation. 3857 2GHKR 8KVXKYKTZGZOUT ͙ ͙ :RUG 5HSUHVHQWDWLRQ w1 w2 wJ ͘͘͘ ͘͘͘ (O29:3 ĂƚƚĞŶƚŝŽŶ -)4 _nj _nj _njI ͙ D>W ,QWHUDFWLRQ/HDUQLQJZLWK 6WDWLVWLFDO/DEHO&RRFFXUUHQFH 2GHKRYVKIOLOI :K^Z8KVXKYKTZGZOUT 0XOWLODEHOORVV /DEHOVSHFLILF &RPSRQHQWV h1 4 h2 4 hC 4 8KIUTYZX[IZOUT -XGVN -)4 2GHKR)UUII[XXKTIK *[GR-XGVN4K[XGR4KZ]UXQLUX 9KSGTZOI/TZKXGIZOUT2KGXTOTM 5HOHDUQLQJZLWK '\QDPLF 5HFRQVWUXFWLRQ *UDSK u1 u2 uC h1 2 h2 hC 2 Figure 1: The architecture of the proposed network LDGN. Firstly, we randomly initialize the label representation C ∈R|C|×dc, and compute the label-aware attention values. Then, we can induce the labelspecific semantic components based on the label guided attention. The formula is as follows: αij = exp hjcT i P j exp hjcT i , (1) ui = X j αijhj , (2) where αij indicates how informative the j-th text feature vector is for the i-th label. ui ∈RD denotes the semantic component related to the label ci in this document. 2.2 Dual Graph Neural Network Interaction Learning with Statistical Label Cooccurrence To capture the mutual interactions between the label-specific semantic components, we build a label graph based on the prior knowledge of label co-occurrence, each node in which correlates to a label-specific semantic component ui. And then we apply a graph neural network to propagate message between nodes. Formally, we define the label graph G = (V, E), where nodes refer to the categories and edges refer to the statistical co-occurrence between nodes (categories). Specifically, we compute the probability between all label pairs in the training set and get the matrix As ∈R|C|×|C|, where As ij denotes the conditional probability of a sample belonging to category Ci when it belongs to category Cj. Then, we utilize GCN (Kipf and Welling, 2017) to learn the deep relationships between labelspecific semantic components guided by the statistical label correlations. GCNs are neural networks operating on graphs, which are capable of enhancing node representations by propagating messages between neighboring nodes. In multi-layer GCN, each GCN layer takes the component representations from previous layer Hl as inputs and outputs enhanced component representations, i.e., Hl+1. The layer-wise propagation rule is as follows: Hl+1 = σ bAsHlWl , (3) where σ (·) denotes LeakyReLU (Maas et al., 2013) activation function. Wl ∈RD×D′ is a transformation matrix to be learned. bA denotes the normalized adjacency matrix, and the normalization method (Kipf and Welling, 2017) is: bA = D−1 2 AD−1 2 , (4) where D is a diagonal degree matrix with entries Dij = ΣjAij Depending on how many convolutional layers are used, GCN can aggregate information only about immediate neighbors (with one convolutional layer) or any nodes at most K-hops neighbors (if K layers are stacked). See (Kipf and Welling, 2017) for more details about GCN. We use a two-layer GCN to learn the interactions between label-specific components. The first layer takes the initialized component representations U ∈R|C|×D in Equation 2 as inputs H0; and the last layer outputs H2 ∈R|C|×D′ with D′ denoting the dimensionality of final node representations. However, the statistical label correlations obtained by training data are incomplete and noisy. 3858 And the co-occurrence patterns between label pairs may form a long-tail distribution. Re-learning with Dynamic Reconstruction Graph To capture the more complete and adaptive interactions between these components, we exploit the above component representations H2 to reconstruct the adjacency graph dynamically, which can make up the deficiency of co-occurrence matrix. And then we re-learn the interactions among the label-specific components guided by the posterior information of dynamic reconstruction graph. Specifically, we apply two 1×1 convolution layers and dot product to get the dynamic reconstruction graph AD as follows: AD = f Wa ∗H2T Wb ∗H2 , (5) where Wa ∈ Rd1×D′ and Wb ∈ Rd1×D′ are the weights of two convolution layers, f is the sigmoid activation function. And then we normalize the reconstruction adjacency matrix as Equation 4, and obtain the normalized adjacency matrix bAD of reconstruction graph. In a similar way as Equation 3, we apply another 2-layer GCN to learn the deep correlations between components with the dynamic reconstruction graph. The first layer of this GCN takes the component representations H2 as inputs, and the last layer outputs the final component representations H4 ∈R|C|×D′. 2.3 Multi-label Text Classification After the above procedures, we concatenate the two types of component representations HO = [H2, H4] and feed it into a fully connected layer for prediction: by = σ(W1HO) , where W1 ∈ R2D′×1 and σ is the sigmoid function. We use y ∈R|C| to represent the ground-truth label of a document, where yi = 0, 1 denotes whether label i appears in the document or not. The proposed model LDGN is trained with the multi-label cross entropy loss: L = C X c=1 yc log (byc) + (1 −yc) log (1 −byc) . (6) 3 Experiment 3.1 Experimental Setup Datasets We evaluate the proposed model on three benchmark multi-label text classification datasets, which are AAPD (Yang et al., 2018), EUR-Lex (Mencia and F¨urnkranz, 2008) and RCV1 (Lewis et al., 2004). The statistics of these three datasets are listed in Table 1. Dataset Ntrain Ntest L L W RCV1 23,149 781,265 101 3.18 259.47 AAPD 54,840 1,000 54 2.41 163.42 EUR-Lex 11,585 3,865 3,954 5.32 1225.2 Table 1: Statistics of the datasets. Ntrain and Ntest denote the number of training and testing samples respectively. L is the total number of classes, L is the average number of labels per sample and W is the average number of words per sample. Evaluation Metric Following the settings of previous work (You et al., 2019; Xiao et al., 2019), we use precision at top K (P@k) and Normalized Discounted Cumulated Gains at top K (nDCG@k) for performance evaluation. The definition of two metrics can be referred to You et al. (2019). Implementation Details For a fair comparison, we apply the same dataset split as previous work (Xiao et al., 2019), which is also the original split provided by dataset publisher (Yang et al., 2018; Mencia and F¨urnkranz, 2008). The word embeddings in the proposed network are initialized with the 300-dimensional word vectors, which are trained on the datasets by Skipgram (Mikolov et al., 2013) algorithm. The hidden sizes of Bi-LSTM and GCNs are set to 300 and 512, respectively. We use the Adam optimization method (Kingma and Ba, 2014) to minimize the cross-entropy loss, the learning rate is initialized to 1e-3 and gradually decreased during the process of training. We select the best parameter configuration based on performance on the validation set and evaluate the configuration on the test set. Our code is available on GitHub1. 3.2 Baselines We compare the proposed model with recent deep learning based methods for MLTC, including seq2seq models, deep embedding models, and label attention based models. And it should be noted that, because of different application scenarios, we did not choose the label tree-based methods and extreme text focused methods as baseline models. • XML-CNN (Liu et al., 2017): a CNN-based 1https://github.com/Makwen1995/LDGN MLTC 3859 Models AAPD EUR-Lex P@1 P@3 P@5 N@3 N@5 P@1 P@3 P@5 N@3 N@5 XML-CNN 74.38 53.84 37.79 71.12 75.93 70.40 54.98 44.86 58.62 53.10 SGM 75.67 56.75 35.65 72.36 75.35 70.45 60.37 43.88 60.72 55.24 DXML 80.54 56.30 39.16 77.23 80.99 75.63 60.13 48.65 63.96 53.60 AttentionXML 83.02 58.72 40.56 78.01 82.31 67.34 52.52 47.72 56.21 50.78 EXAM 83.26 59.77 40.66 79.10 82.79 74.40 61.93 50.98 65.12 59.43 LSAN 85.28 61.12 41.84 80.84 84.78 79.17 64.99 53.67 68.32 62.47 LDGN 86.24 61.95 42.29 83.32 86.85 81.03 67.79 56.36 71.81 66.09 Table 2: Comparisons with state-of-the-art methods on both AAPD and EUR-Lex datasets. The experimental results of all baseline models are directly cited from paper (Xiao et al., 2019). model which uses CNN and a dynamic pooling layer to extract high-level feature for MLTC. • SGM (Yang et al., 2018): a sequence generation model which models label correlations as an ordered sequence. • DXML (Zhang et al., 2018): a deep embedding method which models the feature space and label graph structure simultaneously. • AttentionXML (You et al., 2019): a label treebased deep learning model which uses a probabilistic label tree and multi-label attention to capture informative words in extreme-scale data. • EXAM (Du et al., 2019): a novel framework that leverages the label information to compute the word-level interactions. • LSAN (Xiao et al., 2019): a label-specific attention network model based on self-attention and label-attention mechanism. The SotA model (i.e., LSAN) used BiLSTM model for text representations. For a fair comparison, we also take BiLSTM as text encoder in our model. 3.3 Experimental Results and Analysis Table 2 and Table 3 demonstrate the performance of all the compared methods based on the three datasets. For fair comparison, the experimental results of baseline models are directly cited from previous studies (Xiao et al., 2019). We also bold the best result of each column in all tables. From the Table 2 and Table 3, we can observe that the proposed LDGN outperforms all other Models RCV1 P@1 P@3 P@5 N@3 N@5 XML-CNN 95.75 78.63 54.94 89.89 90.77 SGM 95.37 81.36 53.06 91.76 90.69 DXML 94.04 78.65 54.38 89.83 90.21 AttentionXML 96.41 80.91 56.38 91.88 92.70 EXAM 93.67 75.80 52.73 86.85 87.71 LSAN 96.81 81.89 56.92 92.83 93.43 LDGN 97.12 82.26 57.29 93.80 95.03 Table 3: Comparisons with state-of-the-art methods on the RCV1 dataset. The experimental results of baselines are directly cited from (Xiao et al., 2019). baselines on three datasets. The outstanding results confirm the effectiveness of label-specific semantic interaction learning with dual graph neural network, which include global statistical patterns and local dynamic relations. It is observed that the performance of XMLCNN is worse than other comparison methods. The reason is that it only exploits the text content of documents for classification but ignores the label correlations which have been proven very important for multi-label classification problem. The label tree-based model AttentionXML performs better than the seq2seq method (SGM) and the deep embedding method (DXML). Although both DXML and SGM employ a label graph or an ordered sequence to model the relationship between labels, they ignore the interactions between labels and document content. And AttentionXML uses multi-label attention which can focus on the most relevant parts in content and extract different semantic information for each label. Compared with other label attention based 3860 70 75 80 85 90 95 PSP@1 PSP@3 PSP@5 LSAN LDGN (a) RCV1 74 76 78 80 82 84 86 88 90 PSP@1 PSP@3 PSP@5 LSAN LDGN (b) AAPD 42 44 46 48 50 52 54 56 PSP@1 PSP@3 PSP@5 LSAN LDGN (c) EUR-Lex Figure 2: Performance on tail labels. methods (AttentionXML, EXAM), LSAN performs the best because it takes the semantic correlations between document content and label text into account simultaneously, which exploits an adaptive fusion to integrate self-attention and label-attention mechanisms to learn the labelspecific document representation. In conclusion, the proposed network LDGN outperforms sequence-to-sequence models, deep embedding models, and label attention based models, and the metrics P@k and nDCG@k of multi-label text classification obtain significant improvement. Specifically, on the AAPD dataset, LDGN increases the P@1 of the LSAN method (the best baseline) from 85.28% to 86.24%, and increases nDCG@3 and nDCG@5 from 80.84% to 83.33%, 84.78% to 86.85% , respectively. On the EUR-Lex dataset, the metric P@1 is boosted from 79.17% to 81.03%, and P@5 and nDCG@5 are increased from 53.67% to 56.36%, 62.47% to 66.09%, respectively. On the RCV1 dataset, the P@k of our model increased by 0.3% at average, and LDGN achieves 1% and 1.6% absolute improvement on nDCG@3, 5 compared with LSAN. The improvements of the proposed LDGN model demonstrate that the semantic interaction learning with joint global statistical relations and local dynamic relations are generally helpful and effective, and LDGN can capture the deeper correlations between categories than LSAN. 3.4 Ablation Test We perform a series of ablation experiments to examine the relative contributions of dual graphbased semantic interactions module. To this end, LDGN is compared with its three variants:(1)S: Graph-based semantic interactions only with statistical label co-occurrence; (2)D: Graph-based semantic interactions only with dynamic reconstruction graph; (3)no-G:Removing the dual graph 82 83 84 85 86 87 88 P@1 N@5 S D no-G S+D (a) AAPD 60 65 70 75 80 85 P@1 N@5 S D no-G S+D (b) EUR-Lex Figure 3: Ablation test of LDGN on two datasets. neural network. For a fair comparison, both S and D use 4-layer GCN which is the same as LDGN. As presented in Figure 3, S and D perform better than no-G, which demonstrates that exploring either statistical relations or dynamic relations can correctly capture the effective semantic interactions between label-specific components. D performs better than S, indicating the model with local dynamic relations is adaptive to data and has better stability and robustness, which also shows that the model with local dynamic relations can capture semantic dependencies more effectively and accurately. The performance of S+D (i.e., LDGN) combining two aspect relations obtains significant improvement, which shows dynamic relations can make up the deficiency of statistical co-occurrence and correct the bias of global correlations. Thus, it is necessary to explore their joint effects to further boost the performance. 3.5 Performance on tail labels In order to prove the effectiveness of the proposed LDGN in alleviating the tail labels problem, we evaluate the performance of LDGN by propensity scored precision at k (PSP@k), which is calcu3861 smart grid digitalization power grid visionary acceptation model energy management users engaged producing energy consuming systems aware energy demand response network dynamically varying prices natural question smart grid reality distribution grid updated assume positive answer question lower layers medium low voltage change previous analyzed samples dutch distribution grid previous considered evolutions synthetic topologies modeled studies complex systems technological domains previous paper extra step defining methodology evolving existing physical power grid smart grid model laying foundations decision support system utilities governmental organizations evolution strategies apply dutch distribution grid Figure 4: The Visualization of label attention weights. (The attention weights of ’physics.soc’ for words are shaded in blue, and the attention scores of class CS.CY and CS.CE are shaded in green and yellow color respectively. Darker color represents higher weight score.) lated as follow: PSP@k = 1 k k X l=1 yrank(l) Prank(l) , (7) where Prank(l) is the propensity score (Jain et al., 2016) of label rank(l). Figure 2 shows the results of LDGN and LSAN on three datasets. As shown in Figure 2(a), Figure 2(b) and Figure 2(c), the proposed LDGN performs better in predicting tail labels than the LSAN model (the best baseline) on three datasets. Specifically, on the RCV1 dataset, LDGN achieves 0.97% and 1.35% absolute improvement in term of PSP@3 and PSP@5 compared with LSAN. On the AAPD dataset, the PSP@k increased by at least 0.63% up to 0.90%. And on the EUR-Lex dataset, LDGN achieves 1.94%, 3.64% and 4.93% absolute improvement on PSP@1, 3, 5 compared with LSAN. The reason for the improvement in the EUR-Lex dataset is more obvious is that the semantic interactions learning is more useful to capture related information in the case of a large number of labels. The results prove that LDGN can effectively alleviate the problem of predicting tail labels. 3.6 Case Study To further verify the effectiveness of our label attention module and dual graph neural network in LDGN, we present a typical case and visualize the attention weights on the document words and the similarity scores between label-specific components. We show a test sample from original AAPD dataset, and the document belongs to three categories, ‘Physics and Society’ (physics.soc), ‘Computers and Society’ (cs.cy) and ‘Computational Engineering, Finance, and Science’ (cs.ce). Visualization of Attention We can observe from the Figure 4 that different labels focus on different parts in the document text, and each label has its own concerned words. For example, Figure 5: The Visualization of two adjacency matrices of dual GNN. Darker color represents higher weight. the more important parts in the ‘physics.soc’ category are ‘digitalization power grid’, ‘energy management’. And the words that the ‘cs.ce’ category focuses on are ‘consuming systems’, ‘varying prices’, ‘laying foundations’, ‘lower ’ and etc. For class ‘cs.cy’, the concerned words are ‘samples dutch distribution’, ‘evolutions’ and ‘topologies’. The corresponding related words of the three categories can reflect the semantics of the categories. Visualization of Interactions To gain a clearer view of the importance of our dual graph-based interactions learning module, we display two 3862 heatmaps in Figure 5 to visualize the partial graph structure of dual GCN. The edge weights shown in the heatmaps are obtained by global label cooccurrence and local dynamic relations (i.e., computed by Equation 5), respectively. As presented in heatmaps, different relations between categories are captured by dual GCN. In global statistical relations, ‘cs.cy’ is highly linked with ‘physics.soc’ and wrong label ‘nlin.ao’, while the true label ‘cs.ce’ is isolated. And in local dynamic relations, ‘cs.cy’ is more related to ‘cs.ce’, and the correlations between wrong label ‘nlin.ao’ and true labels are reduced. This demonstrates that local dynamic relations can capture the latent relations that do not appear in global relations, and correct the bias of global correlations. 4 Related Work Multi-label Text Classification The existing methods for MLTC mainly focus on learning enhanced document representation (Liu et al., 2017) and modeling label dependency (Nam et al., 2017; Yang et al., 2018; Tsai and Lee, 2019) to improve the classification performance. With the wide application of neural network methods for text representation, some innovative models have been developed for this task, which include traditional deep learning methods and Seq2Seq based methods. Liu et al. (2017) employed CNNs and dynamic pooling to learn the text representation for MLTC. However, they treated all words equally and cannot explored the informative words in documents. The Seq2Seq methods, such as MLC2Seq (Nam et al., 2017) and SGM (Yang et al., 2018), employed a RNN to encode the input text and an attention based RNN decoder to generate predicted labels sequentially. Although they used attention mechanism to capture the informative words in text content, these models cannot distinguish similar labels well. There is a big reason that most of them neglect the semantic connections between labels and document, and learn the same document representations for different labels. Recently, some studies (You et al., 2019; Xiao et al., 2019; Du et al., 2019) have used attention mechanism to explore the interactions between words and labels, and learned a labelspecific document representation for classification. These methods have obtained promising results in MLTC, which shows the importance of exploring semantic connections. However, they did not further study the interactions between labelspecific semantic components which can help to predict low-frequency labels. To handle these issues, a common way to explore the semantic interactions between labelspecific parts in document, is to utilize the label graph based on statistical co-occurrences. MLC with Label Graph In order to capture the deep correlations of labels in a graph structure, many researches in image classification apply node embedding and graph neural network models to the task of multi-label image classification. Lee et al. (2018) incorporated knowledge graphs for describing the relationships between labels, and the information propagation can model the dependencies between seen and unseen labels for multi-label zero-shot learning. Chen et al. (2019) learned label representations with prior label correlation matrix in GCN, and mapped the label representations to inter-dependent classifiers, which achieved superior performance. However, there were few related approaches for multi-label classification of text. Zhang et al. (2018) established an explicit label cooccurrence graph to explore label embedding in low-dimension latent space. Furthermore, the statistical label correlations obtained by training data are incomplete and noisy. And the co-occurrence patterns between label pairs may form a long-tail distribution. Thus, our goal is to find a way to explore the complete and adaptive interactions among labelspecific semantic components more accurately. 5 Conclusion In this paper, we propose a graph-based network LDGN to capture the semantic interactions related to corresponding labels, which jointly exploits global statistical patterns and local dynamic relations to derive complete and adaptive dependencies between different label-specific semantic parts. We first exploit multi-label attention to extract the label-specific semantic components from documents. Then, we employ GCN to learn component representations using label co-occurrences to guide the information propagation among components. After that, we use the learned component representations to compute the adjacency graph dynamically and re-learn with GCN based on the reconstruction graph. Extensive experiments con3863 ducted on three public datasets show that the proposed LDGN model outperforms other state-ofthe-art models on multi-label text classification task and also demonstrates much higher effectiveness to alleviate the tail label problem. In the future, we will improve the proposed model in efficiency, for example we could construct a dynamic graph for a few samples rather than each sample. And besides, we will explore more information about labels for MLC classification. Acknowledgement We gratefully thank the anonymous reviewers for their insightful comments. This research is supported by the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant No. XDC02060400. References Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. 2019. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5177–5186. Cunxiao Du, Zhaozheng Chen, Fuli Feng, Lei Zhu, Tian Gan, and Liqiang Nie. 2019. Explicit interaction model towards text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6359–6366. Mohammed Jabreel and Antonio Moreno. 2019. A deep learning-based approach for multi-label emotion classification in tweets. Applied Sciences, 9(6):1123. Himanshu Jain, Yashoteja Prabhu, and Manik Varma. 2016. Extreme multi-label loss functions for recommendation, tagging, ranking & other missing label applications. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 935–944. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Thomas N Kipf and Max Welling. 2017. Semisupervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR). Chung-Wei Lee, Wei Fang, Chih-Kuan Yeh, and YuChiang Frank Wang. 2018. Multi-label zero-shot learning with structured knowledge graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1576–1585. David D Lewis, Yiming Yang, Tony G Rose, and Fan Li. 2004. Rcv1: A new benchmark collection for text categorization research. Journal of machine learning research, 5(Apr):361–397. Xin Li, Haoran Xie, Yanghui Rao, Yanjia Chen, Xuebo Liu, Huan Huang, and Fu Lee Wang. 2016. Weighted multi-label classification model for sentiment analysis of online news. In 2016 International Conference on Big Data and Smart Computing (BigComp), pages 215–222. IEEE. Jingzhou Liu, Wei-Cheng Chang, Yuexin Wu, and Yiming Yang. 2017. Deep learning for extreme multi-label text classification. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 115–124. Shuhua Monica Liu and Jiun-Hung Chen. 2015. A multi-label classification based approach for sentiment classification. Expert Systems with Applications, 42(3):1083–1093. Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. 2013. Rectifier nonlinearities improve neural network acoustic models. In Proc. icml, volume 30, page 3. Eneldo Loza Mencia and Johannes F¨urnkranz. 2008. Efficient pairwise multilabel classification for largescale problems in the legal domain. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 50–65. Springer. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems, pages 3111–3119. Jinseok Nam, Eneldo Loza Menc´ıa, Hyunwoo J Kim, and Johannes F¨urnkranz. 2017. Maximizing subset accuracy with recurrent neural networks in multilabel classification. In Advances in neural information processing systems, pages 5413–5423. Che-Ping Tsai and Hung-Yi Lee. 2019. Order-free learning alleviating exposure bias in multi-label classification. arXiv preprint arXiv:1909.03434. Yaqi Wang, Shi Feng, Daling Wang, Ge Yu, and Yifei Zhang. 2016. Multi-label chinese microblog emotion classification via convolutional neural network. In Asia-Pacific Web Conference, pages 567–580. Springer. Lin Xiao, Xin Huang, Boli Chen, and Liping Jing. 2019. Label-specific document representation for multi-label text classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 466–475. 3864 Pengcheng Yang, Xu Sun, Wei Li, Shuming Ma, Wei Wu, and Houfeng Wang. 2018. Sgm: sequence generation model for multi-label classification. arXiv preprint arXiv:1806.04822. Ronghui You, Zihan Zhang, Ziye Wang, Suyang Dai, Hiroshi Mamitsuka, and Shanfeng Zhu. 2019. Attentionxml: Label tree-based attention-aware deep model for high-performance extreme multi-label text classification. In Advances in Neural Information Processing Systems, pages 5820–5830. Wenjie Zhang, Junchi Yan, Xiangfeng Wang, and Hongyuan Zha. 2018. Deep extreme multi-label learning. In Proceedings of the 2018 ACM on International Conference on Multimedia Retrieval, pages 100–107.
|
2021
|
298
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3865–3880 August 1–6, 2021. ©2021 Association for Computational Linguistics 3865 TAN-NTM: Topic Attention Networks for Neural Topic Modeling Madhur Panwar2∗†, Shashank Shailabh3∗†, Milan Aggarwal1∗, Balaji Krishnamurthy1 Media and Data Science Research Labs, Adobe1 Birla Institute of Technology and Science, Pilani (BITS Pilani), India2 Indian Institute of Technology Kanpur (IIT Kanpur), India3 [email protected], [email protected] Abstract Topic models have been widely used to learn text representations and gain insight into document corpora. To perform topic discovery, most existing neural models either take document bag-of-words (BoW) or sequence of tokens as input followed by variational inference and BoW reconstruction to learn topic-word distribution. However, leveraging topic-word distribution for learning better features during document encoding has not been explored much. To this end, we develop a framework TAN-NTM, which processes document as a sequence of tokens through a LSTM whose contextual outputs are attended in a topic-aware manner. We propose a novel attention mechanism which factors in topic-word distribution to enable the model to attend on relevant words that convey topic related cues. The output of topic attention module is then used to carry out variational inference. We perform extensive ablations and experiments resulting in ∼ 9 - 15 percentage improvement over score of existing SOTA topic models in NPMI coherence on several benchmark datasets - 20Newsgroups, Yelp Review Polarity and AGNews. Further, we show that our method learns better latent document-topic features compared to existing topic models through improvement on two downstream tasks: document classification and topic guided keyphrase generation. 1 Introduction Topic models (Steyvers and Griffiths, 2007) have been popularly used to extract abstract topics which occur commonly across documents in a corpus. Each topic is interpreted as a group of semantically coherent words that represent a common concept. In addition to gaining insights from unstructured texts, topic models have been used in several tasks ∗equal contribution †work done during summer internship at Adobe of practical importance such as learning text representations for document classification (Nan et al., 2019), keyphrase extraction (Wang et al., 2019b), understanding reviews for e-commerce recommendations (Jin et al., 2018), semantic similarity detection between texts (Peinelt et al., 2020) etc. Early works on topic discovery include statistical methods such as Latent Semantic Analysis (Deerwester et al., 1990), Latent Dirichlet Allocation (LDA) (Blei et al., 2003) which approximates each topic as a probability distribution over word vocabulary (known as topic-word distribution) and performs approximate inference over documenttopic and topic-word distributions through Variational Bayes. This was followed by Markov Chain Monte Carlo (MCMC) (Andrieu et al., 2003) based inference algorithm - Collapsed Gibbs sampling (Griffiths and Steyvers, 2004). These methods require an expensive iterative inference step which has to be performed for each document. This was circumvented through introduction of deep neural networks and Variational Autoencoders (VAE) (Kingma and Welling, 2013), where variational inference can be performed in single forward pass. Neural variational inference topic models (Miao et al., 2017; Ding et al., 2018; Srivastava and Sutton, 2017) commonly convert a document to Bagof-Words (BoW) determined on the basis of frequency count of each vocabulary token in the document. The BoW input is processed through an MLP followed by variational inference which samples a latent document-topic vector. A decoder network then reconstructs original BoW using latent document-topic vector through topic-word distribution (TWD). VAE based neural topic models can be categorised on the basis of prior enforced on latent document-topic distribution. Methods such as NVDM (Miao et al., 2016), NTM-R (Ding et al., 2018), NVDM-GSM (Miao et al., 2017) use the Gaussian prior. NVLDA and ProdLDA (Srivastava 3866 and Sutton, 2017) use approximation to the Dirichlet prior which enables model to capture the fact that a document stems from a sparse set of topics. However, improving document encoding in topic models in order to capture document distribution and semantics better has not been explored much. In this work, we build upon VAE based topic model and propose a novel framework TAN-NTM: Topic Attention Networks for Neural Topic Modeling which process the sequence of tokens in input document through an LSTM (Hochreiter and Schmidhuber, 1997) whose contextual outputs are attended using Topic-Word Distribution (TWD). We hypothesise that TWD (being learned by the model) can be factored in the attention mechanism (Bahdanau et al., 2014) to enable the model to attend on the tokens which convey topic related information and cues. We perform separate attention for each topic using its corresponding word probability distribution and obtain the topic-wise context vectors. The learned word embeddings and TWD are used to devise a mechanism to determine topic weights representing the proportion of each topic in the document. The topic weights are used to aggregate topic-wise context vectors. The composed context vector is then used to perform variational inference followed by the BoW decoding. We perform extensive ablations to compare TAN-NTM variants and different ways of composing the topicwise context vectors. For evaluation, we compute commonly used NPMI coherence (Aletras and Stevenson, 2013) which measures the extent to which most probable words in a topic are semantically related to each other. We compare our TAN-NTM model with several state-of-the-art topic models (statistical (Blei et al., 2003; Griffiths and Steyvers, 2004), neural VAE (Srivastava and Sutton, 2017; Wu et al., 2020) and non-variational inference based neural model (Nan et al., 2019)) outperforming them on three benchmark datasets of varying scale and complexity: 20Newsgroups (20NG) (Lang, 1995), Yelp Review Polarity and AGNews (Zhang et al., 2015). We verify that our model learns better document feature representations and latent document-topic vectors by achieving a higher document classification accuracy over the baseline topic models. Further, topic models have previously been used to improve supervised keyphrase generation (Wang et al., 2019b). We show that TAN-NTM can be adapted to modify topic assisted keyphrase generation achieving SOTA performance on StackExchange and Weibo datasets. Our contributions can be summarised as: • We propose a document encoding framework for topic modeling which leverages the topicword distribution to perform attention effectively in a topic aware manner. • Our proposed model achieves better NPMI coherence (∼9-15 percentage improvement over the scores of existing best topic models) on various benchmark datasets. • We show that the topic guided attention results in better latent document-topic features achieving a higher document classification accuracy than the baseline topic models. • We show that our topic model encoder can be adapted to improve the topic guided supervised keyphrase generation achieving improved performance on this task. 2 Related Work Development of neural networks has paved path for Variational Autoencoders (VAE) (Kingma and Welling, 2013) which enables performing Variational Inference (VI) efficiently. The VAE-based topic models use a prior distribution to approximate the posterior for latent document-topic space and compute the Evidence Lower Bound (ELBO) using the reparametrization trick. Since our work is based on variational inference, we use ProdLDA and NVLDA (Srivastava and Sutton, 2017) as baselines for comparison. The Dirichlet distribution has been commonly considered as a suitable prior on the latent document-topic space since it captures the property that a document belongs to a sparse subset of topics. However, in order to enforce the Dirichlet prior, VAE methods have to resort to approximations of the Dirichlet distribution. Several works have proposed solutions to impose the Dirichlet prior effectively. Rezaee and Ferraro (2020) enforces Dirichlet prior using VI without reparametrization trick through word-level topic assignments. Some works address the sparsitysmoothness trade-off in dirichlet distribution by factoring dirichlet parameter vector as a product of two vectors (Burkhardt and Kramer, 2019). Wasserstein Autoencoders (WAE) (Tolstikhin et al., 2017) have led to the development of non-variational inference based topic model: Wasserstein-LDA (WLDA) which minimizes the wasserstein distance, a 3867 type of Optimal Transport (OT) distance, by leveraging distribution matching to the Dirichlet prior. We compare our work with W-LDA as a baseline. Zhao et al. (2021) proposed an OT based topic model which directly calculates topic-word distribution without a decoder. Adversarial Topic Model (ATM) (Wang et al., 2019a) was proposed based on GAN (Generative Adversarial Network) (Goodfellow et al., 2014) but it cannot infer document-topic distribution. A major advantage of W-LDA over ATM is distribution matching in document-topic space. Bidirectional Adversarial Topic model (BAT) (Wang et al., 2020) employs a bilateral transformation between document-word and document-topic distribution, while Hu et al. (2020) uses CycleGAN (Zhu et al., 2017) for unsupervised transfer between documentword and document-topic distribution. Hierarchical topic models (Viegas et al., 2020) utilize relationships among the latent topics. Supervised topic models have been explored previously where the topic model is trained through human feedback (Kumar et al., 2019) or with a task specific network simultaneously such that topic extraction is guided through task labels (Pergola et al., 2019; Wang and Yang, 2020). Card et al. (2018) leverages document metadata but without metadata their method is same as ProdLDA which is our baseline. Topic modeling on document networks has been done leveraging relational links between documents (Zhang and Lauw, 2020; Zhou et al., 2020). However our problem setting is completely different, we extract topics from documents in unsupervised way where document links/metadata/labels either don’t exist or are not used to extract the topics. Some very recent works use pre-trained BERT (Devlin et al., 2019) either to leverage improved text representations (Bianchi et al., 2020; Sia et al., 2020) or to augment topic model through knowledge distillation (Hoyle et al., 2020a). Zhu et al. (2020) and Dieng et al. (2020) jointly train words and topics in a shared embedding space. However, we train topic-word distribution as part of our model, embed it using word embeddings being learned and use resultant topic embeddings to perform attention over sequentially processed tokens. iDocNade (Gupta et al., 2019) is an autoregressive topic model for short texts utilizing pre-trained embeddings as distributional prior. However, it attains poorer topic coherence than ProdLDA and GNBNTM as shown in Wu et al. (2020). Some works have attempted to use other prior distributions such as Zhang et al. (2018) uses the Weibull prior, Thibaux and Jordan (2007) uses the beta distribution. Gamma Negative BinomialNeural Topic Model (GNB-NTM) (Wu et al., 2020) is one of the recent neural variational topic models which attempt to combine VI with mixed counting models. Mixed counting models can better model hierarchically dependent and over-dispersed random variables while implicitly introducing nonnegative constraints in topic modeling. GNB-NTM uses reparameterization of Gamma distribution and Gaussian approximation of Poisson distribution. We use their model as a baseline for our work. Topic models have been used with sequence encoders such as LSTM in applications like user activity modeling (Zaheer et al., 2017). Dieng et al. (2016) employs an RNN to detect stop words and merges its output with document-topic vector for next word prediction. Gururangan et al. (2019) uses a VAE pre-trained through topic modeling to perform text classification. We perform document classification and compare our model’s accuracy with the accuracy of VAE based and other topic models. LTMF (Jin et al., 2018) combines text features processed through an LSTM with a topic model for review based recommendations. Fundamentally different from these, we use topic-word distribution to attend on sequentially processed tokens via novel topic guided attention for performing variational inference, learning better document-topic features and improving topic modeling. A key application of topic models is supervised keyphrase generation. Some of the existing neural keyphrase generation methods include SEQ-TAG (Zhang et al., 2016) based on sequence tagging, SEQ2SEQ-CORR (Chen et al., 2018) based on seq2seq model without copy mechanism and SEQ2SEQ-COPY (Meng et al., 2017) which additionally uses copy mechanism. TopicAware Keyphrase Generation (TAKG) (Wang et al., 2019b) is a seq2seq based neural keyphrase generation framework for social media language. TAKG uses a neural topic model in Miao et al. (2017) and a keyphrase generation (KG) module which is conditioned on latent document-topic vector from the topic model. We adapt our proposed topic model to TAKG to improve keyphrase generation and discuss it in detail later in the Experiments section. 3868 Figure 1: A-E: Architecture of TAN-NTM showing flow of document processing through it. Document, being embedded using embedding layer, is processed by LSTM, yielding hidden states on which TAN attends in a topic aware manner. The resultant context vector is used to perform variational inference and processed through a BoW decoder as in VAEs. Attention Module E (zoomed in view of C) computes the blocks in the mentioned order 1-6. 3 Background LDA is a generative statistical model and assumes that each document is a distribution over a fixed number of topics (say K) and that each topic is a distribution of words over the entire vocabulary. LDA proposes an iterative process of document generation where for each document d, we draw a topic distribution θ from Dirichlet(α) distribution. For each word in d at index i, we sample a topic ti from Multinomial(θ) distribution. wi is sampled from p(wi|ti, β) distribution which is a multinomial probability conditioned on topic ti. Given the document corpus and the parameters α and β, we need the joint probability distribution of a topic mixture θ, a set of K topics t, and a set of n words w. This is given analytically by an intractable integral. The solution is to use Variational Inference wherein this problem is converted into an optimization problem for finding various parameters that minimize the KL divergence between the prior and the posterior distribution. This idea is leveraged at scale by the use of Variational Autoencoders. The encoder processes BoW vector of the document xbow by an MLP (Multi Layer Perceptron) which then forks into two independently trainable layers to yield zµ & zlog σ2. Then a re-parametrization trick is employed to sample the latent vector z from a logistic-normal distribution (resulting from an approximation of Dirichlet distribution). This is essential since backpropagation through a sampling node is infeasible. z is then used by decoder’s single dense layer D to yield the reconstructed BoW xrec. The objective function has two terms: (a) Kullback–Leibler (KL) Divergence Term - to match the variational posterior over latent variables with the prior and (b) Reconstruction Term - categorical cross entropy loss between xbow & xrec. LNTM = DKL(p(z) || q(z|x)) −Eq(z|x)[p(x|z)] Our methodology improves upon the document encoder and introduces a topic guided attention whose output is used to sample z. We use the same formulation of decoder as used in ProdLDA. 4 Methodology In this section, we describe the details of our framework where we leverage the topic-word distribution to perform topic guided attention over tokens in a document. Given a collection C with |C| documents {x1, x2, .., x|C|}, we process each document x into BoW vector xbow ∈R|V | and as a token sequence xseq, where V represents the vocabulary. As shown in step A in figure 1, each word wj ∈xseq is embedded as ej ∈RE through an embedding layer E ∈R|V |×E (E = Embedding Dimension) initialised with GloVe (Pennington et al., 2014). The embedded sequence {ej}|x| j=1, 3869 where |x| is the number of tokens in x, is processed through a sequence encoder LSTM (Hochreiter and Schmidhuber, 1997) to obtain the corresponding hidden states hj ∈RH and cell states sj ∈RH (step B in figure 1): hj, sj = fLSTM(ej, (hj−1, sj−1)) where H is LSTM’s hidden size. We construct a memory bank M = ⟨h1, h2, ..., h|x|⟩which is then used to perform topic-guided attention (step C in figure 1). The output vector of the attention module is used to derive prior distribution parameters zµ & zlog σ2 (as in VAE) through two linear layers. Using the re-parameterisation trick, we sample the latent document-topic vector z, which is then given as input to BoW decoder linear layer D that outputs the reconstructed BoW xrec (step D in figure 1). Objective function is same as in VAE setting, involving a reconstruction loss term between xrec & xbow and KL divergence between the prior (laplace approximation to Dirichlet prior as in ProdLDA) and posterior. We now discuss the details of our Topic Attention Network. 4.1 TAN: Topic Attention Network We intend the model to attend on document words in a manner such that the resultant attention is distributed according to the semantics of the topics relevant to the document. We hypothesize that this can enable the model to encode better document features while capturing the underlying latent document-topic representations. The topic-word distribution Tw represents the affinity of each topic towards words in the vocabulary (which is used to interpret the semantics of each topic). Therefore, we factor Tw ∈RK×|V | into the attention mechanism, where K denotes the number of topics. The topic-aware attention encoder and topic-word distribution influence each other during training which consequently results in convergence to better topics as discussed in detail in Experiments section. Specifically, we perform attention on document sequence of tokens for each topic using the embedded representation of the topics TE ∈RK×E: TE = TwE, [topic embeddings] Tw = softmax(D), [topic-word distribution] where D ∈RK×V is the decoder layer which is used to reconstruct xbow from the sampled latent document-topic representation z as the final step D in Figure 1. The topic embeddings are then used to determine the attention alignment matrix A ∈R|x|×K between each topic k ∈{1, 2, ..., K} and words in the document such that: Ajk = exp(score((TE)k, hj)) P|x| j′=1 exp(score((TE)k, hj′)) , score((TE)k, hj) = vA⊤tanh(WA[(TE)k;hj]) where vA ∈RP , WA ∈RP×(E+H), (TE)k ∈ RE is the embedded representation of the kth topic and ; is the concatenation operation. We then determine topic-wise context vector corresponding to each topic as: CT = |x| X j=1 Aj ⊗hj, [topic-wise context matrix] where ⊗denotes outer product. Note that Aj ∈RK (jth row of matrix A) is a K - dimensional vector and hj is a H - dimensional vector, therefore Aj ⊗hj for each j yields a matrix of order K × H, hence CT ∈RK×H. The final aggregated context vector c is computed as a weighted average over all rows of CT (each row representing each topic specific context vector) with document-topic proportion vector td as weights: c = K X k=1 (td)i(CT)k where, (td)k is a scalar, (CT)k ∈RH denotes the kth row of matrix CT & td is the documenttopic distribution which signifies the topic proportions in a document. To compute it, we first normalize the document BoW vector xbow and embed it using the embedding matrix E, followed by multiplication with topic embedding TE ∈RK×E: xnorm = xbow P|V | i=1(xbow)i , [normalized BoW] xemb = x⊤ normE, [document embedding] td = softmax(TE xemb), [document-topic dist.] where xnorm ∈R|V |, xemb ∈RE & td ∈RK. The context vector c is the output of our topic guided attention module which is then used for sampling the latent documents-topic vector followed by the BoW decoding as done in traditional VAE based topic models. 3870 We call this framework as Weighted-TAN or WTAN where the context vector c is a weighted sum of topic-wise context vectors. We also propose another model called Top-TAN or T-TAN where we use context vector of the topic with largest proportion in td as c. It has been experimentally observed that doing so yields a model which generates more coherent topics. First, we find the index m of most probable topic in td. The context vector c is then the row corresponding to index m in matrix CT. 5 Experiments 5.1 Datasets 1. Topic Quality: We evaluate and compare quality of our proposed topic model on three benchmark datasets - 20Newsgroups (20NG)1 (Lang, 1995), AGNews (Zhang et al., 2015) and Yelp Review Polarity (YRP)2 - which are of varying complexity and scale in terms of number of documents, vocabulary size and average length of text after preprocessing3. Table 1 summarises statistics related to these datasets used for evaluating topics quality. Dataset # Train # Test vocab avg.doc.len. 20NG 11259 7488 1995 88.06 AGNews 96000 7600 27881 22.72 YRP 447873 38000 20001 54.46 Table 1: Datasets used for evaluating topic quality 2. Keyphrase Generation: Neural Topic Model (NTM) has been used to improve the task of supervised keyphrase generation (Wang et al., 2019b). To further highlight the efficacy of our proposed encoding framework in providing better document-topic vectors, we modify encoder module of NTM with our proposed TAN-NTM and compare the performance on StackExchange and Weibo Datasets4. 5.2 Implementation and Training Details Documents in AGNews are padded upto a maximum length of 50, while those in 20NG and YRP are padded upto 200 tokens. Documents with longer lengths are truncated. These values were chosen such that ∼80 −99% of all documents in each dataset were included without truncation. We 1Data link for 20NG dataset 2Data link for AGNews and YRP datasets 3We provide our detailed preprocessing steps in Appendix A.1 and release processed data to standardise it. 4The dataset details can be found in the baseline paper use batch size of 100, Adam Optimizer (Kingma and Ba, 2015) with β1 = 0.99, β2 = 0.999 and ϵ = 10−8 and train each model for 200 epochs. For all models except T-TAN, learning rate was fixed at 0.002 ([0.001, 0.003], 5)5. T-TAN converges relatively faster than other models, therefore for smooth training, we decay its learning rate every epoch using exponential staircase scheduler with initial learning rate = 0.002 and decay rate = 0.96. The number of topics K = 50, a value widely used in literature. We perform hyper-parameter tuning manually to determine the hidden dimension value of various layers: E = 200 ([100, 300], 5), H = 450 ([300, 900], 10) and P = 350 ([10, 400], 10). The weight matrices of all dense layers are Xavier initialized, while bias terms are initialized with zeros. All our proposed models and baselines are trained on a machine with 32 virtual CPUs, single NVIDIA Tesla V 100 GPU and 240 GB RAM. 5.3 Comparison with baselines We compare our TAN-NTM with various baselines in table 2 that can be enumerated as (please refer to introduction and related work for their details): 1) LDA (C.G.): Statistical method (McCallum, 2002) which performs LDA using collapsed Gibbs6 sampling. 2) ProdLDA and 3) NVLDA (Srivastava and Sutton, 2017): Neural Variational Inference methods which use approximation to Dirichlet prior7. 4) W-LDA (Nan et al., 2019) which is a non variational inference based neural model using wassestein autoencoder8. 5) NB-NTM and 6) GNB-NTM: Methods using negative binomial and gamma negative binomial distribution as priors for topic discovery9(Wu et al., 2020) respectively. We could not compare with other methods whose official error-free source code is not publicly available yet. We train and evaluate the baseline methods on same data as used for our method using NPMI coherence10 (Aletras and Stevenson, 2013). It computes the semantic relatedness between top L words in a given topic through determining similarity between their word embeddings trained over the 5V ([a, b], t) means t values from [a, b] range tried for this hyper-parameter, of which V yielded best NPMI coherence. 6https://pypi.org/project/lda/ 7Code for ProdLDA and NVLDA 8https://github.com/awslabs/w-lda 9We thank authors for providing code and parameter info. 10Repo used to calculate NPMI. Please refer to Appendix B for a detailed discussion on choice of evaluation metric. 3871 Method 20NG AGNews YRP LDA(C.G) 0.139 0.202 0.114 NVLDA 0.2 0.216 0.165 ProdLDA 0.268 0.322 0.165 W-LDA 0.227 0.262 0.25 NB-NTM 0.165 0.31 0.224 GNB-NTM 0.206 0.312 0.241 W-TAN (ours) 0.261 0.327 0.232 T-TAN (ours) 0.296 0.369 0.272 Table 2: NPMI coherence (determined using top 10 words of each topic) comparison on 50 topics between baselines and our proposed W-TAN and T-TAN on different datasets. It can be seen that T-TAN achieves significantly better scores on all the datasets. corpus used for topic modeling and reports average over topics. For W-LDA, we refer to their original paper to select dataset specific hyper-parameter values while training the model. As can be seen in table 2, our proposed T-TAN model performs significantly better than previous topic models uniformly on all datasets achieving a better NPMI (measured on a scale of -1 to 1) by a margin of 0.028 (10.44%) on 20NG, 0.047 (14.59%) on AGNews and 0.022 (8.8%) on YRP, where percentage improvements are determined over the best baseline score. Even though W-TAN does not uniformly performs better than all baselines on all datasets, it achieves better score than all baselines on AGNews and performs comparably on remaining two datasets. For a more exhaustive comparison, we also evaluate our model’s performance on 20NG dataset (which is the common dataset with GNB-NTM (Wu et al., 2020)) using the NPMI metric from GNB-NTM’s code. The NPMI coherence of our model using their criteria is 0.395 which is better than GNB-NTM’s score of 0.375 (as reported in their paper). However, we would like to highlight that GNB-NTM’s computation of NPMI metric uses relaxed window size, whereas the metric used by us (Lau et al., 2014) uses much stricter window size while determining word co-occurrence counts within a document. Lau et al. (2014) is a much more common and widely used way of computing the NPMI coherence and evaluating topic models. 5.3.1 Document Classification In addition to evaluating our framework in terms of topic coherence, we also compare it with the baselines on the downstream task of document classification. Topic models have been used as text feature extractors to perform classification (Nan et al., 2019). We analyse the quality of encoded document representations and predictive capacity of latent document-topic features generated by our model and compare it with existing topic models11. We train the topic model setting number of topics to 50 and freeze its weights. The trained topic model is then used to infer latent document-topic features. We then separately train a single layer linear classifier through cross entropy loss on the training split using the document-topic vectors as input and Adam optimizer at a learning rate of 0.01. Method 20NG AGNews YRP LDA(C.G.) 51.29 84.78 86.85 ProdLDA 21.33 82.65 77.73 NTM-R 43.34 85.67 86.16 W-LDA 43.08 85.29 85.63 NB-NTM 57.38 86.67 87.51 GNB-NTM 57.16 85.34 84.55 T-TAN (ours) 60.44 88.1 87.38 T-TAN (ours) 64.36 89.78 88.9 (context vector) Table 3: Comparison of accuracy between different topic models on document classification. We perform two experiments with T-TAN: using document-topic vector (2nd to last row) and context vector (last row). We report classification accuracy on the test split of 20NG, AGNews and YRP datasets (comprising of 20, 4 and 2 classes respectively) in Table 3. The document-topic features provided by TTAN achieve best accuracy on AGNews (1.43% improvement over most performant baseline) with most significant improvement of 3.06% on 20NG which shows our model learns better document features. T-TAN performs almost the same as the best baseline on YRP. Further, to analyse the predictive performance of top topic attention based context vector, we use it instead of latent documenttopic vector to perform classification which further boosts accuracy leading to an improvement of ∼6.9% on 20NG, ∼3.1% on AGNews and ∼1.3% on YRP datasets over the baselines. 5.3.2 Running Time Analysis We compare the running time of our method with baselines in terms of average time taken (in seconds) for performing a forward pass through the 11Our aim is to analyse document-topic features among topic models only and not to compare with other non-topic model based generic text classifiers. 3872 model, where the average is taken over 10000 passes. Our TAN-NTM (implemented in tensorflow) takes 0.087s, 0.027s and 0.093s on 20NG, AGNews and YRP datasets respectively. Since TAN-NTM processes the input documents as a sequence of tokens through an LSTM, its running time is proportional to the document lengths which vary according to the dataset. The running time for baseline methods are: ProdLDA - 0.012s (implemented in tensorflow), W-LDA - 0.003s (implemented in mxnet) and GNB-NTM - 0.003s (implemented in pytorch). For baseline methods, we have used their original code implementations. We found that the running time of baseline models is independent of the dataset. This is because they use the Bag-of-Words (BoW) representation of the documents. The sequential processing in TAN-NTM is the reason for increased running time of our models compared to the baselines. In the case of AGNews, since the documents are of lesser lengths than 20NG and YRP, the running time of our TANNTM is relatively less for AGNews. Further, the running time of other ablation variants (introduced in section 5.4) of our method on 20NG, AGNews and YRP datasets respectively are: 1) only LSTM 0.083s, 0.033s and 0.091s ; 2) vanilla attn - 0.088s, 0.037s and 0.095s. 5.4 Ablation Studies In this section, we compare the performance of different variants of our model namely, 1) only LSTM: final hidden state is used to derive sampling parameters zµ & zlog σ2, 2) vanilla attn: final hidden state (w/o topic-word distribution) is used as query to perform attention (Bahdanau et al., 2014) on LSTM outputs such that context vector z is used for VI, 3) W-TAN: Weighted Topic Attention Network, 4) T-TAN: Top Topic Attention Network and 5) T-TAN w/o (without) GloVe: embedding layer in T-TAN is randomly initialised. Table 4 compares the topic coherence scores of these different ablation methods on 20NG, AGNews and YRP. As can be seen, applying attention performs better than simple LSTM model. The weighted TAN performs better than vanilla attention model, however, T-TAN uniformly provides the best coherence scores across all the datasets compared to all other methods. This shows that performing attention corresponding to the most prominent topic in a document results in more coherent topics. Further, we perform an ablation to study the effect of using pre-trained embeddings for T-TAN where it can be seen using Glove for initialising word embeddings results in improved NPMI as compared to training T-TAN initialised with random uniform embeddings (T-TAN w/o GloVe)12. Method 20NG AGNews YRP only LSTM 0.247 0.202 0.092 vanilla attn 0.289 0.244 0.18 W-TAN 0.261 0.327 0.232 T-TAN 0.296 0.369 0.272 T-TAN w/o GloVe 0.274 0.344 0.248 Table 4: Comparison of NPMI coherence between ablation variants of our method for K=50 topics. 5.5 Qualitative Analysis To verify performance of T-TAN qualitatively, we display few topics generated by ProdLDA and TTAN on AGNews in Figure 2. ProdLDA achieves best score among baselines on AGNews. Consider comparison 1 in Figure 2: ProdLDA produces four topics corresponding to space, mixing them with nuclear weapons, while T-TAN produces two separate topics for both of these concepts. In second comparison, we see that ProdLDA has problems distinguishing between closely related topics (football, olympics, cricket) and mixes them while TTAN produces three coherent topics. Figure 2: Two comparisons of corresponding topics (one topic per line) from ProdLDA and T-TAN. Words having similar meaning are highlighted in same colour. The topics of ProdLDA are inter-mixed and incoherent while those of T-TAN are unmixed and coherent. 5.6 TAKG: Topic Aware Keyphrase Generation We further analyse the impact of our proposed framework on another downstream task where the 12We also trained embeddings from scratch for other variants but coherence score remained unaffected. 3873 StackExchange Weibo Method F1@3 F1@5 MAP F1@1 F1@3 MAP TAKG (baseline) 32.931 28.731 34.925 34.584 24.309 40.994 TAKG with W-TAN (ours) 33.521 29.802 35.929 35.616 25.651 42.68 TAKG with T-TAN (ours) 33.15 29.118 35.26 34.813 24.65 41.261 Table 5: F1@k and MAP (Mean average precision) comparison between baseline (TAKG) and our proposed topic model based encoder for topic guided supervised keyphrase generation. The metrics measure overlap between ground truth and top K generated keyphrases factoring in rank of keyphrases generated through beam search. task specific model is assisted by the topic model and both can be trained in an end-to-end manner. For this, we discuss TAKG (Wang et al., 2019b) and how our proposed topic model encoder can be adapted to achieve better performance on supervised keyphrase generation from textual posts. TAKG13 comprises of two sub-modules: (1) a topic model based on NVDM-GSM (as discussed in Introduction) using BoW as input to the encoder and (2) a Seq2Seq based model for keyphrase generation. Both modules have an encoder and a decoder of their own. Keyphrase generation module uses sequence input which is processed by bidirectional GRU (Cho et al., 2014) to encode input sequence. The keyphrase generation decoder uses unidirectional GRU which attends on encoder outputs and takes the latent document-topic vector from the topic model as input in a differentiable manner. Since topic model trains slower than keyphrase generation module, the topic model is warmed up for some epochs separately and then jointly trained with keyphrase generation. Please refer to original paper (Wang et al., 2019b) for more details. We adapted our proposed topic model framework by changing the architecture of encoder in the topic model of TAKG, replacing it with W-TAN and T-TAN. The change subsequently results in better latent document-topic representation depicted by better performance on keyphrase generation as shown in Table 5 where the improved topic model encoding framework results in ∼1-2% improvement in F1 and MAP (mean average precision) on StackExchange and Weibo datasets compared to TAKG. Here, even though TAKG with T-TAN performs marginally better than the baseline, TAKG with W-TAN uniformly performs much better. 6 Conclusion In this work, we propose Topic Attention Network based Neural Topic Modeling framework: TAN13We use their code and data (link) to conduct experiments. NTM to discover topics in a document corpus by performing attention on sequentially processed tokens in a topic guided manner. Attention is performed effectively by factoring Topic-word distribution (TWD) into attention mechanism. We compare different variants of our method through ablations and conclude that processing tokens sequentially without attention or applying attention without TWD gives inferior performance. Our TAN-NTM model generates more coherent topics compared to state-of-the-art topic models on several benchmark datasets. Our model encodes better latent document-topic features as validated through better performance on document classification and supervised keyphrase generation tasks. As future work, we would like to explore our framework with other sequence encoders such as Transformers, BERT etc. for topic modeling. References Nikolaos Aletras and Mark Stevenson. 2013. Evaluating topic coherence using distributional semantics. In Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013)–Long Papers, pages 13–22. Christophe Andrieu, Nando De Freitas, Arnaud Doucet, and Michael I Jordan. 2003. An introduction to mcmc for machine learning. Machine learning, 50(1-2):5–43. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. Federico Bianchi, Silvia Terragni, and Dirk Hovy. 2020. Pre-training is a hot topic: Contextualized document embeddings improve topic coherence. ArXiv, abs/2004.03974. Steven Bird, Ewan Klein, and Edward Loper. 2009. Natural Language Processing with Python. O’Reilly Media. 3874 David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022. S. Burkhardt and S. Kramer. 2019. Decoupling sparsity and smoothness in the dirichlet variational autoencoder topic model. J. Mach. Learn. Res., 20:131:1– 131:27. Dallas Card, Chenhao Tan, and Noah A. Smith. 2018. Neural models for documents with metadata. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2031–2040, Melbourne, Australia. Association for Computational Linguistics. Jonathan Chang, Sean Gerrish, Chong Wang, Jordan Boyd-graber, and David Blei. 2009. Reading tea leaves: How humans interpret topic models. In Advances in Neural Information Processing Systems, volume 22, pages 288–296. Curran Associates, Inc. Jun Chen, Xiaoming Zhang, Yu Wu, Zhao Yan, and Zhoujun Li. 2018. Keyphrase generation with correlation constraints. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4057–4066, Brussels, Belgium. Association for Computational Linguistics. Kyunghyun Cho, Bart van Merri¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724– 1734, Doha, Qatar. Association for Computational Linguistics. Scott Deerwester, Susan T. Dumais, George W. Furnas, Thomas K. Landauer, and Richard Harshman. 1990. Indexing by latent semantic analysis. Journal of the American Society for Information Science, 41(6):391–407. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Adji B. Dieng, Francisco J. R. Ruiz, and David M. Blei. 2020. Topic modeling in embedding spaces. Transactions of the Association for Computational Linguistics, 8:439–453. Adji B Dieng, Chong Wang, Jianfeng Gao, and John Paisley. 2016. Topicrnn: A recurrent neural network with long-range semantic dependency. arXiv preprint arXiv:1611.01702. Ran Ding, Ramesh Nallapati, and Bing Xiang. 2018. Coherence-aware neural topic modeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 830– 836, Brussels, Belgium. Association for Computational Linguistics. Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 2672–2680. Curran Associates, Inc. Thomas L Griffiths and Mark Steyvers. 2004. Finding scientific topics. Proceedings of the National academy of Sciences, 101(suppl 1):5228–5235. Pankaj Gupta, Yatin Chaudhary, Florian Buettner, and Hinrich Sch¨utze. 2019. Document informed neural autoregressive topic models with distributional prior. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6505–6512. Suchin Gururangan, T. Dang, D. Card, and Noah A. Smith. 2019. Variational pretraining for semisupervised text classification. In ACL. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural Comput., 9(8):1735–1780. Alexander Miserlis Hoyle, Pranav Goel, and P. Resnik. 2020a. Improving neural topic models using knowledge distillation. ArXiv, abs/2010.02377. Alexander Miserlis Hoyle, Pranav Goel, and Philip Resnik. 2020b. Improving Neural Topic Models using Knowledge Distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1752–1771, Online. Association for Computational Linguistics. Xuemeng Hu, Rui Wang, Deyu Zhou, and Yuxuan Xiong. 2020. Neural topic modeling with cycleconsistent adversarial training. In EMNLP. S. Ioffe and Christian Szegedy. 2015. Batch normalization: Accelerating deep network training by reducing internal covariate shift. ArXiv, abs/1502.03167. Mingmin Jin, Xin Luo, Huiling Zhu, and Hankz Hankui Zhuo. 2018. Combining deep learning and topic modeling for review understanding in context-aware recommendation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1605–1614, New Orleans, Louisiana. Association for Computational Linguistics. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. CoRR, abs/1412.6980. 3875 Diederik P Kingma and Max Welling. 2013. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114. Varun Kumar, Alison Smith-Renner, Leah Findlater, K. Seppi, and Jordan L. Boyd-Graber. 2019. Why didn’t you listen to me? comparing user control of human-in-the-loop topic models. In ACL. Ken Lang. 1995. Newsweeder: Learning to filter netnews. In Machine Learning Proceedings 1995, pages 331–339. Elsevier. Jey Han Lau, David Newman, and Timothy Baldwin. 2014. Machine reading tea leaves: Automatically evaluating topic coherence and topic model quality. In Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 530–539, Gothenburg, Sweden. Association for Computational Linguistics. Andrew Kachites McCallum. 2002. Mallet: A machine learning for language toolkit. http://mallet. cs. umass. edu. Rui Meng, Sanqiang Zhao, Shuguang Han, Daqing He, Peter Brusilovsky, and Yu Chi. 2017. Deep keyphrase generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 582–592, Vancouver, Canada. Association for Computational Linguistics. Yishu Miao, Edward Grefenstette, and Phil Blunsom. 2017. Discovering discrete latent topics with neural variational inference. volume 70 of Proceedings of Machine Learning Research, pages 2410–2419, International Convention Centre, Sydney, Australia. PMLR. Yishu Miao, Lei Yu, and Phil Blunsom. 2016. Neural variational inference for text processing. volume 48 of Proceedings of Machine Learning Research, pages 1727–1736, New York, New York, USA. PMLR. Feng Nan, Ran Ding, Ramesh Nallapati, and Bing Xiang. 2019. Topic modeling with Wasserstein autoencoders. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6345–6381, Florence, Italy. Association for Computational Linguistics. Nicole Peinelt, Dong Nguyen, and Maria Liakata. 2020. tBERT: Topic models and BERT joining forces for semantic similarity detection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7047–7055, Online. Association for Computational Linguistics. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Gabriele Pergola, Lin Gui, and Yulan He. 2019. Tdam: a topic-dependent attention model for sentiment analysis. Inf. Process. Manag., 56. Radim ˇReh˚uˇrek and Petr Sojka. 2010. Software Framework for Topic Modelling with Large Corpora. In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks, pages 45–50, Valletta, Malta. ELRA. Mehdi Rezaee and F. Ferraro. 2020. A discrete variational recurrent topic model without the reparametrization trick. ArXiv, abs/2010.12055. Suzanna Sia, Ayush Dalmia, and Sabrina J. Mielke. 2020. Tired of topic models? clusters of pretrained word embeddings make for fast and good topics too! In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1728–1736, Online. Association for Computational Linguistics. Akash Srivastava and Charles Sutton. 2017. Autoencoding variational inference for topic models. arXiv preprint arXiv:1703.01488. Nitish Srivastava, Geoffrey E. Hinton, A. Krizhevsky, Ilya Sutskever, and R. Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res., 15:1929–1958. Mark Steyvers and Tom Griffiths. 2007. Probabilistic topic models. Handbook of latent semantic analysis, 427(7):424–440. Romain Thibaux and Michael I. Jordan. 2007. Hierarchical beta processes and the indian buffet process. volume 2 of Proceedings of Machine Learning Research, pages 564–571, San Juan, Puerto Rico. PMLR. Ilya Tolstikhin, Olivier Bousquet, Sylvain Gelly, and Bernhard Schoelkopf. 2017. Wasserstein autoencoders. arXiv preprint arXiv:1711.01558. Felipe Viegas, Washington Cunha, Christian Gomes, Antˆonio Pereira, Leonardo Rocha, and Marcos Goncalves. 2020. CluHTM - semantic hierarchical topic modeling based on CluWords. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8138–8150, Online. Association for Computational Linguistics. Rui Wang, Xuemeng Hu, Deyu Zhou, Yulan He, Yuxuan Xiong, Chenchen Ye, and Haiyang Xu. 2020. Neural topic modeling with bidirectional adversarial training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 340–350, Online. Association for Computational Linguistics. Rui Wang, Deyu Zhou, and Yulan He. 2019a. Atm: Adversarial-neural topic model. Information Processing & Management, 56(6):102098. 3876 X. Wang and Y. Yang. 2020. Neural topic model with attention for supervised learning. In AISTATS. Yue Wang, Jing Li, Hou Pong Chan, Irwin King, Michael R. Lyu, and Shuming Shi. 2019b. Topicaware neural keyphrase generation for social media language. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2516–2526, Florence, Italy. Association for Computational Linguistics. Jiemin Wu, Yanghui Rao, Zusheng Zhang, Haoran Xie, Qing Li, Fu Lee Wang, and Ziye Chen. 2020. Neural mixed counting models for dispersed topic discovery. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6159–6169, Online. Association for Computational Linguistics. M. Zaheer, Amr Ahmed, and Alex Smola. 2017. Latent lstm allocation: Joint clustering and non-linear dynamic modeling of sequence data. In ICML. Ce Zhang and Hady W. Lauw. 2020. Topic modeling on document networks with adjacent-encoder. In AAAI. Hao Zhang, B. Chen, D. Guo, and M. Zhou. 2018. Whai: Weibull hybrid autoencoding inference for deep topic modeling. arXiv: Machine Learning. Qi Zhang, Yang Wang, Yeyun Gong, and Xuanjing Huang. 2016. Keyphrase extraction using deep recurrent neural networks on twitter. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 836–845, Austin, Texas. Association for Computational Linguistics. Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in neural information processing systems, pages 649–657. He Zhao, Dinh Phung, Viet Huynh, Trung Le, and Wray Buntine. 2021. Neural topic model via optimal transport. In International Conference on Learning Representations. Deyu Zhou, Xuemeng Hu, and Rui Wang. 2020. Neural topic modeling by incorporating document relationship graph. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3790–3796, Online. Association for Computational Linguistics. Jun-Yan Zhu, T. Park, Phillip Isola, and Alexei A. Efros. 2017. Unpaired image-to-image translation using cycle-consistent adversarial networks. 2017 IEEE International Conference on Computer Vision (ICCV), pages 2242–2251. Li-Xing Zhu, Yulan He, and Deyu Zhou. 2020. A neural generative model for joint learning topics and topic-specific word embeddings. Transactions of the Association for Computational Linguistics, 8:471– 485. 3877 Appendices A Further Implementation Details A.1 Preprocessing For 20NG dataset, we used its preprocessed version downloaded from ProdLDA’s (Srivastava and Sutton, 2017) repository14, whereas AGNews and YRP datasets were downloaded from this15 link. These two datasets contain train.csv and test.csv files. The csv files of YRP contain a document body only, whereas the csv files for AGNews contain a document title as well as a document body. For uniformity, we concatenate the title and body in the csv files of AGNews and keep it as a single field. The documents from train.csv and test.csv are then read into train and test lists which are passed to PREPROCESS function of Algorithm 1 for preprocessing. Stepwise working of Algorithm 1 is expained in the following points: • Before invoking the PREPROCESS function, we initialize the data sampler by a fixed seed so that preprocessing yields the same result when run multiple times. • For each dataset, we randomly sample tr size documents (as mentioned in Table 6) from the train list in step 2. These values of tr size are taken from Table 1 of W-LDA paper (Nan et al., 2019). Note that # Train in Table 1 represents the number of training documents after preprocessing. Of the tr size documents, some documents may be removed during preprocessing, therefore # Train may be less than tr size. • In steps 3 through 8, we prune the train and test documents by invoking the PRUNE DOC function from Algorithm 2. First, we remove the control characters from the documents viz. ‘\n’, ‘\t’, and ‘\r’ (For YRP, we additionally remove ‘\\t’, ‘\\n’, and ‘\\r’). Next, we remove the numeric tokens16 from the documents, convert them to lowercase and lemmatize each of their tokens using the 14Data link for 20NG dataset 15Data link for AGNews and YRP datasets 16Fully numeric tokens e.g. ‘1487’, ‘1947’, etc. are removed, whereas partially numeric tokens e.g. ‘G47’, ‘DE1080’, etc. are retained. NLTK’s (Bird et al., 2009) WordNetLemmatizer. Finally, we remove punctuations17 and tokens containing any non-ASCII character. • In steps 9 through 15, we construct the vocabulary vocab, which is a mapping of each token to its occurrence count among the pruned training documents tr pruned. We only count a token if it is not an English stopword18 and its length is between 3 and 15 (inclusive). • Steps 16 through 19 filter the vocab by removing tokens whose total occurrence count is less than num below or whose occurrence count per training document is greater than fr abv, where the values of num below and fr abv are taken from Table 6. For YRP, we follow the W-LDA paper (Nan et al., 2019) and restrict its vocab to only contain top 20, 000 most occurring tokens. • Steps 20 through 24 construct the token-toindex map w2idx by mapping each token in vocab to an index starting from 1. Next, we map the padding token to index 0 (Step 25). • The final step in the preprocessing is to encode the train and test documents by mapping each of their tokens to corresponding indices according to w2idx. This is done by the ENCODE function of Algorithm 2 which is invoked in steps 26 and 27. Dataset tr size num below fr abv AGNews 96000 3 0.7 YRP 448000 20 0.7 Table 6: Parameters used for preprocessing the AGNews and YRP datasets. 17Any of the following 32 characters is regarded as a punctuation !”#$%&’()*+,-./:;<=>?@[\]ˆ `{|}∼ 18Gensim’s ( ˇReh˚uˇrek and Sojka, 2010) list of English stopwords is used. 3878 Algorithm 1 Pseudocode for preprocessing AGNews and YRP datasets. 1: function PREPROCESS(train, test) 2: train ←train.sample(tr size) 3: tr pruned ←[] ▷empty list 4: te pruned ←[] ▷empty list 5: for document d in train do 6: tr pruned.append(PRUNE DOC(d)) 7: for document d in test do 8: te pruned.append(PRUNE DOC(d)) 9: vocab ←mapping of each token to 0 10: num doc ←len(tr pruned) 11: for document d in tr pruned do 12: for token t in d do 13: if t /∈stopwords and 14: len(t) ∈[3, 15] then 15: vocab[t]←vocab[t] +1 16: for token t in vocab do 17: if vocab[t] < num below or 18: vocab[t]/num doc > fr abv then 19: vocab[t].remove(t) 20: i ←1 21: w2idx ←empty map 22: for token t in vocab do 23: w2idx[t]= i 24: i ←i + 1 25: w2idx[0]←PAD 26: trD ←ENCODE(tr pruned, w2idx) 27: teD ←ENCODE(te pruned, w2idx) 28: return trD, teD, w2idx A.2 Learning Rate Scheduler As mentioned in section 5.2, we use a learning rate scheduler while training T-TAN. The rate decay follows the following equation: lrate = init rate ∗decay rate j train step decay steps k This is an exponential staircase function which enables decrease in learning rate every epoch during training. We initialize the learning rate by init rate = 0.002 and use decay rate = 0.96. train step is a Algorithm 2 Pseudocode for pruning the document and encoding it given a token-to-index mapping. 1: function PRUNE DOC(doc) 2: doc ←rm control(doc) 3: doc ←rm numeric(doc) 4: doc ←lowercase(doc) 5: doc ←lemmatize(doc) 6: doc ←rm punctuations(doc) 7: doc ←rm non ASCII(doc) 8: return doc 9: function ENCODE(doc list, w2idx) 10: encDocList ←[] 11: for document d in doc list do 12: ecDoc ←[] 13: for token t in d do 14: ecDoc.append(w2idx[t]) 15: encDocList.append(ecDoc) 16: return encDocList global counter of training steps and decay steps = #train docs batch size is the number of training steps taken per epoch. Therefore, effectively, the rate remains constant for all training steps in an epoch and decreases exponentially as per the above equation once the epoch completes. A.3 Regularization We employ two types of regularization during training: • Dropout: We apply dropout (Srivastava et al., 2014) to z with the rate of Pdrop = 0.6 before it is processed by the decoder for reconstruction. • Batch Normalization (BN): We apply a BN (Ioffe and Szegedy, 2015) to the inputs of decoder layer and to the inputs of layers being trained for zµ & zlog σ2, with ϵ = 0.001 and decay = 0.999. B Evaluation Metrics Topic models have been evaluated using various metrics namely perplexity, topic coherence, topic uniqueness etc. However, due to the absence of a gold standard for the unsupervised task of topic modeling, all of that metrics have received criticism by the community. Therefore, a consensus on the best metric has not been reached so far. Perplexity has been found to be negatively correlated to 3879 topic quality and human judgements (Chang et al., 2009). This work presents experimental results which show that in some cases models with higher perplexity were preferred by human subjects. Topic Uniqueness (Nan et al., 2019) quantifies the intersection among topic words globally. However, it also suffers from drawbacks and often penalizes a model incorrectly (Hoyle et al., 2020b). Firstly, it does not account for ranking of intersected words in the topics. Secondly, it fails to distinguish between the following two scenarios: 1) When the intersected words in one topic are all present in a second topic (signifying strong similarity i.e. these two topics are essentially identical) and, 2) When the intersected words of one topic are spread across all the other topics (signifying weak similarity i.e. the topics are diffused). The first is a problem related to uniqueness among topics while second is a problem related to word intrusion in topics. (Chang et al., 2009) conducted experiments with human subjects on two tasks: word intrusion and topic intrusion. Word intrusion measures the presence of those words (called intruder words) which disagree with the semantics of the topic. Topic intrusion measures the presence of those topics (called intruder topics) which do not represent the document corpus appropriately. These are better estimates of human judgement of topic models in comparison to perplexity and uniqueness. However, since these metrics rely on human feedback, they cannot be widely used for unsupervised evaluation. Further, topic uniqueness unfairly penalizes cases when some words are common between topics, however other uncommon words in those topics change the context as well as topic semantics as also discussed in (Hoyle et al., 2020b). According to the work of (Lau et al., 2014), measuring the normalized pointwise mutual information (NPMI) between all the word pairs in a set of topics agrees with human judgements most closely. This is called the NPMI Topic Coherence in the literature and is widely used for the evaluation of topic models. We therefore adopt this metric in our work. Since the effectiveness of a topic model actually depends on the topic representations that it extracts from the documents, we report the performance of our model on two downstream tasks: document classification and keyphrase generation (which use these topic representations) for a better and holistic evaluation and comparison. Would a pilot know that one of their crew is armed? The Federal Flight Deck Officer page on Wikipedia says this: Under the FFDO program, flight crew members are authorized to use firearms. A flight crew member may be a pilot, flight engineer or navigator assigned to the flight. To me, it seems like this would be crucial information for the PIC to know, if their flight engineer (for example) was armed; but on the flip-side of this, the engineer might want to keep that to himself if he’s with a crew he hasn’t flown with before. Is there a guideline on whether an FFDO should inform the crew that he’s armed? GT: security, crew, ffdo TAKG: faa regulations, ffdo, flight training, firearms, far TAKG + W-TAN: ffdo, crew, flight controls, crewed spaceflight, security Do the poisons in “Ode on Melancholy” have deeper meaning? In ”Ode on Melancholy”, Keats uses the images of three poisons in the first stanza: Wolf’s bane, nightshade, and yew-berries. Are these poisons simply meant to connote death/suicide, or might they have a deeper purpose? GT: poetry, meaning, john keats TAKG: the keats, meaning, poetry, ode, melancholy keats TAKG + W-TAN: poetry, meaning, the keats, john keats, greek literature Table 7: Two randomly selected posts (title in bold) from StackExchange dataset with ground truth (GT) and top 5 keyphrases predicted by TAKG with and without W-TAN, denoted as TAKG + W-TAN & TAKG respectively. Keyphrases generated with WTAN are closer to the ground truth in terms of both prediction and ranking. C Qualitative Analysis C.1 Key Phrase Predictions We saw the quantitative improvement in results in Table 5 when we used W-TAN as the topic model 3880 with TAKG. In Table 7, we display some posts from StackExchange dataset with ground truth keyphrases and top 5 predictions by TAKG with and without W-TAN. We observe that using WTAN improves keyphrase generation qualitatively. The first post in Table 7 inquires if a flight officer should inform the pilot in command (PIC) about him being armed or not. For this post, TAKG alone only predicts one ground truth keyphrase correctly and misses ‘security’ and ‘crew’. However, when TAKG is used with W-TAN, it gets all three ground truth keyphrases, two of which are its top 2 predictions as well. The second post is inquiring about a possible deeper meaning of three poisons in a poem by John Keats. TAKG alone predicts two of the ground truth keyphrases correctly but assigns them larger ranks and it misses ‘john keats’. When TAKG is used with W-TAN, it gets all three ground truth keyphrases and its top 2 keyphrases are assigned the exact same rank as they have in the ground truth. This hints that using W-TAN with TAKG improves the prediction as well as ranking of the generated keyphrases compared to using TAKG alone.
|
2021
|
299
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 29–40 August 1–6, 2021. ©2021 Association for Computational Linguistics 29 Engage the Public: Poll Question Generation for Social Media Posts Zexin Lu1 Keyang Ding1 Yuji Zhang1 Jing Li1∗Baolin Peng2 Lemao Liu3 1Department of Computing, The Hong Kong Polytechnic University, HKSAR, China 2Microsoft Research, Redmond, WA 3Tencent AI Lab, Shenzhen, China 1{zexin.lu, keyang.ding, yu-ji.zhang}@connect.polyu.hk [email protected] [email protected] [email protected] Abstract This paper presents a novel task to generate poll questions for social media posts. It offers an easy way to hear the voice from the public and learn from their feelings to important social topics. While most related work tackles formally-written texts (e.g., exam papers), we generate poll questions for short and colloquial social media messages exhibiting severe data sparsity. To deal with that, we propose to encode user comments and discover latent topics therein as contexts. They are then incorporated into a sequence-to-sequence (S2S) architecture for question generation and its extension with dual decoders to additionally yield poll choices (answers). For experiments, we collect a large-scale Chinese dataset from Sina Weibo containing over 20K polls. The results show that our model outperforms the popular S2S models without exploiting topics from comments and the dual decoder design can further benefit the prediction of both questions and answers. Human evaluations further exhibit our superiority in yielding high-quality polls helpful to draw user engagements. 1 Introduction Social media is a crucial outlet for people to exchange ideas, share viewpoints, and keep connected with the world. It allows us to hear the public voice for decision making and better understanding our society. Nevertheless, for the silent majority, they tend to read others’ messages instead of voicing their own opinions with words, possibly because of the introvert personality, busy schedule, and others. How shall we better engage them into the discussions and learn from their thoughts? In this work, we present a novel application to automatically generate a poll question for a social media post. It will encourage public users, especially those reluctant to comment with words, to ∗Jing Li is the corresponding author. [P1]: ...B站市值超过爱奇艺(The market value of B site exceeds iQiyi)... [Q1]: 你们平时常用那个app看视频？(Which app do you usually use to watch videos?) [A1]: 腾讯视频(Tencent Video); 优酷(Youku); 爱奇艺 (iQiyi); B站(B site) [P2]: ...理性分析一下赵粤和希林娜依高：希 林vocal确实厉害，但是...舞蹈实力有点不够看；赵 粤呢舞蹈厉害...但是唱歌实力较弱些... (A rational analysis of Akira and Curley G: Curley’s vocal is indeed great, but ... her dancing is not that good; Akira dances well ... but her singing is weaker...) [Q2]: 谁更适合当c位？(Who should take the center position?) [A2]: 赵粤(Akira); 希林娜依高(Curley G) Figure 1: Example polls from Sina Weibo. Pi, Qi, and Ai (i = 1, 2) refer to the i-th source post, its poll question, and the corresponding poll choices (answers). Different choices are separated by the “;”. Italic words in “()” are the English translation of the original Chinese texts on their left. In the source posts, we fold the words irrelevant to polls in “...” for easy reading. input their reflections via voting. For example, the statistics of our dataset show that 13K users on average engaged in a poll compared with 173 commented to a post. For a better illustration of the task, Figure 1 shows two example poll questions on Sina Weibo1, henceforth Weibo, a popular Chinese microblog. The goal of our task is to output an opinion question, such as Q1 and Q2, and invite other users to engage in the discussion to a source post (e.g., P1 and P2); poll choices (answers like A1 and A2) can be produced together to allow easy public engagement (via voting). To date, most progress made in question generation is built upon the success of encoder-decoder frameworks (Du et al., 2017). Despite of the extensive efforts made in this line (Sun et al., 2018; Yao et al., 2018; Chai and Wan, 2020; Sun et al., 2020), most previous work focus on the processing of formally-written texts, such as exam questions 1weibo.com 30 in reading comprehension tests. The existing methods are therefore suboptimal to handle social media languages with short nature and informal styles, which might present challenges to make sense of the source posts and decide what to ask. For example, from the limited words in P1, it is hard to capture the meanings of “B站” (B site) and “爱奇 艺” (iQiyi) as video apps, which is nevertheless crucial to predict Q1. Moreover, the question itself, being in social media fashion, is likely to contain fresh words, such as “c位” (center position) in Q2, which may further hinder the models’ capability to predict the poll questions in social media style. To tackle these challenges, we first enrich the short contexts of source posts with other users’ comments; a neural topic model is employed to discover topic words therein and help identify the key points made in source posts. It is based on the assumption that the salient words in a source post are likely to be echoed in its comments (Wang et al., 2019b), potentially useful to learn the map from posts to poll questions. For example, the core words in Q1 — “app” and “视频” (video) — co-occur frequently in the comments with “B站” (B site) and “爱奇艺” (iQiyi), which may help the model to link their meanings together. The topic representations are then incorporated into a sequence-to-sequence (S2S) architecture to decode poll questions word by word. Furthermore, we extend the basic S2S to a version with dual decoders to generate questions and answers in a multi-task learning setting and further exploit their correlations. For example, modeling answers in A2 might help indicate that P2 centers around “赵粤” (Akira) and “希林娜依高” (Curley G), two celebrities. To the best of our knowledge, this work is the first to study poll questions on social media, where their interactions among answer choices, source posts, and reader users’ comments are comprehensively explored. As a pilot study over social media polls, we also contribute the very first dataset containing around 20K Weibo polls associated with their source posts and user comments.2 We believe our dataset, being the first of its kind, will largely benefit the research on social media polls and how they help promote the public engagements. On our dataset, we first compare the model performance on poll question generation in terms of automatic evaluation and human evaluation. The 2Our dataset and code are publicly available in https://github.com/polyusmart/Poll-Question-Generation automatic evaluation results show that the latent topics learned from the first few pieces of user comments is already helpful — they result in our models’ significantly better performance than the S2S baselines and their trendy extensions proposed for other tasks. For example, our full model achieves 38.24 ROUGE-1 while S2S with RoBERTa (Liu et al., 2019) yields 34.08. Human evaluation further demonstrates our models’ capability to generate poll questions relevant to the source post, fluent in language, and particularly engaging to draw user attentions for discussions. We then quantify models’ sensitivities to the length of varying source posts and poll questions, where the scores of our model are consistently better. Next, we find our model exhibits an increasing trend in predicting poll questions that will engage more comments in the future, which suggests the potential helpfulness of comments to indicate engaging questions. At last, the performance of dual decoder designs are discussed and it is shown that joint prediction of questions and their answers can benefit both tasks. 2 Study Design 2.1 Task Formulation Our major input is a social media post (i.e., source post) and the main output a poll question that continue the senses of the source post and encourage public users to voice opinions. For each question, possible answer choices (i.e., answers) may also be yielded as a side product to enable participants to easily input their thoughts. To enrich the contexts of source posts, their reply messages (i.e., user comments) are also encoded as external features. 2.2 Data Description Here we describe the dataset we collect to empirically study social media polls. Data Collection. Weibo allows users to create polls, asking questions to the public and inviting others to share their thoughts via voting. It enables the construction of a dataset with user-generated polls. At the beginning, we gathered around 100K random Weibo posts, whereas less than 0.1% of them contain polls. The sparse distribution of polls presents the challenge to scale up the dataset. To deal with that, we looked in to the sampled polls and draw two interesting points: first, many polls carry trendy hashtags (user-annotated topic labels like #COVID19) to draw user attentions; second, a user who once created a poll is likely to do it again. 31 Post Comment Qs Ans Choice Voter Num Len Num Len Len Num Len Num 20,252 54.0 173 16.9 11.0 3.4 5.9 13,004 Table 1: Statistics of our dataset. Num: number; Num: average number per post. Len: average count of words per post; Qs: question; Ans: answer. Inspired by these observations, we first obtained the popular hashtags since Nov 2019.3 Then, we gathered the posts under the hashtag through the Weibo search API, from which the ones containing polls are picked out.4 Next, we examined the authors of these polls and access their posting history to gather more polls they created from Weibo user timeline API.5 Afterwards, for each post, we crawled its comments via the comment API.6 Finally, 20,252 polls were obtained from 1,860 users. Data Analysis. The statistics of the dataset is displayed in Table 1. As can be seen, comments are shorter than posts, probably because users tend to put more efforts in crafting original posts than replying to others and hence comments may be relatively nosier than original posts; both questions and answers are short, which follow the fashion of user-generated contents on social media. To further investigate the data sparsity in social media contents, we sample some texts from LDC news corpus (formally-written texts) (Ahtaridis et al., 2012) — the samples contain the same token number as our social media texts. Our corpus’s vocabulary size and entropy are 24,884 and 7.46, while those for news corpus are 9,891 and 5.98. This suggests the sparsity of social media data. We also observe that each post exhibits more voters than comments, implying that users may prefer to voice opinions via voting, which is easier than commenting with words. We further analyze the effects of polls on user engagements and draw an interesting finding. For the same author, their posts with polls exhibit 1.65, 22.2, and 1.80 times comments, likes, and reposts on average compared to posts without polls.7 This implies that adding polls indeed help to draw user engagements to a post. 3https://open.weibo.com/wiki/Trends/en 4https://open.weibo.com/wiki/C/2/ search/statuses/limited 5https://open.weibo.com/wiki/C/2/ statuses/user_timeline_batch 6https://open.weibo.com/wiki/2/ comments/show 7For each author, we additionally sample 500 posts without polls for comparison. (a) Choice Number Statistics (b) Topic Categories Figure 2: The left figure shows the count of polls over varying choice number in their answers (x-axis: choice number; y-axis: vote count). The right one displays the distribution of the polls’ topic categories. For each poll, there are less than 4 answer choices on average. To further characterize that, Figure 2(a) shows the count of polls over varying numbers of answer choices appearing in them and the statistics suggest that most users are not willing to craft over 5 poll choices, which, interestingly, exhibit similar statistics in exam questions. In addition, we probe into what types of topics are more likely to contain polls. To that end, we examined source posts with hashtags and manually categorized the hashtags into 11 topics. Figure 2(b) shows the poll distribution over topics. Most polls fall in “social events” category, which mostly concern public emergency and in our dataset tremendous posts focus on the outbreak of COVID-19. There are also a large proportion of polls concern entertainment topics such as celebrities and TV shows, probably initiated for advertising purpose. 3 Poll Question Generation Framework This section introduces our framework with two variants: one based on a basic S2S (single decoder) and the other is its extension with dual decoders to predict poll questions and answer choices in a multitask learning setting. The model architecture of the dual decoder model is shown in Figure 3. 3.1 Source Posts and Comments Encoding Following the common practice in S2S (Du et al., 2017), we encode a source post P in the form of word sequence ⟨w1, w2, ..., w|P|⟩, where |P| is the number of words in the post. For user comments C, bag of words (BOW) representations are employed for topic modeling, henceforth Cbow over BoW vocabulary. More details are provided below. Source Post Encoding. To encode the post sequence P, a bidirectional gated recurrent unit (BiGRU) (Cho et al., 2014) is adopted. For the i-th word wi ∈P, we first convert it into an embedding vector νi, which is later processed into hidden 32 Figure 3: The architecture of the dual decoder S2S (sequence-to-sequence) model to jointly generate questions and answers. It contains a neural topic model for context modeling (in the bottom), a sequence encoder fed with the source post (in the center), and two sequence decoders to handle the output, where the left one predicts questions (Q) and the right answers (A). states in the forward (−→ hi) and backward (←− hi) directions, respectively. They are then concatenated as hi = [−→ hi; ←− hi] and sequentially put into a memory bank M = ⟨h1, h1, ..., h|P|⟩, which will be further delivered to decoders for their attentive retrieval. User Comments Modeling. Considering the noisy nature of user comments, latent topics are employed to recognize the salient contents therein. They are explored based on word statistics and represented as clusters of words tending to co-occur in the comments of some posts (probably concerning similar topics), such as the names of video apps in Figure 1. In topic modeling, we assume there are K topics and each topic k is represented with a topic-word distribution over the BoW vocabulary. A post P has a topic mixture θ, which is learned from the words appearing in its comments Cbow. Our topic learning methods (from comments) are inspired by the neural topic model (NTM) based on variational auto-encoder (VAE) (Miao et al., 2017; Zeng et al., 2018), which allows the end-to-end training of NTM with other modules in an unified neural architecture. It employs an encoder and a decoder to resemble the data reconstruction process of the comment words in BoW. Concretely, the input Cbow is first encoded into prior parameters µ and σ using neural perceptrons. Then, through Gaussian transformation, they are applied to draw a latent variable: z = N(µ, σ2), which is further taken to produce the topic composition of comments (θ) with softmax transformation. At last, the decoder reconstructs comments and produces a BOW vector C′ bow (conditioned on the latent topic θ) through another neural perception. 3.2 Poll Decoding Here we further describe how we generate questions (and answers in the dual decoders settings) with the encoded source posts and comments. Question Generation. To handle the output of a question Q, the corresponding decoder (i.e., question decoder) is formed with a uni-directional GRU and fed with the memory bank M from source post encoding and the topic distribution θ from user comment modeling. The words in Q are predicted sequentially with the following formula: Pr(Q | P, Cbow) = |q| Y j=1 Pr (qj | q<j, M, θ) (1) where qj means the j-th word in Q and q<j refers to Q’s predicted word sequence from slot 1 to j −1. To leverage comment modeling results in the decoding, we incorporate θ into the attention weights (defined below) over source posts and concentrate on topic words therein for question generation. αij = exp (fα (hi, sj, θ)) P|P | i′=1 exp (fα (hi′, sj, θ)) (2) sj is the GRU decoder’s j-th hidden states and: fα (hi, sj, θ) = vT α tanh (Wα [hi; sj; θ] + bα) (3) In addition, we adopt copy mechanism (See et al., 2017) to allow the generated questions to contain the keywords from the source posts: pj = λj · pgen + (1 −λj) · pcopy (4) pgen refers to the likelihood to generate a word while pcopy is the extractive distribution derived from the attention weights over the source input. The soft switcher λj ∈[0, 1] can determine whether to copy a word or generate a new one in aware of the comments’ topics: λj = sigmoid (Wλ [uj; sj; tj; θ] + bλ) (5) tj is the context vector (weighted sum) of the attention to predict the Q’s j-th word, whose embedding is uj. Wλ and bλ are both learnable parameters. 33 Answer Generation. To further explore the relations between questions (Q) and answers (A), we “replicate” the question decoder’s architecture and form another decoder to handle answer generation (answer decoder). The answer choices are concatenated to form an answer sequence and neighboring choices are separated with a special token “<sep>”. The answer decoder also adopts the same topic-aware attentions (Eq. 2) as the question decoder (denoted as βij here) and copy mechanisms (Eq. 4) to be able to put topic words from the source into the answer choices, such as “赵粤” (Akira) and “希林娜依高” (Curley G) in Figure 1. Question decoder and answer decoder work together in a dual decoders setting, whose parameters are updated simultaneously to exploit the essential correlations of poll questions and their answers. 3.3 Model Training This subsection describes how we jointly train the neural topic model (henceforth NTM) for comment modeling and the decoders for question and answer generation with multi-task learning. The loss function for NTM is defined as: LNT M = DKL(p(z) || q(z | C)) −Eq(z|C)[p(C|z)] (6) The C above refers to Cbow. The first term is the KL divergence loss and the second is the reconstruction loss in VAE. For question generation, the loss is: LQG = − N X n=1 log (Pr (Qn | Pn, θn)) (7) N is the number of training samples; Qn, Pn, and θn are the target poll question, source post, and topic distribution of the n-th training sample. Answer generation loss LAG is defined similarly. The training loss of the entire model are defined as: L = LNTM + γQ · LQG + γA · LAG (8) where γQ and γA balance the weights over NTM and the two decoders. 4 Experimental Setup Data Preprocessing. First, we removed meta data (e.g., author’s locations and emoji labels) and replaced links, mentions (@username), and digits with generic tags “URL”, “MENT”, and “DIGIT”. Then, for some poll questions echoed in the source posts, we took them away for fair experiments. Next, an open-source toolkit jieba is employed for Chinese word segmentation.8 Afterwards, we filtered out stop words and for the remaining, we maintained two vocabularies with the most frequent 50K words for sequences (input and output) and another 100K words for BoW. Finally, comments are capped at the first 100 words to examine poll question generation with the early comments and their potential to draw future user engagements. In evaluations, we split our data into 80% for training, 10% for validation and 10% for test. Baselines and Comparisons. For baselines, we first consider the basic S2S (Sutskever et al., 2014) (i.e., BASE); also compared are the S2S with pretrained models from the BERT family — tiny ERINE (Sun et al., 2019) (i.e., ERINE), BERT (Devlin et al., 2019) (i.e., BERT), and RoBERTa (Liu et al., 2019) (i.e., ROBERTA), which were implemented with the paddle hub platform9. For all S2S with pre-trained models, their pre-trained parameters were further fine-tuned on our training data. Then, we consider the following S2S extensions with copy mechanism (i.e., COPY) (Meng et al., 2017), topic modeling from posts (i.e., TOPIC) (Wang et al., 2019a), and bidirectional attentions over posts and comments (i.e., CMT (BIATT)) (Wang et al., 2019b). All of them were proposed for keyphrase generation tasks and set up following their original papers. For our models, we consider two variants — CMT (NTM) in the single decoder archetecture and its dual decoder version DUAL DEC.10 Model Settings. All the hyperparameters are tuned on the validation set via grid search. For NTM, it is pre-trained for 50 epochs before joint training and afterwards different modules take turns to update parameters. We adopt two-layers bidirectional GRU to build source post encoder and one-layer unidirectional GRU question and answer decoders. The hidden size of each GRU is 300. 8https://github.com/fxsjy/jieba 9https://www.paddlepaddle.org.cn/hub 10We also finetuned BERT with our models yet cannot observe much performance gain. It is because NTM is able to learn essential features from the input and BERT cannot provide additional benefits. Another possible reason is that social media BERT is unavailable in Chinese and that trained on out-domain data (e.g., news) might not fit well with Weibo languages. Large-scale Weibo data might be acquired for continue pre-training (Gururangan et al., 2020), which is beyond the scope of this paper and will be explored in future work. 34 For a word embedding, the size is set to 150 and randomly initialized. In training, we apply Adam optimizer with initial learning rate as 1e-3, gradient clipping as 1.0, and early-stopping strategy adopted. The weights to trade off losses in multitask learning is set to γQ = γA = 1 (Eq. 8). Evaluation Metrics. We adopt both automatic measures and human ratings for evaluations. For the former, we examine two popular metrics for language generation tasks — ROUGE (Lin, 2004) and BLEU (Papineni et al., 2002). For the latter, human annotators rates with 4 point Likert scale (i.e., {0, 1, 2, 3}) and over three criteria are considered: the relevance to the source posts (relevance), how fluent the generated language reads (fluency), the attractiveness degree of the questions in drawing people’s engagements (engagingness). 5 Experimental Results In this section, we first show the main comparison results on poll question generation involving both automatic evaluations and human ratings (in §5.1). Then, model sensitivity to varying lengths of source posts and poll questions are discussed in §5.2, followed by the analyses of models’ capability to handle poll questions exhibiting varying degrees of user engagements (§5.3). Next, §5.4 discusses the performance of dual decoders that jointly generate questions and answers. A case study is presented at last (in §5.5) to interpret the sample outputs. 5.1 Comparison on Poll Question Generation We first show the comparison results on poll question generation, where we will discuss automatic evaluations and human ratings in turn below. Automatic Evaluations. Table 2 reports the automatic measured results on question generation. As can be seen, our task is challenging and basic S2S performs poorly. Pre-trained models from the BERT family can offer some help though limited. It is probably because the pre-training data is from other domains (e.g., news and online encyclopedia), where the representations learned cannot fully reflect the styles of social media languages. We then observe copy mechanism and latent topics (learn from posts) are both useful, where the former allows the keyword extracted from the post to form a question while the latter further helps find topic words to be copied. On the contrary, user MODEL ROUGE-1 ROUGE-L BLEU-1 BLEU-3 S2S Baselines BASE 21.62±0.7 20.64±0.7 20.35±0.7 2.11±0.5 +ERNIE 29.62±0.5 27.82±0.4 21.66±0.5 3.25±0.4 +BERT 33.62±1.2 31.57±1.1 24.43±0.7 4.54±0.4 +ROBERTA 34.08±1.3 31.98±1.2 24.88±1.0 4.85±0.5 S2S Extensions +COPY 35.13±0.4 33.20±0.4 30.27±0.4 7.95±0.3 +TOPIC 36.65±0.6 34.70±0.6 31.11±0.5 8.66±0.5 +CMT (BIATT) 27.74±0.4 26.21±0.4 23.97±0.3 4.15±0.2 Our Models +CMT (NTM) 37.95±0.4 35.97±0.3 32.07±0.2 8.89±0.3 +DUAL DEC 38.24±0.3 36.14±0.3 32.27±0.4 9.04±0.3 Table 2: Main comparison results for poll question generation. The underlined scores are the best in each column. Average scores are before ± and the numbers after are the standard deviation over 5 runs initialized with different seeds. Our models CMT (NTM) and DUAL DEC significantly outperforms all the other comparison models (paired t-test; p-value < 0.05). comments, though able to provide useful information, are noisy (also implied by Table 1). So, it is important to encode the comments in an appropriate way — CMT (NTM) captures salient topic features from the comments and performs much better than CMT (BIATT), which might be hindered by the noise and exhibit the second worst results. In addition, we notice DUAL DEC slightly outperforms its single decoder variant CMT(NTM), though the gain is small. To better examine their prediction results, we conduct human evaluations. Human Ratings. Here we sampled 400 source posts (and their outputs), and invited four native Chinese speakers to rate the poll questions in a 4 point Likert scale — 0 for extremely bad, 1 for bad, 2 for good, and 3 for extremely good — without knowing where the results come from. Each annotator reviews 100 samples and one’s assignments vary with others’ and Table 3 shows the average ratings over the four annotators. All the models are rated worse than the gold standard, which means automatic poll question generation still has a long way to go. We also observe that models with latent topics exhibit relatively better relevance. This may be because topic models allow the capture of salient contents from the input and detail injection to the output. Besides, CMT (NTM) and DUAL DEC perform the best in engagingness, probably because user comments and poll answers might provide implicit clues (e.g., fresh words) helpful to predict engaging questions. For fluency, BASE outperforms our models by a small margin, as it tends to yield short and generic questions, such as “你怎么看” (What’s your viewpoint?) based on our observation. More35 Relevance Fluency Engagingness Gold Standard 2.79 2.84 2.74 BASE 1.26 2.14 1.35 ROBERTA 1.33 1.06 0.96 TOPIC 1.81 1.66 1.50 CMT (NTM) 1.91 1.67 1.55 DUAL DEC 2.02 1.87 1.67 Table 3: Average human ratings. Higher scores indicate better results. DUAL DEC exhibits good potential generate questions likely to draw user engagements. over, we measure the length of questions generated by BASE and DUAL (our full model) and find that 11.0% questions generated by BASE contain less than 5 words whereas the number for DUAL is only 1.6%. This again demonstrates our potential to generate longer questions with richer details. 5.2 Effects of Post and Question Length We further quantify the question generation results over varying lengths of source posts and poll questions and show the corresponding ROUGE-1 scores in Figure 4. Here, we compare BASE and ROBERTA, TOPIC, and our CMT (NTM).11 Figure 4: ROUGE-1 scores (y-axis) over varying length (word count in x-axis) of source posts (on the left) and poll questions (on the right). For both subfigures, the bars from the left to right shows the results of BASE, ROBERTA, TOPIC, and CMT (NTM). Post length seems not to affect much on the models’ performance, probably attributed to the length limitation in Weibo — even the relatively longer posts contain limited words. On the contrary, for the question length, the two S2S baselines both exhibit obvious performance drops when generating long questions, while TOPIC and CMT (NTM) perform steadily. This suggests that latent topics, either captured from posts or comments, may have the potential to enrich questions with detailed descriptions, and hence can better tackle long questions. Nevertheless, CMT (NTM) presents consistently better ROUGE-1 in diverse scenarios. 11In §5.2 and §5.3, we experiment in the single decoder settings so as to focus on the quality of generated questions. We will further discuss the dual decoders in §5.4. 5.3 Polls Questions vs. User Engagements As shown in the human ratings (§5.1), comments might help to generate engaging poll questions. For a further discussion, Figure 5 shows the ROUGE-1 of ROBERTA, TOPIC, and CMT (NTM) in handling questions for polls that later engage varying user comment numbers. Interestingly, CMT (NTM) performs better when predicting questions that engage more comments at the end. This means that early comments might provide useful clues for models to distinguish attractive questions with the potential to draw more public engagements in the future. Lacking the ability to learn from comments, TOPIC exhibits relatively more stable trends. Figure 5: Model performance in handling polls that result in varying comment numbers (x-axis). Yaxis: ROUGE-1. Bars from left to right represent ROBERTA, TOPIC, and CMT (NTM). 5.4 Discussion on Dual Decoders The previous two subsections are discussed in the single decoder setting and here we further examine the effectiveness to jointly predict questions and answers. BASE, COPY, TOPIC, and CMT (NTM) with single and dual decoders are discussed. We first compare question generation results and Figure 6 shows the ROUGE-1 scores. It is seen that dual decoders can boost the results of BASE and COPY, implying that questions and answers are indeed related and exploiting their interactions can successfully bring performance gain. However, we cannot observe large-margin improvements in TOPIC and CMT (NTM), probably because many words in answers, such as “赵粤” (Akira) and “希 林娜依高” (Curley G) in Figure 1, are also topic words that can be discovered with topic models. Therefore, jointly generating answers only provides limited help to their question generation results. Then, we analyze how the multitask learning ability of dual decoders influence the prediction of poll answers. Table 4 displays the comparison results with pipeline models that sequentially generate questions and then answers. By examining the pipeline results, we first find that source posts are 36 Figure 6: ROUGE-1 scores of BASE, COPY, TOPIC, and CMT (NTM) from left to right. For each model, left bars (in blue) shows them in single decoder setting while the right bars (in orange) dual decoders. MODEL ROUGE-1 ROUGE-L BLEU-1 BLEU-3 Pipeline Models QS ONLY (PRED) 26.65±0.2 25.09±0.2 22.50±0.8 4.27±0.5 QS ONLY (GOLD) 25.51±0.5 24.17±0.4 22.43±0.3 3.76±0.3 PT+QS (PRED) 31.29±0.6 29.18±0.5 26.35±0.1 8.15±0.3 PT+QS (GOLD) 31.78±0.6 29.63±0.6 26.39±0.6 8.14±0.3 Dual Decoders BASE 24.68±0.7 22.59±0.5 21.38±0.3 3.22±0.4 +COPY 30.03±0.5 28.02±0.5 25.55±0.5 8.28±0.3 +TOPIC 30.56±0.8 28.49±0.8 26.00±0.5 8.26±0.4 +CMT (NTM) 31.72±0.7 29.54±0.7 26.55±0.2 8.65±0.2 Table 4: The comparison results of models with dual decoders (on the bottom half) and pipeline models (on the top). For the pipeline models, we first produce questions (QS) using CMT (NTM), from which we further generate answers with the S2S model. QS ONLY is fed with QS only while PT+QS the concatenated sequence of posts (PT) and QS. In the training of answer generation, PRED means the predicted questions are employed as input while for GOLD, we adopt gold standard questions (they are assumed to be unavailable for test). helpful in answer generation, which results in the outperformance of PT+QS over QS ONLY. Besides, answer generation trained with predicted questions or the gold standards do not make much difference. Gold standard questions might exhibit higher quality while predicted questions may better fit the tests (answer choices should be predicted without knowing the human-crafted questions). For dual decoders, CMT (NTM) still performs the best, implying that latent topics from user comments can also contribute to better prediction of poll answers. In comparison with the best pipeline model (PT+QS), the scores from CMT (NTM) are competitive, though the dual decoder allows endto-end training and is easier to be used (with less manual efforts in model training and application). 5.5 Case Study To provide more insights, we further take the two Weibo posts in Figure 1 as the input cases and examine the output of varying models in Table 5.12 Unsurprisingly, BASE tends to yield generic questions as limited features are encoded from the noisy source. ROBERTA sometimes produces repeated words (e.g., its output to P1), hindering its capability to generate fluent language (also indicated by Table 3). This is possibly caused by the overfitting problem as RoBERTa might rely on large-scale in-domain data for fine-tuning. We also find that modeling topics and user comments may enable the output to contain trendy wordings, making it more engaging, such as “c位” (center point) in CMT (NTM)’s output question for P2 and the names of many new video apps in DUAL DEC’s generated answer choices for P1. Furthermore, the dual decoders might learn the cohesive relations between questions and answers, such as the Akira and Curley G occurring in both the generated questions and answer choices (P2). 6 Related Work Our work is in the line with question generation, where most prior efforts focus on how to ask good exam questions given an article and the pre-defined answers. Some adopt manually-crafted rules or features (Labutov et al., 2015; Dhole and Manning, 2020; Fabbri et al., 2020), largely relying on the labor-intensive process for rule design or feature engineering. To simplify the training, automatic feature learning hence becomes popular. For example, Chali and Hasan (2015) first employs a Bayesian model to learn topic features and then leverages them to yield questions. These pipeline methods require the expertise involvement to manually customize the model inference algorithms, while our neural network design allows end-to-end training of topic modeling and question generation. Recently, S2S-based question generation architecture has demonstrated promising results (Du et al., 2017; Chai and Wan, 2020). To better encode the input, researchers adopt successful training design from other tasks, such as self-attention mechanism (Zhao et al., 2018; Scialom et al., 2019), language model pre-training (Pan et al., 2019), variational inference (Yao et al., 2018), and reinforcement learning (Yuan et al., 2017; Pan et al., 2019). Heuristic features, e.g., the answers’ positions in the article (Zhou et al., 2017; Sun et al., 2018; 12Here we analyze the case with two examples while similar observations can be drawn from many output cases. More cases will be discussed in Figure 6 (in the Appendix). 37 BASE 你会看吗(Would you watch) ROBERTA 你平时喜欢哪个视频频频(Which videooooo do you usually like) TOPIC 你平时常用哪个视频(Which video do you usually use) CMT (NTM) 你平时在哪个视频网站(Which video site are you on) DUAL DEC 你平时用哪个视频app (Which video app do you usually use) >bili 哔哩(Bilibili); 爱奇艺(iQiyi); 腾 讯视频(Tencent Video); 芒果tv (Mango TV); 优酷(Youku); 其他评论区补充 (Comment with other choices) BASE 你觉得谁的表现更强(Who do you think is better) ROBERTA 你觉得谁更好(Who do you think is better) TOPIC 你觉得谁出道了(Who do you think debuted) CMT (NTM) 你觉得谁更适合c位(Who do you think is more suitable for the center position) DUAL DEC 你觉得赵粤和希林娜依高谁更可 (Who do you prefer, Akira or Curley G) >赵粤(Akira); 希林娜依高(Curley G) Table 5: Questions generated for the source posts in Figure 1: P1 (top) and P2 (bottom). For DUAL DEC (i.e., CMT (NTM) with dual decoders), the question is followed by the answer in the next row. Kim et al., 2019; Liu, 2020) are sometimes considered. For question decoding, certain constraints are added to control the generation, such as some aspects to be contained (Hu et al., 2018), varying levels of difficulty (Gao et al., 2018) and specificity (Cao et al., 2019). We are also related with previous work handling the generation of questions and answers in a multitask learning setting (Wang et al., 2017; Tang et al., 2017; Sun et al., 2020). Nonetheless, none of the aforementioned research concerns poll questions and answers on social media, which exhibit very different language styles compared with any existing studies and has not been extensively explored. 7 Conclusion We have presented a novel task to generate social media poll questions. User comments encoded with a neural topic model are leveraged in a S2S framework; dual decoder architecture is further adopted to explore the interactions between questions and answers. Extensive experiments on a large-scale dataset newly collected from Weibo have demonstrated the effectiveness of our proposed model. Acknowledgments This work was partially done when Zexin Lu was an intern at Tencent AI Lab under CCF-Tencent Rhino-Bird Young Faculty Open Research Fund (R-ZDCJ). The research is also supported by NSFC Young Scientists Fund (62006203) and PolyU internal funds (1-BE2W, 4-ZZKM, and 1-ZVRH). The authors would like to thank Lida Li, Yue Wang, Yubo Zhang, Zhe Wang, and anonymous reviewers from ACL-IJCNLP 2021 for their insightful suggestions on various aspects of this work. Ethical Considerations The task will not pose ethical problems. First, the polls are open access to the public users (so as to collect their opinions). Second, Weibo allows any users to report suspicious cases with ethical concerns and the reported contents will be removed immediately. Third, the polls are running in an anonymous way to protect the privacy of voters. The dataset is collected through the official APIs of Weibo and is consistent with the Weibo terms of use. We also manually examined the data to ensure the following points. First, we conduct data anonymization and manually examined the data to ensure there are no privacy and ethical concerns, e.g., personal information, toxic language, and hate speech. In the generated polls, we didn’t spot any cases that might have the concern. Second, the involved Weibo users are all public ones. To that end, we automatically filtered out personal users without the official confirmation of Weibo (the confirmed public users can be identified with a “VIP” tag). The user list is manually checked again to mitigate the ethical concern. For the annotation, we recruited part-time research assistants to work with the pay 15.7 USD/hour and at most 20 hours per week. References Eleftheria Ahtaridis, Christopher Cieri, and Denise DiPersio. 2012. LDC language resource database: Building a bibliographic database. In Proceedings of the Eighth International Conference on Language Resources and Evaluation, LREC 2012, Istanbul, Turkey, May 23-25, 2012, pages 1723–1728. European Language Resources Association (ELRA). Yang Trista Cao, Sudha Rao, and Hal Daum´e III. 2019. Controlling the specificity of clarification question generation. In Proceedings of the 2019 Workshop on Widening NLP@ACL 2019, Florence, Italy, July 28, 2019, pages 53–56. Association for Computational Linguistics. Zi Chai and Xiaojun Wan. 2020. Learning to ask more: Semi-autoregressive sequential question generation 38 under dual-graph interaction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 510, 2020, pages 225–237. Association for Computational Linguistics. Yllias Chali and Sadid A. Hasan. 2015. Towards topic-to-question generation. Comput. Linguistics, 41(1):1–20. Kyunghyun Cho, Bart van Merrienboer, C¸ aglar G¨ulc¸ehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014, October 25-29, 2014, Doha, Qatar, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1724–1734. ACL. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics. Kaustubh D. Dhole and Christopher D. Manning. 2020. Syn-qg: Syntactic and shallow semantic rules for question generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 752–765. Association for Computational Linguistics. Xinya Du, Junru Shao, and Claire Cardie. 2017. Learning to ask: Neural question generation for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 August 4, Volume 1: Long Papers, pages 1342–1352. Association for Computational Linguistics. Alexander R. Fabbri, Patrick Ng, Zhiguo Wang, Ramesh Nallapati, and Bing Xiang. 2020. Templatebased question generation from retrieved sentences for improved unsupervised question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4508–4513. Association for Computational Linguistics. Yifan Gao, Jianan Wang, Lidong Bing, Irwin King, and Michael R. Lyu. 2018. Difficulty controllable question generation for reading comprehension. CoRR, abs/1807.03586. Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 8342–8360. Association for Computational Linguistics. Wenpeng Hu, Bing Liu, Jinwen Ma, Dongyan Zhao, and Rui Yan. 2018. Aspect-based question generation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Workshop Track Proceedings. OpenReview.net. Yanghoon Kim, Hwanhee Lee, Joongbo Shin, and Kyomin Jung. 2019. Improving neural question generation using answer separation. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 6602–6609. AAAI Press. Igor Labutov, Sumit Basu, and Lucy Vanderwende. 2015. Deep questions without deep understanding. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 889– 898. The Association for Computer Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Bingran Liu. 2020. Neural question generation based on seq2seq. In Proceedings of the 2020 5th International Conference on Mathematics and Artificial Intelligence, pages 119–123. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. arXiv preprint, abs/1907.11692. Rui Meng, Sanqiang Zhao, Shuguang Han, Daqing He, Peter Brusilovsky, and Yu Chi. 2017. Deep keyphrase generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 582–592. Association for Computational Linguistics. Yishu Miao, Edward Grefenstette, and Phil Blunsom. 2017. Discovering discrete latent topics with neural variational inference. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 2410–2419. PMLR. 39 Boyuan Pan, Hao Li, Ziyu Yao, Deng Cai, and Huan Sun. 2019. Reinforced dynamic reasoning for conversational question generation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 2114–2124. Association for Computational Linguistics. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, July 6-12, 2002, Philadelphia, PA, USA, pages 311–318. ACL. Thomas Scialom, Benjamin Piwowarski, and Jacopo Staiano. 2019. Self-attention architectures for answer-agnostic neural question generation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6027–6032. Association for Computational Linguistics. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 August 4, Volume 1: Long Papers, pages 1073–1083. Association for Computational Linguistics. Xingwu Sun, Jing Liu, Yajuan Lyu, Wei He, Yanjun Ma, and Shi Wang. 2018. Answer-focused and position-aware neural question generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3930– 3939. Association for Computational Linguistics. Yibo Sun, Duyu Tang, Nan Duan, Tao Qin, Shujie Liu, Zhao Yan, Ming Zhou, Yuanhua Lv, Wenpeng Yin, Xiaocheng Feng, Bing Qin, and Ting Liu. 2020. Joint learning of question answering and question generation. IEEE Trans. Knowl. Data Eng., 32(5):971–982. Yu Sun, Shuohuan Wang, Yu-Kun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. ERNIE: enhanced representation through knowledge integration. CoRR, abs/1904.09223. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 3104–3112. Duyu Tang, Nan Duan, Tao Qin, and Ming Zhou. 2017. Question answering and question generation as dual tasks. CoRR, abs/1706.02027. Tong Wang, Xingdi Yuan, and Adam Trischler. 2017. A joint model for question answering and question generation. CoRR, abs/1706.01450. Yue Wang, Jing Li, Hou Pong Chan, Irwin King, Michael R. Lyu, and Shuming Shi. 2019a. Topicaware neural keyphrase generation for social media language. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 2516–2526. Association for Computational Linguistics. Yue Wang, Jing Li, Irwin King, Michael R. Lyu, and Shuming Shi. 2019b. Microblog hashtag generation via encoding conversation contexts. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 1624–1633. Association for Computational Linguistics. Kaichun Yao, Libo Zhang, Tiejian Luo, Lili Tao, and Yanjun Wu. 2018. Teaching machines to ask questions. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden, pages 4546–4552. ijcai.org. Xingdi Yuan, Tong Wang, C¸ aglar G¨ulc¸ehre, Alessandro Sordoni, Philip Bachman, Saizheng Zhang, Sandeep Subramanian, and Adam Trischler. 2017. Machine comprehension by text-to-text neural question generation. In Proceedings of the 2nd Workshop on Representation Learning for NLP, Rep4NLP@ACL 2017, Vancouver, Canada, August 3, 2017, pages 15–25. Association for Computational Linguistics. Jichuan Zeng, Jing Li, Yan Song, Cuiyun Gao, Michael R. Lyu, and Irwin King. 2018. Topic memory networks for short text classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3120– 3131. Association for Computational Linguistics. Yao Zhao, Xiaochuan Ni, Yuanyuan Ding, and Qifa Ke. 2018. Paragraph-level neural question generation with maxout pointer and gated self-attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3901–3910. Association for Computational Linguistics. Qingyu Zhou, Nan Yang, Furu Wei, Chuanqi Tan, Hangbo Bao, and Ming Zhou. 2017. Neural question generation from text: A preliminary study. In Natural Language Processing and Chinese Computing - 6th CCF International Conference, NLPCC 2017, Dalian, China, November 8-12, 2017, Proceedings, volume 10619 of Lecture Notes in Computer Science, pages 662–671. Springer. 40 [Post]: #2020百大最美女星#刘亦菲和迪丽热巴都上榜啦！！！都是天然美女啊～两个人一个人演过电影版 的三生三世，一个演过剧版的三生三世。(#100 Most Beautiful Women in the World 2020# Liu Yifei and Dilraba Dilmurat are both on the list!!! Both of them are natural beauties˜One of them played in the movie Eternal Love while the other played in its TV series version) [Question]: 谁的颜让你心动呢(Whose face makes you heart flip) [Answer]: 刘亦菲(Liu Yifei); 迪丽热巴(Dilraba Dilmurat) [Base]: 你最喜欢谁(Who do you like the best) [RoBERTa]: 你更喜欢谁(Who do you prefer) [Topic]: 你更喜欢哪一个(Which one do you prefer) [Cmt(NTM)]: 你更喜欢谁的造型(Whose look do you prefer) [DualDec]: 你觉得谁更有cp感(Who do you think is better coupled with the leading man) >刘亦菲(Liu Yifei); 迪丽热巴(Dilraba Dilmurat) [Post]: 有意见建议同性婚姻合法化写入民法典(Some people suggest that same-sex marriage be legalized into the Civil Code) [Question]: 你支持同性恋结果合法化吗(Do you support the legalization of same-sex marriage) [Answer]: 同意(Agree); 不同意(Disagree) [Base]: 你怎么看(What do you think) [RoBERTa]: 你支持同性结婚化吗(Do you support the same-sex marriage) [Topic]: 你支持同性恋合法化吗(Do you support the legalization of homosexuality) [Cmt(NTM)]: 你支持同性恋婚姻合法化吗(Do you support the legalization of the same-sex marriage) [DualDec]: 你支持同性恋婚姻合法化吗(Do you support the legalization of the same-sex marriage) >支持(Support); 不支持(Objection) [Post]: #瑞幸咖啡伪造交易22亿# 在否认业绩造假两个月后，瑞幸今日盘前发布公告：内部调查显示， 从2019年第二季度到2019年第四季度与虚假交易相关的总销售金额约为22亿元。于是，#瑞幸暴跌# 。(#Ruixing Coffee forged 2.2 billion transactions# Two months after denying fraud, Luckin released an announcement before the market today: An internal investigation showed that total sales related to invalid transactions from the second quarter of 2019 to the fourth quarter of 2019 amounted to about 2.2 billion Yuan. Consequently, #Luckin Coffee stock plummet#) [Question]: 你还会喝瑞幸咖啡吗(Will you still drink Luckin coffee) [Answer]: 会，我券还没用完呢(Yes. I still have the coupons to use); 不会，没券就不喝(No. No coupon, no coffee.); 从来就没有喝过(I’ve never drunk the coffee there); 不管如何都是死忠粉(Die-hard fan no matter what) [Base]: 你会买iphone 吗(Would you buy an iphone) [RoBERTa]: 你喝过瑞幸咖啡吗(Have you ever drunk Luckin coffee) [Topic]: 你会买瑞幸咖啡吗(Would you buy Luckin coffee) [Cmt(NTM)]: 你觉得瑞幸咖啡合理吗(Do you think Luckin Coffee is reasonable) [DualDec]: 你还会买瑞幸咖啡吗(Will you still buy Luckin coffee) >会(Yes); 不会(No); 看情况(It depends) [Post]: 杨丽萍因为没有结婚生孩子，过着与花草舞蹈为伴的生活，被网友diss是一个失败的范例，真正的女 人应该要儿孙满堂，才是幸福的。(Yang Liping, who has no marriage or children, lives a life with flowers and dancing. However, she has been ridiculed by netizens and viewed as a typical loser — a real woman should have a large family of children and grandchildren to live in happiness.) [Question]: 如何定义成功女性(How to define a successful woman) [Answer]: 事业有成(Success in career); 儿孙满堂(Have children and grandchildren); 家庭事业双丰收(Success in family and career); 充实的灵魂(Interesting soul) [Base]: 你觉得哪种行为有问题(What kind of behavior do you think is problematic) [RoBERTa]: 女女是女人是女人是什么(What is woman is woman) [Topic]: 你觉得结婚应该定义成功吗(Do you think marriage should come to define success) [Cmt(NTM)]: 你怎么看待成功的女性杨丽萍(How do you think of the successful woman Yang Liping) [DualDec]: 你觉得如何定义成功女性(How would you define successful women) > 应该(Should); 不支持(Objection); 评论区补充(Add more details in comments) [Post]: #杨幂魏大勋恋情实锤# 杨幂魏大勋恋情再次被实锤，现在已经成了圈子内外不是秘密的秘密了。 (#Smoking gun of Yang Mi and Wei Daxun# Yang Mi and Wei Daxun’s love affair has been verified again, and it has now become a secret inside and outside the circle.) [Question]: 你看好杨幂魏大勋的恋情吗(Are you optimistic about Yang Mi’s romantic relationship with Wei Daxun) [Answer]: 看好(Optimistic); 不看好(Pessimistic); 有波折终能修成正果(There will be twists and turns but the ending will be good) [Base]: 你觉得这个做法怎么样(What do you think of this approach) [RoBERTa]: 你觉得魏魏勋勋恋爱吗(Do you think Wei Wei Xun Xun is in love) [Topic]: 你觉得谁更渣(Who do you think is more scummy) [Cmt(NTM)]: 你怎么看待这恋情的(What do you think of the romantic relationship) [DualDec]: 你觉得杨幂魏大勋有必要吗(Do you think Yang Mi and Daxun Wei are necessary to do so) >杨幂(Yang Mi); 魏大勋(Wei Daxun); 都不喜欢(Do not like either of them); 吃瓜(I’m an onlooker) Table 6: Five additional cases. One block refers to one case, including its source post (Post), ground truth question (Question) and answer (Answer), followed by and the results generated by varying models (model names are in []). For answers, different choices are separated by “;” and the outputs of DualDec appear after a >. Italic words in “()” are the English translation of the original Chinese texts on their left.
|
2021
|
3
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 351–365 August 1–6, 2021. ©2021 Association for Computational Linguistics 351 PASS: Perturb-and-Select Summarizer for Product Reviews Nadav Oved ∗ Technion - Israel Institute of Technology Haifa, Israel [email protected] Ran Levy Amazon Tel Aviv, Israel [email protected] Abstract The product reviews summarization task aims to automatically produce a short summary for a set of reviews of a given product. Such summaries are expected to aggregate a range of different opinions in a concise, coherent and informative manner. This challenging task gives rise to two shortcomings in existing work. First, summarizers tend to favor generic content that appears in reviews for many different products, resulting in template-like, less informative summaries. Second, as reviewers often disagree on the pros and cons of a given product, summarizers sometimes yield inconsistent, self-contradicting summaries. We propose the PASS system (Perturb-and-Select Summarizer) that employs a large pre-trained Transformer-based model (T5 in our case), which follows a few-shot fine-tuning scheme. A key component of the PASS system relies on applying systematic perturbations to the model’s input during inference, which allows it to generate multiple different summaries per product. We develop a method for ranking these summaries according to desired criteria, coherence in our case, enabling our system to almost entirely avoid the problem of selfcontradiction. We compare our system against strong baselines on publicly available datasets, and show that it produces summaries which are more informative, diverse and coherent.1 1 Introduction Online shopping has become a popular form of purchasing goods even before the most recent acceleration due to the COVID-19 pandemic. As ecommerce websites strive to make the shopping process more useful and enjoyable for customers, many interesting challenges arise. One challenge deals with how to surface opinions from product ∗Completed during an internship at Amazon. 1Summaries generated by PASS are available at: https: //registry.opendata.aws/ reviews in a concise yet reliable fashion. The research community has addressed this challenge early on, starting from the work of (Hu and Liu, 2004) which defined the task of mining and summarizing customer reviews. More recent advancements have relied on modern deep learning models trained on large collections of unannotated customer reviews (Brazinskas et al., 2020b,a). Our first observation relates to the summaries generated by CopyCat (Brazinskas et al., 2020b) and FewSum (Brazinskas et al., 2020a), two of these SOTA systems, which tend to mix generic statements such as “Would recommend this product to anyone” along with more informative content such as “The sound quality is good” (see Table 6 in Appendix B for examples of such generated summaries). Due to the emphasis of summarization systems on conciseness, we maintain that generic content should be used sparingly. Additionally, even if the content is not extremely generic, customers may perceive summaries as less useful if they tend to repeat themselves across products. In order to estimate the similarity between summaries generated for different products, we devise the Set-Pairwise-ROUGE metric (henceforth denoted as SPR), that computes the average ROUGE (Lin, 2004b) scores of summaries for two different products, across all product pairs. Using this metric we show that human written reference summaries are indeed far more diverse than their system generated counterparts, i.e. the SPR of reference summaries is significantly lower. We henceforth denote the notion of cross product diversity of summaries as CPDiversity. Large pre-trained Transformer-based (Vaswani et al., 2017) models such as OpenAI’s GPT-3 (Brown et al., 2020), Google’s T5 (Raffel et al., 2020), PEGASUS (Zhang et al., 2020a), and Facebook’s BART (Lewis et al., 2020) have made com352 pelling advancements on a host of NLG tasks, including abstractive text summarization. In this work we wish to leverage such models for product reviews summarization, aiming to generally improve the quality of generated summaries, and specifically in terms of their diversity across different products. While we aim to generate humanlike texts, care has to be taken with respect to their correctness. Indeed, concerns have been raised regarding the factual consistency of abstractive summaries, i.e., whether the facts conveyed in the summary agree with the source text (Cao et al., 2018; Kryscinski et al., 2019; Maynez et al., 2020). Our second observation relates to this issue of factual consistency in the context of product reviews summarization. Our task not only faces the risk of models hallucinating incorrect information, as in traditional abstractive text summarization, but also the risk of generating self-contradicting summaries which are not caused by model hallucinations. The latter can occur when the source documents contradict one another. This situation is quite likely because reviews may disagree on some product aspects or even disagree entirely. For example, review A states a machine is “easy to operate” vs. review B which states it “requires trial and error” (see more examples in Table 7 in Appendix B). In this unique setup, factual consistency is undefined and instead we wish to measure a different characteristic: the self-consistency of the summary. To the best of our knowledge this issue has not been analyzed in the past and in some sense it renders the task ill-defined because it’s not clear whether the summary is supposed to convey a range of possibly contradicting opinions about the product or the majority opinion. From here on, we shall assume that a summary has to convey the majority opinion of the reviews and do so in a selfconsistent manner. Our proposed method starts by fine-tuning a strong pre-trained language model for product reviews summarization in a few-shot setup. We then employ an input perturbation method that drops k reviews out of the input and concatenates the remaining reviews in random order. This process, denoted as LkO, short for leave k out, produces notable variation between candidate summaries, which increases the model’s output diversity.2 Once we have produced a set of candidate 2Diversity here is between candidate summaries for the summaries, we essentially cast our original summary generation problem as a ranking problem. This approach gives us the choice over what kind of summary we are interested in as the final output, i.e. choosing our ranking criteria. As mentioned above, our main concern in this work is producing self-consistent summaries. Instead of basing our ranking solely on this criterion, we train a more general coherence summary ranker using human annotated coherence scores (Fabbri et al., 2021). Finally, for each product, we select the top ranked summary as the system’s output. We compare our method against strong baselines, comprised of systems introduced in previous work on multi-document opinion summarization, and a T5 language model fine-tuned for abstractive text summarization. We evaluate each over 3 dimensions, of which relevance and coherence are commonly used in summarization (Dang, 2005), and our newly introduced metric for CP-Diversity. We demonstrate that our method produces high quality summaries which are more informative, diverse and coherent. In summary, the main contributions of this work are: (1) highlight two shortcomings of existing product reviews summarizers, namely low CPDiversity and self-inconsistency, and propose a dedicated metric for the former. (2) Propose a method that leverages strong pre-trained models that improve the CP-Diversity while significantly reducing the risk of self-inconsistencies. 2 Related Work Product Review Summarization. Product review summarization is a form of multi-document summarization in which a set of product reviews for a single product serves as the document cluster to be summarized. A common approach for product review summarization, which centers the summary around a set of extracted aspects and their respective sentiment, is termed aspect-based summarization (Hu and Liu, 2004; Kansal and Toshniwal, 2014; Wu et al., 2016; Angelidis and Lapata, 2018; Coavoux et al., 2019). As in traditional summarization, there are two inherently different requirements for the task, a simplified one, in which the goal is to provide an extractive output, i.e., a list of sentences extracted from the review set, or a more advanced one, in which the goal is to provide an abstracsame product, not to be confused with CP-Diversity. 353 tive output, i.e., generated content not restricted to use the same wording of the source set. Extractive summarization include earlier works such as (Carenini et al., 2006; Lerman et al., 2009; Xiong and Litman, 2014). More recently, (Tan et al., 2017) suggested a novel generative topic aspect sentiment model, while (Angelidis et al., 2021) suggested a novel system able to extract both general and aspect-specific summaries. As for abstractive summarization, recent advances on pre-training neural networks were explored in the context of product reviews in unsupervised and few-shot learning schemes which led to promising results (Chu and Liu, 2019; Brazinskas et al., 2020b,a; Suhara et al., 2020; Amplayo et al., 2021). Evaluating Summarization Systems. Evaluation of summarization systems is usually performed utilizing a mix of automatic metrics and human ratings. Among the automated metrics, probably the most well-known is the ROUGE family of scores (Lin, 2004b) that measures ngram overlap between generated summaries and corresponding reference summaries. Many other metrics that aim to quantify how well generated summaries align with reference summaries have been proposed, such as BLEU (Papineni et al., 2002), METEOR (Lavie and Agarwal, 2007), ROUGE-WE (Ng and Abrecht, 2015) and BertScore (Zhang et al., 2020b) to name a few. Unfortunately, such metrics alone do not tell the whole story and recently several works observed that a new requirement is necessary in order to ensure that facts from the summary agree with the source document (Cao et al., 2018; Kryscinski et al., 2019; Maynez et al., 2020). This requirement is usually known as factual consistency. As for human ratings, those are usually obtained across several dimensions of summary quality. The DUC 2005 task (Dang, 2005) suggested the following 5 dimensions: Grammaticality, Non-redundancy, Referential clarity, Focus and Structure, and Coherence. In the context of product reviews summarization (Brazinskas et al., 2020a) use the standard ROUGE-1/2/L metrics as well human comparative judgments on 5 dimensions: Fluency, Coherence, Non-Redundancy, Informativeness and Sentiment. To the best of our knowledge the issues of selfconsistency and diversity across products were not directly analyzed before. 3 Perturb-and-Select Summarizer In this section, we propose a system that employs a large pre-trained Transformer-based model (T5) in a few-shot fine-tuning scheme for multiple reviews abstractive summarization. We aim to leverage the inherent diversity between reviews for a given product to our advantage, by applying systematic perturbations to the model’s input during inference. This allows our fine-tuned model to generate multiple different candidate summaries per product, exhibiting variability both in the content being surfaced as well as in the phrasing of said content. We develop a ranking mechanism for selecting the best candidate summary according to desired criteria, which in our case is coherence. We provide an end-to-end diagram of the PASS Summarizer’s components in Figure 1. 3.1 Fine-tuning T5 for Summary Generation PASS relies on a pre-trained T5 language model, which we fine-tuned on a small publicly available dataset for product reviews summarization (Brazinskas et al., 2020a). We follow a similar fine-tuning scheme for abstractive text summarization to the one presented in (Raffel et al., 2020) with the exception that we concatenate the multiple reviews into a single input text as a preprocessing step. As the dataset contains multiple reference summaries per product, we repeat our training process for each reference summary using the same (concatenated) input text. 3.2 Candidate Summary Generation In light of the natural diversity existing between product reviews, we explore a modeling approach which allows for such diversity to emerge in our summarizer’s output as well. We do this by manipulating the model’s input, sampling which reviews to use each time, in a way that allows for increasing the relative prevalence of certain reviews over others. We also re-shuffle the reviews before concatenation to ensure the model is not affected by their internal order. Note that prior attempts have been made to directly manipulate the content within the reviews (Amplayo and Lapata, 2020) a path that we do not explore here. Our intervention method guarantees that each review’s correctness, integrity and meaning are preserved. Since it only affects the subset of reviews being used and their order of concatenation, this increases the potential for diversity (per product and across products) 354 Figure 1: A diagram of the PASS components, with an example for a collection of reviews of size d = 4, k = 1. emerging from the input’s content, without compromising its linguistic quality. LkO Input Perturbation Method. Given a set of d reviews R = {r1, ..., rd} for a product p, our perturbation method iterates over A(R) the set of all possible subests of size d −k in R, A(R) = S S ⊂R, |S| = d −k, 1 ≤k < d . Given a subset S ∈A(R) we concatenate its reviews in random order, and feed the concatenated text into our fine-tuned T5 summarizer, which generates a candidate summary c. We repeat this step for all S ∈A(R), resulting in a set of generated candidate summaries which we denote as C = {c1, ..., cm}, m = d k . This process, denoted as LkO, short for leave-k-out, produces notable variation between candidate summaries (see Table 8 in Appendix B for examples), and allows for different content and aspects to emerge in the summaries, which were less likely to have surfaced otherwise. We found that this perturbation approach produces higher variation across candidate summaries when applying it on the model’s input only during the inference stage, not during training. Our method produces multiple perturbed versions of a given input while its references remain the same. If applied during training, this might encourage the model to fit a larger range of input features to a smaller set of outputs. We are interested in the opposite effect - we would like to encourage higher output variation as a function of input diversity. Note that when dealing with large review sets, achieving diversity does not require iterating over all subsets in A(R). For such scenarios, we recommend constructing a fixed number (m) of randomly sampled review subsets, so long as m is sufficiently large. In our experiments we employ the full LkO input perturbation method, since standard datasets focus on relatively small review sets.3 An alternative method for increasing novelty and variability in the output of a generative language model, is to directly intervene in its decoding algorithm, e.g., Beam Search (Vijayakumar et al., 2016; Cibils et al., 2018). Note that this will not have the same effect as our proposed approach. First, since beam search is a decoding algorithm, it only has access to the underlying language model, and is completely separated from the model’s input. Second, beam search’s mechanism is fixed to make local word-by-word decisions, before the complete summary is revealed. Finally, our approach guarantees that given a set of input texts, at least one candidate output will not be influenced at all by a specific input text (or more if k > 1). For example, if a set of 4 reviews contains 3 reviews discussing price, and 1 review discussing quality, our method guarantees that at least 1 candidate summary will be generated solely based on the first three (discussing price). Furthermore, our method increases the probability for a summary to mention both price and quality, when a review discussing price is left out. 3.3 Candidate Summary Ranking Once a set of candidate summaries are generated per product, we have essentially cast our summary generation problem as a summary ranking problem. This allows us to retrieve a summary, which ranks best out of a diverse set of candidates, according to desired, interpretable criteria. 3A few recent works attempt to explicitly address this issue (Shapira and Levy, 2020; Angelidis et al., 2021). 355 As mentioned in Section 1, our main concern is producing CP-diverse yet self-consistent and coherent summaries. Since our input perturbation method generates multiple candidate summaries, we are now left with the task of ranking this set by coherence. We would like the ranking process to filter out self-contradicting, incoherent or inconsistent candidates (by assigning low rank) and to promote well-formed, coherent candidates to the top of the list. To achieve this, we train a classifier that receives two summaries as input and decides whether the first summary is more coherent than the second or the opposite. The classifier can also decide that both summaries are equally coherent. Using such a classifier, we can obtain a partial ranking of the reviews by running all pairwise comparisons and count the number of times each summary was better than the summary it was paired with. Pairwise Summary Classifier. We train a model to classify a pair of summaries for coherence, by fine-tuning a pre-trained T5 model for pairwise text classification. Given a pair of summaries, the model is required to classify them as either: summary A is more coherent, summary B is more coherent, or A and B are equivalent in terms of coherence. A pair of summaries can often be considered equivalent when judging them according to specific criteria, stemming from the natural fact that often more than one summary can be considered correct or good. Indeed it has been shown that several reference summaries are needed for reliable evaluation showing that there is more than one truth (Lin, 2004a). Since this model is used as a comparator for ranking candidate summaries, we are especially sensitive to specific types of classification errors. If the model mistakenly classifies a summary to be more coherent than the other while the opposite is true, we consider this a critical classification error. This type of error could be detrimental to the validity of the ranking process, therefore we aim to minimize its rate. While other types of errors also reduce the classifier’s accuracy, we consider a mistake where the model classifies two summaries to be equivalent when in truth one is more coherent than the other, as less harmful for ranking purposes. Ranking Method. Our proposed ranking method iterates over all possible pairs of candidate summaries for a given product, and counts how many times each candidate was classified by the coherence pairwise classifier (our primary comparator), as more coherent than its counterpart. As a tie-breaking, secondary comparator, we train an additional pairwise summary classifier, to classify which candidate is more fluent, out of a pair of given candidates. We select the top ranked candidate as the final output summary for each product. 4 Experimental Setup 4.1 Data We utilize a recent publicly available Amazon product reviews summarization dataset (Brazinskas et al., 2020a) for fine-tuning the T5 model which underlines the PASS system and for evaluating the LkO input perturbation method, both in isolation and as part of the end-to-end PASS system. The dataset contains product reviews and reference summaries for 60 products on Amazon. Each product has 8 reviews and 3 reference summaries written by crowd source workers. We follow the dataset splits to the training, development and test sets provided by the authors of the dataset. While we mainly focus on product reviews summarization, we include the Yelp business reviews summarization dataset (also from (Brazinskas et al., 2020a)) in our end-to-end evaluation for the sake of completeness. The Yelp dataset contains business reviews and reference summaries for 100 businesses. For training and evaluating the pairwise coherence classifier, we utilize a public dataset of human annotated summaries (Fabbri et al., 2021), generated by 16 modern text summarization models for 100 news articles (1600 examples in total) from the CNN/DailyMail dataset (Hermann et al., 2015). Each summary was rated (on a scale of 1 to 5) across 4 dimensions: coherence, consistency, fluency and relevance, by 5 independent crowd source workers and 3 independent experts (8 annotations in total). We chose to use the experts’ annotations only, as they are considered to be more accurate and reliable for coherence and fluency (Fabbri et al., 2021). We construct a pairwise version of this dataset, by creating summary pairs from all 16 model outputs for each of the 100 news stories, along with their annotation scores for each metric respectively. We split the dataset according to news stories, by randomly sampling 20 stories for the test set, 16 stories for the develop356 ment set and the rest are used for the training set. Given a pair of summaries (a, b), their respective average expert rating, (ra, rb) and a threshold parameter ϵ, we define the label for that pair as: label(a, b) = A, if ra −rb ≥ϵ B, if rb −ra ≥ϵ E, otherwise where E denotes the case where both summaries are equivalent, A denotes that summary a is better than b and B denotes the opposite. To ensure that our training data is invariant to a pair’s internal order, we create examples for all (a, b) and (b, a) pairs in the training set. 4.2 Experimental Details Fine-tuning T5 for Summary Generation. We fine-tune a T5-Base model (220M parameters (Raffel et al., 2020)) for abstractive text summarization as described in 3.1 on the training set, and tune its hyperparameters on the development set. We train for maximum 20 epochs while employing a standard early stopping mechanism (Falcon, 2019) based on the development set’s average loss per epoch. We fine-tune a separate model for the Amazon and Yelp datasets. Hyperparameters and further details can be found in Appendix A. LkO Input Perturbation. We experiment with the LkO method described in Section 3.2 with k ∈ {1, 2, 3, 4, 5} on the development set. For the endto-end system we choose k = 2 aiming to obtain high output diversity while limiting computation complexity, and avoiding the risk of dropping a majority of the reviews (k > 4) each time. We provide evaluation details in 5.1. Pairwise Summary Classifier. We train two T5-Base models to classify which summary is better, one in terms of coherence, to be used as our ranking method’s primary comparator, and one in terms of fluency to break ties. We experimented with different values for ϵ ∈{0.25, 0.5, 0.75, 1.0}, and chose ϵ = 0.5 for the coherence classifier and ϵ = 0.25 for the fluency classifier. The choice of ϵ was based on dataset statistics per metric and evaluation of each model’s performance on the development set. Baselines. We compare the PASS system to four baselines: COPYCAT (Brazinskas et al., 2020b) is an unsupervised reviews summarizer that is trained to generate a review given other reviews for the same product. The authors suggest a novelty mechanism that controls the extent to which the summary deviates from the inputs. FEWSUM (Brazinskas et al., 2020a) is a fewshot reviews summarizer that builds upon the ideas of CopyCat but also conditions the model on certain linguistic properties such as writing style. T5 is the pre-trained T5-base language model which was not fine-tuned. We do not report results for this model, as it consistently performed worst. T5-FT is the fine-tuned T5-base model described above. We do not report results for MEANSUM (Chu and Liu, 2019) since it was consistently outperformed by FEWSUM (Brazinskas et al., 2020a). 5 Evaluation 5.1 Candidate Summary Generation Recall that our main objective for generating candidate summaries is to encourage output diversity. Hence, we would like to verify that our perturbation method, LkO, produces sufficiently diverse candidates for a given product. In order to measure textual diversity between candidate summaries for a given product, we need to devise a diversity metric. We propose the SPR metric (shorthand for Set-Pairwise-ROUGE) which measures the opposite of diversity, i.e., the average lexical similarity across pairs of summaries from a given set. We base SPR on ROUGE F1 scores for any ngram level, therefore SPR-1 relies on ROUGE-1 F1 scores and so on. SPR Formal Definition. For a given set of summaries S = {s1, ..., sn}, we define the set of all pairs from S as P(S) = {si, sj} si ∈S, sj ∈ S, i ̸= j . We then define the set-pairwise-rouge (SPR) metric as: SPR(S) = 1 |P(S)| · X {si,sj}∈P(S) ROUGE(si, sj) Note that SPR is a general metric of diversity, applicable to an arbitrary set of summaries. Therefore, it can be applied to measure both IPDiversity (in-product diversity, as we do here) and CP-Diversity (cross-product diversity, as we do in Section 5.3). For clarity, we shall denote IP-SPR when measuring IP-Diversity and CP-SPR when measuring CP-Diversity with SPR. Figure 2 depicts a box plot of the IP-SPR-2 scores for k ranging from 1 to 5. We observe 357 Dataset System Length R-1 R-2 R-L CP-SPR-1 CP-SPR-2 CP-SPR-L Coherence Amazon CopyCat 33.45 27.85 4.77 18.86 36.29 14.12 29.52 – FewSum 52.50 33.56 7.16 21.49 34.54 10.61 23.93 -0.200 T5-FT 52.75 37.07 9.68 23.47 25.56 3.32 17.38 -0.050 PASS 47.75 37.43 8.02 23.34 25.79 2.63 17.38 0.150 Gold 49.82 – – – 19.48 1.61 13.00 0.100 Yelp FewSum 52.9 37.29 9.92 22.76 40.82 17.09 30.34 0.050 T5-FT 40.58 38.72 10.26 24.47 38.93 13.05 29.55 -0.250 PASS 52.15 36.91 8.12 23.09 30.88 6.35 21.33 0.200 Gold 49.81 – – – 24.41 2.80 15.98 0.000 Table 1: End-to-End results on the Amazon (top) and Yelp (bottom) test sets. R stands for average ROUGE F1 scores with reference summaries, CP-SPR for Set-Pairwise-ROUGE scores measuring CP-Diversity and Coherence for Best-Worst Scaling scores, which range from -1 (unanimously worst) to +1 (unanimously best), on a crowdsourced human evaluation task. the biggest drop in similarity (increase in diversity) between k = 1 and k = 2. While we aim to increase diversity, we are also mindful of the increase in runtime as k grows. Additionally, we would like to avoid sampling out a majority of reviews (k > 4), since the risk of generating a summary with minority view or low informativeness also increases with k. Indeed, as shown in Figure 3, which depicts a similar box plot but this time of the ROUGE-2 scores against the reference summaries, the variance increases with k and the worst-case ROUGE-2 score decreases with k. Figure 2: IP-SPR-2 scores (measuring IP-Diversity) box plot, for all pairs of candidate summaries generated with LkO input perturbation method for k = 1, ..., 5. While diversity is certainly not the only aspect for evaluating generated summaries, we explore other dimensions in the following sections. 5.2 Candidate Summary Ranking The pairwise summary classifiers can be evaluated directly using human scores from (Fabbri et al., 2021) after adapting them to our ternary classification task. Figure 4 depicts the confusion matrix for Figure 3: ROUGE-2 F1 scores box plot, for all candidate summary sets generated with LkO input perturbation method for k = 1, ..., 5. our coherence classifier. We observe that the estimated probability of a critical error (choosing A over B or B over A) is very low, 0.05, while at the same time the overall accuracy of 0.61 is reasonably high compared to 0.33 and 0.36 achieved by the random and majority (always predicts that A and B are equally coherent) baselines respectively. Applying the classifier to a set of 28 candidates per product, yields a single top ranking candidate for 70% of products in the Amazon test set. To further break ties, we utilize the fluency classifier as a secondary comparator. See Figure 10 in Appendix C for a similar confusion matrix for the fluency classifier. Again, the probability for a critical error is very low, 0.0125, while the overall accuracy is 0.67. After applying fluency as a tie breaker, we find that all products in the Amazon test set have a unique top ranking summary. The training data for both classifiers comes from a domain (News Articles) which is different from our main dataset’s domain (Product Re358 views). We hypothesize that coherence and fluency are linguistic properties that are not heavily tied with the domain, since they relate to a summary’s overall collective and individual sentence quality (Dang, 2005). Indeed, our results show (see Table 2) that PASS benefited from this data despite the risk of a possible domain shift.4 Figure 4: Confusion matrix for the Coherence Pairwise Classifier. 5.3 End-to-End System We evaluate our end-to-end system across 3 dimensions. The first, informativeness, is traditionally evaluated using the ROUGE-1/2/L F1 measures (Lin, 2004b) and we follow suit. The second dimension, which subsumes the self-consistency issue, is coherence. To this end, we conducted a crowdsourced human evaluation task, which compares between the generated summaries of 4 different summarization systems, including our proposed PASS system. We used Best-Worst Scaling (Louviere and Woodworth, 1991; Louviere et al., 2015; Kiritchenko and Mohammad, 2016, 2017) to compute each system’s score as the difference between the percentage of times it was selected as best, and the percentage of times it was selected as worst (Orme, 2009). This is inline with prior work on product review summarization (Brazinskas et al., 2020b,a). As for our third dimension, recall that we would like our system to generate diverse summaries across different products, a notion that we denoted as CP-Diversity. Lacking an existing metric, we use our previously defined SPR-1/2/L measure on the set of final (top-ranked) summaries across all test set products. 4While we did not find evidence suggesting a domain shift, it is an aspect we leave for further investigation in future work. Table 1 reports results for all 3 dimensions. For the Amazon dataset (top table), we observe that PASS outperforms the baselines in coherence and CP-Diversity while keeping a comparable informativeness to the next best system, T5-FT. The only exception being ROUGE-2 in which T5-FT outperforms PASS which could be explained by the somewhat longer summaries it generates. Interestingly, in CP-Diversity, the performance of PASS is closer to human performance than to CopyCat and FewSum but there’s still room to make the summaries even more diverse. For the sake of completeness and following previous work (Chu and Liu, 2019; Brazinskas et al., 2020b,a) we report results on business reviews from the Yelp dataset in the bottom of Table 1. Recall that our key goals were to avoid generating summaries containing crude coherence (CE) and self-consistency (SCE) errors (see Table 3 for examples of such errors). In order to evaluate these directly, both authors independently marked each of the summaries generated by FewSum, T5FT and PASS for the Amazon test set as having a crude error or not, for both types of errors. Table 2 reports the ratios of crude errors per system, considering cases where at least one annotator (I) and both annotators (II) marked as crude. We measured the level of agreement between the two annotators by calculating Cohen’s Kappa coefficients (Cohen, 1960) for each annotation task, which resulted in κCE = 0.571 and κSCE = 0.779. System CE-I CE-II SCE-I SCE-II FewSum 0.50 0.34 0.3 0.25 T5-FT 0.38 0.25 0.3 0.2 PASS 0.19 0.09 0.05 0.00 Table 2: Ratios of crude coherence (CE) and selfconsistency (SCE) errors for each system on the Amazon test set. I/II refer to cases where at least one/both annotators marked the summary as having an error. Finally, for a qualitative impression we provide in Table 4 an example of the systems’ outputs for a product from the Amazon test set. 6 Conclusion In this work we highlight two shortcomings of existing product reviews summarization systems, namely low CP-Diversity and self-inconsistency. We propose the SPR metric to quantify cross prod359 Tights. These tights are very comfortable and durable. They can be worn with ballet slippers or sandals. The color is beautiful and the fabric is soft. They will last a long time. They are great for transitioning from ballet to ballet. Purse. This purse is not as cute as it looks in the picture. It is very small and will not hold a lot of stuff. It would be a great purse if it was a little bigger but it would have been nice to have a purse that would hold more than one purse. Protein Bar. These bars are a great snack bar. They taste good and have a good amount of protein. They do not have a lot of protein in them so they are not as sweet as some protein bars, but for the price, they are well worth it. Tank Top. This tank top is well made, fits well, and is comfortable to wear. The only thing is that it runs a little small, so order a size up from what you normally wear. Other than that, it’s a great top. It’s well made and it looks like it will last a long time. Love it! Table 3: Example of summaries generated by T5-FT and FewSum models for different products in the Amazon test set, which contain crude errors (CE) and selfconsistency errors (SCE). uct similarity of summaries and demonstrate that indeed, humans summaries are far more diverse than system generated summaries. To overcome this issue we rely on stronger pre-trained models such as the recent T5 model which significantly improves the CP-Diversity. However, the second problem still remains and even intensifies as without the safety net of generic content, the risk of incoherent or even self-contradicting text is substantial. To this end, we propose the Perturb and Select summarizer (PASS). In the first step, PASS applies systematic perturbations to the input texts in a way that allows the T5 model to generate multiple summary candidates that sufficiently differ from one another. Given such a set of diverse summaries, PASS applies a trained ranker to smartly select a promising candidate in terms of coherence. Finally, we show that the resulting PASS system, outperforms SOTA models in the domain of product reviews in terms of informativeness, CP-Diversity and coherence. When comparing to a fine-tuned T5 model PASS outperforms it in coherence and CP-Diversity, while maintaining comparable performance for informativeness. PASS. These Reeboks are great for supporting a high arch and are lightweight and comfortable. They come in a variety of colors and sizes, and are ideal for walking or biking. They are also flexible and well made. T5-FT. These Reeboks are a great choice for those with wide feet. They run true to size and the colors are great. They are lightweight and comfortable, yet they are flexible and flexible. They are recommended for people with wide feet. They are also very popular for running and casual wear. FewSum. These running shoes are great! They fit true to size and are very comfortable to run around in. They are light weight and have great support. They run a little on the narrow side, so make sure to order a half size larger than normal. CopyCat. I love these shoes. They are light weight and comfortable to wear. I have worn them for several months now and they are holding up well. I would recommend them to anyone looking for a comfortable shoe. Table 4: Example of summaries generated by PASS, T5-FT, FewSum and CopyCat systems for the same sports shoes reviews. In future work we plan to investigate the Perturb-and-Select framework in order to promote summaries with a plethora of desired linguistic characteristics, other than coherence. We shall further explore ways of extending this framework to employ other input perturbation methods and experiment with scenarios of larger scale input. In addition, we plan to further investigate our proposed SPR evaluation metric for lexical diversity, by studying its correlation with human judgments. Lastly, we believe our proposed framework and evaluation metric may be applicable to other domains of opinion or news summarization. Acknowledgements We would like to thank Hila Gonen, Iftah Gamzu and anonymous reviewers, who helped improve the draft with their invaluable comments and insight. 360 References Reinald Kim Amplayo, Stefanos Angelidis, and Mirella Lapata. 2021. Unsupervised opinion summarization with content planning. In AAAI. Reinald Kim Amplayo and Mirella Lapata. 2020. Unsupervised opinion summarization with noising and denoising. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1934–1945, Online. Association for Computational Linguistics. Stefanos Angelidis, Reinald Kim Amplayo, Yoshihiko Suhara, Xiaolan Wang, and Mirella Lapata. 2021. Extractive opinion summarization in quantized transformer spaces. Trans. Assoc. Comput. Linguistics, 9:277–293. Stefanos Angelidis and Mirella Lapata. 2018. Summarizing opinions: Aspect extraction meets sentiment prediction and they are both weakly supervised. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3675–3686. Arthur Brazinskas, Mirella Lapata, and Ivan Titov. 2020a. Few-shot learning for opinion summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 4119–4135. Association for Computational Linguistics. Arthur Brazinskas, Mirella Lapata, and Ivan Titov. 2020b. Unsupervised opinion summarization as copycat-review generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 5151–5169. Association for Computational Linguistics. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Ziqiang Cao, Furu Wei, Wenjie Li, and Sujian Li. 2018. Faithful to the original: Fact aware neural abstractive summarization. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 4784–4791. AAAI Press. Giuseppe Carenini, Raymond T. Ng, and Adam Pauls. 2006. Multi-document summarization of evaluative text. In EACL 2006, 11st Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of the Conference, April 37, 2006, Trento, Italy. The Association for Computer Linguistics. Eric Chu and Peter J. Liu. 2019. Meansum: A neural model for unsupervised multi-document abstractive summarization. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 1223–1232. PMLR. Andr´e Cibils, Claudiu Musat, Andreea Hossmann, and Michael Baeriswyl. 2018. Diverse beam search for increased novelty in abstractive summarization. CoRR, abs/1802.01457. Maximin Coavoux, Hady Elsahar, and Matthias Gall´e. 2019. Unsupervised aspect-based multi-document abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 42–47. Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37–46. Hoa Trang Dang. 2005. Overview of duc 2005. In Proceedings of the document understanding conference, volume 2005, pages 1–12. Alexander R. Fabbri, Wojciech Kryscinski, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir R. Radev. 2021. Summeval: Reevaluating summarization evaluation. Trans. Assoc. Comput. Linguistics, 9:391–409. WA Falcon. 2019. Pytorch lightning. GitHub. Note: https://github.com/PyTorchLightning/pytorchlightning, 3. Karl Moritz Hermann, Tom´as Kocisk´y, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 1693–1701. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the Tenth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Seattle, Washington, USA, August 22-25, 2004, pages 168–177. ACM. 361 Hitesh Kansal and Durga Toshniwal. 2014. Aspect based summarization of context dependent opinion words. In 18th International Conference in Knowledge Based and Intelligent Information and Engineering Systems, KES 2014, Gdynia, Poland, 15-17 September 2014, volume 35 of Procedia Computer Science, pages 166–175. Elsevier. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Svetlana Kiritchenko and Saif Mohammad. 2017. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465–470. Svetlana Kiritchenko and Saif M. Mohammad. 2016. Capturing reliable fine-grained sentiment associations by crowdsourcing and best–worst scaling. In Proceedings of The 15th Annual Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL), San Diego, California. Wojciech Kryscinski, Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Neural text summarization: A critical evaluation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 540– 551, Hong Kong, China. Association for Computational Linguistics. Alon Lavie and Abhaya Agarwal. 2007. METEOR: An automatic metric for MT evaluation with high levels of correlation with human judgments. In Proceedings of the Second Workshop on Statistical Machine Translation, pages 228–231, Prague, Czech Republic. Association for Computational Linguistics. Kevin Lerman, Sasha Blair-Goldensohn, and Ryan T. McDonald. 2009. Sentiment summarization: Evaluating and learning user preferences. In EACL 2009, 12th Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of the Conference, Athens, Greece, March 30 - April 3, 2009, pages 514–522. The Association for Computer Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Chin-Yew Lin. 2004a. Looking for a few good metrics: Automatic summarization evaluation - how many samples are enough? In Proceedings of the Fourth NTCIR Workshop on Research in Information Access Technologies Information Retrieval, Question Answering and Summarization, NTCIR-4, National Center of Sciences, Tokyo, Japan, June 2-4, 2004. National Institute of Informatics (NII). Chin-Yew Lin. 2004b. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Jordan J Louviere, Terry N Flynn, and Anthony Alfred John Marley. 2015. Best-worst scaling: Theory, methods and applications. Cambridge University Press. Jordan J Louviere and George G Woodworth. 1991. Best-worst scaling: A model for the largest difference judgments. Technical report, Working Paper. Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906–1919, Online. Association for Computational Linguistics. Jun-Ping Ng and Viktoria Abrecht. 2015. Better summarization evaluation with word embeddings for ROUGE. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1925–1930, Lisbon, Portugal. Association for Computational Linguistics. B. Orme. 2009. Maxdiff analysis : Simple counting , individual-level logit , and hb. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, page 311–318, USA. Association for Computational Linguistics. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch´e-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024–8035. Curran Associates, Inc. 362 Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1–140:67. Ori Shapira and Ran Levy. 2020. Massive multidocument summarization of product reviews with weak supervision. CoRR, abs/2007.11348. Yoshihiko Suhara, Xiaolan Wang, Stefanos Angelidis, and Wang-Chiew Tan. 2020. OpinionDigest: A simple framework for opinion summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5789– 5798. Jiaxing Tan, Alexander Kotov, Rojiar Pir Mohammadiani, and Yumei Huo. 2017. Sentence retrieval with sentiment-specific topical anchoring for review summarization. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, CIKM 2017, Singapore, November 06 - 10, 2017, pages 2323–2326. ACM. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 49, 2017, Long Beach, CA, USA, pages 5998–6008. Ashwin K. Vijayakumar, Michael Cogswell, Ramprasaath R. Selvaraju, Qing Sun, Stefan Lee, David J. Crandall, and Dhruv Batra. 2016. Diverse beam search: Decoding diverse solutions from neural sequence models. CoRR, abs/1610.02424. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020a. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Thomas Wolf, Quentin Lhoest, Patrick von Platen, Yacine Jernite, Mariama Drame, Julien Plu, Julien Chaumond, Clement Delangue, Clara Ma, Abhishek Thakur, Suraj Patil, Joe Davison, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angie McMillan-Major, Simon Brandeis, Sylvain Gugger, Franc¸ois Lagunas, Lysandre Debut, Morgan Funtowicz, Anthony Moi, Sasha Rush, Philipp Schmidd, Pierric Cistac, Victor Muˇstar, Jeff Boudier, and Anna Tordjmann. 2020b. Datasets. GitHub. Note: https://github.com/huggingface/datasets, 1. Haibing Wu, Yiwei Gu, Shangdi Sun, and Xiaodong Gu. 2016. Aspect-based opinion summarization with convolutional neural networks. In 2016 International Joint Conference on Neural Networks, IJCNN 2016, Vancouver, BC, Canada, July 24-29, 2016, pages 3157–3163. IEEE. Wenting Xiong and Diane Litman. 2014. Empirical analysis of exploiting review helpfulness for extractive summarization of online reviews. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1985–1995, Dublin, Ireland. Dublin City University and Association for Computational Linguistics. Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J. Liu. 2020a. PEGASUS: pre-training with extracted gap-sentences for abstractive summarization. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 11328– 11339. PMLR. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020b. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. 363 A PASS Implementation Details and Hyperparameters All models were implemented with the PyTorch (Paszke et al., 2019) deep learning framework, utilizing the T5 (Raffel et al., 2020) pre-trained model and tokenizer implementations from HuggingFace’s Transformers (Wolf et al., 2020a) library, evaluation metrics from HuggingFace’s Datasets (Wolf et al., 2020b) library and PyTorch Lightning (Falcon, 2019) as a model training framework. A.1 T5 Fine-Tuned Summarizer We fine-tune a pre-trained T5-Base model (220M parameters (Raffel et al., 2020)) for product reviews summarization (an abstractive text summarization task) on the training set, employing the Adam optimizer (Kingma and Ba, 2015) with weight decay (Loshchilov and Hutter, 2019). We train the model for a maximum of 20 epochs on a single NVIDIA Tesla V100 GPU, while employing a standard early stopping mechanism (Falcon, 2019) based on the development set’s average loss per epoch. We employ a standard beam search decoding algorithm during inference for generating text. We tune the model’s hyperparameters on the development set, and provide a list of the final model’s tuned hyperparametrs along with the range of values tested during tuning. Hyperparameters T5 Encoder • Max input sequence length = 512 tokens • Training batch size = 8, [8, 12, 16] • Evaluation batch size = 12, [8, 12, 16] Adam Optimizer • Learning rate = 3e−4, [1e−4, 3e−4, 5e−4] • ϵ = 1e −8, [1e −8, 3e −8, 5e −8] • Weight decay: 0.0 • Number of warmup steps: 0 • Gradient accumulation steps = 2, [1, 2, 4] • Max gradient norm = 1.0 T5 Decoder • Max output sequence length = 128 tokens • Min output sequence length = 16 tokens • Beam size = 2, [2, 3, 4] • Length penalty = 2, [1, 2, 3] • Repetition penalty = 2, [1, 2, 3] LkO Input Perturbation (PASS system only) • k = 2, [1, 2, 3, 4, 5] A.2 Pairwise Summary Classifiers For each pairwise summary classifier (coherence, fluency), we fine-tune a pre-trained T5-Base model (220M parameters (Raffel et al., 2020)) for abstractive text summarization task on the respective training set employing the Adam optimizer (Kingma and Ba, 2015) with weight decay (Loshchilov and Hutter, 2019). We train for a maximum of 20 epochs on a single NVIDIA Tesla V100 GPU, while employing a standard early stopping mechanism (Falcon, 2019) based on the development set’s average loss per epoch. We employ a standard greedy decoding algorithm during inference for generating the class label. We tune the model’s hyperparameters on the development set, and provide a list of the final model’s tuned hyperparametrs along with the range of values tested during tuning. Hyperparameters Dataset • Coherence scores difference threshold ϵ = 0.5, [0.25, 0.5, 0.75, 1.0] • Fluency scores difference threshold ϵ = 0.25, [0.25, 0.5, 0.75, 1.0] T5 Encoder • Max input sequence length = 400 tokens • Training batch size = 16, [8, 12, 16] • Evaluation batch size = 16, [8, 12, 16] Adam Optimizer • Learning rate = 1e−4, [1e−4, 3e−4, 5e−4] • ϵ = 1e −8, [1e −8, 3e −8, 5e −8] • Weight decay: 0.0 • Number of warmup steps: 0 • Gradient accumulation steps = 4, [1, 2, 4] • Max gradient norm = 1.0 T5 Decoder • Max output sequence length = 2 tokens • Min output sequence length = 2 tokens B Summary Examples We provide examples for output summaries generated by the different summarization systems discussed in the main paper. Each example qualitatively highlights a different aspect by which we evaluate the quality of a summary, or identify its shortcomings. 364 PASS. This camera is good to have as a first camera before investing in a DSLR. The quality of the pictures is great, and the camera is easy to use. It takes some time to learn about the features and settings, but overall it’s a great camera. T5-FT. This camera is a great camera for taking professional photos. It is easy to use and takes excellent pictures. The low light feature is outstanding and will be helpful in museums and other venues where flash is not allowed. The battery is constantly malfunctioning making the camera unusable. The on off button is also malfinctioning. FewSum. This camera is a great camera for the price. It takes great pictures and is easy to use. The only drawback is that the battery life is not as good as the camera that comes with the camera. It would be nice if it had a battery life to last longer. Overall, it’s a good camera. CopyCat. This is a great camera for the price. It is easy to set up and use. The only downside is that it takes a while to learn how to use it, but it’s not a problem. Table 5: Example of summaries generated by PASS, T5-FT (Raffel et al., 2020), FewSum (Brazinskas et al., 2020a) and CopyCat (Brazinskas et al., 2020b) systems for the same reviews for a digital camera. Travel Sound Conditioner. This is a great product for the price. The sound quality is good and the sound is good. The only problem is that it is not loud enough for a small room. It is loud enough to drown out background noise, but not very loud. Overall, it’s a good product and would recommend it to anyone. Motion Sickness Tablets. This is a great product at a great price. It is easy to use and easy to take. The pills are easy to swallow and do not take up a lot of space. The price is great for a product that will last a long time. Would recommend this product to anyone who suffers from nausea or sickness. Digital Camera. This camera is a great camera for the price. It takes great pictures and is easy to use. The only drawback is that the battery life is not as good as the camera that comes with the camera. It would be nice if it had a battery life to last longer. Overall, it’s a good camera. Table 6: Example of similar summaries generated by FewSum (Brazinskas et al., 2020a) for three different products. Review 1. The machine is very tricky. It requires some trial and errors to make it work right. I do not like to put oil in the dough; however, it appears to me that without oil it is impossible to make tortilla or chapatti. It is useless for me. Review 2. Fun and easy to use! Took me one batch to get my technique worked out, but it was very simple, easy to follow directions. Easy clean up too! I would recommend this to anyone looking for an electric tortilla maker! Summary. This tortilla maker is a great option for making tortillas but it does require some trial and error to make it work right. It requires some trial and error to make it work right. Yes, you should grill them after cooking to get the toasted look. It is easy to use and very easy to clean up. Table 7: Example of a self-contradicting summary generated by our fine-tuned T5 (T5-FT) model. Candidate 1. These NuGo bars are high quality and they come in a variety of flavors and sizes which make them perfect for serving as a snack or as a replacement for processed foods. They are low glycemic and have a smooth, vanilla-like texture which makes them very good. Candidate 2. These NuGo bars are high quality and they come in a variety of flavors and sizes which makes them ideal for snacking on the go. The taste is great and the nutritional value is great as well. Although they can be a little sweet, they are not too sweet. Candidate 3. These NuGo bars are high quality and they come in a variety of flavors and sizes. They are low glycemic and have a great taste. While they may be sweet, they can also have a chalky or barky texture. These are great for replacing junk food with healthy snacks. Candidate 4. These NuGo bars are high quality and they taste great. They are low glycemic, and they contain no added sugar or artificial flavors. These are great for a healthy snack or for a quick breakfast. Candidate 5. These NuGo bars are very good quality and they come in a variety of flavors. They are high in calories and fiber, and are great for snacking on the go. They are often a bit chewy, but they are definitely worth the money. Table 8: Example of 5 candidate summaries (out of 28) generated by PASS for the same product with L2O input perturbation. 365 C Evaluation Figures We provide figures which extend those appearing in the Evaluation section of the main paper. C.1 Candidate Summary Generation Figure 5: Length box plot for all candidate summary sets generated with LkO input perturbation method for k = 1, .., 5. Figure 6: ROUGE-1 box plot for all candidate summary sets generated with LkO input perturbation method for k = 1, .., 5. Figure 7: SPR-1 box plot for all pairs of candidate summaries generated with LkO input perturbation method for k = 1, .., 5. Figure 8: ROUGE-L box plot for all candidate summary sets generated with LkO input perturbation method for k = 1, .., 5. Figure 9: SPR-L box plot for all pairs of candidate summaries generated with LkO input perturbation method for k = 1, .., 5. C.2 Candidate Summary Ranking Figure 10: Confusion matrix for the Fluency Pairwise Classiifer.
|
2021
|
30
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3881–3895 August 1–6, 2021. ©2021 Association for Computational Linguistics 3881 Cross-language Sentence Selection via Data Augmentation and Rationale Training Yanda Chen1, Chris Kedzie1, Suraj Nair2, Petra Galuˇsˇc´akov´a2, Rui Zhang3, Douglas W. Oard2, Kathleen McKeown1 1Columbia University, 2University of Maryland, 3Penn State University [email protected], {kedzie, kathy}@cs.columbia.edu {srnair, petra, oard}@umd.edu, [email protected] Abstract This paper proposes an approach to crosslanguage sentence selection in a low-resource setting. It uses data augmentation and negative sampling techniques on noisy parallel sentence data to directly learn a cross-lingual embedding-based query relevance model. Results show that this approach performs as well as or better than multiple state-of-theart machine translation + monolingual retrieval systems trained on the same parallel data. Moreover, when a rationale training secondary objective is applied to encourage the model to match word alignment hints from a phrase-based statistical machine translation model, consistent improvements are seen across three language pairs (EnglishSomali, English-Swahili and English-Tagalog) over a variety of state-of-the-art baselines. 1 Introduction Sentence-level query relevance prediction is important for downstream tasks such as query-focused summarization and open-domain question answering; accurately pinpointing sentences containing information that is relevant to the query is critical to generating a responsive summary/answer (e.g., Baumel et al. (2016, 2018)). In this work, we focus on sentence-level query relevance prediction in a cross-lingual setting, where the query and sentence collection are in different languages and the sentence collection is drawn from a low-resource language. Our approach enables English speakers (e.g., journalists) to find relevant information expressed in local sources (e.g., local reaction to the pandemic and vaccines in Somalia). While we can use machine translation (MT) to translate either the query or each sentence into a common language, and then use a monolingual Information Retrieval (IR) system to find relevant sentences, work on Probabilistic Structured Queries (PSQ) (Darwish and Oard, 2003) has shown that the performance of such MT+IR pipelines is hindered by errors in MT. As is well known, complete translation of the sentence collection is not necessary. Inspired by previous work (Vuli´c and Moens, 2015), we go a step further and propose a simple cross-lingual embedding-based model that avoids translation entirely and directly predicts the relevance of a query-sentence pair (where the query and sentence are in different languages). For training, we treat a sentence as relevant to a query if there exists a translation equivalent of the query in the sentence. Our definition of relevance is most similar to the lexical-based relevance used in Gupta et al. (2007) and Baumel et al. (2018) but our query and sentence are from different languages. We frame the task as a problem of finding sentences that are relevant to an input query, and thus, we need relevance judgments for query-sentence pairs. Our focus, however, is on low-resource languages where we have no sentence-level relevance judgments with which to train our query-focused relevance model. We thus leverage noisy parallel sentence collections previously collected from the web. We use a simple data augmentation and negative sampling scheme to generate a labeled dataset of relevant and irrelevant pairs of queries and sentences from these noisy parallel corpora. With this synthetic training set in hand, we can learn a supervised cross-lingual embedding space. While our approach is competitive with pipelines of MT-IR, it is still sensitive to noise in the parallel sentence data. We can mitigate the negative effects of this noise if we first train a phrase-based statistical MT (SMT) model on the same parallel sentence corpus and use the extracted word alignments as additional supervision. With these alignment hints, we demonstrate consistent and significant improvements over neural and statistical MT+IR (Niu et al., 2018; Koehn et al., 2007; Heafield, 2011), 3882 three strong cross-lingual embedding-based models (Bivec (Luong et al., 2015), SID-SGNS (Levy et al., 2017), MUSE (Lample et al., 2018)), a probabilistic occurrence model (Xu and Weischedel, 2000), and a multilingual pretrained model XLMRoBERTa (Conneau et al., 2020). We refer to this secondary training objective as rationale training, inspired by previous work in text classification that supervises attention over rationales for classification decisions (Jain and Wallace, 2019). To summarize, our contributions are as follows. We (i) propose a data augmentation and negative sampling scheme to create a synthetic training set of cross-lingual query-sentence pairs with binary relevance judgements, and (ii) demonstrate the effectiveness of a Supervised Embedding-based Cross-Lingual Relevance (SECLR) model trained on this data for low-resource sentence selection tasks on text and speech. Additionally, (iii) we propose a rationale training secondary objective to further improve SECLR performance, which we call SECLR-RT. Finally, (iv) we conduct training data ablation and hubness studies that show our method’s applicability to even lower-resource settings and mitigation of hubness issues (Dinu and Baroni, 2015; Radovanovi´c et al., 2010). These findings are validated by empirical results of experiments in a low-resource sentence selection task, with English queries over sentence collections of text and speech in Somali, Swahili, and Tagalog. 2 Related Work Query-focused Sentence Selection Sentencelevel query relevance prediction is important for various downstream NLP tasks such as queryfocused summarization (Baumel et al., 2016, 2018; Feigenblat et al., 2017) and open-domain question answering (Chen et al., 2017; Dhingra et al., 2017; Kale et al., 2018). Such applications often depend on a sentence selection system to provide attention signals on which sentences to focus upon to generate a query-focused summary or answer a question. Cross-language Sentence Selection A common approach to cross-language sentence selection is to use MT to first translate either the query or the sentence to the same language and then perform standard monolingual IR (Nie, 2010). The risk of this approach is that errors in translation cascade to the IR system. As an alternative to generating full translations, PSQ (Darwish and Oard, 2003) uses wordalignments from SMT to obtain weighted query term counts in the passage collection. In other work, Xu and Weischedel (2000) use a 2-state hidden Markov model (HMM) to estimate the probability that a passage is relevant given the query. Cross-lingual Word Embeddings Crosslingual embedding methods perform cross-lingual relevance prediction by representing query and passage terms of different languages in a shared semantic space (Vuli´c and Moens, 2015; Litschko et al., 2019, 2018; Joulin et al., 2018). Both supervised approaches trained on parallel sentence corpora (Levy et al., 2017; Luong et al., 2015) and unsupervised approaches with no parallel data (Lample et al., 2018; Artetxe et al., 2018) have been proposed to train cross-lingual word embeddings. Our approach differs from previous cross-lingual word embedding methods in two aspects. First, the focus of previous work has mostly been on learning a distributional word representation where translation across languages is primarily shaped by syntactic or shallow semantic similarity; it has not been tuned specifically for cross-language sentence selection tasks, which is the focus of our work. Second, in contrast to previous supervised approaches that train embeddings directly on a parallel corpus or bilingual dictionary, our approach trains embeddings on an artificial labeled dataset augmented from a parallel corpus and directly represents relevance across languages. Our data augmentation scheme to build a relevance model is inspired by Boschee et al. (2019), but we achieve significant performance improvement by incorporating rationale information into the embedding training process and provide detailed comparisons of performance with other sentence selection approaches. Trained Rationale Previous research has shown that models trained on classification tasks sometimes do not use the correct rationale when making predictions, where a rationale is a mechanism of the classification model that is expected to correspond to human intuitions about salient features for the decision function (Jain and Wallace, 2019). Research has also shown that incorporating human rationales to guide a model’s attention distribution can potentially improve model performance on classification tasks (Bao et al., 2018). Trained rationales have also been used in neural MT (NMT); incorporat3883 ing alignments from SMT to guide NMT attention yields improvements in translation accuracy (Chen et al., 2016). 3 Methods We first describe our synthetic training set generation process, which converts a parallel sentence corpus for MT into cross-lingual query-sentence pairs with binary relevance judgements for training our SECLR model. Following that, we detail our SECLR model and finish with our method for rationale training with word alignments from SMT. 3.1 Training Set Generation Algorithm Relevant query/sentence generation. Assume we have a parallel corpus of bilingual sentence pairs equivalent in meaning. Let (E, S) be one such sentence pair, where E is in the query language (in our case, English) and S is in the retrieval collection language (in our case, low-resource languages). For every unigram q in E that is not a stopword, we construct a positive relevant sample by viewing q as a query and S as a relevant sentence. Because sentences E and S are (approximately) equivalent in meaning, we know that there likely exists a translation equivalent of q in the sentence S and so we label the (q, S) pair as relevant (i.e. r = 1). For example, one English-Somali sentence pair is E=“true president gaas attend meeting copenhagen”, S=“ma runbaa madaxweyne gaas baaqday shirka copenhegan” (stopwords removed). By extracting unigrams from E as queries, we generate the following positive examples: (q=“true”, S, r = 1), (q=“president”, S, r = 1), (q=“gaas”, S, r = 1), ..., (q=“copenhagen”, S, r = 1). We generate the positive half of the training set by repeating the above process for every sentence pair in the parallel corpus. We limit model training to unigram queries since higher order ngrams appear fewer times and treating them independently reduces the risk of over-fitting. However, our model processes multi-word queries during evaluation, as described in Section 3.2. Irrelevant query/sentence generation. Since learning with only positive examples is a challenging task, we opt to create negative examples, i.e. tuples (q, S, r = 0), via negative sampling. For each positive sample (q, S, r = 1), we randomly select another sentence pair (E′, S′) from the parallel corpus. We then check whether S′ is relevant to q or not. Note that both the query q and sentence E′ are in the same language, so checking whether q or a synonym can be found in E′ is a monolingual task. If we can verify that there is no direct match or synonym equivalent of q in E′ then by transitivity it is unlikely there exists a translation equivalent in S′, making the pair (q, S′) a negative example. To account for synonymy when we check for matches, we represent q and the words in E′ with pretrained word embeddings. Let wq, wq′ ∈Rd be the embeddings associated with q and the words q′ ∈E′. We judge the pair (q, S′) to be irrelevant (i.e. r = 0) if: max q′∈E′ cos-sim(wq, wq′) ≤λ1 where λ1 is a parameter. We manually tuned the relevance threshold λ1 on a small development set of query-sentence pairs randomly generated by the algorithm, and set λ1 = 0.4 to achieve highest label accuracy on the development set. If (q, S′) is not relevant we add (q, S′, r = 0) to our synthetic training set, otherwise we re-sample (E′, S′) until a negative sample is found. We generate one negative sample for each positive sample to create a balanced dataset. For example, if we want to generate a negative example for the positive example (q=“meeting”, S=“ma runbaa madaxweyne gaas baaqday shirka copenhegan”, r = 1), we randomly select another sentence pair (E′=“many candidates competing elections one hopes winner”, S′=“musharraxiin tiro badan sidoo u tartamaysa doorashada wuxuuna mid kasta rajo qabaa guusha inay dhinaciisa ahaato”) from the parallel corpus. To check whether q=“meeting” is relevant to S′, by transitivity it suffices to check whether q=“meeting” or a synonym is present in E′, a simpler monolingual task. If q is irrelevant to S′, we add (q, S′, r = 0) as a negative example. 3.2 Cross-Lingual Relevance Model We propose SECLR, a model that directly makes relevance classification judgments for queries and sentences of different languages without MT as an intermediate step by learning a cross-lingual embedding space between the two languages. Not only should translation of equivalent words in either language map to similar regions in the embedding space, but dot products between query and sentence words should be correlated with the probability of relevance. We assume the training set generation process (Section 3.1) provides us with a corpus of n query-sentence pairs along 3884 with their corresponding relevance judgements, i.e. D = {(qi, Si, ri)}|n i=1. We construct a bilingual vocabulary V = VQ ∪VS and associate with it a matrix W ∈Rd×|V| where wx = W·,x is the word embedding associated with word x ∈V. When the query is a unigram q (which is true by design in our training data D), we model the probability of relevance to a sentence S as: p(r = 1|q, S; W) = σ max s∈S w⊺ qws where σ denotes the logistic sigmoid (σ(x) = 1/ (1 + exp(−x))). In our evaluation setting, the query is very often a phrase Q = [q1, . . . , q|Q|]. In this case, we require all query words to appear in a sentence in order for a sentence to be considered as relevant. Thus, we modify our relevance model to be: p(r = 1|Q, S; W) = σ min q∈Q max s∈S w⊺ qws Our only model parameter is the embedding matrix W which is initialized with pretrained monolingual word embeddings and learned via minimization of the cross entropy of the relevance classification task: Lrel = −log p(r|q, S; W) 3.3 Guided Alignment with Rationale Training We can improve SECLR by incorporating additional alignment information as a secondary training objective, yielding SECLR-RT. Our intuition is that after training, the word ˆs = arg maxs∈S w⊺ swq should correspond to a translation of q. However, it is possible that ˆs simply co-occurs frequently with the true translation in our parallel data but its association is coincidental or irrelevant outside the training contexts. We use alignment information to correct for this. We run two SMT word alignment models, GIZA++ (Och and Ney, 2003) and Berkeley Aligner (Haghighi et al., 2009), on the orginal parallel sentence corpus. The two resulting alignments are concatenated as in Zbib et al. (2019) to estimate a unidirectional probabilistic word translation matrix A ∈[0, 1]|VQ|×|VS|, such that A maps each word in the query language vocabulary to a list of document language words with different probabilities, i.e. Aq,s is the probability of translating q to s and P s∈VS Aq,s = 1. For each relevant training sample, i.e. (q, S, r = 1), we create a rationale distribution ρ ∈[0, 1]|S| which is essentially a re-normalization of possible query translations found in S and represents our intuitions about which words s ∈S that q should be most similar to in embedding space, i.e. ρs = Aq,s P s′∈S Aq,s′ . for s ∈S. We similarly create a distribution under our model, α ∈[0, 1]|S|, where αs = exp (w⊺ qws) P s′∈S exp (w⊺ qws′) for s ∈S. To encourage α to match ρ, we impose a Kullback–Leibler (KL) divergence penalty, denoted as: Lrat = KL(ρ∥α) to our overall loss function. The total loss for a single positive sample then will be a weighted sum of the relevance classification objective and the KL divergence penalty, i.e. L = Lrel + λ2Lrat where λ2 is a relative weight between the classification loss and rationale similarity loss. Note that we do not consider rationale loss for the following three types of samples: negative samples, positive samples where the query word is not found in the translation matrix, and positive samples where none of the translations of the query in the matrix are present in the source sentence. 4 Experiments 4.1 Dataset Generation from Parallel Corpus The parallel sentence data for training our proposed method and all baselines includes the parallel data provided in the BUILD collections of both the MATERIAL1 and LORELEI (Christianson et al., 2018) programs for three low resource languages: Somali (SO), Swahili (SW), and Tagalog (TL) (each paired with English). Additionally, we include in our parallel corpus publicly available resources from OPUS (Tiedemann, 2012), and lexicons mined from Panlex (Kamholz et al., 2014) and Wiktionary.2 Statistics of these parallel corpora and augmented data are shown in Table 1 and Table 2, respectively. Other preprocessing details are in Appendix A. 1https://www.iarpa.gov/index.php/ research-programs/material 2https://dumps.wikimedia.org/ 3885 EN-SO EN-SW EN-TL # sents. 69,818 251,928 232,166 EN tkn. 1,827,826 1,946,556 2,553,439 LR tkn. 1,804,428 1,848,184 2,682,076 Table 1: Parallel corpus statistics; “EN tkn.” refers to number of English tokens in the parallel corpus; “LR tkn.” refers to number of low-resource tokens (Somali, Swahili, Tagalog) in the parallel corpus. Lang. Pair Augmented Dataset Size EN-SO 1,649,484 EN-SW 2,014,838 EN-TL 2,417,448 Table 2: Augmented dataset statistics; “augmented dataset size” refers to total number of positive and negative query-sentence samples in the augmented dataset. 4.2 Query Sets and Evaluation Sets We evaluate our sentence-selection model on English (EN) queries over three collections in SO, SW, and TL recently made available as part of the IARPA MATERIAL program. In contrast to our training data which is synthetic, our evaluation datasets are human-annotated for relevance between real-world multi-domain queries and documents. For each language there are three partitions (Analysis, Dev, and Eval), with the former two being smaller collections intended for system development, and the latter being a larger evaluation corpus. In our main experiments we do not use Analysis or Dev for development and so we report results for all three (the ground truth relevance judgements for the TL Eval collection have not been released yet so we do not report Eval for TL). See Table 3 for evaluation statistics. All queries are text. The speech documents are first transcribed with an ASR system (Ragni and Gales, 2018), and the 1-best ASR output is used in the sentence selection task. Examples of the evaluation datasets are shown in Appendix B. We refer readers to Rubino (2020) for further details about MATERIAL test collections used in this work. While our model and baselines work at the sentence-level, the MATERIAL relevance judgements are only at the document level. Following previous work on evaluation of passage retrieval, we aggregate our sentence-level relevance scores to obtain document-level scores (Kaszkiel and Zobel, 1997; Wade and Allan, 2005; Fan et al., 2018; Inel et al., 2018; Akkalyoncu Yilmaz et al., 2019). Given a document D = [S1, . . . , S|D|], which is a sequence of sentences, and a query Q, following Liu and Croft (2002) we assign a relevance score by: ˆr = max S∈D p(r = 1|Q, S; W) 4.3 Experiment Settings We initialize English word embeddings with word2vec (Mikolov et al., 2013), and initialize SO/SW/TL word embeddings with FastText (Grave et al., 2018). For training we use a SparseAdam (Kingma and Ba, 2015) optimizer with learning rate 0.001. The hyperparameter λ2 in Section 3.3 is set to be 3 so that Lrel and λ2Lrat are approximately on the same scale during training. More details on experiments are included in Appendix C. 4.4 Baselines Cross-Lingual Word Embeddings. We compare our model with three other cross-lingual embedding methods, Bivec (Luong et al., 2015), MUSE (Lample et al., 2018), and SID-SGNS (Levy et al., 2017). Bivec and SID-SGNS are trained using the same parallel sentence corpus as the dataset generation algorithm used to train SECLR; thus, Bivec and SID-SGNS are trained on parallel sentences while SECLR is trained on query-sentence pairs derived from that corpus. We train MUSE with the bilingual dictionary from Wiktionary that is used in previous work (Zhang et al., 2019). The SO-EN, SW-EN and TL-EN dictionaries have 7633, 5301, and 7088 words respectively. Given embeddings W ′ from any of these methods, we compute sentence level relevance scores similarly to our model but use the cosine similarity: p(r = 1|Q, S; W ′) = min q∈Q max s∈S cos-sim(w′ s, w′ q) since these models are optimized for this comparison function (Luong et al., 2015; Lample et al., 2018; Levy et al., 2017). Document aggregation scoring is handled identically to our SECLR models (see Section 4.2). MT+IR. We also compare to a pipeline of NMT (Niu et al., 2018) with monolingual IR and a pipeline of SMT 3 with monolingual IR. Both MT systems are trained on the same parallel sentence 3We used Moses (Koehn et al., 2007) and KenLM for the language model (Heafield, 2011). 3886 Lang. Analysis Dev Eval #Q #T #S #Q #T #S #Q #T #S Somali 300 338 142 300 482 213 1300 10717 4642 Swahili 300 316 155 300 449 217 1300 10435 4310 Tagalog 300 291 171 300 460 244 / / / Table 3: MATERIAL dataset statistics: “#Q” refers to the number of queries; “#T” refers to the number of text documents; “#S” refers to the number of speech documents. There is no Tagalog Eval dataset. Somali Swahili Analysis Dev Eval Analysis Dev Eval Method T S T S T S T S T S T S Bivec 19.6 16.2 15.0 12.0 4.2 4.5 23.9 22.7 21.9 21.6 6.2 4.8 SID-SGNS 25.5 24.3 22.2 16.0 10.2 9.1 38.8 36.3 33.7 30.3 16.2 13.6 MUSE 9.9 9.9 10.3 16.5 1.9 2.0 27.8 24.5 27.3 28.8 9.5 8.1 NMT+IR 18.8 12.5 21.1 13.4 9.4 8.4 23.7 24.9 26.8 26.7 15.3 11.4 SMT+IR 17.4 11.2 19.1 16.8 9.1 8.3 25.5 28.6 27.1 25.2 15.4 13.3 PSQ 27.0 16.6 25.0 20.7 11.1 8.6 39.0 36.6 38.0 38.6 20.4 13.8 XLM-R 13.9 11.0 10.7 12.4 2.3 2.9 23.3 29.0 20.0 29.7 6.2 7.5 SECLR 27.8 24.4 23.0 17.4 7.7 7.4 43.8 37.9 40.3 38.1 16.0 13.1 SECLR-RT 35.4† 28.4 29.5 22.0 13.1† 11.2† 48.3† 48.1† 39.6 45.4 22.7† 17.7† Table 4: Document-level MAP scores for text (T) and speech (S) for Somali and Swahili. † indicates significance at the p = 0.01 level between SECLR-RT and the best baseline. Analysis Dev Method T S T S Bivec 36.7 41.4 39.6 26.9 SID-SGNS 44.6 43.9 40.9 41.7 MUSE 27.4 26.5 26.0 16.5 NMT+IR 37.7 42.3 32.6 37.5 SMT+IR 44.4 52.7 39.3 35.3 PSQ 51.6 55.0 52.7 44.7 SECLR 46.7 45.0 49.3 33.9 SECLR-RT 61.1 55.5 59.0 45.7 Table 5: Document-level MAP scores for text (T) and speech (S) for Tagalog. data as our SECLR models. The 1-best output from each MT system is then scored with Indri (Strohman et al., 2005) to obtain relevance scores. Details of NMT and SMT systems are included in Appendix C.2. PSQ. To implement the PSQ model of Darwish and Oard (2003), we use the same alignment matrix as in rationale training (see Section 3.3) except that here we normalize the matrix such that ∀s ∈VD, P q∈VQ Aq,s = 1. Additionally, we embed the PSQ scores into a two-state hidden Markov model which smooths the raw PSQ scores with a background unigram language model (Xu and Weischedel, 2000). The PSQ model scores each sentence and then aggregates the scores to document level as in Section 4.2. Multilingual XLM-RoBERTa. We compare our model to the cross-lingual model XLM-RoBERTa (Conneau et al., 2020), which in previous research has been shown to have better performance on lowresource languages than multilingual BERT (Devlin et al., 2019). We use the Hugging Face implementation (Wolf et al., 2019) of XLM-RoBERTa (Base). We fine-tuned the model on the same augmented dataset of labeled query-sentence pairs as the SECLR models, but we apply the XLMRoBERTa tokenizer before feeding examples to the model. We fine-tuned the model for four epochs using an AdamW optimizer (Loshchilov and Hutter, 2019) with learning rate 2 × 10−5. Since XLMRoBERTa is pretrained on Somali and Swahili but not Tagalog, we only compare our models to XLMRoBERTa on Somali and Swahili. 3887 5 Results and Discussion We report Mean Average Precision (MAP) of our main experiment in Table 4 (SO & SW) and Table 5 (TL). Overall, we see that SECLR-RT consistently outperforms the other baselines in 15 out of 16 settings, and in the one case where it is not the best (SW Dev text), SECLR is the best. SECLR-RT is statistically significantly better than the best baseline on all Eval partitions.4 Since Analysis/Dev are relatively small, only three out of 12 Analysis/Dev settings are significant. The differences between SECLR and SECLR-RT can be quite large (e.g., as large as 70.4% relative improvement on SO Eval text), suggesting that the rationale training provides a crucial learning signal to the model. Bivec and MUSE under-perform both of our model variants across all test conditions, suggesting that for the sentence selection task the relevance classification objective is more important than learning monolingual distributional signals. Curiously, SID-SGNS is quite competitive with SECLR, beating it on SO and SW Eval (both modalities) and TL Dev speech (five out of 16 test conditions) and is competitive with the other baselines. Again, the rationale training proves more effective as SID-SGNS never surpasses SECLR-RT. While MT+IR is a competitive baseline, it is consistently outperformed by PSQ across all test conditions, suggesting that in low-resource settings it is not necessary to perform full translation to achieve good sentence selection performance. SMT, PSQ, and SECLR-RT all make use of the same word-alignment information but only SMT generates translations, adding additional evidence to this claim. PSQ and SECLR are close in performance on Analysis and Dev sets with SECLR eking out a slight advantage on seven of 12 Anaylsis/Dev set conditions. On the larger Eval partitions, it becomes clearer that PSQ is superior to SECLR, suggesting that the relevance classification objective is not as informative as word alignment information. The relevance classification and trained rationale objectives capture slightly different information it seems; SECLR-RT, which uses both, out-performs PSQ across all 16 test conditions. 6 Training Data Ablation Study In Section 5, we have shown that SECLR-RT consistently out-performs all baselines across all languages. Since this work targets cross-language sentence selection in a low-resource setting, we perform a training data ablation study to understand how training data size affects effectiveness. We performed the ablation study for our two models SECLR and SECLR-RT, and the two strongest baseline methods PSQ and SID-SGNS. To simulate further the scenario of data scarcity, we sub-sampled our parallel corpus uniformly at random for 5%, 10%, 25%, 50% of the sentence pairs of the original corpus. Each sentence pair in the parallel corpus is sampled with equal probability regardless of sentence length. For consistency, for each sample size, the same sampled parallel corpus is used across all models. The word alignment probability matrix used by PSQ and SECLR-RT is generated from the same sampled corpus. Since we tune the vocabulary size on the Dev set, for fair comparison we only report MAP scores on the Analysis and Eval sets. We plot MAP scores of the four models as a function of percentage of data sampled in Figure 1. Overall, we see that SECLR-RT consistently outperforms other baselines across all sample sizes in 9 out of 10 settings, and in the one case where it does not yield consistent improvement (Tagalog Analysis speech), SECLR-RT achieves comparable performance to PSQ. In the low-resource setting when the sample size is 5% or 10%, SECLR consistently underperforms other models, confirming our observation that SECLR is sensitive to noise and vulnerable to learning co-occurrences of word pairs that are in fact irrelevant. When the sample size is 5% or 10%, PSQ consistently achieves better performance than SID-SGNS and SECLR (although still under-performing SECLR-RT), indicating that alignment-based methods are more robust to noise and especially useful when data is extremely scarce. The fact that SECLR-RT consistently out-performs SECLR by a wide margin for small sample sizes indicates the necessity and effectiveness of incorporating alignment-based information into SECLR to improve the robustness of the model and learn more precise alignments. 4We use a two-tailed paired t-test with Bonferroni correction for multiple comparisons at p < 0.01 for all significance tests. 3888 5 10 25 50 100 Percent of Data Used 10 20 30 MAP Somali Analysis Text 5 10 25 50 100 Percent of Data Used 10 15 20 25 MAP Somali Analysis Speech 5 10 25 50 100 Percent of Data Used 5 10 MAP Somali Eval Text 5 10 25 50 100 Percent of Data Used 2.5 5.0 7.5 10.0 MAP Somali Eval Speech 5 10 25 50 100 Percent of Data Used 20 30 40 50 MAP Swahili Analysis Text 5 10 25 50 100 Percent of Data Used 20 30 40 MAP Swahili Analysis Speech 5 10 25 50 100 Percent of Data Used 5 10 15 20 MAP Swahili Eval Text 5 10 25 50 100 Percent of Data Used 5 10 15 MAP Swahili Eval Speech 5 10 25 50 100 Percent of Data Used 20 30 40 50 60 MAP Tagalog Analysis Text 5 10 25 50 100 Percent of Data Used 20 30 40 50 MAP Tagalog Analysis Speech PSQ SID-SGNS SECLR SECLR-RT Figure 1: Ablation study results of model performances as a function of sub-sampling percentages. Note that the x-coordinate uses the log scale for better illustration of low-resource cases. 7 Alleviating the Hubness Problem In this section, we show that by incorporating alignment information through rationale training, SECLR-RT significantly alleviates the hubness problem present in the trained cross-lingual embedding space produced by SECLR. Previous research on cross-lingual word embeddings has observed that a high-dimensional representation space with a similarity-based metric often induces a hub structure (Dinu and Baroni, 2015). Specifically, in a high-dimensional space (e.g., a cross-lingual word embedding space) defined with a pairwise similarity metric (e.g., cosine similarity), there exist a few vectors that are the nearest neighbors of many other vectors. Such vectors are referred to as “hubs.” The hub structure is problematic in IR since the hub vectors are often wrongly predicted as relevant and similar in meaning to queries that are in fact irrelevant (Radovanovi´c et al., 2010). Let VQ and VS be the embedding spaces for the query and sentence collection languages respectively. We define the size of the neighborhood of embeddings around y ∈VS as Nk(y) = |{x ∈VQ|rx(y) ≤k}| where rx(y) is the rank of y if we order VS by similarity to x from highest to lowest, and k is a Model Somali Swahili Tagalog SECLR 29.36 54.98 43.29 SECLR-RT 6.78 14.73 11.73 Table 6: SN10 scores of SECLR and SECLR-RT respectively on Somali, Swahili and Tagalog. positive integer. A large value of Nk(y) indicates that y is similar to many x ∈VQ, and suggests that y is a likely hub in embedding space. Following Radovanovi´c et al. (2010), we use SN10 = Ey∈VS[(N10(y) −µ)3/σ3] to measure the skewness of the distribution of N10, where µ and σ refer to the mean and standard deviation of N10(y) respectively. Since cosine similarity is more frequently used as the similarity metric in hubness analysis, we re-train SECLR and SECLR-RT by replacing the dot product similarity metric with cosine similarity and still get performance comparable to Table 4 and Table 5. We report SN10 scores for SECLR and SECLRRT respectively in Table 6. We see that SECLRRT consistently has lower SN10 value compared to SECLR on all three languages, indicating that the extra alignment information incorporated with rationale training is helpful in reducing hubness. 3889 8 Conclusion In this work, we presented a supervised crosslingual embedding-based query relevance model, SECLR, for cross-language sentence selection and also applied a rationale training objective to further increase model performance. The resulting SECLR-RT model outperforms a range of baseline methods on a cross-language sentence selection task. Study of data ablation and hubness further indicate our model’s efficacy in handling lowresource settings and reducing hub structures. In future work, we hope to apply our sentence-level query relevance approach to downstream NLP tasks such as query-focused summarization and opendomain question answering. Acknowledgements This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via contract #FA865017-C-9117. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes not withstanding any copyright annotation therein. References Zeynep Akkalyoncu Yilmaz, Wei Yang, Haotian Zhang, and Jimmy Lin. 2019. Cross-Domain Modeling of Sentence-Level Evidence for Document Retrieval. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3490–3496, Hong Kong, China. Association for Computational Linguistics. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. A Robust Self-learning Method for Fully Unsupervised Cross-lingual Mappings of Word Embeddings. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 789–798, Melbourne, Australia. Association for Computational Linguistics. Yujia Bao, Shiyu Chang, Mo Yu, and Regina Barzilay. 2018. Deriving Machine Attention from Human Rationales. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1903–1913, Brussels, Belgium. Association for Computational Linguistics. Tal Baumel, Raphael Cohen, and Michael Elhadad. 2016. Topic Concentration in Query Focused Summarization Datasets. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, February 12-17, 2016, Phoenix, Arizona, USA, pages 2573–2579. AAAI Press. Tal Baumel, Matan Eyal, and Michael Elhadad. 2018. Query Focused Abstractive Summarization: Incorporating Query Relevance, Multi-Document Coverage, and Summary Length Constraints into seq2seq Models. CoRR, abs/1801.07704. Elizabeth Boschee, Joel Barry, Jayadev Billa, Marjorie Freedman, Thamme Gowda, Constantine Lignos, Chester Palen-Michel, Michael Pust, Banriskhem Kayang Khonglah, Srikanth Madikeri, Jonathan May, and Scott Miller. 2019. SARAL: A Low-Resource Cross-Lingual Domain-Focused Information Retrieval System for Effective Rapid Document Triage. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 19–24, Florence, Italy. Association for Computational Linguistics. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to Answer OpenDomain Questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870– 1879, Vancouver, Canada. Association for Computational Linguistics. Wenhu Chen, Evgeny Matusov, Shahram Khadivi, and Jan-Thorsten Peter. 2016. Guided Alignment Training for Topic-Aware Neural Machine Translation. CoRR, abs/1607.01628. Caitlin Christianson, Jason Duncan, and Boyan A. Onyshkevych. 2018. Overview of the DARPA LORELEI Program. Machine Translation, 32(12):3–9. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440– 8451, Online. Association for Computational Linguistics. Kareem Darwish and Douglas W. Oard. 2003. Probabilistic Structured Query Methods. In Proceedings of the 26th Annual International ACM SIGIR Conference on Research and Development in Informaion Retrieval, SIGIR ’03, page 338–344, New York, NY, USA. Association for Computing Machinery. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference 3890 of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Bhuwan Dhingra, Kathryn Mazaitis, and William W. Cohen. 2017. Quasar: Datasets for Question Answering by Search and Reading. CoRR, abs/1707.03904. Georgiana Dinu and Marco Baroni. 2015. Improving Zero-shot Learning by Mitigating the Hubness Problem. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Workshop Track Proceedings. Yixing Fan, Jiafeng Guo, Yanyan Lan, Jun Xu, Chengxiang Zhai, and Xueqi Cheng. 2018. Modeling Diverse Relevance Patterns in Ad-hoc Retrieval. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, SIGIR 2018, Ann Arbor, MI, USA, July 08-12, 2018, pages 375–384. ACM. Guy Feigenblat, Haggai Roitman, Odellia Boni, and David Konopnicki. 2017. Unsupervised QueryFocused Multi-Document Summarization Using the Cross Entropy Method. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR’17, page 961–964, New York, NY, USA. Association for Computing Machinery. Edouard Grave, Piotr Bojanowski, Prakhar Gupta, Armand Joulin, and Tom´as Mikolov. 2018. Learning Word Vectors for 157 Languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation, LREC 2018, Miyazaki, Japan, May 7-12, 2018. European Language Resources Association (ELRA). Surabhi Gupta, Ani Nenkova, and Dan Jurafsky. 2007. Measuring Importance and Query Relevance in Topic-focused Multi-document Summarization. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 193–196, Prague, Czech Republic. Association for Computational Linguistics. Aria Haghighi, John Blitzer, John DeNero, and Dan Klein. 2009. Better Word Alignments with Supervised ITG Models. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 923–931, Suntec, Singapore. Association for Computational Linguistics. Kenneth Heafield. 2011. KenLM: Faster and Smaller Language Model Queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland. Association for Computational Linguistics. Oana Inel, Giannis Haralabopoulos, Dan Li, Christophe Van Gysel, Zolt´an Szl´avik, Elena Simperl, Evangelos Kanoulas, and Lora Aroyo. 2018. Studying Topical Relevance with EvidenceBased Crowdsourcing. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management, CIKM ’18, page 1253–1262, New York, NY, USA. Association for Computing Machinery. Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543–3556, Minneapolis, Minnesota. Association for Computational Linguistics. Armand Joulin, Piotr Bojanowski, Tomas Mikolov, Herv´e J´egou, and Edouard Grave. 2018. Loss in Translation: Learning Bilingual Word Mapping with a Retrieval Criterion. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2979–2984, Brussels, Belgium. Association for Computational Linguistics. Marcin Junczys-Dowmunt. 2018. Dual Conditional Cross-Entropy Filtering of Noisy Parallel Corpora. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 888–895, Belgium, Brussels. Association for Computational Linguistics. S. Kale, A. Kulkarni, R. Patil, Y. Haribhakta, K. Bhattacharjee, S. Mehta, S. Mithran, and A. Kumar. 2018. Open-Domain Question Answering using Feature Encoded Dynamic Coattention Networks. In 2018 International Conference on Advances in Computing, Communications and Informatics (ICACCI), pages 1058–1062. David Kamholz, Jonathan Pool, and Susan M. Colowick. 2014. Panlex: Building a Resource for Panlingual Lexical Translation. In Proceedings of the Ninth International Conference on Language Resources and Evaluation, LREC 2014, Reykjavik, Iceland, May 26-31, 2014, pages 3145–3150. European Language Resources Association (ELRA). Marcin Kaszkiel and Justin Zobel. 1997. Passage Retrieval Revisited. In Proceedings of the 20th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’97, page 178–185, New York, NY, USA. Association for Computing Machinery. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, 3891 Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague, Czech Republic. Association for Computational Linguistics. Guillaume Lample, Alexis Conneau, Marc’Aurelio Ranzato, Ludovic Denoyer, and Herv´e J´egou. 2018. Word Translation without Parallel Data. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. Omer Levy, Anders Søgaard, and Yoav Goldberg. 2017. A Strong Baseline for Learning Cross-Lingual Word Embeddings from Sentence Alignments. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers. Robert Litschko, Goran Glavaˇs, Ivan Vulic, and Laura Dietz. 2019. Evaluating Resource-Lean CrossLingual Embedding Models in Unsupervised Retrieval. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR’19, page 1109–1112, New York, NY, USA. Association for Computing Machinery. Robert Litschko, Goran Glavaˇs, Simone Paolo Ponzetto, and Ivan Vuli´c. 2018. Unsupervised Cross-Lingual Information Retrieval Using Monolingual Data Only. The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval. Xiaoyong Liu and W. Bruce Croft. 2002. Passage Retrieval Based on Language Models. In Proceedings of the Eleventh International Conference on Information and Knowledge Management, CIKM ’02, page 375–382, New York, NY, USA. Association for Computing Machinery. Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Bilingual Word Representations with Monolingual Quality in Mind. In Proceedings of the 1st Workshop on Vector Space Modeling for Natural Language Processing, pages 151–159, Denver, Colorado. Association for Computational Linguistics. Tom´as Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Estimation of Word Representations in Vector Space. In 1st International Conference on Learning Representations, ICLR 2013, Scottsdale, Arizona, USA, May 2-4, 2013, Workshop Track Proceedings. Jian-yun Nie. 2010. Cross-Language Information Retrieval. Synthesis Lectures on Human Language Technologies, 3:1–125. Xing Niu, Michael Denkowski, and Marine Carpuat. 2018. Bi-Directional Neural Machine Translation with Synthetic Parallel Data. Proceedings of the 2nd Workshop on Neural Machine Translation and Generation. Franz Josef Och and Hermann Ney. 2003. A Systematic Comparison of Various Statistical Alignment Models. Computational Linguistics, 29(1):19–51. Milos Radovanovi´c, Alexandros Nanopoulos, and Mirjana Ivanovi´c. 2010. On the Existence of Obstinate Results in Vector Space Models. In Proceedings of the 33rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’10, page 186–193, New York, NY, USA. Association for Computing Machinery. Anton Ragni and Mark Gales. 2018. Automatic Speech Recognition System Development in the “Wild”. In Proc. Interspeech 2018, pages 2217–2221. Prajit Ramachandran, Peter Liu, and Quoc Le. 2017. Unsupervised Pretraining for Sequence to Sequence Learning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 383–391, Copenhagen, Denmark. Association for Computational Linguistics. Carl Rubino. 2020. The Effect of Linguistic Parameters in CLIR Performance. In Proceedings of the workshop on Cross-Language Search and Summarization of Text and Speech (CLSSTS2020), pages 1– 6, Marseille, France. European Language Resources Association. Trevor Strohman, Donald Metzler, Howard Turtle, and W Bruce Croft. 2005. Indri: A Language Modelbased Search Engine for Complex Queries. In Proceedings of the international conference on intelligent analysis, volume 2, pages 2–6. Citeseer. J¨org Tiedemann. 2012. Parallel Data, Tools and Interfaces in OPUS. In Proceedings of the Eighth International Conference on Language Resources and Evaluation, LREC 2012, Istanbul, Turkey, May 2325, 2012, pages 2214–2218. European Language Resources Association (ELRA). Ivan Vuli´c and Marie-Francine Moens. 2015. Monolingual and Cross-Lingual Information Retrieval Models Based on (Bilingual) Word Embeddings. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’15, page 363–372, New York, NY, USA. Association for Computing Machinery. Courtney Wade and James Allan. 2005. Passage Retrieval and Evaluation. Technical report. 3892 Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace’s Transformers: State-of-the-art Natural Language Processing. CoRR, abs/1910.03771. Jinxi Xu and Ralph Weischedel. 2000. Cross-Lingual Information Retrieval Using Hidden Markov Models. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora: Held in Conjunction with the 38th Annual Meeting of the Association for Computational Linguistics - Volume 13, EMNLP ’00, page 95–103, USA. Association for Computational Linguistics. Rabih Zbib, Lingjun Zhao, Damianos Karakos, William Hartmann, Jay DeYoung, Zhongqiang Huang, Zhuolin Jiang, Noah Rivkin, Le Zhang, Richard Schwartz, and John Makhoul. 2019. NeuralNetwork Lexical Translation for Cross-Lingual IR from Text and Speech. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR’19, page 645–654, New York, NY, USA. Association for Computing Machinery. Rui Zhang, Caitlin Westerfield, Sungrok Shim, Garrett Bingham, Alexander Fabbri, William Hu, Neha Verma, and Dragomir Radev. 2019. Improving Low-Resource Cross-lingual Document Retrieval by Reranking with Deep Bilingual Representations. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 3893 A Extra Training Dataset Details When we train SECLR and SECLR-RT via data augmentation, we randomly split the parallel corpus into train set (96%), validation set (3%) and test set (1%). We then use the dataset augmentation technique introduced in Section 3.1 to generate positive and negative samples for each set. Augmenting the dataset upon the split corpus allows us to achieve more independence between train/validation/test set compared to splitting the dataset augmented on the entire parallel corpus. Note that we only use the validation set for early stopping but we do not tune hyperparameters with the validation set. We preprocess the parallel corpus, the query collection and the sentence collection with the Moses toolkit (Koehn et al., 2007). The same preprocessing steps are used for all four languages (English, Somali, Swahili, Tagalog). First, we use Moses puncutation normalizer to normalize the raw text. Second, we use the Moses tokenizer to tokenize the normalized text. Finally, we remove the diacritics in the tokenized text as a cleaning step. B Examples of Evaluation Data In this section we demonstrate some examples from the MATERIAL dataset used for evaluation. Example queries include: “evidence”, “human rights”, “chlorine”, “academy”, “ratify”, “constitution”, “carnage” and “Kenya”. On average only 0.13% of the documents in the Eval collection are relevant to each query, which makes the task hard. Here are two examples from Somali Analysis text. Because the documents are long, here we only include the relevant segment of a long relevant document. In the first example, the English query is “contravention” and the relevant segment of a long relevant document (translated from Somali to English by human) is “the security forces captured military equipment coming into the country illegally.” This segment is relevant to the query because of the word “illegally”. Here is another example where the the English query is “integrity”. The relevant segment of a long relevant document (translated from Somali to English by human) is “Hargeisa (Dawan) - Ahmed Mohamed Diriye (Nana) the member of parliament who is part of the Somaliland house of representatives has accused the opposition parties (Waddani and UCID) of engaging in acts of national destruction, that undermines the existence and sovereignty of the country of Somaliland.” This segment is relevant to the query because of the word “sovereignty”. Since there are multiple ways to translate a word and since MT performance is relatively poor in lowresource settings, the task is far more challenging than a simple lexical match between queries and translated documents. C Extra Experimental Details In this section we include extra implementation and experiment details that are not included in the main paper. Information already included in the main paper are not repeated here for conciseness. C.1 Model and Training Details We train our SECLR and SECLR-RT models on Tesla V100 GPUs. Each model is trained on a single GPU. We report training time of SECLR and SECLR-RT on Somali, Swahili and Tagalog in Table 7. Somali Swahili Tagalog SECLR 77 112 124 SECLR-RT 179 254 319 Table 7: Training time of SECLR and SECLR-RT on Somali, Swahili and Tagalog respectively (in minutes). As is discussed in Section 3.2, the only trainable model parameters of SECLR and SECLR-RT are the word embedding matrices. Thus, SECLR and SECLR-RT have the same number of model parameters. We report the number of trainable parameters of both models on Somali, Swahili and Tagalog in Table 8. Somali Swahili Tagalog # Params. 14.03M 22.31M 21.35M Table 8: Number of trainable model parameters of SECLR/SECLR-RT on Somali, Swahili and Tagalog. “M” stands for million. We used Mean Average Precision (MAP) as the evaluation metric in this work. We use the following implementation to compute MAP: https: //trec.nist.gov/trec_eval/. C.2 MT Baseline Details For NMT we train bidirectional MT systems with a 6-layer Transformer architecture with model size of 3894 Somali Swahili Tagalog Analysis Dev Analysis Dev Analysis Dev Method T S T S T S T S T S T S With LSTM 16.3 14.5 11.9 12.0 27.5 27.0 19.5 25.1 29.7 29.7 23.0 27.1 No LSTM 27.8 24.4 23.0 17.4 43.8 37.9 40.3 38.1 46.7 45.0 49.3 33.9 Table 9: Document-level MAP scores for text (T) and speech (S) of the SECLR model with and without LSTM. Somali Swahili Tagalog Analysis Dev Analysis Dev Analysis Dev Embed. Init. T S T S T S T S T S T S Cross-lingual 35.3 27.5 31.1 23.2 48.8 41.1 42.5 41.6 56.3 51.1 53.8 45.3 Monolingual 35.4 28.4 29.5 22.0 48.3 48.1 39.6 45.4 61.1 55.5 59.0 45.7 Table 10: Document-level MAP scores for text (T) and speech (S) of the SECLR-RT model with monolingual or cross-lingual (SID-SGNS) word embedding initialization. 512, feed-forward network size of 2048, 8 attention heads, and residual connections. We adopt layer normalization and label smoothing. We tie the output weight matrix with the source and target embeddings. We use Adam optimizer with a batch size of 2048 words. We checkpoint models every 1000 updates. Training stops after 20 checkpoints without improvement. During inference, the beam size is set to 5. Our SMT system uses the following feature functions: phrase translation model, distance-based reordering model, lexicalized reordering model, 5-gram language model on the target side, word penalty, distortion, unknown word penalty and phrase penalty. We use backtranslation in earlier versions of MT systems. Following previous work (Niu et al., 2018), we train a bidirectional NMT model that backtranslates source or target monolingual data without an auxiliary model. This backtranslationbased model was the state-of-the-art MT model on Somali and Swahili when the above paper is published. Later, we discover that decoder pretraining with monolingual data achieves better performance compared to backtranslation. The decoder pretraining scheme we use now is most similar to the paper by Ramachandran et al. (2017), where the authors show state-of-the-art results on the WMT English to German translation task with decoder pretraining. There is no WMT benchmark for Somali, Swahili or Tagalog, but we use state-of-the-art techniques in our MT systems. We have also experimented with the bilingual data selection method (Junczys-Dowmunt, 2018). However, this technique does not work well, mostly because lowresource MT systems are not good enough to do scoring. D Extra Experimental Results In this section we include extra experimental results that are not included in the main text due to limited space. D.1 SECLR Architecture Exploration When we are designing the SECLR model, we experiment with adding LSTMs and using the dot product between LSTM hidden states to compute pairwise similarity between the query and the sentence. We report MAP scores of SECLR with LSTM in Table 9. Experimental results show that adding LSTMs reduces model performance consistently across all three languages. We conjecture that in low-resource settings, contextualized models create spurious correlations (Section 3.3). In fact, the XLM-RoBERTa baseline, which captures context effectively via self-attention, also underperforms our SECLR model consistently. 3895 D.2 Word Embeddings Initialization In our SECLR and SECLR-RT models, we initialize word embeddings with monolingual word embeddings in English, Somali, Swahili and Tagalog (Mikolov et al., 2013; Grave et al., 2018). One natural question is whether we can achieve performance improvement if we directly initialize with crosslingual word embeddings. Because SID-SGNS out-performs both Bivec and MUSE consistently by a wide margin (Table 4 and Table 5), in this experiment we initialize SECLR-RT with the crosslingual embeddings produced by SID-SGNS. The results of monolingual and cross-lingual embedding initialization (SID-SGNS) are shown in Table 10. We see that overall monolingual initialization slightly out-performs cross-lingual initialization. Monolingual initialization yields better performance in eight out of 12 Analysis/Dev set conditions and a MAP improvement of 1.7 points when we take the average across Analysis/Dev and all three languages.
|
2021
|
300
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3896–3907 August 1–6, 2021. ©2021 Association for Computational Linguistics 3896 A Neural Model for Joint Document and Snippet Ranking in Question Answering for Large Document Collections Dimitris Pappas1,2 and Ion Androutsopoulos1 1Department of Informatics, Athens University of Economics and Business, Greece [email protected],[email protected] 2Institute for Language and Speech Processing, Research Center ‘Athena’, Greece [email protected] Abstract Question answering (QA) systems for large document collections typically use pipelines that (i) retrieve possibly relevant documents, (ii) re-rank them, (iii) rank paragraphs or other snippets of the top-ranked documents, and (iv) select spans of the top-ranked snippets as exact answers. Pipelines are conceptually simple, but errors propagate from one component to the next, without later components being able to revise earlier decisions. We present an architecture for joint document and snippet ranking, the two middle stages, which leverages the intuition that relevant documents have good snippets and good snippets come from relevant documents. The architecture is general and can be used with any neural text relevance ranker. We experiment with two main instantiations of the architecture, based on POSITDRMM (PDRMM) and a BERT-based ranker. Experiments on biomedical data from BIOASQ show that our joint models vastly outperform the pipelines in snippet retrieval, the main goal for QA, with fewer trainable parameters, also remaining competitive in document retrieval. Furthermore, our joint PDRMM-based model is competitive with BERT-based models, despite using orders of magnitude fewer parameters. These claims are also supported by human evaluation on two test batches of BIOASQ. To test our key findings on another dataset, we modified the Natural Questions dataset so that it can also be used for document and snippet retrieval. Our joint PDRMM-based model again outperforms the corresponding pipeline in snippet retrieval on the modified Natural Questions dataset, even though it performs worse than the pipeline in document retrieval. We make our code and the modified Natural Questions dataset publicly available. 1 Introduction Question answering (QA) systems that search large document collections (Voorhees, 2001; Tsatsaronis et al., 2015; Chen et al., 2017) typically use pipelines operating at gradually finer text granularities. A fully-fledged pipeline includes components that (i) retrieve possibly relevant documents typically using conventional information retrieval (IR); (ii) re-rank the retrieved documents employing a computationally more expensive document ranker; (iii) rank the passages, sentences, or other ‘snippets’ of the top-ranked documents; and (iv) select spans of the top-ranked snippets as ‘exact’ answers. Recently, stages (ii)–(iv) are often pipelined neural models, trained individually (Hui et al., 2017; Pang et al., 2017; Lee et al., 2018; McDonald et al., 2018; Pandey et al., 2019; Mackenzie et al., 2020; Sekuli´c et al., 2020). Although pipelines are conceptually simple, errors propagate from one component to the next (Hosein et al., 2019), without later components being able to revise earlier decisions. For example, once a document has been assigned a low relevance score, finding a particularly relevant snippet cannot change the document’s score. We propose an architecture for joint document and snippet ranking, i.e., stages (ii) and (iii), which leverages the intuition that relevant documents have good snippets and good snippets come from relevant documents. We note that modern web search engines display the most relevant snippets of the top-ranked documents to help users quickly identify truly relevant documents and answers (Sultan et al., 2016; Xu et al., 2019; Yang et al., 2019a). The top-ranked snippets can also be used as a starting point for multi-document query-focused summarization, as in the BIOASQ challenge (Tsatsaronis et al., 2015). Hence, methods that identify good snippets are useful in several other applications, apart from QA. We also note that many neural models for stage (iv) have been proposed, often called QA or Machine Reading Comprehension (MRC) models (Kadlec et al., 2016; Cui et al., 2017; Zhang et al., 2020), but they typically search for answers 3897 only in a particular, usually paragraph-sized snippet, which is given per question. For QA systems that search large document collections, stages (ii) and (iii) are also important, if not more important, but have been studied much less in recent years, and not in a single joint neural model. The proposed joint architecture is general and can be used in conjunction with any neural text relevance ranker (Mitra and Craswell, 2018). Given a query and N possibly relevant documents from stage (i), the neural text relevance ranker scores all the snippets of the N documents. Additional neural layers re-compute the score (ranking) of each document from the scores of its snippets. Other layers then revise the scores of the snippets taking into account the new scores of the documents. The entire model is trained to jointly predict document and snippet relevance scores. We experiment with two main instantiations of the proposed architecture, using POSIT-DRMM (McDonald et al., 2018), hereafter called PDRMM, as the neural text ranker, or a BERT-based ranker (Devlin et al., 2019). We show how both PDRMM and BERT can be used to score documents and snippets in pipelines, then how our architecture can turn them into models that jointly score documents and snippets. Experimental results on biomedical data from BIOASQ (Tsatsaronis et al., 2015) show the joint models vastly outperform the corresponding pipelines in snippet extraction, with fewer trainable parameters. Although our joint architecture is engineered to favor retrieving good snippets (as a near-final stage of QA), results show that the joint models are also competitive in document retrieval. We also show that our joint version of PDRMM, which has the fewest parameters of all models and does not use BERT, is competitive to BERT-based models, while also outperforming the best system of BIOASQ 6 (Brokos et al., 2018) in both document and snippet retrieval. These claims are also supported by human evaluation on two test batches of BIOASQ 7 (2019). To test our key findings on another dataset, we modified Natural Questions (Kwiatkowski et al., 2019), which only includes questions and answer spans from a single document, so that it can be used for document and snippet retrieval. Again, our joint PDRMMbased model largely outperforms the corresponding pipeline in snippet retrieval on the modified Natural Questions, though it does not perform better than the pipeline in document retrieval, since the joint model is geared towards snippet retrieval, i.e., even though it is forced to extract snippets from fewer relevant documents. Finally, we show that all the neural pipelines and joint models we considered improve the BM25 ranking of traditional IR on both datasets. We make our code and the modified Natural Questions publicly available.1 2 Methods 2.1 Document Ranking with PDRMM Our starting point is POSIT-DRMM (McDonald et al., 2018), or PDRMM, a differentiable extension of DRMM (Guo et al., 2016) that obtained the best document retrieval results in BIOASQ 6 (Brokos et al., 2018). McDonald et al. (2018) also reported it performed better than DRMM and several other neural rankers, including PACRR (Hui et al., 2017). Given a query q = ⟨q1, . . . , qn⟩of n query terms (q-terms) and a document d = ⟨d1, . . . , dm⟩ of m terms (d-terms), PDRMM computes contextsensitive term embeddings c(qi) and c(di) from the static (e.g., WORD2VEC) embeddings e(qi) and e(di) by applying two stacked convolutional layers with trigram filters, residuals (He et al., 2016), and zero padding to q and d, respectively.2 PDRMM then computes three similarity matrices S1, S2, S3, each of dimensions n × m (Fig. 1). Each element si,j of S1 is the cosine similarity between c(qi) and c(dj). S2 is similar, but uses the static word embeddings e(qi), e(dj). S3 uses one-hot vectors for qi, dj, signaling exact matches. Three row-wise pooling operators are then applied to S1, S2, S3: max-pooling (to obtain the similarity of the best match between the q-term of the row and any of the d-terms), average pooling (to obtain the average match), and average of k-max (to obtain the average similarity of the k best matches).3 We thus obtain three scores from each row of each similarity matrix. By concatenating row-wise the scores from the three matrices, we obtain a new n × 9 matrix S′ (Fig. 1). Each row of S′ indicates how well the corresponding q-term matched any of the d-terms, using the three different views of the terms (onehot, static, context-aware embeddings). Each row of S′ is then passed to a Multi-Layer Perceptron 1See http://nlp.cs.aueb.gr/publications. html for links to the code and data. 2McDonald et al. (2018) use a BILSTM encoder instead of convolutions. We prefer the latter, because they are faster, and we found that they do not degrade performance. 3We added average pooling to PDRMM to balance the other pooling operators that favor long documents. 3898 Figure 1: PDRMM for document scoring. The same model (with different trained parameters) also scores sentences in the PDRMM+PDRMM pipeline and the joint JPDRMM model (adding the layers of Fig. 2). (MLP) to obtain a single match score per q-term. Each context aware q-term embedding is also concatenated with the corresponding IDF score (bottom left of Fig. 1) and passed to another MLP that computes the importance of that q-term (words with low IDFs may be unimportant). Let v be the vector containing the n match scores of the q-terms, and u the vector with the corresponding n importance scores (bottom right of Fig. 1). The initial relevance score of the document is ˆr(q, d) = vT u. Then ˆr(q, d) is concatenated with four extra features: z-score normalized BM25 (Robertson and Zaragoza, 2009); percentage of q-terms with exact match in d (regular and IDF weighted); percentage of q-term bigrams matched in d. An MLP computes the final relevance r(q, d) from the 5 features. Neural rankers typically re-rank the top N documents of a conventional IR system. We use the same BM25-based IR system as McDonald et al. (2018). PDRMM is trained on triples ⟨q, d, d′⟩, where d is a relevant document from the top N of q, and d′ is a random irrelevant document from the top N. We use hinge loss, requiring the relevance of d to exceed that of d′ by a margin. 2.2 PDRMM-based Pipelines for Document and Snippet Ranking Brokos et al. (2018) used the ‘basic CNN’ (BCNN) of Yin et al. (2016) to score (rank) the sentences of the re-ranked top N documents. The resulting pipeline, PDRMM+BCNN, had the best document and snippet results in BIOASQ 6, where snippets were sentences. Hence, PDRMM+BCNN is a reasonable document and snippet retrieval baseline pipeline. In another pipeline, PDRMM+PDRMM, we replace BCNN by a second instance of PDRMM that scores sentences. The second PDRMM instance Figure 2: Final layers of JPDRMM and JBERT. The input sentence scores are generated by PDRMM (Fig. 1) or BERT (Fig. 3) now applied to document sentences. The document’s score is obtained from the score of its best sentence and external features, and is also used to revise the sentence scores. Training jointly minimizes document and sentence loss. is the same as when scoring documents (Fig. 1), but the input is now the query (q) and a single sentence (s). Given a triple ⟨q, d, d′⟩used to train the document-scoring PDRMM, the sentence-scoring PDRMM is trained to predict the true class (relevant, irrelevant) of each sentence in d and d′ using cross entropy loss (with a sigmoid on r(q, s)). As when scoring documents, the initial relevance score ˆr(q, s) is combined with extra features using an MLP, to obtain r(q, s). The extra features are now different: character length of q and s, number of shared tokens of q and s (with/without stop-words), sum of IDF scores of shared tokens (with/without stop-words), sum of IDF scores of shared tokens divided by sum of IDF scores of q-terms, number of shared token bigrams of q and s, BM25 score of s against the sentences of d and d′, BM25 score of the document (d or d′) that contained s. The two PDRMM instances are trained separately. 2.3 Joint PDRMM-based Models for Document and Snippet Ranking Given a document d with sentences s1, . . . , sk and a query q, the joint document/snippet ranking version of PDRMM, called JPDRMM, processes separately each sentence si of d, producing a relevance score r(q, si) per sentence, as when PDRMM scores sentences in the PDRMM+PDRMM pipeline. The highest sentence score maxi r(q, si) is concatenated (Fig. 2) with the extra features that are used when PDRMM ranks documents, and an MLP produces the document’s score.4 JPDRMM then revises the sentence scores, by concatenating the score of each sentence with the document score 4We also tried alternative mechanisms to obtain the document score from the sentence scores, including average of k-max sentence scores and hierarchical RNNs (Yang et al., 2016), but they led to no improvement. 3899 and passing each pair of scores to a dense layer to compute a linear combination, which becomes the revised sentence score. Notice that JPDRMM is mostly based on scoring sentences, since the main goal for QA is to obtain good snippets (almost final answers). The document score is obtained from the score of the document’s best sentence (and external features), but the sentence scores are revised, once the document score has been obtained. We use sentence-sized snippets, for compatibility with BIOASQ, but other snippet granularities (e.g., paragraph-sized) could also be used. JPDRMM is trained on triples ⟨q, d, d′⟩, where d, d′ are relevant and irrelevant documents, respectively, from the top N of query q, as in the original PDRMM; the ground truth now also indicates which sentences of the documents are relevant or irrelevant, as when training PDRMM to score sentences in PDRMM+PDRMM. We sum the hinge loss of d and d′ and the cross-entropy loss of each sentence.5 We also experiment with a JPDRMM version that uses a pre-trained BERT model (Devlin et al., 2019) to obtain input token embeddings (of wordpieces) instead of the more conventional pre-trained (e.g., WORD2VEC) word embeddings that JPDRMM uses otherwise. We call it BJPDRMM if BERT is finetuned when training JPDRMM, and BJPDRMM-NF if BERT is not fine-tuned. In another variant of BJPDRMM, called BJPDRMM-ADAPT, the input embedding of each token is a linear combination of all the embeddings that BERT produces for that token at its different Transformer layers. The weights of the linear combination are learned via backpropagation. This allows BJPDRMM-ADAPT to learn which BERT layers it should mostly rely on when obtaining token embeddings. Previous work has reported that representations from different BERT layers may be more appropriate for different tasks (Rogers et al., 2020). BJPDRMM-ADAPT-NF is the same as BJPDRMM-ADAPT, but BERT is not finetuned; the weights of the linear combination of embeddings from BERT layers are still learned. 2.4 Pipelines and Joint Models Based on Ranking with BERT The BJPDRMM model we discussed above and its variants are essentially still JPDRMM, which in turn invokes the PDRMM ranker (Fig. 1, 2); BERT is used only to obtain token embeddings that are fed 5Additional experiments with JPDRMM, reported in the appendix, indicate that further performance gains are possible by tuning the weights of the two losses. Figure 3: Document scoring with BERT. The same model scores sentences in JBERT (adding the layers of Fig. 2), but with an MLP replacing the final dense layer. to JPDRMM. Instead, in this subsection we use BERT as a ranker, replacing PDRMM. For document ranking alone (when not cosidering snippets), we feed BERT with pairs of questions and documents (Fig. 3). BERT’s top-layer embedding of the ‘classification’ token [CLS] is concatenated with external features (the same as when scoring documents with PDRMM, Section 2.1), and a dense layer again produces the document’s score. We fine-tune the entire model using triples ⟨q, d, d′⟩ with a hinge loss between d and d′, as when training PDRMM to score documents.6 Our two pipelines that use BERT for document ranking, BERT+BCNN and BERT+PDRMM, are the same as PDRMM+BCNN and PDRMM+PDRMM (Section 2.2), respectively, but use the BERT ranker (Fig. 3) to score documents, instead of PDRMM. The joint JBERT model is the same as JPDRMM, but uses the BERT ranker (Fig. 3), now applied to sentences, instead of PDRMM (Fig. 1), to obtain the initial sentence scores. The top layers of Fig. 2 are then used, as in all joint models, to obtain the document score from the sentence scores and revise the sentence scores. Similarly to BJPDRMM, we also experimented with variations of JBERT, which do not fine-tune the parameters of BERT (JBERT-NF), use a linear combination (with trainable weights) of the [CLS] embeddings from all the BERT layers (JBERT-ADAPT), or both (JBERT-ADAPT-NF). 2.5 BM25+BM25 Baseline Pipeline We include a BM25+BM25 pipeline to measure the improvement of the proposed models on conventional IR engines. This pipeline uses the question 6We use the pre-trained uncased BERT BASE of Devlin et al. (2019). The ‘documents’ of the BIOASQ dataset are concatenated titles and abstracts. Most question-document pairs do not exceed BERT’s max. length limit of 512 wordpieces. If they do, we truncate documents. The same approach could be followed in the modified Natural Questions dataset, where ‘documents’ are Wikipedia paragraphs, but we did not experiment with BERT-based models on that dataset. 3900 as a query to the IR engine and selects the Nd documents with the highest BM25 scores.7 The Nd documents are then split into sentences and BM25 is re-computed, this time over all the sentences of the Nd documents, to retrieve the Ns best sentences. 3 Experiments 3.1 Data and Experimental Setup BioASQ data and setup Following McDonald et al. (2018) and Brokos et al. (2018), we experiment with data from BIOASQ (Tsatsaronis et al., 2015), which provides English biomedical questions, relevant documents from MEDLINE/PUBMED8, and relevant snippets (sentences), prepared by biomedical experts. This is the only previous large-scale IR dataset we know of that includes both gold documents and gold snippets. We use the BIOASQ 7 (2019) training dataset, which contains 2,747 questions, with 11 gold documents and 14 gold snippets per question on average. We evaluate on test batches 1–5 (500 questions in total) of BIOASQ 7.9 We measure Mean Average Precision (MAP) (Manning et al., 2008) for document and snippet retrieval, which are the official BIOASQ evaluation measures. The document collection contains approx. 18M articles (concatenated titles and abstracts only, discarding articles with no abstracts) from the MEDLINE/PUBMED ‘baseline’ 2018 dataset. In PDRMM and BCNN, we use the biomedical WORD2VEC embeddings of McDonald et al. (2018). We use the GALAGO10 IR engine to obtain the top N = 100 documents per query. After re-ranking, we return Nd = 10 documents and Ns = 10 sentences, as required by BIOASQ. We train using Adam (Kingma and Ba, 2015). Hyperparameters were tuned on held-out validation data. Natural Questions data and setup Even though there was no other large-scale IR dataset providing multiple gold documents and snippets per question, we needed to test our best models on a second dataset, other than BIOASQ. Therefore we modified the Natural Questions dataset (Kwiatkowski et al., 2019) to a format closer to BIOASQ’s. Each instance of Natural Questions consists of an HTML 7In each experiment, the same IR engine and BM25 hyperparameters are used in all other methods. All BM25 hyperparameters are tuned on development data. 8https://www.ncbi.nlm.nih.gov/pubmed 9BIOASQ 8 (2020) was ongoing during this work, hence we could not use its data for comparisons. See also the discussion of BIOASQ results after expert inspection in Section 3.2. 10www.lemurproject.org/galago.php document of Wikipedia and a question. The answer to the question can always be found in the document as if a perfect retrieval engine were used. A short span of HTML source code is annotated by humans as a ‘short answer’ to the question. A longer span of HTML source code that includes the short answer is also annotated, as a ‘long answer’. The long answer is most commonly a paragraph of the Wikipedia page. In the original dataset, more than 300,000 questions are provided along with their corresponding Wikipedia HTML documents, short answer and long answer spans. We modified Natural Questions to fit the BIOASQ setting. From every Wikipedia HTML document in the original dataset, we extracted the paragraphs and indexed each paragraph separately to an ElasticSearch11 index, which was then used as our retrieval engine. We discarded all the tables and figures of the HTML documents and any question that was answered by a paragraph containing a table. For every question, we apply a query to our retrieval engine and retrieve the first N = 100 paragraphs. We treat each paragraph as a document, similarly to the BIOASQ setting. For each question, the gold (correct) documents are the paragraphs (at most two per question) that were included in the long answers of the original dataset. The gold snippets are the sentences (at most two per question) that overlap with the short answers of the original dataset. We discard questions for which the retrieval engine did not manage to retrieve any of the gold paragraphs in its top 100 paragraphs. We ended up with 110,589 questions and 2,684,631 indexed paragraphs. Due to lack of computational resources, we only use 4,000 questions for training, 400 questions for development, and 400 questions for testing, but we make the entire modified Natural Questions dataset publicly available. Hyper-parameters were again tuned on held-out validation data. All other settings were as in the BIOASQ experiments. 3.2 Experimental Results BioASQ results Table 1 reports document and snippet MAP scores on the BIOASQ dataset, along with the trainable parameters per method. For completeness, we also show recall at 10 scores, but we base the discussion below on MAP, the official measure of BIOASQ, which also considers the ranking of the 10 documents and snippets BIOASQ allows participants to return. The Oracle re-ranks the N 11www.elastic.co/products/elasticsearch 3901 Method Params Doc. MAP (%) Snip. MAP (%) Doc. Recall@10(%) Snip. Recall@10(%) BM25 +BM25 4 6.86 4.29 48.65 4.93 PDRMM+BCNN 21.83k 7.47 5.67 52.97 12.43 PDRMM+PDRMM 11.39k 7.47 9.16 52.97 18.43 JPDRMM 5.79k 6.69 15.72 53.68 18.83 BERT+BCNN 109.5M 8.79 6.07 55.73 13.05 BERT+PDRMM 109.5M 8.79 9.63 55.73 19.30 BJPDRMM 88.5M 7.59 16.82 52.21 19.57 BJPDRMM-ADAPT 88.5M 6.93 15.70 48.77 19.38 BJPDRMM-NF 3.5M 6.84 15.77 48.81 17.95 BJPDRMM-ADAPT-NF 3.5M 7.42 17.35 52.12 19.66 JBERT 85M 7.93 16.29 53.44 19.87 JBERT-ADAPT 85M 7.81 15.99 52.94 19.87 JBERT-NF 6.3K 7.90 15.99 52.78 19.64 JBERT-ADAPT-NF 6.3K 7.84 16.53 53.18 19.64 Oracle 0 19.24 25.18 72.67 41.14 Sentence PDRMM 5.68K 6.39 8.73 48.60 18.57 Table 1: Parameters learned, document and snippet MAP on BIOASQ 7, test batches 1–5, before expert inspection. Systems in the 2nd (or 3rd) zone use (or not) BERT. In each zone, best scores shown in bold. In the 2nd and 3rd zones, we underline the results of the best pipeline, the results of JPDRMM, and the best results of the BJPDRMM and JBERT variants. The differences between the underlined MAP scores are statistically significant (p ≤0.01). = 100 documents (or their snippets) that BM25 retrieved, moving all the relevant documents (or snippets) to the top. Sentence PDRMM is an ablation of JPDRMM without the top layers (Fig. 2); each sentence is scored using PDRMM, then each document inherits the highest score of its snippets. PDRMM+BCNN and PDRMM+PDRMM use the same document ranker, hence the document MAP of these two pipelines is identical (7.47). However, PDRMM+PDRMM outperforms PDRMM+BCNN in snippet MAP (9.16 to 5.67), even though PDRMM has much fewer trainable parameters than BCNN, confirming that PDRMM can also score sentences and is a better sentence ranker than BCNN. PDRMM+BCNN was the best system in BIOASQ 6 for both documents and snippets, i.e., it is a strong baseline. Replacing PDRMM by BERT for document ranking in the two pipelines (BERT+BCNN and BERT+PDRMM) increases the document MAP by 1.32 points (from 7.47 to 8.79) with a marginal increase in snippet MAP for BERT+PDRMM (9.16 to 9.63) and a slightly larger increase for BERT+BCNN (5.67 to 6.07), at the expense of a massive increase in trainable parameters due to BERT (and computational cost to pre-train and fine-tune BERT). We were unable to include a BERT+BERT pipeline, which would use a second BERT ranker for sentences, with a total of approx. 220M trainable parameters, due to lack of computational resources. The main joint models (JPDRMM, BJPDRMM, JBERT) vastly outperform the pipelines in snippet extraction, the main goal for QA (obtaining 15.72, 16.82, 16.29 snippet MAP, respectively), though their document MAP is slightly lower (6.69, 7.59, 7.93) compared to the pipelines (7.47, 8.79), but still competitive. This is not surprising, since the joint models are geared towards snippet retrieval (they directly score sentences, document scores are obtained from sentence scores). Human inspection of the retrieved documents and snippets, discussed below (Table 2), reveals that the document MAP of JPDRMM is actually higher than that of the best pipeline (BERT+PDRMM), but is penalized in Table 1 because of missing gold documents. JPDRMM, which has the fewest parameters of all neural models and does not use BERT at all, is competitive in snippet retrieval with models that employ BERT. More generally, the joint models use fewer parameters than comparable pipelines (see the zones of Table 1). Not fine-tuning BERT (-NF variants) leads to a further dramatic decrease in trainable parameters, at the expense of slightly lower document and snippet MAP (7.59 to 6.84, and 16.82 to 15.77, respectively, for BJPDRMM, and similarly for JBERT). Using linear combinations of token embeddings from all BERT layers (-ADAPT variants) harms both document and snippet MAP when fine-tuning BERT, but is beneficial in most cases when not fine-tuning BERT (-NF). The snippet MAP of BJPDRMM-NF increases from 15.77 to 17.35, and document MAP increases from 6.84 to 7.42. A similar increase is observed in the snippet MAP of JBERT-NF (15.99 to 16.53), but MAP decreases (7.90 to 7.84). In the second and third result zones of Table 1, we underline the results of the best pipelines, the results of JPDRMM, and the 3902 results of the best BJPDRMM and JBERT variant. In each zone and column, the differences between the underlined MAP scores are statistically significant (p ≤0.01); we used single-tailed Approximate Randomization (Dror et al., 2018), 10k iterations, randomly swapping in each iteration the rankings of 50% of queries. Removing the top layers of JPDRMM (Sentence PDRMM), clearly harms performance for both documents and snippets. The oracle scores indicate there is still scope for improvements in both documents and snippets. BioASQ results after expert inspection At the end of each BIOASQ annual contest, the biomedical experts who prepared the questions and their gold documents and snippets inspect the responses of the participants. If any of the documents and snippets returned by the participants are judged relevant to the corresponding questions, they are added to the gold responses. This process enhances the gold responses and avoids penalizing participants for responses that are actually relevant, but had been missed by the experts in the initial gold responses. However, it is unfair to use the post-contest enhanced gold responses to compare systems that participated in the contest to systems that did not, because the latter may also return documents and snippets that are actually relevant and are not included in the gold data, but the experts do not see these responses and they are not included in the gold ones. The results of Table 1 were computed on the initial gold responses of BIOASQ 7, before the post-contest revision, because not all of the methods of that table participated in BIOASQ 7.12 In Table 2, we show results on the revised postcontest gold responses of BIOASQ 7, for those of our methods that participated in the challenge. We show results on test batches 4 and 5 only (out of 5 batches in total), because these were the only two batches were all three of our methods participated together. Each batch comprises 100 questions. We also show the best results (after inspection) of our competitors in BIOASQ 7, for the same batches. A first striking observation in Table 2 is that all results improve substantially after expert inspection, i.e., all systems retrieved many relevant documents and snippets the experts had missed. Again, the two joint models (JPDRMM, BJPDRMMNF) vastly outperform the BERT+PDRMM pipeline 12Results without expert inspection can be obtained at any time, using the BIOASQ evaluation platform. Results with expert inspection can only be obtained during the challenge. in snippet MAP. As in Table 1, before expert inspection the pipeline has slightly better document MAP than the joint models. However, after expert inspection JPDRMM exceeds the pipeline in document MAP by almost two points. BJPDRMM-NF performs two points better than JPDRMM in snippet MAP after expert inspection, though JPDRMM performs two points better in document MAP. After inspection, the document MAP of BJPDRMM-NF is also very close to the pipeline’s. Table 2 confirms that JPDRMM is competitive with models that use BERT, despite having the fewest parameters. All of our methods clearly outperformed the competition. Natural Questions results Table 3 reports results on the modified Natural Questions dataset. We experiment with the best pipeline and joint model of Table 1 that did not use BERT (and are computationally much cheaper), i.e., PDRMM+PDRMM and JPDRMM, comparing them to the more conventional BM25+BM25 baseline. Since there are at most two relevant documents and snippets per question in this dataset, we measure Mean Reciprocal Rank (MRR) (Manning et al., 2008), and Recall at top 1 and 2. Both PDRMM+PDRMM and JPDRMM clearly outperform the BM25+BM25 pipeline in both document and snippet retrieval. As in Table 1, the joint JPDRMM model outperforms the PDRMM+PDRMM pipeline in snippet retrieval, but the pipeline performs better in document retrieval. Again, this is unsurprising, since the joint models are geared towards snippet retrieval. We also note that JPDRMM uses half of the trainable parameters of PDRMM+PDRMM (Table 1). No comparison to previous work that used the original Natural Questions is possible, since the original dataset provides a single document per query (Section 3.1). 4 Related Work Neural document ranking (Guo et al., 2016; Hui et al., 2017; Pang et al., 2017; Hui et al., 2018; McDonald et al., 2018) only recently managed to improve the rankings of conventional IR; see Lin (2019) for caveats. Document or passage ranking models based on BERT have also been proposed, with promising results, but most use only simplistic task-specific layers on top of BERT (Yang et al., 2019b; Nogueira and Cho, 2019), similar to our use of BERT for document scoring (Fig. 3). An exception is the work of MacAvaney et al. (2019), who explored combining ELMO (Peters et al., 2018) and BERT (Devlin et al., 2019) with complex neu3903 Before expert inspection After expert inspection Method Document MAP Snippet MAP Document MAP Snippet MAP BERT+PDRMM 7.29 7.58 14.86 15.61 JPDRMM 5.16 12.45 16.55 21.98 BJPDRMM-NF 6.18 13.89 14.65 23.96 Best BIOASQ 7 competitor n/a n/a 13.18 14.98 Table 2: Document and snippet MAP (%) on BIOASQ 7 test batches 4 and 5 before and after post-contest expert inspection of system responses, for methods that participated in BIOASQ 7. We also show the results (after inspection) of the best other participants of BIOASQ 7 for the same batches. Document Retrieval Snippet Retrieval Method MRR Recall@1 Recall@2 MRR Recall@1 Recall@2 BM25+BM25 30.18 16.50 29.75 8.19 3.75 7.13 PDRMM+PDRMM 40.33 28.25 38.50 22.86 13.75 22.75 JPDRMM 36.50 24.50 36.00 26.92 19.00 25.25 Table 3: MRR (%) and recall at top 1 and 2 (%) on the modified Natural Questions dataset. ral IR models, namely PACRR (Hui et al., 2017), DRMM (Guo et al., 2016), KNRM (Dai et al., 2018), CONVKNRM (Xiong et al., 2017), an approach that we also explored here by combining BERT with PDRMM in BJPDRMM and JBERT. However, we retrieve both documents and snippets, whereas MacAvaney et al. (2019) retrieve only documents. Models that directly retrieve documents by indexing neural document representations, rather than re-ranking documents retrieved by conventional IR, have also been proposed (Fan et al., 2018; Ai et al., 2018; Khattab and Zaharia, 2020), but none addresses both document and snippet retrieval. Yang et al. (2019a) use BERT to encode, index, and directly retrieve snippets, but do not consider documents; indexing snippets is also computationally costly. Lee et al. (2019) propose a joint model for direct snippet retrieval (and indexing) and answer span selection, again without retrieving documents. No previous work combined document and snippet retrieval in a joint neural model. This may be due to existing datasets, which do not provide both gold documents and gold snippets, with the exception of BIOASQ, which is however small by today’s standards (2.7k training questions, Section 3.1). For example, Pang et al. (2017) used much larger clickthrough datasets from a Chinese search engine, as well as datasets from the 2007 and 2008 TREC Million Query tracks (Qin et al., 2010), but these datasets do not contain gold snippets. SQUAD (Rajpurkar et al., 2016) and SQUAD v.2 (Rajpurkar et al., 2018) provide 100k and 150k questions, respectively, but for each question they require extracting an exact answer span from a single given Wikipedia paragraph; no snippet retrieval is performed, because the relevant (paragraph-sized) snippet is given. Ahmad et al. (2019) provide modified versions of SQUAD and Natural Questions, suitable for direct snippet retrieval, but do not consider document retrieval. SearchQA (Dunn et al., 2017) provides 140k questions, along with 50 snippets per question. The web pages the snippets were extracted from, however, are not included in the dataset, only their URLs, and crawling them may produce different document collections, since the contents of web pages often change, pages are removed etc. MS-MARCO (Nguyen et al., 2016) was constructed using 1M queries extracted from Bing’s logs. For each question, the dataset includes the snippets returned by the search engine for the top-10 ranked web pages. However the gold answers to the questions are not spans of particular retrieved snippets, but were freely written by humans after reading the returned snippets. Hence, gold relevant snippets (or sentences) cannot be identified, making this dataset unsuitable for our purposes. 5 Conclusions and Future Work Our contributions can be summarized as follows: (1) We proposed an architecture to jointly rank documents and snippets with respect to a question, two particularly important stages in QA for large document collections; our architecture can be used with any neural text relevance model. (2) We instantiated the proposed architecture using a recent neural relevance model (PDRMM) and a BERTbased ranker. (3) Using biomedical data (from BIOASQ), we showed that the two resulting joint models (PDRMM-based and BERT-based) vastly outperform the corresponding pipelines in snippet re3904 trieval, the main goal in QA for document collections, using fewer parameters, and also remaining competitive in document retrieval. (4) We showed that the joint model (PDRMM-based) that does not use BERT is competitive with BERT-based models, outperforming the best BIOASQ 6 system; our joint models (PDRMM- and BERT-based) also outperformed all BIOASQ 7 competitors. (5) We provide a modified version of the Natural Questions dataset, suitable for document and snippet retrieval. (6) We showed that our joint PDRMM-based model also largely outperforms the corresponding pipeline on open-domain data (Natural Questions) in snippet retrieval, even though it performs worse than the pipeline in document retrieval. (7) We showed that all the neural pipelines and joint models we considered improve the traditional BM25 ranking on both datasets. (8) We make our code publicly available. We hope to extend our models and datasets for stage (iv), i.e., to also identify exact answer spans within snippets (paragraphs), similar to the answer spans of SQUAD (Rajpurkar et al., 2016, 2018). This would lead to a multi-granular retrieval task, where systems would have to retrieve relevant documents, relevant snippets, and exact answer spans from the relevant snippets. BIOASQ already includes this multi-granular task, but exact answers are provided only for factoid questions and they are freely written by humans, as in MS-MARCO, with similar limitations. Hence, appropriately modified versions of the BIOASQ datasets are needed. Acknowledgements We thank Ryan McDonald for his advice in earlier stages of this work. References Amin Ahmad, Noah Constant, Yinfei Yang, and Daniel Cer. 2019. ReQA: An evaluation for end-to-end answer retrieval models. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 137–146, Hong Kong, China. Qingyao Ai, Brendan O’Connor, and W. Bruce Croft. 2018. A Neural Passage Model for Ad-hoc Document Retrieval. In Advances in Information Retrieval, Cham. Lisa Bauer, Yicheng Wang, and Mohit Bansal. 2018. Commonsense for generative multi-hop question answering tasks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4220–4230, Brussels, Belgium. George Brokos, Polyvios Liosis, Ryan McDonald, Dimitris Pappas, and Ion Androutsopoulos. 2018. AUEB at BioASQ 6: Document and Snippet Retrieval. In Proceedings of the 6th BioASQ Workshop, pages 30–39, Brussels, Belgium. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer opendomain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870– 1879, Vancouver, Canada. Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. 2017. Attention-overAttention Neural Networks for Reading Comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 593–602, Vancouver, Canada. Zhuyun Dai, Chenyan Xiong, Jamie Callan, and Zhiyuan Liu. 2018. Convolutional neural networks for soft-matching n-grams in ad-hoc search. In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, pages 126– 134, Marina Del Rey, CA. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. 2018. The Hitchhiker’s Guide to Testing Statistical Significance in Natural Language Processing. In Proceedings of the 56th Annual Meeting of the ACL (Volume 1: Long Papers), pages 1383–1392. Matthew Dunn, Levent Sagun, Mike Higgins, V. Ugur G¨uney, Volkan Cirik, and Kyunghyun Cho. 2017. SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine. ArXiv, abs/1704.05179. Yixing Fan, Jiafeng Guo, Yanyan Lan, Jun Xu, Chengxiang Zhai, and Xueqi Cheng. 2018. Modeling Diverse Relevance Patterns in Ad-Hoc Retrieval. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval. Jiafeng Guo, Yixing Fan, Qingyao Ai, and W. Bruce Croft. 2016. A Deep Relevance Matching Model for Ad-hoc Retrieval. In Proceedings of the 25th ACM International on Conference on Information and Knowledge Management, pages 55–64, Indianapolis, Indiana, USA. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE conference 3905 on computer vision and pattern recognition, pages 770–778. Stefan Hosein, Daniel Andor, and Ryan McDonald. 2019. Measuring domain portability and error propagation in biomedical QA. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 686–694, Wurzburg, Germany. Kai Hui, Andrew Yates, Klaus Berberich, and Gerard de Melo. 2017. PACRR: A Position-Aware Neural IR Model for Relevance Matching. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1049–1058, Copenhagen, Denmark. Kai Hui, Andrew Yates, Klaus Berberich, and Gerard de Melo. 2018. Co-PACRR: A context-aware neural IR model for ad-hoc retrieval. In Proceedings of the 11th ACM International Conference on Web Search and Data Mining, pages 279–287, Marina Del Rey, CA. Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. 2016. Text Understanding with the Attention Sum Reader Network. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 908–918, Berlin, Germany. Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT. ArXiv, abs/2004.12832. Tushar Khot, Ashish Sabharwal, and Peter Clark. 2019. What’s missing: A knowledge gap guided approach for multi-hop question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2814–2828, Hong Kong, China. Association for Computational Linguistics. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. CoRR, abs/1412.6980. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: a Benchmark for Question Answering Research. Transactions of the Association of Computational Linguistics. Jinhyuk Lee, Seongjun Yun, Hyunjae Kim, Miyoung Ko, and Jaewoo Kang. 2018. Ranking paragraphs for improving answer recall in open-domain question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 565–569, Brussels, Belgium. Association for Computational Linguistics. Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6086–6096, Florence, Italy. Jimmy Lin. 2019. The Neural Hype and Comparisons Against Weak Baselines. SIGIR Forum, 52(2):40–51. Sean MacAvaney, Andrew Yates, Arman Cohan, and Nazli Goharian. 2019. CEDR: Contextualized Embeddings for Document Ranking. CoRR, abs/1904.07094. Joel Mackenzie, Zhuyun Dai, Luke Gallagher, and Jamie Callan. 2020. Efficiency Implications of Term Weighting for Passage Retrieval, page 1821–1824. Association for Computing Machinery, New York, NY, USA. Christopher D. Manning, Prabhakar Raghavan, and Hinrich Sch¨utze. 2008. Introduction to Information Retrieval. Cambridge University Press. Ryan McDonald, George Brokos, and Ion Androutsopoulos. 2018. Deep Relevance Ranking Using Enhanced Document-Query Interactions. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1849–1860, Brussels, Belgium. Bhaskar Mitra and Nick Craswell. 2018. An Introduction to Neural Information Retrieval. Now Publishers. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. CoRR, abs/1611.09268. Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage Re-ranking with BERT. CoRR, abs/1901.04085. S. Pandey, I. Mathur, and N. Joshi. 2019. Information retrieval ranking using machine learning techniques. In 2019 Amity International Conference on Artificial Intelligence (AICAI), pages 86–92. Liang Pang, Yanyan Lan, Jiafeng Guo, Jun Xu, Jingfang Xu, and Xueqi Cheng. 2017. DeepRank: A New Deep Architecture for Relevance Ranking in Information Retrieval. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management. Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep Contextualized Word Representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2227–2237, New Orleans, Louisiana. 3906 Tao Qin, Tie-Yan Liu, Jun Xu, and Hang Li. 2010. Letor: A benchmark collection for research on learning to rank for information retrieval. Inf. Retrieval. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784–789, Melbourne, Australia. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Stephen Robertson and Hugo Zaragoza. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4):333–389. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8. Apoorv Saxena, Aditay Tripathi, and Partha Talukdar. 2020. Improving multi-hop question answering over knowledge graphs using knowledge base embeddings. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4498–4507, Online. Ivan Sekuli´c, Amir Soleimani, Mohammad Aliannejadi, and Fabio Crestani. 2020. Longformer for MS MARCO Document Re-ranking Task. ArXiv, abs/2009.09392. Md Arafat Sultan, Vittorio Castelli, and Radu Florian. 2016. A Joint Model for Answer Sentence Ranking and Answer Extraction. Transactions of the Association for Computational Linguistics, 4:113–125. G. Tsatsaronis, G. Balikas, P. Malakasiotis, I. Partalas, M. Zschunke, M.R. Alvers, D. Weissenborn, A. Krithara, S. Petridis, D. Polychronopoulos, Y. Almirantis, J. Pavlopoulos, N. Baskiotis, P. Gallinari, T. Artieres, A. Ngonga, N. Heino, E. Gaussier, L. Barrio-Alvers, M. Schroeder, I. Androutsopoulos, and G. Paliouras. 2015. An overview of the BioASQ Large-Scale Biomedical Semantic Indexing and Question Answering Competition. BMC Bioinformatics, 16(138). Ellen M. Voorhees. 2001. The TREC question answering track. Natural Language Engineering, 7(4):361–378. Chenyan Xiong, Zhuyun Dai, Jamie Callan, Zhiyuan Liu, and Russell Power. 2017. End-to-end neural ad-hoc ranking with kernel pooling. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 55–64, Shinjuku, Tokyo, Japan. Peng Xu, Xiaofei Ma, Ramesh Nallapati, and Bing Xiang. 2019. Passage Ranking with Weak Supervsion. arxiv. Wei Yang, Yaxiong Xie, Aileen Lin, Xingyu Li, Luchen Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019a. End-to-End open-Domain Question Answering with BERTserini. CoRR, abs/1902.01718. Wei Yang, Haotian Zhang, and Jimmy Lin. 2019b. Simple Applications of BERT for Ad Hoc Document Retrieval. CoRR, abs/1903.10972. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369–2380, Brussels, Belgium. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical Attention Networks for Document Classification. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Wenpeng Yin, Hinrich Sch¨utze, Bing Xiang, and Bowen Zhou. 2016. ABCNN: Attention-based convolutional neural network for modeling sentence pairs. Transactions of the Association for Computational Linguistics, 4. Zhuosheng Zhang, Jun jie Yang, and Hai Zhao. 2020. Retrospective reader for machine reading comprehension. ArXiv. Appendix Tuning the weights of the two losses and the effect of extra features in JPDRMM In Table 1, all joint models used the sum of the document and snippet loss (L = Ldoc + Lsnip). By contrast, in Table 4 we use a linear combination L = Ldoc+λsnipLsnip and tune the hyper-parameter λsnip ∈{10, 1, 0.1, 0.01}. We also try removing the extra document and/or sentence features (Fig. 1– 3) to check their effect. This experiment was performed only with JPDRMM, which is one of our best joint models and computationally much cheaper than methods that employ BERT. As in Table 1, we use the BIOASQ data, but here we perform a 10-fold cross-validation on the union of the training and development subsets. This is why the results for λsnip = 1 when using both the sentence and document extra features (row 4, in italics) are slightly different than the corresponding JPDRMM results of Table 1 (6.69 and 15.72, respectively). 3907 Sent. Doc. Doc. Snip. Extra Extra λsnip MAP (%) MAP (%) Yes Yes 10 6.23 ± 0.14 14.73 ± 0.32 Yes No 10 1.20 ± 0.14 3.59 ± 0.45 No Yes 10 1.18 ± 0.23 2.19 ± 0.29 Yes Yes 1 6.80 ± 0.07 15.42 ± 0.23 Yes No 1 1.35 ± 0.24 3.77 ± 0.73 No Yes 1 7.35 ± 0.16 14.58 ± 0.88 Yes Yes 0.1 7.85 ± 0.08 17.28 ± 0.26 Yes No 0.1 6.77 ± 0.25 13.86 ± 1.10 No Yes 0.1 7.59 ± 0.12 15.77 ± 0.60 Yes Yes 0.01 7.83 ± 0.07 17.34 ± 0.37 Yes No 0.01 6.61 ± 0.19 12.96 ± 0.29 No Yes 0.01 7.65 ± 0.10 14.24 ± 1.63 Table 4: JPDRMM results on BIOASQ 7 data for tuned weights of the two losses, with and without the extra sentence and document features. The 4th row (in italics) corresponds to the JPDRMM configuration of Table 1, but the results here are slightly different, because we used a 10-fold cross-validation on the training and development data. The MAP scores are averaged over the 10 folds. We also report standard deviations (±). Table 4 shows that further performance gains (6.80 to 7.85 document MAP, 15.42 to 17.34 snippet MAP) are possible by tuning the weights of the two losses. The best scores are obtained when using both the extra sentence and document features. However, the model performs reasonably well even when one of the two types of extra features is removed, with the exception of λsnip = 10. The standard deviations of the MAP scores over the folds of the cross-validation indicate that the performance of the model is reasonably stable. Error Analysis and Limitations We conducted an exploratory analysis of the retrieved snippets in the two datasets. For each dataset, we used the model with the best snippet retrieval performance, i.e., JPDRMM for the modified Natural Questions (Table 3) and BJPDRMM-ADAPTNF for BIOASQ (Table 1). Both models struggle to retrieve the gold sentences when the answer is not explicitly mentioned in them. For example, the gold sentence for the question “What is the most famous fountain in Rome?” of the Natural Questions dataset is: “The Trevi Fountain (Italian: Fontana di Trevi) is a fountain in the Trevi district in Rome, Italy, designed by Italian architect Nicola Salvi and completed by Giuseppe Pannini.” Instead, the top sentence of JPDRMM is the following, which looks reasonably good, but mentions famous fountains (of a particular kind) near Rome. “The most famous fountains of this kind were found in the Villa d’Este, at Tivoli near Rome, which featured a hillside of basins, fountains and jets of water, as well as a fountain which produced music by pouring water into a chamber, forcing air into a series of flute-like pipes.”. To prefer the gold sentence, the model needs to know that Fontana di Trevi is also very famous, but this information is not included in the gold sentence itself, though it is included in the next sentence: “Standing 26.3 metres (86 ft) high and 49.15 metres (161.3 ft) wide, it is the largest Baroque fountain in the city and one of the most famous fountains in the world.” Hence, some form of multi-hop QA (Yang et al., 2018; Bauer et al., 2018; Khot et al., 2019; Saxena et al., 2020) seems to be needed to combine the information that Fontana di Trevi is in Rome (explicitly mentioned in the gold sentence) with information from the next sentence and, more generally, other sentences even from different documents. In the case of the question “What part of the body is affected by mesotheliomia?” of the BIOASQ dataset, the gold sentence is: ‘’Malignant pleural mesothelioma (MPM) is a hard to treat malignancy arising from the mesothelial surface of the pleura.” Instead, the top sentence of BJPDRMM-ADAPT-NF is the following, which contains several words of the question, but not ‘mesothelioma’, which is the most important question term. “For PTs specialized in acute care, geriatrics and pediatrics, the body part most commonly affected was the low back, while for PTs specialized in orthopedics and neurology, the body part most commonly affected was the neck.” In this case, the gold sentence does not explicitly convey that the pleura is a membrane that envelops each lung of the human body and, therefore, a part of the body. Again, this additional information can be found in other sentences.
|
2021
|
301
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3908–3918 August 1–6, 2021. ©2021 Association for Computational Linguistics 3908 W-RST: Towards a Weighted RST-style Discourse Framework Patrick Huber∗, Wen Xiao∗, Giuseppe Carenini Department of Computer Science University of British Columbia Vancouver, BC, Canada, V6T 1Z4 {huberpat, xiaowen3, carenini}@cs.ubc.ca Abstract Aiming for a better integration of data-driven and linguistically-inspired approaches, we explore whether RST Nuclearity, assigning a binary assessment of importance between text segments, can be replaced by automatically generated, real-valued scores, in what we call a Weighted-RST framework. In particular, we find that weighted discourse trees from auxiliary tasks can benefit key NLP downstream applications compared to nuclearity-centered approaches. We further show that real-valued importance distributions partially and interestingly align with the assessment and uncertainty of human annotators. 1 Introduction Ideally, research in Natural Language Processing (NLP) should balance and integrate findings from machine learning approaches with insights and theories from linguistics. With the enormous success of data-driven approaches over the last decades, this balance has arguably and excessively shifted, with linguistic theories playing a less and less critical role. Even more importantly, there are only little attempts made to improve such theories in light of recent empirical results. In the context of discourse, two main theories have emerged in the past: The Rhetorical Structure Theory (RST) (Carlson et al., 2002) and PDTB (Prasad et al., 2008). In this paper, we focus on RST, exploring whether the underlying theory can be refined in a data-driven manner. In general, RST postulates a complete discourse tree for a given document. To obtain this formal representation as a projective consituency tree, a given document is first separated into so called Elementary Discourse Units (or short EDUs), representing clause-like sentence fragments of the input ∗Equal contribution. document. Afterwards, the discourse tree is built by hierarchically aggregating EDUs into larger constituents annotated with an importance indicator (in RST called nuclearity) and a relation holding between siblings in the aggregation. The nuclearity attribute in RST thereby assigns each sub-tree either a nucleus-attribute, indicating central importance of the sub-tree in the context of the document, or a satellite-attribute, categorizing the sub-tree as of peripheral importance. The relation attribute further characterizes the connection between sub-trees (e.g. Elaboration, Cause, Contradiction). One central requirement of the RST discourse theory, as for all linguistic theories, is that a trained human should be able to specify and interpret the discourse representations. While this is a clear advantage when trying to generate explainable outcomes, it also introduces problematic, humancentered simplifications; the most radical of which is arguably the nuclearity attribute, indicating the importance among siblings. Intuitively, such a coarse (binary) importance assessment does not allow to represent nuanced differences regarding sub-tree importance, which can potentially be critical for downstream tasks. For instance, the importance of two nuclei siblings is rather intuitive to interpret. However, having siblings annotated as “nucleus-satellite” or “satellitenucleus” leaves the question on how much more important the nucleus sub-tree is compared to the satellite, as shown in Figure 1. In general, it is unclear (and unlikely) that the actual importance distributions between siblings with the same nuclearity attribution are consistent. Based on this observation, we investigate the potential of replacing the binary nuclearity assessment postulated by RST with automatically generated, real-valued importance scores in a new, Weighted-RST framework. In contrast with previous work that has assumed RST and developed 3909 Figure 1: Document wsj 0639 from the RST-DT corpus with inconsistent importance differences between N-S attributions. (The top-level satellite is clearly more central to the overall context than the lower-level satellite. However, both are similarly assigned the satellite attribution by at least one annotator). Top relation: Annotator 1: N-S, Annotator 2: N-N. computational models of discourse by simply applying machine learning methods to RST annotated treebanks (Ji and Eisenstein, 2014; Feng and Hirst, 2014; Joty et al., 2015; Li et al., 2016; Wang et al., 2017; Yu et al., 2018), we rely on very recent empirical studies showing that weighted “silver-standard” discourse trees can be inferred from auxiliary tasks such as sentiment analysis (Huber and Carenini, 2020b) and summarization (Xiao et al., 2021). In our evaluation, we assess both, computational benefits and linguistic insights. In particular, we find that automatically generated, weighted discourse trees can benefit key NLP downstream tasks. We further show that real-valued importance scores (at least partially) align with human annotations and can interestingly also capture uncertainty in human annotators, implying some alignment of the importance distributions with linguistic ambiguity. 2 Related Work First introduced by Mann and Thompson (1988), the Rhetorical Structure Theory (RST) has been one of the primary guiding theories for discourse analysis (Carlson et al., 2002; Subba and Di Eugenio, 2009; Zeldes, 2017; Gessler et al., 2019; Liu and Zeldes, 2019), discourse parsing (Ji and Eisenstein, 2014; Feng and Hirst, 2014; Joty et al., 2015; Li et al., 2016; Wang et al., 2017; Yu et al., 2018), and text planning (Torrance, 2015; Gatt and Krahmer, 2018; Guz and Carenini, 2020). The RST framework thereby comprehensively describes the organization of a document, guided by the author’s communicative goals, encompassing three components: (1) A projective constituency tree structure, often referred to as the tree span. (2) A nuclearity attribute, assigned to every internal node of the discourse tree, encoding relative importance between the nodes’ sub-trees, with the nucleus expressing primary importance and a satellite signifying supplementary sub-trees. (3) A relation attribute for every internal node describing the relationship between the sub-trees of a node (e.g., Contrast, Evidence, Contradiction). Arguably, the weakest aspect of an RST representation is the nuclearity assessment, which makes a too coarse differentiation between primary and secondary importance of sub-trees. However, despite its binary assignment of importance and even though the nuclearity attribute is only one of three components of an RST tree, it has major implications for many downstream tasks, as already shown early on by Marcu (1999), using the nuclearity attribute as the key signal in extractive summarization. Further work in sentiment analysis (Bhatia et al., 2015) also showed the importance of nuclearity for the task by first converting the constituency tree into a dependency tree (more aligned with the nuclearity attribute) and then using that tree to predict sentiment more accurately. Both of these results indicate that nuclearity, even in the coarse RST version, already contains valuable information. Hence, we believe that this coarsegrained classification is reasonable when manually annotating discourse, but see it as a major point of improvement, if a more fine-grained assessment could be correctly assigned. We therefore explore the potential of assigning a weighted nuclearity attribute in this paper. While plenty of studies have highlighted the important role of discourse for real-world downstream tasks, including summarization, (Gerani et al., 2014; Xu et al., 2020; Xiao et al., 2020), sentiment analysis (Bhatia et al., 2015; Hogenboom et al., 2015; Nejat et al., 2017) and text classification (Ji and Smith, 2017), more critical to our approach is very recent work exploring such connection in the opposite direction. In Huber and Carenini (2020b), we exploit sentiment related information to generate “silver-standard” nuclearity annotated discourse trees, showing their potential on the domain-transfer discourse parsing task. Crucially for our purposes, this approach internally generates real-valued importance-weights for trees. For the task of extractive summarization, we follow our intuition given in Xiao et al. (2020) and Xiao et al. (2021), exploiting the connection be3910 Figure 2: Three phases of our approach to generate weighted RST-style discourse trees. Left and center steps are described in section 3, right component is described in section 4. † = As in Huber and Carenini (2020b), ‡ = As in Marcu (1999), ∗= Sentiment prediction component is a linear combination, mapping the aggregated embedding to the sentiment output. The linear combination has been previously learned on the training portion of the dataset. tween summarization and discourse. In particular, in Xiao et al. (2021), we demonstrate that the selfattention matrix learned during the training of a transformer-based summarizer captures valid aspects of constituency and dependency discourse trees. To summarize, building on our previous work on creating discourse trees through distant supervision, we take a first step towards generating weighted discourse trees from the sentiment analysis and summarization tasks. 3 W-RST Treebank Generation Given the intuition from above, we combine information from machine learning approaches with insights from linguistics, replacing the humancentered nuclearity assignment with real-valued weights obtained from the sentiment analysis and summarization tasks1. An overview of the process to generate weighted RST-style discourse trees is shown in Figure 2, containing the training phase (left) and the W-RST discourse inference phase (center) described here. The W-RST discourse evaluation (right), is covered in section 4. 3.1 Weighted Trees from Sentiment To generate weighted discourse trees from sentiment, we slightly modify the publicly available code2 presented in Huber and Carenini (2020b) by removing the nuclearity discretization component. An overview of our method is shown in Figure 2 (top), while a detailed view is presented in the left and center parts of Figure 3. First (on the left), we train the Multiple Instance Learning (MIL) 1Please note that both tasks use binarized discourse trees, as commonly used in computational models of RST. 2Code available at https://github.com/nlpat/ MEGA-DT model proposed by Angelidis and Lapata (2018) on a corpus with document-level sentiment goldlabels, internally annotating each input-unit (in our case EDUs) with a sentiment- and attention-score. After the MIL model is trained (center), a tuple (si, ai) containing a sentiment score si and an attention ai is extracted for each EDU i. Based on these tuples representing leaf nodes, the CKY algorithm (Jurafsky and Martin, 2014) is applied to find the tree structure to best align with the overall document sentiment, through a bottom-up aggregation approach defined as3: sp = sl ∗al + sr ∗ar al + ar ap = al + ar 2 with nodes l and r as the left and right childnodes of p respectively. The attention scores (al, ar) are here interpreted as the importance weights for the respective sub-trees (wl = al/(al + ar) and wr = ar/(al + ar)), resulting in a complete, normalized and weighted discourse structure as required for W-RST. We call the discourse treebank generated with this approach W-RST-Sent. 3.2 Weighted Trees from Summarization In order to derive weighted discourse trees from a summarization model we follow Xiao et al. (2021)4, generating weighted discourse trees from the selfattention matrices of a transformer-based summarization model. An overview of our method is shown in Figure 2 (bottom), while a detailed view is presented in the left and center parts of Figure 4. We start by training a transformer-based extractive summarization model (left), containing three 3Equations taken from Huber and Carenini (2020b) 4Code available at https://github.com/ Wendy-Xiao/summ_guided_disco_parser 3911 Figure 3: Three phases of our approach. Left/Center: Detailed view into the generation of weighted RST-style discourse trees using the sentiment analysis downstream task. Right: Sentiment discourse application evaluation Figure 4: Three phases of our approach. Left/Center: Detailed view into the generation of weighted RST-style discourse trees using the summarization downstream task. Right: Summarization discourse application evaluation components: (1) A pre-trained BERT EDU Encoder generating EDU embeddings, (2) a standard transformer architecture as proposed in Vaswani et al. (2017) and (3) a final classifier, mapping the outputs of the transformer to a probability score for each EDU, indicating whether the EDU should be part of the extractive summary. With the trained transformer model, we then extract the self-attention matrix A and build a discourse tree in bottom-up fashion (as shown in the center of Figure 4). Specifically, the self-attention matrix A reflects the relationships between units in the document, where entry Aij measures how much the i-th EDU relies on the j-th EDU. Given this information, we generate an unlabeled constituency tree using the CKY algorithm (Jurafsky and Martin, 2014), optimizing the overall tree score, as previously done in Xiao et al. (2021). In terms of weight-assignment, given a sub-tree spanning EDUs i to j, split into child-constituents at EDU k, then max(Ai:k,(k+1):j), representing the maximal attention value that any EDU in the left constituent is paying to an EDU in the right childconstituent, reflects how much the left sub-tree relies on the right sub-tree, while max(A(k+1):j,i:k) defines how much the right sub-tree depends on the left. We define the importance-weights of the left (wl) and right (wr) sub-trees as: wl = max(A(k+1):j,i:k)/(wl + wr) wr = max(Ai:k,(k+1):j)/(wl + wr) In this way, the importance scores of the two subtrees represent a real-valued distribution. In combination with the unlabeled structure computation, we generate a weighted discourse tree for each document. We call the discourse treebank generated with the summarization downstream information W-RST-Summ. 4 W-RST Discourse Evaluation To assess the potential of W-RST, we consider two evaluation scenarios (Figure 2, right): (1) Apply weighted discourse trees to the tasks of sentiment analysis and summarization and (2) analyze the weight alignment with human annotations. 4.1 Weight-based Discourse Applications In this evaluation scenario, we address the question of whether W-RST trees can support downstream tasks better than traditional RST trees with nuclearity. Specifically, we leverage the discourse trees learned from sentiment for the sentiment analysis task itself and, similarly, rely on the discourse trees learned from summarization to benefit the summarization task. 3912 4.1.1 Sentiment Analysis In order to predict the sentiment of a document in W-RST-Sent based on its weighted discourse tree, we need to introduce an additional source of information to be aggregated according to such tree. Here, we choose word embeddings, as commonly used as an initial transformation in many models tackling the sentiment prediction task (Kim, 2014; Tai et al., 2015; Yang et al., 2016; Adhikari et al., 2019; Huber and Carenini, 2020a). To avoid introducing additional confounding factors through sophisticated tree aggregation approaches (e.g. TreeLSTMs (Tai et al., 2015)), we select a simple method, aiming to directly compare the inferred tree-structures and allowing us to better assess the performance differences originating from the weight/nuclearity attribution (see right step in Figure 3). More specifically, we start by computing the average word-embedding for each leaf node leafi (here containing a single EDU) in the discourse tree. leafi = j<|leafi| X j=0 Emb(wordj i)/|leafi| With |leafi| as the number of words in leaf i, Emb(·) being the embedding lookup and wordj i representing word j within leaf i. Subsequently, we aggregate constituents, starting from the leaf nodes (with leafi as embedding constituent ci), according to the weights of the discourse tree. For any two sibling constituents cl and cr of the parent sub-tree cp in the binary tree, we compute cp = cl ∗wl + cr ∗wr with wl and wr as the real-valued weightdistribution extracted from the inferred discourse tree and cp, cl and cr as dense encodings. We aggregate the complete document in bottom-up fashion, eventually reaching a root node embedding containing a tree-weighted average of the leaf-nodes. Given the root-node embedding representing a complete document, a simple Multilayer Perceptron (MLP) trained on the original training portion of the MIL model is used to predict the sentiment of the document. 4.1.2 Summarization In the evaluation step of the summarization model (right of Figure 4), we use the weighted discourse tree of a document in W-RST-Summ to predict its extractive summary by applying an adaptation of the unsupervised summarization method by Marcu (1999). We choose this straightforward algorithm over more elaborate and hyper-parameter heavy approaches to avoid confounding factors, since our aim is to evaluate solely the potential of the weighted discourse trees compared to standard RST-style annotations. In the original algorithm, a summary is computed based on the nuclearity attribute by recursively computing the importance scores for all units as: Sn(u, N) = dN, u ∈Prom(N) S(u, C(N)) s.t. u ∈C(N) otherwise where C(N) represents the child of N, and Prom(N) is the promotion set of node N, which is defined in bottom-up fashion as follows: (1) Prom of a leaf node is the leaf node itself. (2) Prom of an internal node is the union of the promotion sets of its nucleus children. Furthermore, dN represents the level of a node N, computed as the distance from the level of the lowest leaf-node. This way, units in the promotion set originating from nodes that are higher up in the discourse tree are amplified in their importance compared to those from lower levels. As for the W-RST-Summ discourse trees with real-valued importance-weights, we adapt Marcu’s algorithm by replacing the promotion set with realvalued importance scores as shown here: Sw(u, N) = d + wN, N is leaf Sw(u, C(N)) + wN , u ∈C(N) otherwise Once Sn or Sw are computed, the top-k units of the highest promotion set or with the highest importance scores respectively are selected into the final summary. 4.1.3 Nuclearity-attributed Baselines To test whether the W-RST trees are effectively predicting the downstream tasks, we need to generate traditional RST trees with nuclearity to compare against. However, moving from weighted discourse trees to coarse nuclearity requires the introduction of a threshold. More specifically, while “nucleus-satellite” and “satellite-nucleus” assignments can be naturally generated depending on the distinct weights, in order to assign the third “nucleus-nucleus” class, frequently appearing in 3913 Figure 5: Three phases of our approach. Left: Generation of W-RST-Sent/Summ discourse trees. Right: Linguistic evaluation RST-style treebanks, we need to specify how close two weights have to be for such configuration to apply. Formally, we set a threshold t as follows: If: |wl −wr| < t → nucleus-nucleus Else: If: wl > wr → nucleus-satellite Else: If: wl ≤wr → satellite-nucleus This way, RST-style treebanks with nuclearity attributions can be generated from W-RST-Sent and W-RST-Summ and used for the sentiment analysis and summarization downstream tasks. For the nuclearity-attributed baseline of the sentiment task, we use a similar approach as for the W-RST evaluation procedure, but assign two distinct weights wn and ws to the nucleus and satellite child respectively. Since it is not clear how much more important a nucleus node is compared to a satellite using the traditional RST notation, we define the two weights based on the threshold t as: wn = 1 −(1 −2t)/4 ws = (1 −2t)/4 The intuition behind this formulation is that for a high threshold t (e.g. 0.8), the nuclearity needs to be very prominent (the difference between the normalized weights needs to exceed 0.8), making the nucleus clearly more important than the satellite, while for a small threshold (e.g. 0.1), even relatively balanced weights (for example wl = 0.56, wr = 0.44) will be assigned as “nucleus-satellite”, resulting in the potential difference in importance of the siblings to be less eminent. For the nuclearity-attributed baseline for summarization, we directly apply the original algorithm by Marcu (1999) as described in section 4.1.2. However, when using the promotion set to determine which EDUs are added to the summarization, potential ties can occur. Since the discourse tree does not provide any information on how to prioritize those, we randomly select units from the candidates, whenever there is a tie. This avoids exploiting any positional bias in the data (e.g. the lead bias), which would confound the results. 4.2 Weight Alignment with Human Annotation As for our second W-RST discourse evaluation task, we investigate if the real-valued importanceweights align with human annotations. To be able to explore this scenario, we generate weighted tree annotations for an existing discourse treebank (RST-DT (Carlson et al., 2002)). In this evaluation task we verify if: (1) The nucleus in a gold-annotation generally receives more weight than a satellite (i.e. if importance-weights generally favour nuclei over satellites) and, similarly, if nucleus-nucleus relations receive more balanced weights. (2) In accordance with Figure 1, we further explore how well the weights capture the extend to which a relation is dominated by the nucleus. Here, our intuition is that for inconsistent human nuclearity annotations the spread should generally be lower than for consistent annotations, assuming that human misalignment in the discourse annotation indicates ambivalence on the importance of sub-trees. To test for these two properties, we use discourse documents individually annotated by two human annotators and analyze each sub-tree within the doubly-annotated documents with consistent interannotator structure assessment for their nuclearity assignment. For each of the 6 possible interannotator nuclearity assessments, consisting of 3 consistent annotation classes (namely N-N/N-N, NS/N-S and S-N/S-N) and 3 inconsistent annotation classes (namely N-N/N-S, N-N/S-N and N-S/SN)5, we explore the respective weight distribution of the document annotated with the two W-RST tasks – sentiment analysis and summarization (see Figure 5). We compute an average spread sc for each of the 6 inter-annotator nuclearity assessments classes c as: sc = ( j<|c| X j=0 wj l −wj r)/|c| With wj l and wj r as the weights of the left and right child node of sub-tree j in class c, respectively. 5We don’t take the order of annotators into consideration, mapping N-N/N-S and N-S/N-N both onto N-N/N-S. 3914 5 Experiments 5.1 Experimental Setup Sentiment Analysis: We follow our previous approach in Huber and Carenini (2020b) for the model training and W-RST discourse inference steps (left and center in Figure 3) using the adapted MILNet model from Angelidis and Lapata (2018) trained with a batch-size of 200 and 100 neurons in a single layer bi-directional GRU with 20% dropout for 25 epochs. Next, discourse trees are generated using the best-performing heuristic CKY method with the stochastic exploration-exploitation trade-off from Huber and Carenini (2020b) (beam size 10, linear decreasing τ). As word-embeddings in the W-RST discourse evaluation (right in Figure 3), we use GloVe embeddings (Pennington et al., 2014), which previous work (Tai et al., 2015; Huber and Carenini, 2020a) indicates to be suitable for aggregation in discourse processing. For training and evaluation of the sentiment analysis task, we use the 5-class Yelp’13 review dataset (Tang et al., 2015). To compare our approach against the traditional RST approach with nuclearity, we explore the impact of 11 distinct thresholds for the baseline described in §4.1.3, ranging from 0 to 1 in 0.1 intervals. Summarization: To be consistent with RST, our summarizer extracts EDUs instead of sentences from a given document. The model is trained on the EDU-segmented CNNDM dataset containing EDU-level Oracle labels published by Xu et al. (2020). We further use a pre-trained BERT-base (“uncased”) model to generate the embeddings of EDUs. The transformer used is the standard model with 6 layers and 8 heads in each layer (d = 512). We train the extractive summarizer on the training set of the CNNDM corpus (Nallapati et al., 2016) and pick the best attention head using the RST-DT dataset (Carlson et al., 2002) as the development set. We test the trees by running the summarization algorithm in Marcu (1999) on the test set of the CNNDM dataset, and select the top-6 EDUs based on the importance score to form a summary in natural order. Regarding the baseline model using thresholds, we apply the same 11 thresholds as for the sentiment analysis task. Weight Alignment with Human Annotation: As discussed in §4.2, this evaluation requires two parallel human generated discourse trees for every document. Luckily, in the RST-DT corpus pub0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.6 0.8 1 Nucleus Ratio 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 53 54 55 56 Accuracy nuclearity 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 0.6 0.8 1 threshold Nucleus Ratio 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 20 21 22 23 Avg ROUGE Figure 6: Top: Sentiment Analysis accuracy of the WRST model compared to the standard RST framework with different thresholds. Bottom: Average ROUGE score (ROUGE-1, -2 and -L) of the W-RST summarization model compared to different thresholds. Full numerical results are shown in Appendix A. N-N N-S S-N N-N 273 99 41 N-S 694 75 S-N 172 Table 1: Statistics on consistently and inconsistently annotated samples of the 1, 354 structure-aligned subtrees generated by two distinct human annotators. lished by Carlson et al. (2002), 53 of the 385 documents annotated with full RST-style discourse trees are doubly tagged by a second linguist. We use the 53 documents containing 1, 354 consistent structure annotations between the two analysts to evaluate the linguistic alignment of our generated W-RST documents with human discourse interpretations. Out of the 1, 354 structure-aligned subtrees, in 1, 139 cases both annotators agreed on the nuclearity attribute, while 215 times a nuclearity mismatch appeared, as shown in detail in Table 1. 5.2 Results and Analysis The results of the experiments on the discourse applications for sentiment analysis and summarization are shown in Figure 6. The results for 3915 Sent N-N N-S S-N N-N -0.228 (106) -0.238 (33) -0.240 (19) N-S -0.038 (325) -0.044 (22) S-N -0.278 (115) Summ N-N N-S S-N N-N 0.572 (136) 0.604 (42) 0.506 (25) N-S 0.713 (418) 0.518 (36) S-N 0.616 (134) Table 2: Confusion Matrices based on human annotation showing the absolute weight-spread using the Sentiment (top) and Summarization (bottom) tasks on 620 and 791 sub-trees aligned with the human structure prediction, respectively. Cell upper value: Absolute weight spread for the respective combination of humanannotated nuclearities. Lower value (in brackets): Support for this configuration. sentiment analysis (top) and summarization (bottom) thereby show a similar trend: With an increasing threshold and therefore a larger number of N-N relations (shown as grey bars in the Figure), the standard RST baseline (blue line) consistently improves for the respective performance measure of both tasks. However, reaching the best performance at a threshold of 0.8 for sentiment analysis and 0.6 for summarization, the performance starts to deteriorate. This general trend seems reasonable, given that N-N relations represent a rather frequent nuclearity connection, however classifying every connection as N-N leads to a severe loss of information. Furthermore, the performance suggests that while the N-N class is important in both cases, the optimal threshold varies depending on the task and potentially also the corpus used, making further task-specific fine-tuning steps mandatory. The weighted discourse trees following our W-RST approach, on the other hand, do not require the definition of a threshold, resulting in a single, promising performance (red line) for both tasks in Figure 6. For comparison, we apply the generated trees of a standard RST-style discourse parser (here the Two-Stage parser by Wang et al. (2017)) trained on the RST-DT dataset (Carlson et al., 2002) on both downstream tasks. The fully-supervised parser reaches an accuracy of 44.77% for sentiment analysis and an average ROUGE score of 26.28 for summarization. While the average ROUGE score Sent N-N N-S S-N N-N ∅-0.36 ∅-0.43 ∅-0.45 N-S ∅+1.00 ∅+0.96 S-N ∅-0.72 Summ N-N N-S S-N N-N ∅-0.13 ∅+0.13 ∅-0.66 N-S ∅+1.00 ∅-0.56 S-N ∅+0.22 Table 3: Confusion Matrices based on human annotation showing the weight-spread relative to the taskaverage for Sentiment (top) and Summarization (bottom), aligned with the human structure prediction, respectively. Cell value: Relative weight spread as the divergence from the average spread across all cells in Table 2. Color: Positive/Negative divergence, ∅= Average value of absolute scores. of the fully-supervised parser is above the performance of our W-RST results for the summarization task, the accuracy on the sentiment analysis task is well below our approach. We believe that these results are a direct indication of the problematic domain adaptation of fully supervised discourse parsers, where the application on a similar domain (Wall Street Journal articles vs. CNN-Daily Mail articles) leads to superior performances compared to our distantly supervised method, however, with larger domain shifts (Wall Street Journal articles vs. Yelp customer reviews), the performance drops significantly, allowing our distantly supervised model to outperform the supervised discourse trees for the downstream task. Arguably, this indicates that although our weighted approach is still not competitive with fully-supervised models in the same domain, it is the most promising solution available for cross-domain discourse parsing. With respect to exploring the weight alignment with human annotations, we show a set of confusion matrices based on human annotation for each W-RST discourse generation task on the absolute and relative weight-spread in Tables 2 and 3 respectively. The results for the sentiment analysis task are shown on the top of both tables, while the performance for the summarization task is shown at the bottom. For instance, the top right cell of the upper confusion matrix in Table 2 shows that for 19 sub-trees in the doubly annotated subset of RST-DT one of the annotators labelled the subtree with a nucleus-nucleus nuclearity attribution, while the second annotator identified it as satellite3916 nucleus. The average weight spread (see §4.2) for those 19 sub-trees is −0.24. Regarding Table 3, we subtract the average spread across Table 2 defined as ∅= P ci∈C (ci)/|C| (with C = {c1, c2, ...c6} containing the cell values in the upper triangle matrix) from each cell value ci and normalize by max = maxci∈C(|ci−∅|), with ∅= −0.177 and max = 0.1396 across the top table. Accordingly, we transform the −0.24 in the top right cell into (−0.24 −avg)/max = −0.45. Moving to the analysis of the results, we find the following trends in this experiment: (1) As presented in Table 2, the sentiment analysis task tends to strongly over-predict S-N (i.e., wl << wr), leading to negative spreads in all cells. In contrast, the summarization task is heavily skewed towards N-S assignments (i.e., wl >> wr), leading to exclusively positive spreads. We believe both trends are consistent with the intrinsic properties of the tasks, given that the general structure of reviews tends to become more important towards the end of a review (leading to increased S-N assignments), while for summarization, the lead bias potentially produces the overall strong nucleus-satellite trend. (2) To investigate the relative weight spreads for different human annotations (i.e., between cells) beyond the trends shown in Table 2, we normalize values within a table by subtracting the average and scaling between [−1, 1]. As a result, Table 3 shows the relative weight spread for different human annotations. Apart from the general trends described in Table 2, the consistently annotated samples of the two linguists (along the diagonal of the confusion matrices) align reasonably. The most positive weight spread is consistently found in the agreed-upon nucleus-satellite case, while the nucleus-nucleus annotation has, as expected, the lowest divergence (i.e., closest to zero) along the diagonal in Table 3. (3) Regarding the inconsistently annotated samples (shown in the triangle matrix above the diagonal) it becomes clear that in the sentiment analysis model the values for the N-N/N-S and N-N/S-N annotated samples (top row in Table 3) are relatively close to the average value. This indicates that, similar to the nucleus-nucleus case, the weights are also ambivalent, with the N-N/NS value (top center) slightly larger than the value for N-N/S-N (top right). The N-S/S-N case for the sentiment analysis model is less aligned with our intuition, showing a strongly negative weightspread (i.e. wl << wr) where we would have expected a more ambivalent result with wl ≈wr (however, aligned with the overall trend shown in Table 2). For summarization, we see a very similar trend with the values for N-N/N-S and N-N/S-N annotated samples. Again, both values are close to the average, with the N-N/N-S cell showing a more positive spread than N-N/S-N. However for summarization, the consistent satellite-nucleus annotation (bottom right cell) seems misaligned with the rest of the table, following instead the general trend for summarization described in Table 2. All in all, the results suggest that the values in most cells are well aligned with what we would expect regarding the relative spread. Interestingly, human uncertainty appears to be reasonably captured in the weights, which seem to contain more fine grained information about the relative importance of sibling sub-trees. 6 Conclusion and Future Work We propose W-RST as a new discourse framework, where the binary nuclearity assessment postulated by RST is replaced with more expressive weights, that can be automatically generated from auxiliary tasks. A series of experiments indicate that W-RST is beneficial to the two key NLP downstream tasks of sentiment analysis and summarization. Further, we show that W-RST trees interestingly align with the uncertainty of human annotations. For the future, we plan to develop a neural discourse parser that learns to predict importance weights instead of nuclearity attributions when trained on large W-RST treebanks. More longer term, we want to explore other aspects of RST that can be refined in light of empirical results, plan to integrate our results into state-of-the-art sentiment analysis and summarization approaches (e.g. Xu et al. (2020)) and generate parallel W-RST structures in a multi-task manner to improve the generality of the discourse trees. Acknowledgments We thank the anonymous reviewers for their insightful comments. This research was supported by the Language & Speech Innovation Lab of Cloud BU, Huawei Technologies Co., Ltd and the Natural Sciences and Engineering Research Council of Canada (NSERC). Nous remercions le Conseil de recherches en sciences naturelles et en g´enie du Canada (CRSNG) de son soutien. 3917 References Ashutosh Adhikari, Achyudh Ram, Raphael Tang, and Jimmy Lin. 2019. Rethinking complex neural network architectures for document classification. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4046– 4051. Stefanos Angelidis and Mirella Lapata. 2018. Multiple instance learning networks for fine-grained sentiment analysis. Transactions of the Association for Computational Linguistics, 6:17–31. Parminder Bhatia, Yangfeng Ji, and Jacob Eisenstein. 2015. Better document-level sentiment analysis from RST discourse parsing. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2212–2218. Lynn Carlson, Mary Ellen Okurowski, and Daniel Marcu. 2002. RST discourse treebank. Linguistic Data Consortium, University of Pennsylvania. Vanessa Wei Feng and Graeme Hirst. 2014. A lineartime bottom-up discourse parser with constraints and post-editing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 511–521. Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artificial Intelligence Research, 61:65–170. Shima Gerani, Yashar Mehdad, Giuseppe Carenini, Raymond T Ng, and Bita Nejat. 2014. Abstractive summarization of product reviews using discourse structure. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1602–1613. Luke Gessler, Yang Janet Liu, and Amir Zeldes. 2019. A discourse signal annotation system for rst trees. In Proceedings of the Workshop on Discourse Relation Parsing and Treebanking 2019, pages 56–61. Grigorii Guz and Giuseppe Carenini. 2020. Towards domain-independent text structuring trainable on large discourse treebanks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 3141–3152. Alexander Hogenboom, Flavius Frasincar, Franciska De Jong, and Uzay Kaymak. 2015. Using rhetorical structure in sentiment analysis. Commun. ACM, 58(7):69–77. Patrick Huber and Giuseppe Carenini. 2020a. From sentiment annotations to sentiment prediction through discourse augmentation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 185–197. Patrick Huber and Giuseppe Carenini. 2020b. MEGA RST discourse treebanks with structure and nuclearity from scalable distant sentiment supervision. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7442–7457. Yangfeng Ji and Jacob Eisenstein. 2014. Representation learning for text-level discourse parsing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 13–24. Yangfeng Ji and Noah A Smith. 2017. Neural discourse structure for text categorization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 996–1005. Shafiq Joty, Giuseppe Carenini, and Raymond T Ng. 2015. CODRA: A novel discriminative framework for rhetorical analysis. Computational Linguistics, 41(3). Dan Jurafsky and James H Martin. 2014. Speech and language processing, volume 3. Pearson London. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751. Qi Li, Tianshi Li, and Baobao Chang. 2016. Discourse parsing with attention-based hierarchical neural networks. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 362–371. Yang Liu and Amir Zeldes. 2019. Discourse relations and signaling information: Anchoring discourse signals in rst-dt. Proceedings of the Society for Computation in Linguistics, 2(1):314–317. William C Mann and Sandra A Thompson. 1988. Rhetorical structure theory: Toward a functional theory of text organization. Text, 8(3):243–281. Daniel Marcu. 1999. Discourse trees are good indicators of importance in text. Advances in automatic text summarization, 293:123–136. Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, C¸ a˘glar Gu`I‡lc¸ehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence RNNs and beyond. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 280–290. Association for Computational Linguistics. Bita Nejat, Giuseppe Carenini, and Raymond Ng. 2017. Exploring joint neural model for sentence level discourse parsing and sentiment analysis. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 289–298. 3918 Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The penn discourse treebank 2.0. LREC. Rajen Subba and Barbara Di Eugenio. 2009. An effective discourse parser that uses rich linguistic information. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 566–574. Association for Computational Linguistics. Kai Sheng Tai, Richard Socher, and Christopher D Manning. 2015. Improved semantic representations from tree-structured long short-term memory networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1556–1566. Duyu Tang, Bing Qin, and Ting Liu. 2015. Document modeling with gated recurrent neural network for sentiment classification. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1422–1432. Mark Torrance. 2015. Understanding planning in text production. Handbook of writing research, pages 1682–1690. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 6000–6010. Yizhong Wang, Sujian Li, and Houfeng Wang. 2017. A two-stage parsing method for text-level discourse analysis. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 184–188. Wen Xiao, Patrick Huber, and Giuseppe Carenini. 2020. Do we really need that many parameters in transformer for extractive summarization? discourse can help! In Proceedings of the First Workshop on Computational Approaches to Discourse, pages 124– 134. Wen Xiao, Patrick Huber, and Giuseppe Carenini. 2021. Predicting discourse trees from transformer-based neural summarizers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4139–4152, Online. Association for Computational Linguistics. Jiacheng Xu, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020. Discourse-aware neural extractive text summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5021–5031. Association for Computational Linguistics. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical attention networks for document classification. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 1480–1489. Nan Yu, Meishan Zhang, and Guohong Fu. 2018. Transition-based neural rst parsing with implicit syntax features. In Proceedings of the 27th International Conference on Computational Linguistics, pages 559–570. Amir Zeldes. 2017. The GUM corpus: Creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581–612. A Numeric Results The numeric results of our W-RST approach for the sentiment analysis and summarization downstream tasks presented in Figure 6 are shown in Table 4 below, along with the threshold-based approach, as well as the supervised parser. Approach Sentiment Summarization Accuracy R-1 R-2 R-L Nuclearity with Threshold t = 0.0 53.76 28.22 8.58 26.45 t = 0.1 53.93 28.41 8.69 26.61 t = 0.2 54.13 28.64 8.85 26.83 t = 0.3 54.33 28.96 9.08 27.14 t = 0.4 54.44 29.36 9.34 27.51 t = 0.5 54.79 29.55 9.50 27.68 t = 0.6 54.99 29.78 9.65 27.90 t = 0.7 55.07 29.57 9.45 27.74 t = 0.8 55.32 29.18 9.08 27.32 t = 0.9 54.90 28.11 8.29 26.35 t = 1.0 54.15 26.94 7.60 25.27 Our Weighted RST Framework weighted 54.76 29.70 9.58 27.85 Supervised Training on RST-DT supervised 44.77 34.20 12.77 32.09 Table 4: Results of the W-RST approach compared to threshold-based nuclearity assignments and supervised training on RST-DT.
|
2021
|
302
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3919–3931 August 1–6, 2021. ©2021 Association for Computational Linguistics 3919 ABCD: A Graph Framework to Convert Complex Sentences to a Covering Set of Simple Sentences Yanjun Gao and Ting-Hao (Kenneth) Huang and Rebecca J. Passonneau Pennsylvania State University {yug125,txh710,rjp49}@psu.edu Abstract Atomic clauses are fundamental text units for understanding complex sentences. Identifying the atomic sentences within complex sentences is important for applications such as summarization, argument mining, discourse analysis, discourse parsing, and question answering. Previous work mainly relies on rulebased methods dependent on parsing. We propose a new task to decompose each complex sentence into simple sentences derived from the tensed clauses in the source, and a novel problem formulation as a graph edit task. Our neural model learns to Accept, Break, Copy or Drop elements of a graph that combines word adjacency and grammatical dependencies. The full processing pipeline includes modules for graph construction, graph editing, and sentence generation from the output graph. We introduce DeSSE, a new dataset designed to train and evaluate complex sentence decomposition, and MinWiki, a subset of MinWikiSplit. ABCD achieves comparable performance as two parsing baselines on MinWiki. On DeSSE, which has a more even balance of complex sentence types, our model achieves higher accuracy on the number of atomic sentences than an encoder-decoder baseline. Results include a detailed error analysis. 1 Introduction Atomic clauses are fundamental text units for understanding complex sentences. The ability to decompose complex sentences facilitates research that aims to identify, rank or relate distinct predications, such as content selection in summarization (Fang et al., 2016; Peyrard and Eckle-Kohler, 2017), labeling argumentative discourse units in argument mining (Jo et al., 2019) or elementary discourse units in discourse analysis (Mann and Thompson, 1986; Burstein et al., 1998; Demir et al., 2010), or extracting atomic propositions for question answering (Pyatkin et al., 2020). In this work, Orig Sokuhi was born in Fujian and was ordained at 17. SS1 Sokuhi was born in Fujian. SS2 Sokuhi was ordained at 17. Figure 1: Example of a complex sentence (Orig) rewritten as two simple sentences (SS1, SS2). Underlined words in the source are preserved in the same order in the two outputs, the conjunction and (red font) is dropped, and the subject Sokuhi (blue font) is copied to the second simple sentence. we propose a new task to decompose complex sentences into a covering set of simple sentences, with one simple output sentence per tensed clause in the source sentence. We focus on tensed clauses rather than other constituents because they are syntactically and semantically more prominent, thus more essential in downstream tasks like argument mining, summarization, and question answering. The complex sentence decomposition task we address has some overlap with related NLP algorithms, but each falls short in one or more respects. Elementary discourse unit (EDU) segmentation segments source sentences into a sequence of non-overlapping spans (Carlson et al., 2003; Wang et al., 2018). The output EDUs, however, are not always complete clauses. Text simplification rewrites complex sentences using simpler vocabulary and syntax (Zhang and Lapata, 2017). The output, however, does not preserve every tensed clause in the original sentence. The split-and-rephrase (SPRP) task aims to rewrite complex sentences into sets of shorter sentences, where an output sentence can be derived from non-clausal constituents in the source (Narayan et al., 2017). In contrast to the preceding methods, we convert each tensed clause in a source sentence, including each conjunct in a conjoined VP, into an independent simple sentence. Unlike EDU segmentation, a belief verb and its that-complement do not lead to two output units. Unlike text simplification, no propositions in the source are omitted from the output. Unlike SPRP, a phrase that lacks a tensed verb in the source cannot 3920 lead to a distinct sentence in the output. Figure 1 shows an example complex sentence (Orig) with conjoined verb phrases and its rewrite into two simple sentences (SSs). Observe that besides producing two sentences from one, thus breaking the adjacency between words, words inside the verb phrases (underlined in the figure) remain in the same linear order in the output; the single subject Sokuhi in the source is copied to the more distant verb phrase. Finally, the connective and is dropped. We find that most rewrites of complex sentences into simple sentences that preserve the one-to-one mapping of source tensed predication with target simple sentence involve similar operations. Building on these observations, we propose a neural model that learns to Accept, Break, Copy or Drop elements of a special-purpose sentence graph that represents word adjacency and grammatical dependencies, so the model can learn based on both kinds of graph proximity. We also introduce DeSSE (Decomposed Sentences from Students Essays), a new annotated dataset to support our task. The rest of the paper presents two evaluation datasets, our full pipeline, and our ABCD model. Experimental results show that ABCD achieves comparable or better performance than baselines. 1 2 Related Work Related work falls largely into parsing-based methods, neural models that rewrite, and neural segmenters. Gao et al. (2019) propose a decomposition parser (DCP) that extracts VP constituents and clauses from complex sentences as part of a summarization evaluation tool. Niklaus et al. (2019a) present a system (DisSim) based on parsing to extract simple sentences from complex ones. Jo et al. (2020) propose seven rules to extract complete propositions from parses of complex questions and imperatives for argumentation mining. Though performance of these methods depends on parser quality, they often achieve very good performance. We include two whose code is available (DCP, DisSim) among our baselines. SPRP models are based on encoder-decoder architectures, and the output is highly depending on the training corpus. Aharoni and Goldberg (2018) present a Copy-augmented network (Copy512) based on (Gu et al., 2016) that encour1ABCD is available at https://github.com/ serenayj/ABCD-ACL2021. ages the model to copy most words from the original sentence to the output. As it achieves improvement over an earlier encoder-decoder SPRP model (Narayan et al., 2017), we include Copy512 among our baselines. Finally, recent neural EDU segmenters (Wang et al., 2018; Li et al., 2018) achieve state-of-the-art performance on a discourse relation corpus, RSTDT (Carlson et al., 2003). As they do not output complete sentences, we do not include any among our baselines. Our ABCD model leverages the detailed information captured by parsing methods, and the powerful representation learning of neural models. As part of a larger pipeline that converts input sentences to graphs, ABCD learns to predict graph edits for a post processor to execute. 3 Datasets Here we present DeSSE, a corpus we collected for our task, and MinWiki, a modification of an existing SPRP corpus (MinWikiSplit (Niklaus et al., 2019b)) to support our aims. We also give a brief description of differences in their distributions. Neural models are heavily biased by the distributions in their training data (Niven and Kao, 2019), and we show that DeSSE has a more even balance of linguistic phenomena. 3.1 DeSSE DeSSE is collected in an undergraduate social science class, where students watched video clips about race relations, and wrote essays in a blog environment to share their opinions with the class. It was created to support analysis of student writing, so that different kinds of feedback mechanisms can be developed regarding sentence organization. Students have difficulty with revision to address lack of clarity in their writing (Kuhn et al., 2016), such as non-specific uses of connectives, run on sentences, repetitive statements and the like. These make DeSSE different from corpus with expert written text, such as Wikipedia and newspaper. The annotation process is unique in that it involves identifying where to split a source complex sentence into distinct clauses, and how to rephrase each resulting segment as a semantically complete simple sentence, omitting any discourse connectives. It differs from corpora that identify discourse units within sentences, such as RST-DT (Carlson et al., 2003) and PTDB (Prasad et al., 2008), because 3921 • Orig: (I believe that talking about race more in a civil way can only improve our society), || but I can see why other people may have a different opinion. • Rephrase 1: I believe that talking about race more in a civil way can only improve our society. • Rephrase 2: I can see why other people may have a different opinion. Figure 2: An original sentences from DeSSE with an intrasentential connective (but), a verb that takes a clausal argument. The annotation first splits the sentence (at ||), then rephrases each segment into a simple sentence, dropping the connective. Dataset Disc. VPWh- & Restric. thatConn. Conj. Rel. Cl. Rel. Cl. comp. MinWiki 58% 36% 10% 26% 0% DeSSE 66% 22% 32% 34% 24% Table 1: Prevalence of five linguistic phenomena in 50 randomly selected examples per dataset. Categories are not mutually exclusive. clauses are explicitly rewritten as simple sentences. It differs from split-and-rephrase corpora such as MinWikiSplit, because of the focus in DeSSE on rephrased simple sentences that have a one-to-one correspondence to tensed clauses in the original complex sentence. DeSSE is also used for connective prediction tasks, as in (Gao et al., 2021).2 We perform our task on Amazon Mechanical Turk (AMT). In a series of pilot tasks on AMT, we iteratively designed annotation instructions and an annotation interface, while monitoring quality. Figure 2 illustrates two steps in the annotation: identification of n split points between tensed clauses, and rephrasing the source into n+1 simple clauses, where any connectives are dropped. The instructions ask annotators to focus on tensed clauses occurring in conjoined or subordinate structures, relative clauses, parentheticals, and conjoined verb phrases, and to exclude gerundive phrases, infintival clauses, and clausal arguments of verbs. The final version of the instructions describes the two annotation steps, provides a list of connectives, and illustrates a positive and negative example.3 The training and tests sets contains 12K and 790 examples, respectively. 3.2 MinWikiSplit MinWikiSplit has 203K complex sentences and their rephrased versions (Niklaus et al., 2019b). 2DeSSE and MinWiki are available at https:// github.com/serenayj/DeSSE. 3In step 2, the interface checked for any remaining connectives, to warn annotators. Details about the interface and quality control are included in appendix A. It is built from WikiSplit, a text simplification dataset derived from Wikipedia revision histories (Narayan et al., 2017), modified to focus on minimal propositions that cannot be further decomposed. It was designed for simplifying complex sentences into multiple simple sentences, where the simple sentences can correspond to a very wide range of structures from the source sentences, such as prepositional or adjectival phrases. To best utilize this corpus for our purposes, we selected a subsample where the number of tensed verb phrases in the source sentences matches the number of rephrased propositions. The resulting MinWiki corpus has an 18K/1,075 train/test split. 3.3 Linguistic phenomena Table 1 presents prevalence of syntactic patterns characterizing complex sentences in the two datasets. Four are positive examples of one-to-one correspondence of tensed clauses in the source with simple sentences in the rephrasings: discourse connectives (Disc. Conn.), VP-conjunction, clauses introduced by wh- subordinating conjunctions (e.g., when, whether, how) combined with non-restrictive relative clauses (wh- & Rel. Cl.), and restrictive relative clauses (Restric. Rel. Cl.). The sixth column (negative examples) covers clausal arguments, which are often that-complements of verbs that express belief, speaking, attitude, emotion, and so on. MinWiki has few of the latter, presumably due to the genre difference between opinion essays (DeSSE) and Wikipedia (MinWiki). 4 Problem Formulation We formulate the problem of converting complex sentences into covering sets of simple sentences as a graph segmentation problem. Each sentence is represented as a Word Relation Graph (WRG), a directed graph constructed from each input sentence with its dependency parse. Every word token and its positional index becomes a WRG vertex. For every pair of words, one or more edges are added as follows: a neighbor edge that indicates that the pair of words are linearly adjacent; a dependency edge that shows every pair of words connected by a dependency relation, adding critical grammatical relations, such as subject. Figure 3 shows an example sentence and a simplified version of its WRG (edge directions are not shown, for readability). Vertices are labeled with word-index pairs in red font, and edges are labeled 3922 Figure 3: Example complex sentence (Orig), ground truth output (SS 1 and SS 2), and WRG (best seen in color; edge directions and punctuation omitted for readability). Vertices are word tokens and their indices, edges are neighbor (ngbh) and/or dependency relations. Dashed lines represent edges to Break, the green curved line represents an edge to Copy, the open circle node for and-6 is for Drop, and all other parts of the graph get Accept. At bottom left is a fragment of the corresponding Edge Triple Set. as ngbh for neighboring words, or with the tags corresponding to their dependency relations, such as nsubj between Sokuhi-1 and ordained-13. An edge can have both types of relation, e.g. neighbor and dependency for was-12 and ordained-13. The graph is stored as an Edge Triple Set, a set of triples with (source node, target node, label) representing each pair of words connected by an edge, as shown in Figure 3, bottom left. Given a sentence and its WRG, our goal is to decompose the graph into n connected components (CC) where each CC is later rewritten as an output simple sentence. To perform the graph decomposition, decisions are made on every edge triple.We define four edit types: • Accept: retain the triple in the output • Break: break the edge between a pair of words • Copy: copy a target word into a CC • Drop: delete the word from the output CCs A training example consists of an input sentence, and one or more output sentences. If the input sentence is complex, the ground truth output consists of multiple simple sentences. The next section presents the ABCD pipeline. Two initial modules construct the WRG graphs for each input sentence, and the ABCD labels for the Edge Triple Sets based on the ground truth output. A neural model learns to assign ABCD labels to input WRG graphs, and a final graph segmenter generates simple sentences from the labeled WRG graphs. Details about the neural model are in the subsequent section. 5 System Overview The full processing pipeline consists of five major components, as shown in Figure 4. Three preprocessing modules handle the WRG graph construction, conversion of graph triples to vectors, and creation of distant supervision labels for the graph. The fourth component is the ABCD neural model that learns to label a WRG graph, which is described in section 6. The last part of the pipeline is a post-processing module to segment WRG graphs based on the labels learned by the ABCD model, and to map each graph segment to a simple sentence. Graph Constructor The first module in the system is a Graph Constructor that converts an input sentence and its dependency parse into a collection of vertices and edges. It is used during training and inference. It first extracts words and their indices from the input sentences of the training examples for the vertices of each WRG graph. A directed edge and ngbh label is assigned to all pairs of adjacent words. A directed edge and label is also assigned to every governing and dependent word pair (cf. Figure 3). Edge Triples DB The Edge Triples DB, which is used during training and inference, creates vector representations for the input Edge Triples Sets for each training instance, using latent representations learned by an encoder component of the ABCD model. Using the word indices, a function maps the source and target words from every triple into its hidden representation learned by the encoder, and the triple’s edge label is converted into a one-hot encoding with dimension d. For an edge triples set with m triples, the source and target word hidden states are each stacked into an m × h matrix, and the one-hot vectors for edge labels are stacked into an m × d matrix. These three source, target, edge matrices that represent an Edge Triple Set are then fed into an attention layer, as discussed in section 6. Distant Supervision Label Creator The expected supervision for our task is the choice of edit type for each triple, where the ground truth consists of pairs of an input sentence, and one or more output simple sentences. We use distant supervision where we automatically create edit labels for each triple based on the alignment between the original input sentence and the set of output simple sentences. In the Distant Supervision Label Creator 3923 Figure 4: ABCD system overview during training (top) and inference (bottom). module, for every triple, we check the following conditions: if the edge is a ”neighbor” relation, and both source and target words are in the same output simple sentence, we mark this pair with edit type A; if the source and target words of a triple occur in different output simple sentences, the corresponding edit is B; if the source and target are in the same output simple sentence, and the only edge is a dependency label (meaning that they are not adjacent in the original sentence), we mark this pair as C; finally, if a word is not in any output simple sentence, we mark the corresponding type as D. Graph Segmenter This module segments the graph into connected components using predicted edits, and generates the output sentences, as part of the inference pipeline. There are four stages consisting of: graph segmentation, traversal, subject copying, and output rearranging. In the graph segmentation stage, the module first performs actions on every triple per the predicted edit: if the edit is A, no action is taken; if the edit is B, the edge between the pair of words is dropped; given C, the edge is dropped, and the edge triple is stored in a temporary list for later retrieval; if the edit is D, the target word is dropped from the output graphs. After carrying out the predicted edits, we run a graph traversal algorithm on modified edge triples to find all CCs, using a modified version of the Depth-First-Search algorithm with linear time proposed in (Tarjan, 1972; Nuutila and SoisalonSoininen, 1994). For each CC, the vertices are kept and the edges are dropped. Then we enter the subject copying stage: for each source, target pair in the temporary list mentioned earlier, we copy the word to the CC containing the target. Finally for every CC, we arrange all words in their linear order by indices, and output a simple sentence. Figure 5: Architecture for ABCD model. 6 Neural Model The ABCD model consists of three neural modules depicted in Figure 5: a sentence encoder to learn a hidden representation for the input sentence, a self-attention layer to generate attention scores on every edge label, and a classifier that generates a predicted distribution over the four edit types, based on the word’s hidden representation, the edge label representation, and the attention scores. 6.1 Sentence Representation The sentence representation module has two components: a word embedding look up layer based on GloVe (Pennington et al., 2014), and a bidirectional LSTM (Hochreiter and Schmidhuber, 1997) (see Figure 5). Given an input sentence length l, and the hidden state dimension M, the output from this module is l × M. For a word with index i in the input sentence, we generate its hidden representation hi such that it combines the hidden states from forward and backward LSTMs, with 3924 hi ∈RM. A positional encoding function is added to the word embeddings. We found this particularly helpful in our task, presumably because the same word type at different positions might have different relations with other words, captured by distinct learned representations. Our experiments compare biLSTM training from scratch to use of BERT (Devlin et al., 2019), to see if pre-trained representations are helpful. To utilize the learned word representations in the context of the relational information captured in the WRG graph, we send the sentence representation to the Edge Triple DB and extract representations hi and hj for the source and target words, based on indices i and j. A one-hot vector with dimensionality N encodes relations between pairs of source and target words; each edge triple is thus converted into three vectors: hsrc, htgt, drel. We take positionwise summation over all one hot vectors if there is more than one label on an edge. 6.2 Edge Self-Attention Attention has been useful for many NLP tasks. In our model, we adapt the multi-head self attention mechanism (Vaswani et al., 2017) to learn importance weights on types of edit operations, as shown in the middle green block in Figure 5. Given m edge triples, we first stack all source vectors hsrc into a matrix Hsrc, and operate the same way on htgt and drel to obtain Htgt and Drel, such that Hsrc, Htgt ∈Rm×M, and Drel ∈Rm×N. These three matrices are the input to self-attention. For every head of the multi-head attention, we first obtain a feature representation with the three parameters V, K, Q mapping to sources, targets and relations, respectively, then compute a co-efficient e with a learnable parameter W e as follows: e = LeakyRelu(W e(V Hsrc; KHtgt; QDrel)) (1) where e ∈Rm×1. Then we compute the attention scores by taking a softmax over e: head = softmax(e) (2) Finally, we concatenate all head attentions together, and pass them through a linear layer to learn the relations between heads, and generate the final attention scores: α = W(concat((head1, head2, . . .)) (3) α ∈Rm×1. The attention scores are sent to the next module to help the classifier make its decision. Dataset A B C D MinWiki 85.23% 4.58% 3.60% 6.57% DeSSE 74.77% 2.39% 5.62% 17.21% MinWiki 0.0167 0.3533 0.4164 0.2135 DeSSE 0.0200 0.6266 0.2658 0.0876 Table 2: Distributions (Top) and inverse class weights (Bottom) for the four edit labels on both MinWiki and DeSSE datasets. 6.3 Edit Classification The last component of our neural model is a classifier, as shown at the right of Figure 5. To aggregate the feature representation from the previous layer, we first concatenate the three matrices Hsrc, Htgt, Drel into one representation, and multiply the attention scores as follows: H′ = α(Hsrc; Htgt, Drel) (4) An MLP layer then takes H′ as its input and generates the output distribution over the four edit types for each edge triple: OutM = Softmax(MLP(H′)) (5) where OutM ∈Rm×4. As an alternative to MLP, we also investigated a bilinear classifier, which has proved efficient in capturing fine-grained differences in features for classification task (Dozat and Manning, 2017). The bilinear layer first takes Hsrc and Htgt as input and generates transposed bilinear features : outputbi = H⊺ srcW AHtgt + b (6) where W A, b are learnable parameters. Then we sum the bilinear features with the MLP decisions and apply softmax on the result to get the final distribution over the four edit labels: OutB = Softmax(outputbi + MLP(H′)) (7) where OutB ∈Rm×4. We use cross entropy loss between predicted edit types and gold edit types created from distant supervision (see above). 6.4 Training The class balance for our task is highly skewed: the frequency of class A is much higher than the other three classes, as shown in the top portion of Table 2. To mitigate the impact on training, we adopt the inverse class weighting for cross entropy loss introduced in (Huang et al., 2016). With this weighting, loss is weighted heavily towards rare classes, which forces the model to learn more about 3925 the rare cases. Table 2 shows the weights for four edit labels on both datasets. On MinWiki, A occurs the most and has the lowest weights as 0.0167, a sharp contrast to B,C,D. On DeSSE, both A and D occur frequently while B and C have lower frequency with higher weights, at 0.6266 and 0.2658. DeSSE has fewer B, and more C and D than MinWiki. From this perspective, MinWiki is “simpler” than DeSSE because there are fewer edits on rewriting the sentences. This might be due to the different distributions of linguistic phenomena in the two datasets (see Table 1). In the next section, we will show that ABCD shows stronger improvements on complicated edits. Training details are in the appendix. 7 Experiments We carry out two intrinsic evaluations of ABCD performance on MinWiki and DeSSE. Section 7.1 presents an intrinsic evaluation of ABCD variants on edit prediction, with error analysis and ablation studies. Section 7.2 compares the best ABCD model with several baselines on the quality of output propositions. We discuss evaluation metrics in section 7.3. Results show that ABCD models show consistently good performance compared to other baseline models on both datasets. 7.1 Intrinsic Evaluation on Edit Prediction We report F1 scores on all four edit types from ABCD and its model variants. We compare two classifiers as mentioned in previous sections and investigate the difference between using biLSTM and BERT with fine-tuning, to see if pre-trained knowledge is useful for the task. Table 3 presents results on MinWiki and DeSSE from the four model settings. All models perform better on MinWiki than DeSSE, and biLSTM+bilinear shows the best performance on both, with F1 scores of 0.82 and 0.67 on MinWiki and DeSSE respectively. Presumably this reflects the greater linguistic diversity of DeSSE shown in Table 1. The lower performance from BERT variants indicates the pre-trained knowledge is not helpful. Among the four edit types, all models have high F1 scores on A across datasets, high F1 on C for MinWiki, but not on DeSSE. B and D show lower scores, and all four models report lower F1 on B than D on both datasets. To examine the significant drop on B and D from MinWiki to DeSSE, Table 4 presents error analysis on pairs of gold labels and predictions for B and D, using predictions from biLSTM+mlp. The model does poorly on B in both datasets, compared with predictions of 36.1% for A on MinWiki, on on DeSSE, 27.42% for A and 15.18% for C. The model has high agreement on D from MinWiki, but predicts 42.63% A on DeSSE. We suspect that improved feature representation could raise performance; that is, pairs of words and their relations might be a weak supervision signal for B and D. We conducted an ablation study on the inverse class weights mentioned in section 6 on MinWiki. After removing the weights, the model fails to learn other classes and only predicts A due to the highly imbalanced label distributions, which demonstrates the benefit of weighting the loss function. We also ablate positional encoding which leads to F1 scores of 0.90 for A, 0.51 for C, and 0 for both B and D, indicating the importance of positional encoding. 7.2 Intrinsic Evaluation of Output Sentences For baselines, we use Copy512 and DisSim, which both report performance on Wikisplit in previous work. We also include DCP, which relies on three rules applied to token-aligned dependency and constituency parses: DCPvp extracts clauses with tensed verb phrases; DCPsbar extracts SBAR subtrees from constituency trees; DCPrecur recursively applies the preceding rules. For evaluation, we use BLEU with four-grams (BL4) (Papineni et al., 2002) and BERTScore (BS) (Zhang et al., 2019). We also include descriptive measures specific to our task. To indicate whether a model retains roughly the same number of words as the source sentence in the target output, we report average number of tokens per simple sentence (#T/SS). To capture the correspondence between the number of target simple sentences in the ground truth and model predictions, we use percentage of samples where the model predicts the correct number of simple sentences (Match #SS). BL4 captures the 4-gram alignments between candidate and reference word strings, but fails to assess similarity of latent meaning. BS applies token-level matching through contextualized word embeddings, therefore evaluates candidates on their word meanings. For each example, we first align each simple sentence in the ground truth with a prediction, compute the pairwise BL4 and BS scores, and take the average as the score for the example. A predicted output sentence with no 3926 Category MinWiki DeSSE biLSTM BERT biLSTM BERT mlp bilinear mlp bilinear mlp bilinear mlp bilinear A 0.98 0.98 0.93 0.86 0.91 0.88 0.88 0.87 B 0.48 0.48 0.41 0.36 0.34 0.42 0.31 0.28 C 0.99 0.99 0.95 0.98 0.89 0.78 0.89 0.55 D 0.80 0.84 0.39 0.75 0.49 0.54 0.45 0.45 All 0.78 0.82 0.72 0.74 0.66 0.67 0.63 0.57 Table 3: Performance (F1) of our model and its variants on MinWiki (N=1075) and DeSSE (N=790). Data Gold Predicted A B C D Minwiki B 36.10 48.33 5.59 9.98 D 14.01 0.14 0.46 85.38 DeSSE B 27.42 46.62 15.18 10.76 D 42.63 3.44 5.08 48.84 Table 4: Percentage (%) of count of predicted labels where gold labels are B and D from ABCD biLSTM+mlp. correspondent in the ground truth, or a ground truth sentence with no correspondent in the predicted, will add 0 to the numerator and 1 to the denominator of this average. Table 5 presents results from the baselines and our ABCD best variant, biLSTM with two classifiers. None of the models surpasses all others on both datasets. All models show lower performance on DeSSE than MinWiki, again an indication that DeSSE is more challenging. On MinWiki, ABCD is competitive with Copy512, the best performing model, with a narrow gap on Match#SS (0.65%) and BLEU4 (4.58). On DeSSE, ABCD BL4 and BS surpass all baselines. ABCD performance is 2.34% less than DCPrecur on Match #SS, but biLSTM+mlp output sentences have an average length of 8.85, which is closer to the gold average length of 9.07, in contrast to much longer output from DCPrecur of 14.16. To summarize, ABCD achieves competitive results on both datasets. 7.3 Error Analysis While Table 4 presents error analysis on predictions of B that lead to an incorrect number of outputs, here we examine test sentences from both datasets where the prediction and ground truth have the same number of outputs. Table 6 shows the total number of examples for MinWiki (1,075) and for the positive examples in DeSSE (DeSSEpos, 521). The M columns for each dataset give the number of examples where the number of targets in the ground truth matches the number of targets predicted by the model. On MinWiki, ABCD has marginally better BL4 and BS scores than Copy512, but Copy512 has 7 more cases with the correct number of outputs. For DeSSE, we restrict attention to the positive examples (MinWiki has no negative examples), because Copy512 and ABCD perform equally well on the negative examples. By the BL4 and BS scores on DeSSEpos, Copy512 appears to perform much better than ABCD, but these scores are on 20 out of 521 examples (3.8%). Although ABCD’s scores are lower, it produces the correct number of output sentences in 47.4% of cases for the mlp, and 48.1% for the bilin. Figure 6 shows three complex sentences from DeSSE with the annotated rewriting, and predicted propositions from Copy512 and ABCD mlp. Copy512 correctly decomposes only one of the examples and copies the original input on the other two samples. On the one example where Copy produces two simple sentences, it alters the sentence meaning by replacing the word “genetics” with the word “interesting”. This exposes a drawback of encoder-decoder architectures on the proposition identification task, that is, the decoder can introduce words that are not in the input sentence, therefore failing to preserve the original meaning. In contrast, ABCD shows good performance on all three sentences by producing the same number of simple sentences as in the annotated rewriting. Especially for the third sentence, which contains an embedded clause, “which has been the main mission since 9/11”, the first proposition written by the annotator is not grammatically correct, and the subject of the second proposition is a pronoun it, referring to the semantic subject Our main mission. Nonetheless, ABCD generates two propositions, both of which are grammatically correct and meaning preserving. 8 Discussion In this section, we discuss limitations of ABCD to guide future work. The first limitation is the low performance of ABCD on B. We observe that in DeSSE, some annotators did not break the sentences appropriately. We randomly selected 50 samples, and found 13 out of 50 (26%) examples 3927 Group Model MinWiki DeSSE #T Match BLEU4 BERTSc #T Match BLEU4 BERTSc /SS #SS(%) /SS #SS(%) Parsing DisSim 8.50 68.46 64.20 94.42 9.59 40.00 37.89 89.54 DCPvp 14.82 45.49 28.80 64.50 15.99 42.40 47.25 60.18 DCPsbar 19.07 17.49 19.35 49.07 17.24 44.81 48.02 59.89 DCPrecur 16.30 67.90 31.78 58.08 14.16 55.63 34.44 61.37 Encoder-decoder COPY 9.37 79.26 80.96 95.96 18.13 36.20 45.91 88.71 ABCD biLSTM mlp 9.37 78.61 75.80 92.91 8.85 53.29 53.42 90.23 bilin 9.53 76.72 76.38 90.28 8.10 52.66 41.57 94.78 Table 5: Performance of baselines and our models on Minwiki test set (N=1075, #T/SS = 10.03), and DeSSE test set (N=790, #T/SS =9.07). We report numbers of token per propositions (#T/SS), number of input sentences that have match number of output between prediction and ground truth in percentage (Match #SS%), BLEU with four-gram and BERTScore. Orig He did not do anything wrong, yet he was targeted and his family was murdered. Human He did not do anything wrong. || He was targeted. || His family was murdered. Copy He did not do anything wrong, he was targeted and his family was murdered. ABCD He did not do anything wrong.|| he was targeted. || his family was murdered. Orig I guess I always knew it was genetics but I didnt know why our features are the way that they are. Human I guess I always knew it was genetics. || I didnt know why our features are the way that they are. Copy I guess I always knew it was interesting.|| I didnt know why our features are the way that they are. ABCD I guess I always knew it was genetics.|| I didnt know why our features are the way that they are. Orig Our main mission, which has been the main mission since 9/11 is to eliminate terrorism wherever it may exist. Human Our main mission, which has been the main mission since 9/11.|| It is to eliminate terrorism wherever it may exist. Copy same as Orig ABCD Our main mission has been the main mission. || mission is to eliminate terrorism wherever it may exist. Figure 6: Three input complex sentences (Orig) from DeSSE, with the annotated rewriting (Human), and the predicted propositions from Copy and ABCD. MinWiki (N=1075) DeSSEpos(N=521) M BL4 BS M BL4 BS Copy 852 88.81 97.16 20 92.48 98.66 mlp 845 89.59 97.21 247 78.49 95.73 bilin 825 92.00 96.94 251 74.25 98.21 Table 6: Performance of Copy512 and our ABCD biLSTM models on all positive samples from MinWiki and DeSSE test set. Columns show the raw count of complex sentences where prediction has correct number of outputs (M), BL4 and BS. where annotators add breaks to rewrite NPs and infinitives as clauses. This introduces noise into the data. Another reason of lower performance on B might be attributed to the current design of ABCD that neglects sequential relations among all words. Among all edge triples where it fails to assign B, 67% and 27.42% are with ngbh relations on MinWiki and DeSSE, respectively. Two possibilities for improving performance to investigate are enhancements to the information in the WRG graph, and re-formulating the problem into sequence labeling of triples. The second limitation pertains mainly to DeSSE. In the training data, 34.7% of sentences have OOV words. For example, we noticed that annotators sometimes introduced personal pronouns (e.g.he/she/they) in their rewrites of VPconjunction, instead of copying the subjects, or they substituted a demonstrative pronoun (e.g.this/these) for clausal arguments. This could be addressed by expanding the edit types to include the ability to INSERT words from a restricted insertion vocabulary. Nevertheless, our model has a small performance gap with Copy512 on MinWiki, and outperforms the baselines on DeSSE. A third issue is whether ABCD would generalize to other languages. We expect ABCD would perform well on European languages with existing dependency and constituency parsers, and with an annotated dataset. 9 Conclusion We presented a new task to decompose complex sentences into simple ones, along with DeSSE, a new dataset designed for this task. We proposed the neural ABCD model to predict four edits operations on sentence graphs, as part of a larger pipeline from our graph-edit problem formulation. ABCD performance comes close to or outperforms the parsing-based and encoder-decoder baselines. Our work selectively integrates modules to capitalize on the linguistic precision of parsing-based methods, and the expressiveness of graphs for encoding different aspects of linguistic structure, while still capitalizing on the power of neural networks for representation learning. 3928 References Roee Aharoni and Yoav Goldberg. 2018. Split and rephrase: Better evaluation and stronger baselines. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 719–724. Jill Burstein, Karen Kukich, Susanne Wolff, Chi Lu, and Martin Chodorow. 1998. Enriching automated essay scoring using discourse marking. In Discourse Relations and Discourse Markers. Lynn Carlson, Daniel Marcu, and Mary Ellen Okurowski. 2003. Building a discourse-tagged corpus in the framework of rhetorical structure theory. In Current and new directions in discourse and dialogue, pages 85–112. Springer. Seniz Demir, Sandra Carberry, and Kathleen F McCoy. 2010. A discourse-aware graph-based contentselection framework. In Proceedings of the 6th International Natural Language Generation Conference. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. Yimai Fang, Haoyue Zhu, Ewa Muszy´nska, Alexander Kuhnle, and Simone Teufel. 2016. A propositionbased abstractive summariser. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 567–578. Chris Fournier. 2013. Evaluating text segmentation using boundary edit distance. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1702–1712, Sofia, Bulgaria. Association for Computational Linguistics. Yanjun Gao, Ting-Hao Huang, and Rebecca J. Passonneau. 2021. Learning clause representation from dependency-anchor graph for connective prediction. In Proceedings of the Fifteenth Workshop on GraphBased Methods for Natural Language Processing (TextGraphs-15), pages 54–66, Mexico City, Mexico. Association for Computational Linguistics. Yanjun Gao, Chen Sun, and Rebecca J. Passonneau. 2019. Automated pyramid summarization evaluation. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 404–418, Hong Kong, China. Association for Computational Linguistics. Jiatao Gu, Zhengdong Lu, Hang Li, and Victor OK Li. 2016. Incorporating copying mechanism in sequence-to-sequence learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1631–1640. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Chen Huang, Yining Li, Chen Change Loy, and Xiaoou Tang. 2016. Learning deep representation for imbalanced classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5375–5384. Yohan Jo, Elijah Mayfield, Chris Reed, and Eduard Hovy. 2020. Machine-aided annotation for finegrained proposition types in argumentation. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 1008–1018. Yohan Jo, Jacky Visser, Chris Reed, and Eduard Hovy. 2019. A cascade model for proposition extraction in argumentation. In Proceedings of the 6th Workshop on Argument Mining, pages 11–24, Florence, Italy. Association for Computational Linguistics. Deanna Kuhn, Laura Hemberger, and Valerie Khait. 2016. Tracing the development of argumentive writing in a discourse-rich context. Written Communication, 33(1):92–121. Jing Li, Aixin Sun, and Shafiq Joty. 2018. SegBot: a generic neural text segmentation model with pointer network. In Proceedings of the 27th International Joint Conference on Artificial Intelligence (IJCAI), pages 4166–4172. William C Mann and Sandra A Thompson. 1986. Relational propositions in discourse. Discourse processes, 9(1):57–90. Shashi Narayan, Claire Gardent, Shay Cohen, and Anastasia Shimorina. 2017. Split and rephrase. In EMNLP 2017: Conference on Empirical Methods in Natural Language Processing, pages 617–627. Christina Niklaus, Matthias Cetto, Andr´e Freitas, and Siegfried Handschuh. 2019a. DisSim: A discourseaware syntactic text simplification framework for English and German. In Proceedings of the 12th International Conference on Natural Language Generation, pages 504–507, Tokyo, Japan. Association for Computational Linguistics. 3929 Christina Niklaus, Andr´e Freitas, and Siegfried Handschuh. 2019b. MinWikiSplit: A sentence splitting corpus with minimal propositions. In Proceedings of the 12th International Conference on Natural Language Generation, pages 118–123, Tokyo, Japan. Association for Computational Linguistics. Timothy Niven and Hung-Yu Kao. 2019. Probing neural network comprehension of natural language arguments. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4658–4664. Esko Nuutila and Eljas Soisalon-Soininen. 1994. On finding the strongly connected components in a directed graph. Information processing letters, 49(1):9–14. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Maxime Peyrard and Judith Eckle-Kohler. 2017. Supervised learning of automatic pyramid for optimization-based multi-document summarization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1084–1094. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind K Joshi, and Bonnie L Webber. 2008. The penn discourse treebank 2.0. In LREC. Citeseer. Valentina Pyatkin, Ayal Klein, Reut Tsarfaty, and Ido Dagan. 2020. Qadiscourse-discourse relations as qa pairs: Representation, crowdsourcing and baselines. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2804–2819. Robert Tarjan. 1972. Depth-first search and linear graph algorithms. SIAM journal on computing, 1(2):146–160. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. Yizhong Wang, Sujian Li, and Jingfeng Yang. 2018. Toward fast and accurate neural discourse segmentation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 962–967. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. Xingxing Zhang and Mirella Lapata. 2017. Sentence simplification with deep reinforcement learning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 584–594. 3930 A Annotation Instruction in DeSSE Here we present the instructions for annotators, as shown by Figure 7. Figure 7: Instruction for DeSSE annotation The instructions illustrate the two phases of annotation. The annotator first chooses whether to add one or more split points to an input sentence, where the word after a split point represents the first word of a new segment. Once an annotator has identified the split points (first page of the AMT interface, shown as Figure 8), a second page of the interface appears. Figure 9 shows the second view when annotators rewrite the segments. Every span of words defined by split points (or the original sentence if no split points), appears in its own text entry box for the annotator to rewrite. Annotators cannot submit if they remove all the words from a text entry box. They are instructed to rewrite each text span as a complete sentence, and to leave out the discourse connectives. Several kinds of auto-checking and warnings are applied in the interface to ensure quality. If a rewrite contains a discourse connective, a warning box pops up asking if they should drop the discourse connective before submitting it. A warning box will show up if annotators use vocabulary outside the original sentence. To prevent annotators from failing to rewrite, we monitored the output, checking for cases where they submitted the text spans with no rewriting. Annotators were prohibited to submit if the interface detected an empty Figure 8: Interface of splitting the sentence Figure 9: Interface of rewriting the segments from Figure 8 into complete sentences rewrite box or the total lengths of the rewrites are too short compared to the source sentence. We warned annotators by email that if they failed to produce complete sentences in the rewrite boxes, they would be blocked. Some annotators were blocked, but most responded positively to the warnings. B Quality control in DeSSE To test the clarity of instruction and interface, the initial 500 sentences were used for evaluating the task quality, each labeled by three turkers (73 turkers overall), using three measures of consistency, all in [0,1]. Average pairwise boundary similarity (Fournier, 2013), a very conservative measure of whether annotators produce the same number of segments with boundaries at nearly the same locations, was 0.55. Percent agreement on number of output substrings was 0.80. On annotations with the same number of segments, we measured the average Jaccard score (ratio of set intersection to set union) of words in segments from different annotators, which was 0.88, and words from rephrasings, which was 0.73. With all metrics close to 1, and boundary similarity above 0.5, we concluded 3931 quality was already high. During the actual data collection, quality was higher because we monitored quality on a daily basis and communicated with turkers who had questions. C Experiment Settings We trained our model on a Linux machine with four Nvidia RTX 2080 Ti GPUs. We conducted grid search for the hyper-parameters, with learning rage in the range of [1e-2, 1e-5] (step size 0.0005), weight decay between [0.90, 0.99], hidden size [200, 800] (step size 200). Final parameters are set with Adam optimizer and learning rate at 1e −4, weight decay 0.99, embedding dropout at 0.2, maximum epoch as 100 with early stop. We use GloVe 100 dimension vectors, hidden size of network as 800. We set the number of heads in self-attention as 4, corresponding to the four edit types. With batch size 64, it takes about 6 hours to train MinWiki and 4 hours for DeSSE. For BERT fine-tuning, we use 1e −4 learning rate, weight decay at 0.99.
|
2021
|
303
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3932–3945 August 1–6, 2021. ©2021 Association for Computational Linguistics 3932 Which Linguist Invented the Lightbulb? Presupposition Verification for Question-Answering Najoung Kim†,∗, Ellie Pavlickφ,δ, Burcu Karagol Ayanδ, Deepak Ramachandranδ,∗ †Johns Hopkins University φBrown University δGoogle Research [email protected] {epavlick,burcuka,ramachandrand}@google.com Abstract Many Question-Answering (QA) datasets contain unanswerable questions, but their treatment in QA systems remains primitive. Our analysis of the Natural Questions (Kwiatkowski et al., 2019) dataset reveals that a substantial portion of unanswerable questions (∼21%) can be explained based on the presence of unverifiable presuppositions. Through a user preference study, we demonstrate that the oracle behavior of our proposed system—which provides responses based on presupposition failure—is preferred over the oracle behavior of existing QA systems. Then, we present a novel framework for implementing such a system in three steps: presupposition generation, presupposition verification, and explanation generation, reporting progress on each. Finally, we show that a simple modification of adding presuppositions and their verifiability to the input of a competitive end-to-end QA system yields modest gains in QA performance and unanswerability detection, demonstrating the promise of our approach. 1 Introduction Many Question-Answering (QA) datasets including Natural Questions (NQ) (Kwiatkowski et al., 2019) and SQuAD 2.0 (Rajpurkar et al., 2018) contain questions that are unanswerable. While unanswerable questions constitute a large part of existing QA datasets (e.g., 51% of NQ, 36% of SQuAD 2.0), their treatment remains primitive. That is, (closed-book) QA systems label these questions as Unanswerable without detailing why, as in (1): (1) a. Answerable Q: Who is the current monarch of the UK? System: Elizabeth II. ∗Corresponding authors, †Work done at Google b. Unanswerable Q: Who is the current monarch of France? System: Unanswerable. Unanswerability in QA arises due to a multitude of reasons including retrieval failure and malformed questions (Kwiatkowski et al., 2019). We focus on a subset of unanswerable questions—namely, questions containing failed presuppositions (background assumptions that need to be satisfied). Questions containing failed presuppositions do not receive satisfactory treatment in current QA. Under a setup that allows for Unanswerable as an answer (as in several closed-book QA systems; Figure 1, left), the best case scenario is that the system correctly identifies that a question is unanswerable and gives a generic, unsatisfactory response as in (1-b). Under a setup that does not allow for Unanswerable (e.g., open-domain QA), a system’s attempt to answer these questions results in an inaccurate accommodation of false presuppositions. For example, Google answers the question Which linguist invented the lightbulb? with Thomas Edison, and Bing answers the question When did Marie Curie discover Uranium? with 1896 (retrieved Jan 2021). These answers are clearly inappropriate, because answering these questions with any name or year endorses the false presuppositions Some linguist invented the lightbulb and Marie Curie discovered Uranium. Failures of this kind are extremely noticeable and have recently been highlighted by social media (Munroe, 2020), showing an outsized importance regardless of their effect on benchmark metrics. We propose a system that takes presuppositions into consideration through the following steps (Figure 1, right): 1. Presupposition generation: Which linguist invented the lightbulb? →Some linguist invented the lightbulb. 3933 QA Model Knowledge source Unanswerable QA Model Q: Which linguist invented the lightbulb? Knowledge source Unanswerable Presupposition Generation (S5.1) P: Some linguist invented the lightbulb Q: Which linguist invented the lightbulb? Verification (S5.2) Explanation Generator Explanation Generation (S5.3) End-to-End QA (S6) User study (S4) because ... Figure 1: A comparison of existing closed-book QA pipelines (left) and the proposed QA pipeline in this work (right). The gray part of the pipeline is only manually applied in this work to conduct headroom analysis. 2. Presupposition verification: Some linguist invented the lightbulb. →Not verifiable 3. Explanation generation: (Some linguist invented the lightbulb, Not verifiable) →This question is unanswerable because there is insufficient evidence that any linguist invented the lightbulb. Our contribution can be summarized as follows: • We identify a subset of unanswerable questions—questions with failed presuppositions—that are not handled well by existing QA systems, and quantify their role in naturally occurring questions through an analysis of the NQ dataset (S2, S3). • We outline how a better QA system could handle questions with failed presuppositions, and validate that the oracle behavior of this proposed system is more satisfactory to users than the oracle behavior of existing systems through a user preference study (S4). • We propose a novel framework for handling presuppositions in QA, breaking down the problem into three parts (see steps above), and evaluate progress on each (S5). We then integrate these steps end-to-end into a competitive QA model and achieve modest gains (S6). 2 Presuppositions Presuppositions are implicit assumptions of utterances that interlocutors take for granted. For example, if I uttered the sentence I love my hedgehog, it is assumed that I, the speaker, do in fact own a hedgehog. If I do not own one (hence the presupposition fails), uttering this sentence would be inappropriate. Questions may also be inappropriate in the same way when they contain failed presuppositions, as in the question Which linguist invented the lightbulb?. Presuppositions are often associated with specific words or syntactic constructions (‘triggers’). We compiled an initial list of presupposition triggers based on Levinson (1983: 181–184) and Van der Sandt (1992),1 and selected the following triggers based on their frequency in NQ (» means ‘presupposes’): • Question words (what, where, who...): Who did Jane talk to? » Jane talked to someone. • Definite article (the): I saw the cat » There exists some contextually salient, unique cat. • Factive verbs (discover, find out, prove...): I found out that Emma lied. » Emma lied. • Possessive ’s: She likes Fred’s sister. » Fred has a sister. • Temporal adjuncts (when, during, while...): I was walking when the murderer escaped from prison. » The murderer escaped from prison. • Counterfactuals (if + past): I would have been happier if I had a dog. » I don’t have a dog. Our work focuses on presuppositions of questions. We assume presuppositions project from 1We note that it is a simplifying view to treat all triggers under the banner of presupposition; see Karttunen (2016). 3934 Cause of unanswerability % Example Q Comment Unverifiable presupposition 30% what is the stock symbol for mars candy Presupposition ‘stock symbol for mars candy exists’ fails Reference resolution failure 9% what kind of vw jetta do i have The system does not know who ‘i’ is Retrieval failure 6% when did the salvation army come to australia Page retrieved was Safe Schools Coalition Australia Subjectivity 3% what is the perfect height for a model Requires subjective judgment Commonsensical 3% where does how to make an american quilt take place Document contains no evidence that the movie took place somewhere, but it is commonsensical that it did Actually answerable 8% when do other cultures celebrate the new year The question was actually answerable given the document Not a question/Malformed question 3% where do you go my lovely full version Not an actual question Table 1: Example causes of unanswerability in NQ. % denotes the percentage of questions that both annotators agreed to be in the respective cause categories. wh-questions—that is, presuppositions (other than the presupposition introduced by the interrogative form) remain constant under wh-questions as they do under negation (e.g., I don’t like my sister has the same possessive presupposition as I like my sister). However, the projection problem is complex; for instance, when embedded under other operators, presuppositions can be overtly denied (Levinson 1983: 194). See also Schlenker (2008), Abrusán (2011), Schwarz and Simonenko (2018), Theiler (2020), i.a., for discussions regarding projection patterns under wh-questions. We adopt the view of Strawson (1950) that definite descriptions presuppose both existence and (contextual) uniqueness, but this view is under debate. See Coppock and Beaver (2012), for instance, for an analysis of the that does not presuppose existence and presupposes a weaker version of uniqueness. Furthermore, we currently do not distinguish predicative and argumental definites. Presuppositions and unanswerability. Questions containing failed presuppositions are often treated as unanswerable in QA datasets. An example is the question What is the stock symbol for Mars candy? from NQ. This question is not answerable with any description of a stock symbol (that is, an answer to the what question), because Mars is not a publicly traded company and thus does not have a stock symbol. A better response would be to point out the presupposition failure, as in There is no stock symbol for Mars candy. However, statements about negative factuality are rarely explicitly stated, possibly due to reporting bias (Gordon and Van Durme, 2013). Therefore, under an extractive QA setup as in NQ where the answers are spans from an answer source (e.g., a Wikipedia article), it is likely that such questions will be unanswerable. Our proposal is based on the observation that the denial of a failed presupposition (¬P) can be used to explain the unanswerability of questions (Q) containing failed presuppositions (P), as in (2). (2) Q: Who is the current monarch of France? P: There is a current monarch of France. ¬P: There is no such thing as a current monarch of France. An answer that refers to the presupposition, such as ¬P, would be more informative compared to both Unanswerable (1-b) and an extractive answer from documents that are topically relevant but do not mention the false presupposition. 3 Analysis of Unanswerable Questions First, to quantify the role of presupposition failure in QA, two of the authors analyzed 100 randomly selected unanswerable wh-questions in the NQ development set.2 The annotators labeled each question as presupposition failure or not presupposition failure, depending on whether its unanswerability could be explained by the presence of an unverifiable presupposition with respect to the associated document. If the unanswerability could not be explained in terms of presupposition failure, the annotators provided a reasoning. The Cohen’s κ for inter-annotator agreement was 0.586. We found that 30% of the analyzed questions could be explained by the presence of an unverifiable presupposition in the question, considering only the cases where both annotators were in agreement (see Table 1).3 After adjudicating the reasoning about unanswerability for the nonpresupposition failure cases, another 21% fell into cases where presupposition failure could be partially informative (see Table 1 and Appendix A for details). The unverifiable presuppositions were 2The NQ development set provides 5 answer annotations per question—we only looked at questions with 5/5 Null answers here. 3wh-questions constitute ∼69% of the NQ development set, so we expect the actual portion of questions with presupposition failiure-based explanation to be ∼21%. 3935 Question: where can i buy a japanese dwarf flying squirrel Simple unanswerable This question is unanswerable. Presupposition failure-based This question is unanswerable because we could not verify that you can buy a Japanese Dwarf Flying Squirrel anywhere. Extractive explanation This question is unanswerable because it grows to a length of 20 cm (8 in) and has a membrane connecting its wrists and ankles which enables it to glide from tree to tree. DPR rewrite After it was returned for the second time, the original owner, referring to it as “the prodigal gnome", said she had decided to keep it and would not sell it on Ebay again. Table 2: Systems (answer types) compared in the user preference study and examples. 18% 15% 0% 65% S1: Extractive S2: No Explanation System 1 is Better System 2 is Better Both are Good Both are Bad 39% 8% 2% 49% S1: Karpukhin (2020) S2: No Explanation 37% 10% 3% 49% S1: Karpukhin (2020) S2: Extractive 74% 0% 2%23% S1: Presup. S2: No Explanation 57% 9%7% 24% S1: Presup. S2: Extractive 41% 26% 6% 24% S1: Presup. S2: Karpukhin (2020) Figure 2: Results of the user preference study. Chart labels denote the two systems being compared (S1 vs. S2). triggered by question words (19/30), the definite article the (10/30), and a factive verb (1/30). 4 User Study with Oracle Explanation Our hypothesis is that statements explicitly referring to failed presuppositions can better4 speak to the unanswerability of corresponding questions. To test our hypothesis, we conducted a side-by-side comparison of the oracle output of our proposed system and the oracle output of existing (closedbook) QA systems for unanswerable questions. We included two additional systems for comparison; the four system outputs compared are described below (see Table 2 for examples): • Simple unanswerable: A simple assertion that the question is unanswerable (i.e., This question is unanswerable). This is the oracle behavior of closed-book QA systems that allow Unanswerable as an answer. • Presupposition failure-based explanation: A denial of the presupposition that is unverifiable from the answer source. This takes the form of either This question is unanswerable because we could not verify that... or ...because it is unclear that... depending on the 4We define better as user preference in this study, but other dimensions could also be considered such as trustworthiness. type of the failed presupposition. See Section 5.3 for more details. • Extractive explanation: A random sentence from a Wikipedia article that is topically related to the question, prefixed by This question is unanswerable because.... This system is introduced as a control to ensure that length bias is not in play in the main comparison (e.g., users may a priori prefer longer, topicallyrelated answers over short answers). That is, since our system, Presupposition failurebased explanation, yields strictly longer answers than Simple unanswerable, we want to ensure that our system is not preferred merely due to length rather than answer quality. • Open-domain rewrite: A rewrite of the nonoracle output taken from the demo5 of Dense Passage Retrieval (DPR; Karpukhin et al. 2020), a competitive open-domain QA system. This system is introduced to test whether presupposition failure can be easily addressed by expanding the answer source, since a single Wikipedia article was used to determine presupposition failure. If presupposition failure is a problem particular only to closed-book systems, a competitive open-domain system would suffice to address this issue. While the outputs compared are not oracle, this system 5http://qa.cs.washington.edu:2020/ 3936 Question (input) Template Presupposition (output) which philosopher advocated the idea of return to nature some __ some philosopher advocated the idea of return to nature when was it discovered that the sun rotates __ the sun rotates when is the year of the cat in chinese zodiac __ exists ‘year of the cat in chinese zodiac’ exists when is the year of the cat in chinese zodiac __ is contextually unique ‘year of the cat in chinese zodiac’ is contextually unique what do the colors on ecuador’s flag mean __ has __ ‘ecuador’ has ‘flag’ Table 3: Example input-output pairs of our presupposition generator. Text in italics denotes the part taken from the original question, and the plain text is the part from the generation template. All questions are taken from NQ. has an advantage of being able to refer to all of Wikipedia. The raw output was rewritten to be well-formed, so that it was not unfairly disadvantaged (see Appendix B.2). Study. We conducted a side-by-side study with 100 unanswerable questions. These questions were unanswerable questions due to presupposition failure, as judged independently and with high confidence by two authors.6 We presented an exhaustive binary comparison of four different types of answers for each question (six binary comparisons per question). We recruited five participants on an internal crowdsourcing platform at Google, who were presented with all binary comparisons for all questions. All comparisons were presented in random order, and the sides that the comparisons appeared in were chosen at random. For each comparison, the raters were provided with an unanswerable question, and were asked to choose the system that yielded the answer they preferred (either System 1 or 2). They were also given the options Both answers are good/bad. See Appendix B.1 for additional details about the task setup. Results. Figure 2 shows the user preferences for the six binary comparisons, where blue and gray denote preferences for the two systems compared. We find that presupposition-based answers are preferred against all three answer types with which they were compared, and prominently so when compared to the oracle behavior of existing closedbook QA systems (4th chart, Presup. vs. No Explanation). This supports our hypothesis that presupposition failure-based answers would be more satisfactory to the users, and suggests that building a QA system that approaches the oracle behavior of our proposed system is a worthwhile pursuit. 6Hence, this set did not necessarily overlap with the randomly selected unanswerable questions from Section 3; we wanted to specifically find a set of questions that were representative of the phenomena we address in this work. 5 Model Components Given that presupposition failure accounts for a substantial proportion of unanswerable questions (Section 3) and our proposed form of explanations is useful (Section 4), how can we build a QA system that offers such explanations? We decompose this task into three smaller sub-tasks: presupposition generation, presupposition verification, and explanation generation. Then, we present progress towards each subproblem using NQ.7 We use a templatic approach for the first and last steps. The second step involves verification of the generated presuppositions of the question against an answer source, for which we test four different strategies: zero-shot transfer from Natural Language Inference (NLI), an NLI model finetuned on verification, zero-shot transfer from fact verification, and a rulebased/NLI hybrid model. Since we used NQ, our models assume a closed-book setup with a single document as the source of verification. 5.1 Step 1: Presupposition Generation Linguistic triggers. Using the linguistic triggers discussed in Section 2, we implemented a rulebased generator to templatically generate presuppositions from questions. See Table 3 for examples, and Appendix C for a full list. Generation. The generator takes as input a constituency parse tree of a question string from the Berkeley Parser (Petrov et al., 2006) and applies trigger-specific transformations to generate the presupposition string (e.g., taking the sentential complement of a factive verb). If there are multiple triggers in a single question, all presuppositions corresponding to the triggers are generated. Thus, a single question may have multiple presuppositions. See Table 3 for examples of input questions and output presuppositions. 7Code and data will be available at https://github.com/google-research/ google-research/presup-qa 3937 How good is our generation? We analyzed 53 questions and 162 generated presuppositions to estimate the quality of our generated presuppositions. This set of questions contained at least 10 instances of presuppositions pertaining to each category. One of the authors manually validated the generated presuppositions. According to this analysis, 82.7% (134/162) presuppositions were valid presuppositions of the question. The remaining cases fell into two broad categories of error: ungrammatical (11%, 18/162) or grammatical but not presupposed by the question (6.2%, 10/162). The latter category of errors is a limitation of our rulebased generator that does not take semantics into account, and suggests an avenue by which future work can yield improvements. For instance, we uniformly apply the template ‘A’ has ‘B’8 for presuppositions triggered by ’s. While this template works well for cases such as Elsa’s sister » ‘Elsa’ has ‘sister’, it generates invalid presuppositions such as Bachelor’s degree » #‘Bachelor’ has ‘degree’. Finally, the projection problem is another limitation. For example, who does pip believe is estella’s mother has an embedded possessive under a nonfactive verb believe, but our generator would nevertheless generate ‘estella’ has ‘mother’. 5.2 Step 2: Presupposition Verification The next step is to verify whether presuppositions of a given question is verifiable from the answer source. The presuppositions were first generated using the generator described in Section 5.1, and then manually repaired to create a verification dataset with gold presuppositions. This was to ensure that verification performance is estimated without a propagation of error from the previous step. Generator outputs that were not presupposed by the questions were excluded. To obtain the verification labels, two of the authors annotated 462 presuppositions on their binary verifiability (verifiable/not verifiable) based on the Wikipedia page linked to each question (the links were provided in NQ). A presupposition was labeled verifiable if the page contained any statement that either asserted or implied the content of the presupposition. The Cohen’s κ for inter-annotator agreement was 0.658. The annotators reconciled the disagreements based on a post-annotation dis8We used a template that puts possessor and possessee NPs in quotes instead of using different templates depending on posessor/possessee plurality (e.g., A __ has a __/A __ has __/__ have a __/__ have __). cussion to finalize the labels to be used in the experiments. We divided the annotated presuppositions into development (n = 234) and test (n = 228) sets.9 We describe below four different strategies we tested. Zero-shot NLI. NLI is a classification task in which a model is given a premise-hypothesis pair and asked to infer whether the hypothesis is entailed by the premise. We formulate presupposition verification as NLI by treating the document as the premise and the presupposition to verify as the hypothesis. Since Wikipedia articles are often larger than the maximum premise length that NLI models can handle, we split the article into sentences and created n premise-hypothesis pairs for an article with n sentences. Then, we aggregated these predictions and labeled the hypothesis (the presupposition) as verifiable if there are at least k sentences from the document that supported the presupposition. If we had a perfect verifier, k = 1 would suffice to perform verification. We used k = 1 for our experiments, but k could be treated as a hyperparameter. We used ALBERT-xxlarge (Lan et al., 2020) finetuned on MNLI (Williams et al., 2018) and QNLI (Wang et al., 2019) as our NLI model. Finer-tuned NLI. Existing NLI datasets such as QNLI contain a broad distribution of entailment pairs. We adapted the model further to the distribution of entailment pairs that are specific to our generated presuppositions (e.g., Hypothesis: NP is contextually unique) through additional finetuning (i.e., finer-tuning). Through crowdsourcing on an internal platform, we collected entailment labels for 15,929 (presupposition, sentence) pairs, generated from 1000 questions in NQ and 5 sentences sampled randomly from the corresponding Wikipedia pages. We continued training the model fine-tuned on QNLI on this additional dataset to yield a finer-tuned NLI model. Finally, we aggregated per-sentence labels as before to get verifiability labels for (presupposition, document) pairs. Zero-shot FEVER. FEVER is a fact verification task proposed by Thorne et al. (2018). We formulate presupposition verification as a fact verification task by treating the Wikipedia article as the evidence source and the presupposition as the claim. While typical FEVER systems have a docu9The dev/test set sizes did not exactly match because we kept presuppositions of same question within the same split, and each question had varying numbers of presuppositions. 3938 Model Macro F1 Acc. Majority class 0.44 0.78 Zero-shot NLI (ALBERT MNLI + Wiki sentences) 0.50 0.51 Zero-shot NLI (ALBERT QNLI + Wiki sentences) 0.55 0.73 Zero-shot FEVER (KGAT + Wiki sentences) 0.54 0.66 Finer-tuned NLI (ALBERT QNLI + Wiki sentences) 0.58 0.76 Rule-based/NLI hybrid (ALBERT QNLI + Wiki presuppositions) 0.58 0.71 Rule-based/NLI hybrid (ALBERT QNLI + Wiki sentences + Wiki presuppositions) 0.59 0.77 Finer-tuned, rule-based/NLI hybrid (ALBERT QNLI + Wiki sentences + Wiki presuppositions) 0.60 0.79 Table 4: Performance of verification models tested. Models marked with ‘Wiki sentence’ use sentences from Wikipedia articles as premises, and ‘Wiki presuppositions’, generated presuppositions from Wikipedia sentences. ment retrieval component, we bypass this step and directly perform evidence retrieval on the article linked to the question. We used the Graph Neural Network-based model of Liu et al. (2020) (KGAT) that achieves competitive performance on FEVER. A key difference between KGAT and NLI models is that KGAT can consider pieces of evidence jointly, whereas with NLI, the pieces of evidence are verified independently and aggregated at the end. For presuppositions that require multihop reasoning, KGAT may succeed in cases where aggregated NLI fails—e.g., for uniqueness. That is, if there is no sentence in the document that bears the same uniqueness presupposition, one would need to reason over all sentences in the document. Rule-based/NLI hybrid. We consider a rulebased approach where we apply the same generation method described in Section 5 to the Wikipedia documents to extract the presuppositions of the evidence sentences. The intended effect is to extract content that is directly relevant to the task at hand—that is, we are making the presuppositions of the documents explicit so that they can be more easily compared to presuppositions being verified. However, a naïve string match between presuppositions of the document and the questions would not work, due to stylistic differences (e.g., definite descriptions in Wikipedia pages tend to have more modifiers). Hence, we adopted a hybrid approach where the zero-shot QNLI model was used to verify (document presupposition, question presupposition) pairs. Results. Our results (Table 4) suggest that presupposition verification is challenging to existing models, partly due to class imbalance. Only the model that combines finer-tuning and rule-based document presuppositions make modest improvement over the majority class baseline (78% → 79%). Nevertheless, gains in F1 were substantial for all models (44% →60% in best model), showing that these strategies do impact verifiability, albeit with headroom for improvement. QNLI provided the most effective zero-shot transfer, possibly because of domain match between our task and the QNLI dataset—they are both based on Wikipedia. The FEVER model was unable to take advantage of multihop reasoning to improve over (Q)NLI, whereas using document presuppositions (Rulebased/NLI hybrid) led to gains over NLI alone. 5.3 Step 3: Explanation Generation We used a template-based approach to explanation generation: we prepended the templates This question is unanswerable because we could not verify that... or ...because it is unclear that... to the unverifiable presupposition (3). Note that we worded the template in terms of unverifiability of the presupposition, rather than asserting that it is false. Under a closed-book setup like NQ, the only ground truth available to the model is a single document, which leaves a possibility that the presupposition is verifiable outside of the document (except in the rare occasion that it is refuted by the document). Therefore, we believe that unverifiability, rather than failure, is a phrasing that reduces false negatives. (3) Q: when does back to the future part 4 come out Unverifiable presupposition: there is some point in time that back to the future part 4 comes out Simple prefixing: This question is unanswerable because we could not verify that there is some point in time that back to the future part 4 comes out. 3939 Model Average F1 Long answer F1 Short answer F1 Unans. Acc Unans. F1 ETC (our replication) 0.645 0.742 0.548 0.695 0.694 + Presuppositions (flat) 0.641 0.735 0.547 0.702 0.700 + Verification labels (flat) 0.645 0.742 0.547 0.687 0.684 + Presups + labels (flat) 0.643 0.744 0.544 0.702 0.700 + Presups + labels (structured) 0.649 0.743 0.555 0.703 0.700 Table 5: Performance on NQ development set with ETC and ETC augmented with presupposition information. We compare our augmentation results against our own replication of Ainslie et al. (2020) (first row). For the user study (Section 4), we used a manual, more fluent rewrite of the explanation generated by simple prefixing. In future work, fluency is a dimension that can be improved over templatic generation. For example, for (3), a fluent model could generate the response: This question is unanswerable because we could not verify that Back to the Future Part 4 will ever come out. 6 End-to-end QA Integration While the 3-step pipeline is designed to generate explanations for unanswerability, the generated presuppositions and their verifiability can also provide useful guidance even for a standard extractive QA system. They may prove useful both to unanswerable and answerable questions, for instance by indicating which tokens of a document a model should attend to. We test several approaches to augmenting the input of a competitive extractive QA system with presuppositions and verification labels. Model and augmentation. We used Extended Transformer Construction (ETC) (Ainslie et al., 2020), a model that achieves competitive performance on NQ, as our base model. We adopted the configuration that yielded the best reported NQ performance among ETC-base models.10 We experiment with two approaches to encoding the presupposition information. First, in the flat model, we simply augment the input question representation (token IDs of the question) by concatenating the token IDs of the generated presuppositions and the verification labels (0 or 1) from the ALBERT QNLI model. Second, in the structured model (Figure 4), we take advantage of the global input layer of ETC that is used to encode the discourse units of large documents like paragraphs. Global tokens attend (via self-attention) to all tokens of their in10The reported results in Ainslie et al. (2020) are obtained using a custom modification to the inference procedure that we do not incorporate into our pipeline, since we are only interested in the relative gains from presupposition verification. ternal text, but for other text in the document, they only attend to the corresponding global tokens. We add one global token for each presupposition, and allow the presupposition tokens to only attend to each other and the global token. The value of the global token is set to the verification label (0 or 1). Metrics. We evaluated our models on two sets of metrics: NQ performance (Long Answer, Short Answer, and Average F1) and Unanswerability Classification (Accuracy and F1).11 We included the latter because our initial hypothesis was that sensitivity to presuppositions of questions would lead to better handling of unanswerable questions. The ETC NQ model has a built-in answer type classification step which is a 5-way classification between {Unanswerable, Long Answer, Short Answer, Yes, No}. We mapped the classifier outputs to binary answerability labels by treating the predicted label as Unanswerable only if its logit was greater than the sum of all other options. Results and Discussion Table 5 shows that augmentations that use only the presuppositions or only the verification labels do not lead to gains in NQ performance over the baseline, but the presuppositions do lead to gains on Unanswerability Classification. When both presuppositions and their verifiability are provided, we see minor gains in Average F1 and Unanswerability Classification.12 For Unanswerability Classification, the improved accuracy is different from the baseline at the 86% (flat) and 89% (structured) confidence level using McNemar’s test. The main bottleneck of our model is the quality of the verification labels used for augmentation (Table 4)—noisy labels limit the capacity of the QA model to attend to the augmentations. While the gain on Unanswerability Classification is modest, an error analysis suggests that 11Here, we treated ≥4 Null answers as unanswerable, following the definition in Kwiatkowski et al. (2019). 12To contextualize our results, a recently published NQ model (Ainslie et al., 2020) achieved a gain of around ∼2%. 3940 the added presuppositions modulate the prediction change in our best-performing model (structured) from the baseline ETC model. Looking at the cases where changes in model prediction (i.e., Unanswerable (U) ↔Answerable (A)) lead to correct answers, we observe an asymmetry in the two possible directions of change. The number of correct A →U cases account for 11.9% of the total number of unanswerable questions, whereas correct U → A cases account for 6.7% of answerable questions. This asymmetry aligns with the expectation that the presupposition-augmented model should achieve gains through cases where unverified presuppositions render the question unanswerable. For example, given the question who played david brent’s girlfriend in the office that contains a false presupposition David Brent has a girlfriend, the structured model changed its prediction to Unanswerable from the base model’s incorrect answer Julia Davis (an actress, not David Brent’s girlfriend according to the document: . . . arrange a meeting with the second woman (voiced by Julia Davis)). On the other hand, such an asymmetry is not observed in cases where changes in model prediction results in incorrect answers: incorrect A →U and U →A account for 9.1% and 9.2%, respectively. More examples are shown in Appendix F. 7 Related Work While presuppositions are an active topic of research in theoretical and experimental linguistics (Beaver, 1997; Simons, 2013; Schwarz, 2016, i.a.,), comparatively less attention has been given to presuppositions in NLP (but see Clausen and Manning (2009) and Tremper and Frank (2011)). More recently, Cianflone et al. (2018) discuss automatically detecting presuppositions, focusing on adverbial triggers (e.g., too, also...), which we excluded due to their infrequency in NQ. Jeretic et al. (2020) investigate whether inferences triggered by presuppositions and implicatures are captured well by NLI models, finding mixed results. Regarding unanswerable questions, their importance in QA (and therefore their inclusion in benchmarks) has been argued by works such as Clark and Gardner (2018) and Zhu et al. (2019). The analysis portion of our work is similar in motivation to unanswerability analyses in Yatskar (2019) and Asai and Choi (2020)—to better understand the causes of unanswerability in QA. Hu et al. (2019); Zhang et al. (2020); Back et al. (2020) consider answerability detection as a core motivation of their modeling approaches and propose components such as independent no-answer losses, answer verification, and answerability scores for answer spans. Our work is most similar to Geva et al. (2021) in proposing to consider implicit assumptions of questions. Furthermore, our work is complementary to QA explanation efforts like Lamm et al. (2020) that only consider answerable questions. Finally, abstractive QA systems (e.g., Fan et al. 2019) were not considered in this work, but their application to presupposition-based explanation generation could be an avenue for future work. 8 Conclusion Through an NQ dataset analysis and a user preference study, we demonstrated that a significant portion of unanswerable questions can be answered more effectively by calling out unverifiable presuppositions. To build models that provide such an answer, we proposed a novel framework that decomposes the task into subtasks that can be connected to existing problems in NLP: presupposition identification (parsing and text generation), presupposition verification (textual inference and fact verification), and explanation generation (text generation). We observed that presupposition verification, especially, is a challenging problem. A combination of a competitive NLI model, finer-tuning and rule-based hybrid inference gave substantial gains over the baseline, but was still short of a fully satisfactory solution. As a by-product, we showed that verified presuppositions can modestly improve the performance of an end-to-end QA model. In the future, we plan to build on this work by proposing QA systems that are more robust and cooperative. For instance, different types of presupposition failures could be addressed by more fluid answer strategies—e.g., violation of uniqueness presuppositions may be better handled by providing all possible answers, rather than stating that the uniqueness presupposition was violated. Acknowledgments We thank Tom Kwiatkowski, Mike Collins, Tania Rojas-Esponda, Eunsol Choi, Annie Louis, Michael Tseng, Kyle Rawlins, Tania Bedrax-Weiss, and Elahe Rahimtoroghi for helpful discussions about this project. We also thank Lora Aroyo for help with user study design, and Manzil Zaheer for pointers about replicating the ETC experiments. 3941 References Márta Abrusán. 2011. Presuppositional and negative islands: a semantic account. Natural Language Semantics, 19(3):257–321. Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. ETC: Encoding long and structured inputs in transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 268–284, Online. Association for Computational Linguistics. Akari Asai and Eunsol Choi. 2020. Challenges in information seeking QA: Unanswerable questions and paragraph retrieval. arXiv:2010.11915. Seohyun Back, Sai Chetan Chinthakindi, Akhil Kedia, Haejun Lee, and Jaegul Choo. 2020. NeurQuRI: Neural question requirement inspector for answerability prediction in machine reading comprehension. In International Conference on Learning Representations. David Beaver. 1997. Presupposition. In Handbook of Logic and Language, pages 939–1008. Elsevier. Andre Cianflone, Yulan Feng, Jad Kabbara, and Jackie Chi Kit Cheung. 2018. Let’s do it “again”: A first computational approach to detecting adverbial presupposition triggers. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2747– 2755, Melbourne, Australia. Association for Computational Linguistics. Christopher Clark and Matt Gardner. 2018. Simple and effective multi-paragraph reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 845–855, Melbourne, Australia. Association for Computational Linguistics. David Clausen and Christopher D. Manning. 2009. Presupposed content and entailments in natural language inference. In Proceedings of the 2009 Workshop on Applied Textual Inference (TextInfer), pages 70–73, Suntec, Singapore. Association for Computational Linguistics. Elizabeth Coppock and David Beaver. 2012. Weak uniqueness: The only difference between definites and indefinites. In Semantics and Linguistic Theory, volume 22, pages 527–544. Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. 2019. ELI5: Long form question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3558–3567, Florence, Italy. Association for Computational Linguistics. Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did Aristotle use a laptop? A question answering benchmark with implicit reasoning strategies. arXiv:2101.02235. Jonathan Gordon and Benjamin Van Durme. 2013. Reporting bias and knowledge acquisition. In Proceedings of the 2013 Workshop on Automated Knowledge Base Construction, AKBC ’13, page 25–30, New York, NY, USA. Association for Computing Machinery. Minghao Hu, Furu Wei, Yuxing Peng, Zhen Huang, Nan Yang, and Dongsheng Li. 2019. Read+ verify: Machine reading comprehension with unanswerable questions. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6529– 6537. Paloma Jeretic, Alex Warstadt, Suvrat Bhooshan, and Adina Williams. 2020. Are natural language inference models IMPPRESsive? Learning IMPlicature and PRESupposition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8690–8705, Online. Association for Computational Linguistics. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769– 6781, Online. Association for Computational Linguistics. Lauri Karttunen. 2016. Presupposition: What went wrong? In Semantics and Linguistic Theory, volume 26, pages 705–731. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452–466. Matthew Lamm, Jennimaria Palomaki, Chris Alberti, Daniel Andor, Eunsol Choi, Livio Baldini Soares, and Michael Collins. 2020. QED: A framework and dataset for explanations in question answering. arXiv:2009.06354. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. ALBERT: A Lite BERT for self-supervised learning of language representations. In International Conference on Learning Representations. Stephen C. Levinson. 1983. Pragmatics, Cambridge Textbooks in Linguistics, pages 181–184, 194. Cambridge University Press. 3942 Zhenghao Liu, Chenyan Xiong, Maosong Sun, and Zhiyuan Liu. 2020. Fine-grained fact verification with kernel graph attention network. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7342–7351, Online. Association for Computational Linguistics. Randall Munroe. 2020. Learning new things from Google. https://twitter.com/xkcd/status/ 1333529967079120896. Accessed: 2021-02-01. Slav Petrov, Leon Barrett, Romain Thibaux, and Dan Klein. 2006. Learning accurate, compact, and interpretable tree annotation. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the Association for Computational Linguistics, pages 433–440, Sydney, Australia. Association for Computational Linguistics. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784– 789, Melbourne, Australia. Association for Computational Linguistics. Rob A. Van der Sandt. 1992. Presupposition projection as anaphora resolution. Journal of Semantics, 9(4):333–377. Philippe Schlenker. 2008. Presupposition projection: Explanatory strategies. Theoretical Linguistics, 34(3):287–316. Bernhard Schwarz and Alexandra Simonenko. 2018. Decomposing universal projection in questions. In Sinn und Bedeutung 22, volume 22, pages 361–374. Florian Schwarz. 2016. Experimental work in presupposition and presupposition projection. Annual Review of Linguistics, 2(1):273–292. Mandy Simons. 2013. Presupposing. Pragmatics of Speech Actions, pages 143–172. Peter F. Strawson. 1950. On referring. Mind, 59(235):320–344. Nadine Theiler. 2020. An epistemic bridge for presupposition projection in questions. In Semantics and Linguistic Theory, volume 30, pages 252–272. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana. Association for Computational Linguistics. Galina Tremper and Anette Frank. 2011. Extending fine-grained semantic relation classification to presupposition relations between verbs. Bochumer Linguistische Arbeitsberichte. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations. Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1112–1122, New Orleans, Louisiana. Association for Computational Linguistics. Mark Yatskar. 2019. A qualitative comparison of CoQA, SQuAD 2.0 and QuAC. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2318–2323, Minneapolis, Minnesota. Association for Computational Linguistics. Zhuosheng Zhang, Junjie Yang, and Hai Zhao. 2020. Retrospective reader for machine reading comprehension. arXiv:2001.09694. Haichao Zhu, Li Dong, Furu Wei, Wenhui Wang, Bing Qin, and Ting Liu. 2019. Learning to ask unanswerable questions for machine reading comprehension. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4238–4248, Florence, Italy. Association for Computational Linguistics. A Additional Causes of Unanswerable Questions Listed below are cases of unanswerable questions for which presupposition failure may be partially useful: • Document retrieval failure: The retrieved document is unrelated to the question, so the presuppositions of the questions are unlikely to be verifiable from the document. • Failure of commonsensical presuppositions: The document does not directly support the presupposition but the presupposition is commonsensical. • Presuppositions involving subjective judgments: verification of the presupposition requires subjective judgment, such as the existence of the best song. 3943 Figure 3: The user interface for the user preference study. • Reference resolution failure: the question contains an unresolved reference such as a pro-form (I, here...) or a temporal expression (next year...). Therefore the presuppositions also fail due to unresolved reference. B User Study B.1 Task Design Figure 3 shows the user interface (UI) for the study. The raters were given a guideline that instructed them to select the answer that they preferred, imagining a situation in which they have entered the given question to two different QA systems. To avoid biasing the participants towards any answer type, we used a completely unrelated, nonsensical example (Q: Are potatoes fruit? System 1: Yes, because they are not vegetables. System 2: Yes, because they are not tomatoes.) in our guideline document. B.2 DPR Rewrites The DPR answers we used in the user study were rewrites of the original outputs. DPR by default returns a paragraph-length Wikipedia passage that contains the short answer to the question. From this default output, we manually extracted the sentencelevel context that fully contains the short answer, and repaired the context into a full sentence if the extracted context was a sentence fragment. This was to ensure that all answers compared in the study were well-formed sentences, so that user preference was determined by the content of the sentences rather than their well-formedness. C Presupposition Generation Templates See Table 6 for a full list of presupposition triggers and templates used for presupposition generation. D Data Collection The user study (Section 4) and data collection of entailment pairs from presuppositions and Wikipedia sentences (Section 5) have been performed by crowdsourcing internally at Google. Details of the user study is in Appendix B. Entailment judgements were elicited from 3 raters for each pair, and majority vote was used to assign a label. Because of class imbalance, all positive labels were kept in the data and negative examples were down-sampled to 5 per document. E Modeling Details E.1 Zero-shot NLI MNLI and QNLI were trained following instructions for fine-tuning on top of ALBERT-xxlarge at https://github.com/google-research/ albert/blob/master/albert_glue_fine_ tuning_tutorial.ipynb with the default settings and parameters. E.2 KGAT We used the off-the-shelf model from https:// github.com/thunlp/KernelGAT (BERT-base). E.3 ETC models For all ETC-based models, we used the same model parameter settings as Ainslie et al. (2020) used for NQ, only adjusting the maximum global input length to 300 for the flat models to accommodate the larger set of tokens from presuppositions. Model selection was done by choosing hyperparameter configurations yielding maximum Average F1. Weight lifting was done from BERT-base instead of RoBERTa to keep the augmentation experiments simple. All models had 109M parameters. All model training was done using the Adam optimizer with hyperparameter sweeps of learning 3944 Question (input) Template Presupposition (output) who sings it’s a hard knock life there is someone that __ there is someone that sings it’s a hard knock life which philosopher advocated the idea of return to nature some __ some philosopher advocated the idea of return to nature where do harry potter’s aunt and uncle live there is some place that __ there is some place that harry potter’s aunt and uncle live what did the treaty of paris do for the US there is something that __ there is something that the treaty of paris did for the US when was the jury system abolished in india there is some point in time that __ there is some point in time that the jury system was abolished in india how did orchestra change in the romantic period __ orchestra changed in the romantic period how did orchestra change in the romantic period there is some way that __ there is some way that orchestra changed in the romantic period why did jean valjean take care of cosette __ jean valjean took care of cosette why did jean valjean take care of cosette there is some reason that __ there is some reason that jean valjean took care of cosette when is the year of the cat in chinese zodiac __ exists ‘year of the cat in chinese zodiac’ exists when is the year of the cat in chinese zodiac __ is contextually unique ‘year of the cat in chinese zodiac’ is contextually unique what do the colors on ecuador’s flag mean __ has __ ‘ecuador’ has ‘flag’ when was it discovered that the sun rotates __ the sun rotates how old was macbeth when he died in the play __ he died in the play who would have been president if the south won the civil war it is not true that __ it is not true that the south won the civil war Table 6: Example input-output pairs of our presupposition generator. Text in italics denotes the part taken from the original question, and the plain text is the part from the generation template. All questions are taken from NQ. Figure 4: The structured augmentation to the ETC model. Qk are question tokens, Pk are presupposition tokens, Sl are sentence tokens, Pv are verification labels, Qid is the (constant) global question token and Sid is the (constant) global sentence token. rates in {3×10−5, 5×10−5} and number of epochs in {3, 5} (i.e., 4 settings). In cases of overfitting, an earlier checkpoint of the run with optimal validation performance was picked. All training was done on servers utilizing a Tensor Processing Unit 3.0 architecture. Average runtime of model training with this architecture was 8 hours. Figure 4 illustrates the structure augmented ETC model that separates question and presupposition tokens that we discussed in Section 6. F ETC Prediction Change Examples We present selected examples of model predictions from Section 6 that illustrate the difference in behavior of the baseline ETC model and the structured, presupposition-augmented model: 1. [Correct Answerable →Unanswerable] NQ Question: who played david brent’s girlfriend in the office Relevant presupposition: David Brent has a girlfriend Wikipedia Article: The Office Christmas specials Gold Label: Unanswerable Baseline label: Answerable Structured model label: Unanswerable Explanation: The baseline model incorrectly predicts arrange a meeting with the second woman (voiced by Julia Davis) as a long answer and Julia Davis as a short answer, inferring that the second woman met by David Brent was his girlfriend. The structured model correctly flips the prediction to Unanswerable, possibly making use of the unverifiable presupposition David Brent has a girlfriend. 2. [Correct Unanswerable →Answerable] NQ Question: when did cricket go to 6 ball overs Relevant presupposition: Cricket went to 6 balls per over at some point Wikipedia Article: Over (cricket) Gold Label: Answerable Baseline label: Unanswerable Structured model label: Answerable Explanation: The baseline model was likely confused because the long answer candidate only mentions Test Cricket, but support for the presupposition came from the sentence Although six was the usual number of balls, it was not always the case, leading the structured model to choose the correct long answer candidate. 3. [Incorrect Answerable →Unanswerable] NQ Question: what is loihi and where does it originate from Relevant presupposition: there is some place that it originates from Wikipedia Article: L¯oihi Seamount Gold Label: Answerable Baseline label: Answerable Structured model label: Unanswerable Explanation: The baseline model finds the correct answer (Hawaii hotspot) but the struc3945 tured model incorrectly changes the prediction. This is likely due to verification error— although the presupposition there is some place that it originates from is verifiable, it was incorrectly labeled as unverifiable. Possibly, the the unresolved it contributed to this verification error, since our verifier currently does not take the question itself into consideration.
|
2021
|
304
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3946–3957 August 1–6, 2021. ©2021 Association for Computational Linguistics 3946 Adversarial Learning for Discourse Rhetorical Structure Parsing Longyin Zhang1,2, Fang Kong1,2∗, Guodong Zhou1,2 1. Institute of Artificial Intelligence, Soochow University, China 2. School of Computer Science and Technology, Soochow University, China [email protected] {kongfang,gdzhou}@suda.edu.cn Abstract Text-level discourse rhetorical structure (DRS) parsing is known to be challenging due to the notorious lack of training data. Although recent top-down DRS parsers can better leverage global document context and have achieved certain success, the performance is still far from perfect. To our knowledge, all previous DRS parsers make local decisions for either bottomup node composition or top-down split point ranking at each time step, and largely ignore DRS parsing from the global view point. Obviously, it is not sufficient to build an entire DRS tree only through these local decisions. In this work, we present our insight on evaluating the pros and cons of the entire DRS tree for global optimization. Specifically, based on recent well-performing top-down frameworks, we introduce a novel method to transform both gold standard and predicted constituency trees into tree diagrams with two color channels. After that, we learn an adversarial bot between gold and fake tree diagrams to estimate the generated DRS trees from a global perspective. We perform experiments on both RST-DT and CDTB corpora and use the original Parseval for performance evaluation. The experimental results show that our parser can substantially improve the performance when compared with previous state-of-the-art parsers. 1 Introduction As the main linguistic theory on discourse rhetorical structure (DRS), Rhetorical Structure Theory (RST) (Mann and Thompson, 1988) describes an article as a discourse tree (DT). As illustrated in Figure 1, each leaf node of the tree corresponds to an Elementary Discourse Unit (EDU), and relevant leaf nodes are connected by relation and nuclearity (nucleus (N) or satellite (S)) tags to form high-layer discourse units (DUs), where the ∗Corresponding author [e1: In fact,] [e2: Budget indicated] [e3: it saw some benefit] [e4: to staying involved in these programs,] [e5: in which renters earn frequent-flier miles] [e6: and fliers can get car-rental discounts.] wsj_2394 e1 e2 e3 e4 e5 e6 Same-Unit (NN) Attribution (NS) List (NN) Elaboration (NS) Elaboration (NS) Figure 1: An example RST-style discourse tree. nucleus is considered more important than the satellite. Since the RST structure can well describe the organization of an article, it has been playing a central role in various down-stream tasks like summarization (Xu et al., 2020), text categorization (Ji and Smith, 2017), and so on. With the release of various discourse corpora, text-level DSR parsing has been drawing more and more attention in the last decade. However, since the corpus annotation is usually time-consuming, existing DRS corpora are much limited in size. For example, the English RST-DT (Carlson et al., 2001) corpus only contains 385 WSJ articles, and the Chinese CDTB (Li et al., 2014b) corpus only contains 500 newswire articles. In this situation, previous studies usually rely on multifarious handengineered features (Hernault et al., 2010; Feng and Hirst, 2014; Ji and Eisenstein, 2014; Li et al., 2014a, 2016; Braud et al., 2017). And all these systems perform DRS parsing in a bottom-up fashion. Until recently, some researchers turn to top-down DRS parsing (Lin et al., 2019; Zhang et al., 2020; Kobayashi et al., 2020) to explore the potential capabilities of data-driven models. Nevertheless, text-level DRS parsing is still challenging and worthy of in-depth exploration. Theoretically, in supervised learning, annotated 3947 (a) (b) e1 e2 e3 n1 Local OPT or e1 e2 e3 Global OPT n2 n1 n2 Figure 2: Local and global optimization of DRS trees. data corpora can provide neural models with specific learning objectives, and the corpus size limitation will weaken the learning of these goals. To mitigate this problem, we researchers need (i) an efficient model to better learn from the limited data and (ii) more high-quality training objectives to enhance the model learning. Existing studies on text-level DRS parsing show that • Compared with bottom-up DRS parsers, recent top-down frameworks can better leverage global document context and have achieved promising results in text-level DRS parsing (Zhang et al., 2020; Kobayashi et al., 2020). • All previous studies produce their DRS parsers with local decisions made at each time step for either bottom-up node composition or top-down split point selection (Figure 2 (a)), and no global decisions are made for the entire DRS structure (Figure 2 (b)). Therefore, it is difficult for them to achieve global optimization. Although some studies (Braud et al., 2017; Mabona et al., 2019) leverage “beam-search” to traverse the solution space to find the optimal parsing route, the algorithms are time-consuming to some extent. Considering the above-mentioned status quo, in this work, we study a global optimization method based on the well-performing top-down parsers. For model structure, we take the top-down parser of Zhang et al. (2020) as our baseline system and make some improvements to it. For global optimization, we first utilize a novel strategy to transform both gold standard and predicted DRS trees into tree diagrams with two color channels. After that, an LSGAN-based adversarial bot is structured between gold and fake tree diagrams as an examiner for global estimation and optimization. Experimental results on the RST-DT and CDTB corpora show that our approaches are effective. 2 Related Work In the literature, previous studies on RST-style DRS parsing mainly consist of two categories, i.e., bottom-up and top-down frameworks. For the first category, early studies on DRS parsing heavily relied on hand-crafted features and linguistic characteristics (Hernault et al., 2010; Joty et al., 2013; Feng and Hirst, 2014). During the past decade, more and more researchers turned to data-driven approaches, and some effective strategies were proposed to adapt to the small-scale data corpora. Among these studies, (Ji and Eisenstein, 2014; Li et al., 2014a, 2016; Mabona et al., 2019) used some trivial features as auxiliaries in their data-driven systems; Braud et al. (2016; 2017) harnessed task supervision from related tasks, alternative views on discourse structures, and crosslingual data to alleviate the data insufficiency problem; Wang et al. (2017) introduced a two-stage parser to first parse a naked tree structure and then determine rhetorical relations for different discourse levels to mitigate data sparsity; Yu et al. (2018) employed both syntax information and discourse boundaries in their transition-based system and achieved good performance. For the second category, some researchers (Lin et al., 2019; Liu et al., 2019; Zhang et al., 2020; Kobayashi et al., 2020) turned to top-down frameworks to tap the potential capabilities of data-driven models. Among them, (Lin et al., 2019; Liu et al., 2019) have achieved certain success in sentencelevel DRS parsing. Nevertheless, due to the longdistance dependency over the discourse, text-level DRS parsing remains challenging. To alleviate this problem, Zhang et al. (2020) proposed a top-down architecture tailored for text-level DRS parsing. Kobayashi et al. (2020) used contextualized word representation and proposed to parse a document in three granularity levels for good performance. In the past decade, GANs have achieved great progress in NLP (Wu et al., 2019; Elazar and Goldberg, 2018; Chen and Chen, 2019; Zou et al., 2020). However, to our knowledge, there is still no research on adversarial learning in DRS parsing so far. In this work, we explore to adversarially train a discriminator to estimate the quality of the entire DRS tree for global optimization. Notably, we propose to transform each DRS tree into a continuous tree diagram, and thus our adversarial method does not suffer from the “discrete data” problem. 3 Baseline Top-Down Architecture In this section, we give a brief introduction to our baseline system, the top-down parser of Zhang et 3948 al. (2020), and make some improvements to it. The parsing process is illustrated in Figure 3. Hierarchical Split Point Encoding. For split point representation1, Zhang et al. (2020) introduced a hierarchical RNN-CNN architecture in their paper. Firstly, they use an attention-based GRU encoder to encode each EDU, obtaining ei. Then, the obtained EDU vectors are fed into another BiGRU for context modeling, as shown in Figure 3. Next, a CNN net with a window size of 2 and a stride size of 1 is built for each window of EDUs in the discourse for split point encoding. To our knowledge, Zhang et al. (2020) produced dummy split points at both ends of a discourse. Since the dummy split points do not participate in the split point selection process, they could be redundant. Here, we try to simplify the parsing procedure with the dummy split points discarded, as shown in Figure 3. Following previous work (Yu et al., 2018; Kobayashi et al., 2020), we also splice the sentence- and paragraph-level boundary feature vectors to the representation of split points to enhance the encoder model. Top-Down Split Point Ranking. After achieving split point representations, an encoder-decoder is used to rank the split points, as shown in Figure 3. During encoding, the previously obtained split point vectors are taken as input to the BiGRU encoder, obtaining H0, . . . , Hn−2. During decoding, a uni-directional GRU with an internal stack is used to control the split point ranking process. Initially, the stack contains only one element, i.e., indexes of the boundary split points in the discourse. Notably, since we do not add dummy split points in this parser, we allow patterns like (τ, τ) to appear in the stack. At the j-th step, the tuple (B, E) is popped from the stack and we enter the concatenated cj = (HB; HE) into the decoder for dj. After that, a biaffine function (Dozat and Manning, 2017) is built between the encoder and decoder outputs for split point ranking. Different from (Zhang et al., 2020), all split points in the interval [B, E] are selectable in this work. At the step j, we calculate the attention score between Hi and dj as: sj,i = HT i Wdj + UHi + V dj + b (1) where W, U, V, b are model parameters and sj,i ∈ 1The split position between any two neighboring EDUs is called the split point. (0,4) (2,4) (0,0) (2,4) (3,4) (4,4) e0 e1 e2 e3 e4 e5 1 0 2 3 4 he0 he1 he2 he3 he4 he5 hs0 hs1 hs2 hs3 hs4 H0 H1 H2 H3 H4 c0 c1 c2 c3 c4 d0 d1 d2 d3 d4 䜭⭘Ҷ ᢰᵟҏ ቡᱟ࣏㜭ᙗṨ⻱ޡᥟᡀ ۿ ᢰᵟᶕሩབྷ㝁䘋㹼䙐ᖡDŽ ԆԜˈԆؙˈĂ 䜭⭘Ҷ ᢰᵟҏ ቡᱟ࣏㜭ᙗṨ⻱ޡᥟ ᡀۿ ᢰᵟᶕሩབྷ㝁䘋㹼䙐ᖡDŽ Figure 3: Neural architecture of the encoder-decoder. Rk denotes the score of the i-th split point over different categories (for split point ranking, k equals 1). With this attention function used, at each time step, split position with the highest score is selected as the split point and the original text span is split into two adjacent text spans. Meanwhile, newly generated text spans with unselected split points are pushed onto the stack for following steps, as shown in Figure 3. In this way, a DRS tree is built after 5 iterations with the split points (1, 0, 2, 3, 4) detected in turn. To our knowledge, Zhang et al. (2020) use three biaffine classifiers in their parser for structure, nuclearity and relation prediction, respectively. Considering the differences between the three learning objectives, using three independent classifiers could weaken the “Full” performance. To alleviate this problem, we combine nuclearity and relation tags into N-R tags and only use two classifiers for DRS parsing. Therefore, for N-R prediction, the category number k equals 41 and 46 for the RSTDT and CDTB corpus respectively. 4 Adversarial Learning for DRS Parsing This section introduces the proposed adversarial learning method which consists of two parts: graphical representation of gold and fake DRS trees and the adversarial model learning process. 4.1 Graphical Representation of DRS Trees In this study, we aim to learn from the entire DRS tree to optimize our model from a global perspective. Usually, our computer understands DRS trees in two ways: either language description or graphical representation. Since tree diagrams can reflect the structural features more intuitively and are easy for machines to understand, we explore graphical representation of DRS trees in this work. For gold standard trees, we propose to transform each tree into multi-pattern matrices which 3949 Convolution-Layer XST R3 R4 -1 -1 R2 R1 R3 0 -1 -1 -1 0 0 0 0 -1 -1 0 0 0 -2 -1 -1 -1 -2 -2 Parse Tree Image Generarion Gold Tree Adversarial Bot Feature Extraction z Feature Extraction Max Pooling Image Generation Reshape Feature x e1 e2 e3 e4 e5 e6 N-R1 N-R3 N-R2 N-R3 N-R4 XNR 0 m=5, n=5 i=3, j=3 i=2, j=0 Figure 4: Graphical representation of DRS structure for adversarial learning of text-level DRS parsing. is similar to a low resolution image with two color channels (i.e., the structure (ST) and nuclearityrelation (NR) channels). Formally, given a DRS tree of height m with n split points, each split point corresponds to a specific non-leaf node in the tree, and we construct two matrices, XST and XNR, of size m × (n + 2) corresponding to the two color channels, as shown in Figure 4. (i) For the ST channel, all the elements in the matrix XST are initialized2 to -2. With the upper left corner of the matrix as the origin of the coordinate axis, given the split point j at the i-th tree layer (top-down direction), we directly set the element at (i-1, j+1) by zero. Besides, if the left span of the split point is an EDU, then we set the element at (i, j) by -1, and the right span is processed in a similar way. With this method, we can recursively construct the tree diagram from top to down. Additionally, some EDU positions are actually shared in the matrix, and this does not affect the understanding of these nodes. For the example in Figure 4, although e2 and e3 share a same position in the ST channel, the following two patterns in the matrix can still reveal an accurate representation of each node: N1 : 0 −2 −2 −1 N2 : −2 0 −1 −2 (2) (ii) For the NR channel, we set the positions representing non-leaf nodes to specific N-R labels and the positions of leaf nodes to −1 and other nonnode positions to zero. For the automatically parsed trees, we directly use our model outputs to build the tree diagram with two color channels, X′ ST and X′ NR. And the 2We set these non-node positions to -2 in two reasons: (i) we apply a log-softmax function to the attention weights for split point ranking with the output ranging (−∞, 0]; (ii) we simply set the non-node positions by -2 to distinguish them from the leaf nodes marked with -1. two matrices of size m × (n + 2) are initialized with zero. (i) For the ST channel, as stated before, a set of attention weights are assigned to the encoder outputs during pointing and a split point is selected according to the weights. Obviously, each split point corresponds to a group of attention weights (after log-softmax). Therefore, we directly add these n-dimensional attention weights of each split point in the i-th tree layer (top-down direction) to the i-th line of X′ ST. Notably, the first and last columns of the matrices are actually placeholders initialized with unlearnable scalars representing leaves or non-node positions, so we only add the split point attention weights to the range from 1 to n in each row. (ii) For the NR channel, we simply replace these elements corresponding to split points in X′ ST with predicted N-R labels3 and other elements keep the same as XNR. Alternatively, only the replaced elements in the matrix X′ NR are learnable, while other positions serve as static features in the image. In this way, the model outputs are also abstracted as a tree diagram with two color channels. Through the above methods, we achieve graphical representation for both gold standard and automatically predicted DRS trees. And the graphical representation can provide our model with a global perspective, which makes the global optimization (Subsection 4.2) of DRS parsing possible. 4.2 Adversarial Model Learning For model learning, we have two goals: (i) learning of DRS parsing at each time step for local optimization and (ii) learning an adversarial bot to evaluate 3Here, we need to map the attention score, sj,i ∈Rk, to a specific N-R label. Since the argmax function does not support gradient calculation, we give an alternative solution: Lj,i = Fsigmoid(wl · sj,i + bl) × K, where K is the number of N-R labels and Lj,i ∈R1 is the learnable N-R label. 3950 the pros and cons of the entire tree for global optimization. For the first goal, we use two negative log-likelihood loss terms to optimize the parsing model. For split point ranking, we use Ls to maximize the probability of correct split point selection at each decoding step. For N-R prediction, given the selected split point, we use Lnr to maximize the probability of correct N-R labeling for the split point. Since the convergence speeds of the two loss terms are different, we add two loss weights before the loss terms to balance the model training as: LDRS = α1Ls + α2Lnr (3) For the second goal, we explore to learn from the entire DRS tree for global optimization. To that end, we produce an adversarial bot in our parser to estimate the generated DRS tree diagrams, as shown in Figure 4. Since the composition and sources of gold and generated tree diagrams are completely different, we use two isomorphic feature extractors to understand the two kinds of images separately. For feature extraction, based on such a 2D image-like representation, we perform convolution on every 3 × (n + 2) window to dig out the structural details of the entire tree: ϱ(f) win = Frelu(w(f) · Xwin + b(f)) (4) Then we perform max-pooling in each nonoverlapping 3 × 1 window for feature extraction, and the resulting matrices are reshaped as ϱ ∈R1×D to serve as the distributed representation of the tree. In this work, we do not just need an excellent discriminator expert in classification, we need the adversarial nets to continuously give feedback to our parsing model even when the generated trees are correctly classified. On this basis, we leverage Least Squares Generative Adversarial Network (LSGAN) (Mao et al., 2017) as our adversarial bot which has proven to perform more stable and face less problem of vanishing gradients than the original GAN. Formally, our adversarial nets consist of two parts: (i) a generative net G to capture the data distribution pz over the training data X and (ii) a discriminative net D to estimate the probability that a sample comes from X rather than pz. On this basis, given the distributed representation of the gold tree x and fake tree z, we formulate the loss functions as follows: min D V (D) = 1 2Ex∼pdata(x)[(D(x) −b)2] +1 2Ez∼pz(z)[(D(G(z)) −a)2] (5) min G V (G) = 1 2Ez∼pz(z)[(D(G(z)) −c)2] (6) Similar to Mao et al. (2017), we set a = 0 and b = c = 1 to make G generate samples as real as possible. Technically, the generator G consists of the parsing model and the feature extractor for fake trees, and the discriminator is an MLP (In: feature size (ϵ), Hidden: ϵ/2, Out: 1) without the sigmoid activation function. Therefore, when learning G, parameters of the parsing model and the feature extractor for fake trees are updated. Likewise, parameters of the discriminator and the feature extractor for real trees are learned when tuning D. At this time, we have a traditional loss term to train the top-down parser at each splitting step and two adversarial loss terms to estimate the entire DRS tree for global optimization. It is worth mentioning that we first optimize the LDRS for 7 epochs to warm up the model parameters, and then the adversarial nets join the training process for global optimization of DRS parsing. 5 Experimentation 5.1 Experimental Settings Datasets. Following our previous work (Zhang et al., 2020), we utilize both the English RST Discourse Treebank (RST-DT) (Carlson et al., 2001) and the Chinese Connective-driven Discourse TreeBank (CDTB) (Li et al., 2014b) as the benchmark corpora for experimentation. Here, we give a brief introduction to the two corpora: • The RST-DT corpus contains 385 news articles (347 for training and 38 for testing) from the Wall Street Journal (WSJ). Following previous work, we randomly select 34 documents from the training corpus as the development corpus for parameter tuning. And we also binarize those non-binary subtrees in RST-DT with right-branching (Sagae and Lavie, 2005) for preprocessing. • The Chinese CDTB corpus is motivated by taking advantages of both the English RST-DT corpus and the PDTB corpus (Prasad et al., 2008). The CDTB corpus annotates each paragraph as a Connective-driven Discourse Tree (CDT). The corpus consists of 500 newswire articles which are further segmented into 2336 paragraphs and 10650 EDUs. The corpus is divided into three parts with 425 articles (2002 CDT trees) for training, 25 articles (105 CDT trees) for validation, and 50 articles (229 CDT trees) for testing. 3951 Metrics. Following previous studies, we measure the performance of bare tree structure (S), tree structure labeled with nuclearity (N), and tree structure labeled with rhetorical relation (R). Recently, the Full (F) indicator is used to estimate the tree structure labeled with both nuclearity and relation categories. However, since current performances on S, N and R are imbalanced, the performance on F is much limited by relation prediction. In other words, the Full score may underestimate the performance in span and nuclearity prediction. In this work, we combine nuclearity and rhetorical relation tags for joint N-R prediction aiming to reduce the uncertainty of the Full measure. Moreover, since RST-Parseval (Marcu, 2000) overestimates the DRS parsing performance to a certain extent, (Morey et al., 2017; Mabona et al., 2019; Zhang et al., 2020; Koto et al., 2021) adopt the original Parseval to reveal the actual performance level of DRS parsing. Following these studies, we also use the original Parseval for evaluation and report the micro-averaged F1 scores by default. Hyper-Parameter Setting. For word representation, we employed the 300D vectors of GloVe (Pennington et al., 2014) and the 1024D vectors of ELMo (Peters et al., 2018) for RST-DT and the 300D vectors of Qiu et al. (2018) (Qiu-W2V) for CDTB, and we did not update these vectors during training. The English POS tags were obtained through the Stanford CoreNLP toolkit (Manning et al., 2014), the Chinese tags were borrowed from Chinese PTB, and all the POS embeddings were optimized during training. For model learning, we used the development set to fine-tune the parameters in Table 1, and the number of parameter search trials was around 20. All the experiments based on the above-mentioned settings were conducted on GeForce RTX 2080Ti GPU, and the codes will be published at https://github.com/ NLP-Discourse-SoochowU/GAN_DP. 5.2 Experimental Results Comparison between different system settings. As stated before, we explore to make possible improvements to the top-down architecture of Zhang et al. (2020). Here, we study the effects of these simplification methods based on our simplified architecture. For clarity, we remove the adversarial learning process in each system, and the results are presented in Table 2. For the RST-DT corpus, the first two rows show that the top-down parser Parameter EN CN POS embedding 30 30 Uni-directional GRU 512 512 BiGRU 256 256 Biaffine-MLP-Split 128 64 Biaffine-MLP-NR 128 128 Boundary feature size 30 Dropout rate 0.2 0.33 Warm up epochs 7 7 Training epochs 20 20 Batch size (DTs) 5 64 Learning rate of D 1e-4 5e-4 Learning rate of other nets 1e-3 1e-3 α1 0.3 0.3 α2 1.0 1.0 Table 1: Fine-tuned hyper-parameters. Systems S N R F EN T2D 70.7 58.3 46.5 45.2 + DS 69.2 57.7 46.1 44.9 + TC 70.6 57.9 46.1 44.4 CN T2D 82.5 57.3 51.7 48.2 + DS 83.2 57.8 52.7 49.0 + DS&TC 85.2 57.3 53.3 45.7 Table 2: Results under different model settings. “T2D” denotes our simplified architecture, which excludes the dummy split points and only uses two classifiers for DRS parsing; “DS” means the dummy split points are used; “TC” means three classifiers are used. performs worse when dummy split points are used, and the decline is obvious in tree structure parsing. Then, we further apply three classifiers to the simplified architecture, and the results (lines 1 and 3) show that the Full score drops by 1.8% for lack of correlation between the three learning goals. For the CDTB corpus, due to the differences in languages and annotation strategies, the situation is quite different. Specifically, lines 4 and 5 show that the top-down parser performs better on all the four indicators when using dummy split points (Zhang et al., 2020). Based on the better-performing parser using “DS”, we further report its performance with three independent classifiers used, and the results (line 6) show that the Full score still drops a lot (6.7%), which suggests the necessity of joint N-R prediction. Considering the above results, in the following, we separately use two sets of model settings for different languages. For English, we build our final model based on the simplified architecture without dummy split points. For Chinese, we build our final model based on the architecture of Zhang et al. (2020). For both systems, we only use two classifiers for DRS parsing. 3952 Systems S N R F EN Final 71.8 59.5 47.0 45.9 - Advers. bot 70.7 58.3 46.5 45.2 CN Final 84.9 58.4 54.5 50.3 - Advers. bot 83.2 57.8 52.7 49.0 Table 3: Comparison on the adversarial bot. Comparison on the adversarial bot. Here, we perform experiments to explore the effects of the adversarial learning approach, and the experimental results are presented in Table 3. For the RST-DT corpus, the results show that our adversarial model setting can improve the performance on all the four indicators, especially in structure and nuclearity prediction. Similarly, the results on the CDTB corpus show that our adversarial method still works much better than the unreinforced parser in structure, relation, and full detection. The overall results indicate that the global optimization method we use is definitely effective, although the effectiveness has not yet reached the level of qualitative change. In fact, as a preliminary attempt for global optimization of DRS parsing, this research still has much room for improvement which deserves further exploration. Comparison with previous studies. In this part, we compare with seven previous state-of-the-art (SOTA) parsers on text-level DRS parsing. Here, we briefly review these studies as follows: • Ji and Eisenstein (2014), a shift-reduce parser with an SVM that is trained by their extracted latent features. In this paper, we compare with the updated version of their parser (designated as “JE2017-updated”) (Morey et al., 2017). • Feng and Hirst (2014), a two-stage greedy parser with linear-chain CRF models and some handengineered features. • Li et al. (2016), an attention-based hierarchical neural model with hand-crafted features used. • Braud et al. (2016), a hierarchical BiLSTM model that leverages information from various sequence prediction tasks. • Braud et al. (2017), a transition-based neural model with both cross-lingual information and hand-crafted features used. • Mabona et al. (2019), a generative model with a beam search algorithm used for DRS parsing. Systems S N R F EN JE2017-updated 64.1 54.2 46.8 46.3 Feng and Hirst (2014) 68.6 55.9 45.8 44.6 Li et al. (2016) 64.5 54.0 38.1 36.6 Braud et al. (2016) 59.5 47.2 34.7 34.3 Braud et al. (2017) 62.7 54.5 45.5 45.1 Mabona et al. (2019) 67.1 57.4 45.5 45.0 Zhang et al. (2020) 67.2 55.5 45.3 44.3 Ours (GloVe) 69.9 57.3 46.3 45.0 Ours (ELMo) 71.8 59.5 47.0 45.9 CN Zhang et al. (2020) 85.2 57.3 53.3 45.7 Zhang et al. (2020)* 84.0 59.0 54.2 47.8 Ours (Qiu-W2V) 84.9 58.4 54.5 50.3 Table 4: Performance comparison with previous work. Results of the first five lines are directly borrowed from (Morey et al., 2017). “*” denotes the updated results based on the strict evaluation metric we use. • Zhang et al. (2020), a top-down neural architecture tailored for text-level DRS parsing. Different from many previous studies, this parser is a pure neural parser without using any additional handcrafted features. For the RST-DT corpus, the results are presented in the upper part of Table 4. From the results, although our previous top-down parser (Zhang et al., 2020) can achieve good results without using handcrafted features, the performance is still far from perfect. Comparing our GloVe-based top-down parser with previous state-of-the-art parsers, our parser performs better than most previous ones due to its ability in leveraging global context and the adversarial learning strategy. Furthermore, comparing the final parser (line 9) with previous work, our ELMo-based parser can further improve the performance on all the four indicators, and the improvements on structure (4.7%) and nuclearity (3.7%) are significant. Obviously, the contextualized word representation can greatly improve the parsing performance, especially in such a task with small-scale data corpora. For the CDTB corpus, we explore to employ a more strict metric4 for performance evaluation and the overall results are presented in the lower part of Table 4. In comparison with previous work, our parser achieves comparable performance in nuclearity and relation prediction and much better results on the other two indicators, which proves the usefulness of the adversarial nets we use. In 4We borrow the strict evaluation method from https: //github.com/NLP-Discourse-SoochowU/t2d_ discourseparser for evaluation in this study, and report the macro-averaged F1-scores for performance. 3953 Systems S N R F EN Koto et al. (2021) 73.1 62.3 51.5 50.3 Ours (XLNet) 76.3 65.5 55.6 53.8 - Advers. bot 76.1 64.4 54.3 52.9 CN Ours (Qiu-W2V) 84.9 58.4 54.5 50.3 Ours (XLNet) 86.6 65.0 62.1 55.4 - Advers. bot 85.8 64.5 60.5 53.7 Table 5: Performance comparison with LMs used. particular, compared with previous parsers, our parser performs significantly better on “F” due to the joint prediction of nuclearity and relation categories. This suggests the robustness of our simplified parser with only two classifiers. Moreover, since the two top-down DRS parsers in the table show similar results on “R”, we speculate that the Chinese rhetorical relation prediction has encountered a bottleneck to some extent, which requires more effort to be invested. Performances based on the SOTA language models. Recently, more and more researchers (Shi et al., 2020; Koto et al., 2021) propose to improve DRS parsing performance through powerful language models (LMs) like Bert (Devlin et al., 2019) and XLNet (Yang et al., 2019). Following these studies, in this work, we perform additional experiments on the XLNet-base models in (Yang et al., 2019) and (Cui et al., 2020) for the RST-DT and CDTB corpus, respectively. For better model integration, we slightly adjust the previously described model architecture5, more specifically, the EDU encoder. We first use a pre-trained LM to encode each entire discourse where each EDU is attached with the [SEP] and [CLS] tokens and then take the LM outputs corresponding to [CLS] as our EDU representation. Moreover, we segment each document according to the maximum length of 768 tokens and encode these text segments one by one to avoid the problem of memory overflow. For the RST-DT corpus, we report the results of the recent Bert-based top-down parser (Koto et al., 2021) for comparison. For the CDTB corpus, we compare with our previously described system based on traditional word vectors, and the overall results are shown in Table 5. From the results we find that our parsers achieve superior results when using the contextualized XLNet for experimentation, which suggests the great effectiveness of pre-trained LMs in such a task with 5Adjusted model parameters are shown in Appendix. Systems UAS LAS Wang et al. (2017)* 61.5 47.8 Yu et al. (2018)* 61.9 48.4 Kobayashi et al. (2020)* 64.9 48.5 Ours (Final) 72.3 57.6 - Advers. bot 71.4 56.5 Table 6: Evaluation on dependency trees. “*” denotes the results are borrowed from (Kobayashi et al., 2020). limited corpus size. Moreover, the ablation study on the adversarial learning strategy further demonstrates the usefulness of our proposed method. It should be noted that we report the performance using LMs in this paper never mean to advocate using pre-trained LMs or blindly pursuing performance improvements in DRS parsing. Sometimes, the rewards generated by the large-scale LMs could be quite different from and much more effective than that generated by language phenomena, which may hinder the study on the relatively shallow (compared with powerful LMs) yet valuable discourse features. With this in mind, it is reasonable to perform ablation study using simple word representation to explore useful discourse features and report the performance on powerful LMs for reference. 5.3 Analysis and Discussion Performance Evaluation of Dependency Trees. Recently, discourse-level dependency structure has attracted more and more attention. Here, we explore whether the proposed global optimization method can improve the RST dependency analysis to some extent. To achieve this, we first convert the predicted DRS trees into dependency trees as Kobayashi et al. (2020) did and then perform evaluation on the converted dependencies labeled (LAS) and unlabeled (UAS) with rhetorical relations, and the results are shown in Table 6. Firstly, lines 1 to 4 show that our parser can greatly outperform previous systems in terms of both UAS and LAS indicators. Secondly, the last two rows show that the global optimization of constituency trees can simultaneously improve the dependency performance, which further proves the usefulness of our proposed adversarial method. Remarkable Progress in DRS Parsing. Compared with Chinese DRS parsing where each paragraph is annotated as a DT, the English parsing with 313 DTs for training is much more challenging. Nevertheless, results in Table 4 and Table 5 show that our parser can largely outperform previous 3954 Systems NN/23% NS/61% SN/16% Ours (GloVe) 43.3 62.9 55.7 Ours (ELMo) 47.8 64.1 58.5 Ours (XLNet) 56.7 67.4 69.6 - Advers. bot 58.8 66.4 66.7 Table 7: Performance on nuclearity detection. /5 /5 5 /5 /5 5 loss loss loss loss step step step step Figure 5: Convergence of our parsing model over different learning rates (LRs). state-of-the-art parsers on “Full”. (i) For nuclearity prediction, we display the results of our parsers on each nuclearity category to explore where the improvement comes from, as shown in Table 7. From the results, it’s obvious that the LM we use plays a big role in nuclearity prediction, and the proposed adversarial method can further improve the performance to a certain extent. (ii) For relation prediction, the classification problem with 18 coarse-grained relation tags (RST-DT) is really a challenge. From the results in Table 4 we can find that the progress in relation prediction is much limited in recent decade for the lack of data. And most of previous state-of-the-art parsers employee a variety of hand-engineered features for good performance. Hopefully, the experimental results in Table 5 show that powerful LMs can free data-driven models from corpus size limitation and thus our XLNet-based parser strongly outperforms JE2017updated (Morey et al., 2017) by 18.8% on “R”. The results of our parsers on each rhetorical relation category are shown in Appendix. Discussion on Adversarial Learning. Similar to previous GAN work, improving the quality of the generated tree images is really a challenge, and the instability of the adversarial learning process is another intractable issue. In order for our model to continuously modify the generated images even when they are correctly classified, we leverage a least squares loss in our system for model learning. To avoid the over-learning of the discriminator, we tune it with a moderate learning rate and parameter scale. Intuitively, the convergence of our model over different learning rates is presented in Figure 5. From the results, as the learning rate of the discriminator increases, the fluctuation of the loss value becomes larger, and it is hard to reduce the generator loss. In these four cases, the first group seems to be more stable and in line with our expectations. Therefore, we set the learning rate to 1e-4 in our systems for experimentation. Notably, we also tried the sigmoid cross entropy loss in this research which performs much worse than the LSGAN we use. For reference, we also present the model convergence over different loss functions in Appendix for reference. 6 Conclusion In this research, we explored a global optimization method based on recent top-down frameworks. Particularly, we proposed a novel strategy to transform both gold standard and predicted DRS trees into tree diagrams with two color channels. On this basis, we produced an LSGAN-based adversarial bot between gold and fake trees for global optimization. Experimental results on two popular corpora showed that our proposed adversarial approach is effective in DRS parsing and has established new state-of-the-art results for both corpora. Acknowledgements Here, the first author (Longyin Zhang) would like to thank his fiancee, Dr. Xin Tan, for her valuable discussion on this research. This work was supported by the National Key R&D Program of China under Grant No. 2020AAA0108600, Projects 61876118 and 61976146 under the National Natural Science Foundation of China and the Priority Academic Program Development of Jiangsu Higher Education Institutions. References Chlo´e Braud, Maximin Coavoux, and Anders Søgaard. 2017. Cross-lingual RST discourse parsing. In EACL, pages 292–304, Valencia, Spain. Association for Computational Linguistics. Chlo´e Braud, Barbara Plank, and Anders Søgaard. 2016. Multi-view and multi-task training of RST discourse 3955 parsers. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1903–1913, Osaka, Japan. The COLING 2016 Organizing Committee. Lynn Carlson, Daniel Marcu, and Mary Ellen Okurovsky. 2001. Building a discourse-tagged corpus in the framework of Rhetorical Structure Theory. In Proceedings of the Second SIGdial Workshop on Discourse and Dialogue. Francine Chen and Yan-Ying Chen. 2019. Adversarial domain adaptation using artificial titles for abstractive title generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2197–2203, Florence, Italy. Association for Computational Linguistics. Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Shijin Wang, and Guoping Hu. 2020. Revisiting pre-trained models for Chinese natural language processing. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 657–668, Online. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Timothy Dozat and Christopher D. Manning. 2017. Deep biaffine attention for neural dependency parsing. In The 5th International Conference on Learning Representations, ICLR2017. Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11–21, Brussels, Belgium. Association for Computational Linguistics. Vanessa Wei Feng and Graeme Hirst. 2014. A lineartime bottom-up discourse parser with constraints and post-editing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 511–521, Baltimore, Maryland. Association for Computational Linguistics. Hugo Hernault, Helmut Prendinger, Mitsuru Ishizuka, et al. 2010. Hilda: A discourse parser using support vector machine classification. Dialogue and Discourse, 1(3). Yangfeng Ji and Jacob Eisenstein. 2014. Representation learning for text-level discourse parsing. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13–24, Baltimore, Maryland. Association for Computational Linguistics. Yangfeng Ji and Noah A. Smith. 2017. Neural discourse structure for text categorization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 996–1005, Vancouver, Canada. Association for Computational Linguistics. Shafiq Joty, Giuseppe Carenini, Raymond Ng, and Yashar Mehdad. 2013. Combining intra- and multisentential rhetorical parsing for document-level discourse analysis. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 486–496, Sofia, Bulgaria. Association for Computational Linguistics. Naoki Kobayashi, Tsutomu Hirao, Hidetaka Kamigaito, Manabu Okumura, and Masaaki Nagata. 2020. Topdown rst parsing utilizing granularity levels in documents. In Association for the Advancement of Artificial Intelligence 2020, AAAI2020. Fajri Koto, Jey Han Lau, and Timothy Baldwin. 2021. Top-down discourse parsing via sequence labelling. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 715–726, Online. Association for Computational Linguistics. Jiwei Li, Rumeng Li, and Eduard Hovy. 2014a. Recursive deep models for discourse parsing. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2061– 2069, Doha, Qatar. Association for Computational Linguistics. Qi Li, Tianshi Li, and Baobao Chang. 2016. Discourse parsing with attention-based hierarchical neural networks. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 362–371, Austin, Texas. Association for Computational Linguistics. Yancui Li, wenhe Feng, jing Sun, Fang Kong, and Guodong Zhou. 2014b. Building chinese discourse corpus with connective-driven dependency tree structure. In Proceedings of EMNLP 2014, pages 2105– 2114. Xiang Lin, Shafiq Joty, Prathyusha Jwalapuram, and M Saiful Bari. 2019. A unified linear-time framework for sentence-level discourse parsing. In ACL, pages 4190–4200, Florence, Italy. Association for Computational Linguistics. Linlin Liu, Xiang Lin, Shafiq Joty, Simeng Han, and Lidong Bing. 2019. Hierarchical pointer net parsing. In EMNLP-IJCNLP, pages 1006–1016, Hong Kong, China. Association for Computational Linguistics. Amandla Mabona, Laura Rimell, Stephen Clark, and Andreas Vlachos. 2019. Neural generative rhetorical structure parsing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference 3956 on Natural Language Processing (EMNLP-IJCNLP), pages 2284–2295, Hong Kong, China. Association for Computational Linguistics. William C Mann and Sandra A Thompson. 1988. Rhetorical structure theory: Toward a functional theory of text organization. Text-Interdisciplinary Journal for the Study of Discourse, 8(3):243–281. Christopher Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In Proceedings of 52nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 55–60, Baltimore, Maryland. Association for Computational Linguistics. X. Mao, Q. Li, H. Xie, R. Y. K. Lau, Z. Wang, and S. P. Smolley. 2017. Least squares generative adversarial networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2813–2821. Daniel Marcu. 2000. The Theory and Practice of Discourse Parsing and Summarization. MIT Press, Cambridge, MA, USA. Mathieu Morey, Philippe Muller, and Nicholas Asher. 2017. How much progress have we made on RST discourse parsing? a replication study of recent results on the RST-DT. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1319–1324, Copenhagen, Denmark. Association for Computational Linguistics. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, Doha, Qatar. Association for Computational Linguistics. Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 2227–2237, New Orleans, Louisiana. Association for Computational Linguistics. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The penn discourse treebank 2.0. In LREC 2008. Yuanyuan Qiu, Hongzheng Li, Shen Li, Yingdi Jiang, Renfen Hu, and Lijiao Yang. 2018. Revisiting correlations between intrinsic and extrinsic evaluations of word embeddings. In CCL & NLP-NABD 2017, pages 209–221. Springer. Kenji Sagae and Alon Lavie. 2005. A classifier-based parser with linear run-time complexity. In Proceedings of the Ninth International Workshop on Parsing Technology, pages 125–132, Vancouver, British Columbia. Association for Computational Linguistics. Ke Shi, Zhengyuan Liu, and Nancy F. Chen. 2020. An end-to-end document-level neural discourse parser exploiting multi-granularity representations. CoRR, abs/2012.11169. Yizhong Wang, Sujian Li, and Houfeng Wang. 2017. A two-stage parsing method for text-level discourse analysis. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 184–188, Vancouver, Canada. Association for Computational Linguistics. Jiawei Wu, Xin Wang, and William Yang Wang. 2019. Self-supervised dialogue learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3857–3867, Florence, Italy. Association for Computational Linguistics. Jiacheng Xu, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020. Discourse-aware neural extractive text summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5021–5031, Online. Association for Computational Linguistics. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. Nan Yu, Meishan Zhang, and Guohong Fu. 2018. Transition-based neural RST parsing with implicit syntax features. In Proceedings of the 27th International Conference on Computational Linguistics, pages 559–570, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Longyin Zhang, Yuqing Xing, Fang Kong, Peifeng Li, and Guodong Zhou. 2020. A top-down neural architecture towards text-level parsing of discourse rhetorical structure. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6386–6395, Online. Association for Computational Linguistics. Wei Zou, Shujian Huang, Jun Xie, Xinyu Dai, and Jiajun Chen. 2020. A reinforced generation of adversarial examples for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3486–3497, Online. Association for Computational Linguistics. Appendix A. Adversarial Model Learning Here, we display the convergence of our models with different loss functions and model settings 3957 applied, as shown in Figure 6. Comparing the first two legends, since the sigmoid cross entropy loss suffers from gradient vanishing, it’s hard for our model to update the generator net, and the generator loss keeps growing up. To avoid the over-learning of the discriminator net, we simplify the original discriminator network from a 3-layer MLP to a linear function, and the results are presented in Figure 6 (c). From the results, it’s really hard to train both generator and discriminator nets, and the adversarial learning in Figure 6 (c) seems to be meaningless for DRS parsing. LOSS LOSS LOSS (a) (b) (c) Figure 6: Figure (a) refers to our final model based on LSGAN; figure (b) refers to our model with the sigmoid cross entropy loss function used; based on figure (b), we use a simplified discriminator in figure (c). B. Results on Different Relation Categories Table 8 and Table 9 present the performances (F1scores) of our systems on each relation category in the RST-DT and CDTB corpora, respectively. C. Configurations of the LM-based Systems For better model integration, we slightly tuned the model hyper-parameters to adapt to the LM-based systems. For RST-DT, we set the LRs of all the nets to 1e-4, the hidden size of BiGRU to 384, the hidden size of uni-directional GRU to 768, and the batch size to 1 to suit the NVIDIA Tesla P40 Type-ratio% GloVe ELMo XLNet Elaborate-30.4 47.9 48.8 60.4 Joint-15.1 36.3 39.2 49.4 Attribution-11.7 77.9 83.0 86.7 Same-unit-10.9 70.3 71.9 75.9 Contrast-5.8 34.5 27.0 42.6 Explanation-3.8 11.3 16.1 21.7 Background-3.4 23.0 20.8 27.8 Temporal-3.0 15.4 15.5 34.6 Cause-2.9 3.7 7.7 18.5 Evaluation-2.2 4.1 0.0 10.5 Enablement-2.2 54.7 42.0 66.7 Comparison-1.7 12.5 12.9 36.7 Topic-change-1.6 7.7 11.1 40.0 Textual-org-1.3 20.0 28.6 53.3 Condition-1.2 42.1 29.0 62.5 Topic-comment-1.0 0.0 0.0 8.3 Manner-means-0.8 33.3 32.1 44.0 Summary-0.8 47.8 44.0 50.0 Table 8: Results on the RST-DT corpus. “ratio” means the proportion of each category label in the corpus. Type-ratio% Qiu-W2V XLNet 并列/ Same-unit-47.8 80.2 88.0 解说/ Explanation-12.6 50.0 60.7 因果/ Cause-9.4 32.5 55.9 顺承/ Consequent-7.1 4.1 58.1 目的/ Purpose-4.6 48.5 58.5 例证/ Example-3.4 10.5 34.5 总分/ Overall-branch-3.2 75.0 73.9 评价/ Evaluation-3.1 26.7 56.3 转折/ Contrast-2.7 69.0 75.0 背景/ Background-1.8 0.0 36.4 条件/ Condition-1.0 0.0 16.7 假设/ Suppose-1.0 0.0 66.7 递进/ Progressive-0.9 0.0 0.0 对比/ Comparison-0.8 0.0 40.0 推断/ Deduce-0.5 0.0 0.0 让步/ Concession-0.2 0.0 0.0 Table 9: Results on the CDTB corpus. GPU memory. For CDTB, we set the LRs of the discriminator, LM, and other nets to 5e-4, 1e-4, and 2e-5, respectively. We trained the LM-based systems for around 30 rounds and the other system settings remained the same as the aforementioned non-LM-based systems.
|
2021
|
305
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3958–3969 August 1–6, 2021. ©2021 Association for Computational Linguistics 3958 Exploring Discourse Structures for Argument Impact Classification Xin Liu1∗ Jiefu Ou1 Yangqiu Song1 Xin Jiang2 1Department of CSE, the Hong Kong University of Science and Technology 2Huawei Noah’s Ark Lab [email protected] [email protected] [email protected] [email protected] Abstract Discourse relations among arguments reveal logical structures of a debate conversation. However, no prior work has explicitly studied how the sequence of discourse relations influence a claim’s impact. This paper empirically shows that the discourse relations between two arguments along the context path are essential factors for identifying the persuasive power of an argument. We further propose DISCOC to inject and fuse the sentence-level structural discourse information with contextualized features derived from large-scale language models. Experimental results and extensive analysis show that the attention and gate mechanisms that explicitly model contexts and texts can indeed help the argument impact classification task defined by Durmus et al. (2019), and discourse structures among the context path of the claim to be classified can further boost the performance. 1 Introduction It is an interesting natural language understanding problem to identify the impact and the persuasiveness of an argument in a conversation. Previous works have shown that many factors can affect the persuasiveness prediction, ranging from textual and argumentation features (Wei et al., 2016), style factors (Baff et al., 2020), to the traits of source or audience (Durmus and Cardie, 2018, 2019; ShmueliScheuer et al., 2019). Discourse relations, such as Restatement and Instantiation, among arguments reveal logical structures of a debate conversation. It is natural to consider using the discourse structure to study the argument impact. Durmus et al. (2019) initiated a new study of the influence of discourse contexts on determining argument quality by constructing a new dataset Kialo. ∗This work was done when Xin Liu was an intern at Huawei Noah’s Ark Lab. Thesis: Physical torture of prisoners is an acceptable interrogation tool. S1: Torture can help force prisoners to reveal information that could prevent attacks and save lives. O1: Torture is ineffective at getting prisoners to reveal desired information. O2: If torture is allowed, then it could easily be misused or performed in excess. S2: The knowledge that torture is acceptable and may be applied is in and of itself a strong incentive for prisoners to cooperate with their captors. S3: Interrogators and prison guards could torture prisoners solely to fulfill their own sadistic desires or out of a motivation for personal revenge. Support Support Oppose Support Oppose Result Result Contrast Result Not Impactful Impactful Impactful Restatement, Instantiation Figure 1: Example of an argument tree from Kialo. Stances, impact labels, and discourse relations are annotated in orange, red, and violet respectively. As shown in Figure 1, it consists of arguments, impact labels, stances where every argument is located in an argument tree for a controversial topic. They argue contexts reflect the discourse of arguments and conduct experiments to utilize historical arguments. They find BERT with flat context concatenation is the best, but discourse structures are not easily captured by this method because it is difficult to reflect implicit discourse relations by the surface form of two arguments (Prasad et al., 2008; Lin et al., 2009; Xue et al., 2015; Lan et al., 2017; Varia et al., 2019). Therefore, there is still a gap to study how discourse relations and their sequential structures or patterns affect the argument impact and persuasiveness prediction. In this paper, we acquire discourse relations for argument pairs with the state-of-the-art classifier for implicit discourse relations. Then we train a BiLSTM whose input is the sequence of discourse relations between two adjacent arguments to predict the last argument’s impact, and the performance is comparable to that of a BiLSTM on raw text. This indicates that a sequence of discourse re3959 lations is one of the essential factors for identifying the persuasive power of an argument. Based on this intuition, we further propose a new model called DISCOC (Discourse Context Oriented Classifier) to explicitly produce discourse-dependent contextualized representations, fuse context representations in long distances, and make predictions. By simple finetuning, our model beats the backbone RoBERTa (Liu et al., 2019) over 1.67% and previous best model BERT over 2.38%. Extensive experiments show that DISCOC results in steady increases when longer context paths with discourse structures, e.g., stances and discourse relations, are provided. On the contrary, encoders with full-range attentions are hard to capture such interactions, and narrow-range attentions cannot handle complex contexts and even become poisoned. Our contributions can be highlighted as follows: 1. To the best of our knowledge, we are the first to explicitly analyze the effect of discourse among contexts and an argument on the persuasiveness. 2. We propose a new model called DISCOC to utilize attentions to imitate recurrent networks for sentence-level contextual representation learning. 3. Fair and massive experiments demonstrate the significant improvement; detailed ablation studies prove the necessities of modules. 4. Last, we discover distinct discourse relation path patterns in a machine learning way and conduct consistent case studies. Code is publicly released at https://github. com/HKUST-KnowComp/DisCOC. 2 Argument Tree Structure 2.1 Overview Kialo dataset is collected by Durmus et al. (2019), which consists of 47,219 argument claim texts from kialo.com for 741 controversial topics and corresponding impact votes. Arguments are organized as tree structures, where a tree is rooted in an argument thesis, and each node corresponds to an argument claim. Along a path of an argument tree, every claim except the thesis was made to either support or oppose its parent claim and propose a viewpoint. As shown in Figure 1, an argument tree is rooted at the thesis “Physical torture of prisoners is an acceptable interrogation tool.”. There is one claim to support this thesis (S1 in green) and one to oppose it (O2 in fuchsia). Moreover, S1 is supported by its child claim S2 and opposed by O1, and S3 holds the same viewpoint of O2. Stance / Impact Train Validation Test Pro 9,158 1,949 1,953 Con 8,695 1,873 1,891 Impactful 3,021 641 646 Medium Impact 1,023 215 207 Not Impactful 1,126 252 255 Table 1: Statistics of stances and impact labels in the training, validation, and test data. 2.2 Claim and Context Path As each claim was put in view of all its ancestral claims and surrounding siblings, the audience evaluated the claim based on how timely and appropriate it is. Therefore, the context information is of most interest to be discussed and researched in the Kialo dataset. We define that a claim denoted as C is the argumentative and persuasive text to express an idea for the audience, and a context path of a claim of length l is the path from the ancestor claim to its parent claim, denoted as (C0, C1, · · · , Cl−1) where Cl−1 is the parent of C. For simplicity, we may use Cl instead of C without causing ambiguity. The longest path of C starts from the thesis. Statistically, the average length of the longest paths is 3.5. 2.3 Argument Stance In a controversial topic, each argument claim except the thesis would have a stance, whether to support or oppose the argument thesis or its parent claim. In Kialo, users need to directly add a stance tag (Pro or Con) to show their agreement or disagreement about the chosen parent argument when they post their arguments. We use si to denote the stance whether Ci is to support or oppose its parent Ci−1 when i ≥1. The statistics of these stances are shown in Table 1. 2.4 Impact Label After reading claims as well as the contexts, users may agree or disagree about these claims. The impact vote for each argument claim is provided by users who can choose from 1 to 5. Durmus et al. (2019) categorize votes into three impact classes (Not Impactful, Medium Impact, and Impactful) based on the agreement and the valid vote numbers to reduce noise. We can see the overall distribution from Table 1. The argument impact classification is defined to predict the impact label y of C given the claim text C and its corresponding context path (C0, C1, · · · , Cl−1). 3960 Discourse Relations Reason Conjunction Contrast Restatement Result Instantiation Chosen Alternative Numbers 6,559 6,421 5,718 5,343 1,355 99 23 Table 2: Statistics of predicted discourse relations. 3 Discourse Structure Analysis 3.1 Argument Impact from the Perspective of Discourse As paths under a controversial topic are strongly related to Comparison (e.g., Contrast), Contingency (e.g., Reason), Expansion (e.g., Restatement), and Temporal (e.g., Succession) discourse relations (Prasad et al., 2008), we model the discourse structures from a view of discourse relations. The first step is to acquire discourse relation annotations. BMGF-RoBERTa (Liu et al., 2020) is the state-of-the-art model proposed to detect implicit discourse relations from raw text. In the following experiments, we use that as our annotation model to predict discourse relation distributions for each adjacent claim pair. Specifically, for a given argument claim Cl and its context path (C0, C1, · · · , Cl−1), we denote pdisco(Cl) = (r1, r2, · · · , rl) as a discourse relation path such that ri ∈R indicates the discourse relation between Ci−1 and Ci when i ≥ 1. In this work, we adopt the 14 discourse relation senses in CoNLL2015 Shared Task (Xue et al., 2015) as R. And we also define the corresponding distributed discourse relation path to be pdist(Cl) = (d1, d2, · · · , dl) such that di = F(Ci−1, Ci) is the predicted discourse relation distribution between claims Ci−1 and Ci (i ≥1) by a predictive model F. In experiments, F is BMGFRoBERTa1. 8 out of 14 relations appear in the predictions, and the statistics of 7 frequent predictions are shown in Table 2. As discourse contexts would affect the persuasive power of claims, we first discover the correlations between impacts and stances as well as correlations between impacts and discourse relations, illustrated in Figure 2. From the label distribution and correlations, we find there are some clear trends: 1) Stances have little influence on argument impact, but discourse relations do. Correlations indicate that it is the contents instead of standpoints that contribute to potential impacts; 2) It is a smart choice to show some examples to convince others 1The official open-source code is at https://github. com/HKUST-KnowComp/BMGF-RoBERTa. We train such a classifier on CoNLL2015 Shared Task training data, and achieve 57.57% accuracy on the test set. Figure 2: Impact label distributions, the correlations between labels and stances, and the correlations between labels and discourse relations. Normalization is applied to the columns. because Instantiation is more relevant to Impactful than any other relations; 3) Similarly, explaining is also helpful to make voices outstanding; 4) Restatement is also positively correlated with Impactful so that we can also share our opinions by paraphrasing others’ viewpoints to command more attention. On the contrary, Chosen Alternative is a risky method because the audience may object. To investigate the role of discourse relations in impact analysis, we design a simple experiment that a single-layer BiLSTM followed by a 2-layer MLP with batch normalization predicts the impact by utilizing the distributed discourse relation path pdist(Cl). For the purposes of comparison and analysis, we build another BiLSTM on the raw text. Each claim has [BOS] and [EOS] tokens to clarify boundaries and we use 300-dim pretrained GloVe word embeddings (Pennington et al., 2014) and remain them fixed. We set different thresholds for context path lengths so that we can control how many discourse relations or contexts are provided. From Figure 3, discourse features can result in comparable performance, especially when longer discourse paths are provided. Instead, the model with raw text gets stuck in complex contexts. 3.2 Discourse Context Oriented Classifier It is generally agreed that the informative context can help understand the text to be classified. However, it is still unclear how to determine whether a context is helpful. One drawback of a broader context is the increasing ambiguity, especially in the scenario of the argument context path from different users like the results shown in Figure 3. Take claims in Figure 1 for example, S1 and O2 give two different consequences to support or oppose 3961 Figure 3: Performance of BiLSTM on discourse relations and BiLSTM on raw text. the thesis. And O1 objects S1 by a contrast conclusion. It is hard to build a connection between the thesis and O1 if S1 is not given because it is challenging to build a connection between “reveal desired information” with “interrogation tool” without a precondition “Torture can help force prisoners to reveal information”. On the contrary, thesis and S2 are still compatible as S2 is also a kind of result. Hence, a recurrent model with the gating mechanism that depicts pair-wise relations and passes to the following texts makes more sense. LSTM has gates to decide whether to remember or forget during encoding, but it cannot handle long-range information with limited memory. Recently, transformer-based encoders have shown remarkable performance in various complicated tasks. These models regard sequences as fully connected graphs to learn the correlations and representations for each token. People assume that transformers can learn whether two tokens are relevant and how strong the correlation is by back-propagation. Table 3 illustrates different possible ways to aggregation context information. Transformer (Vaswani et al., 2017) and BERT (Devlin et al., 2019) adopt full-range attentions while TransformerXL (Dai et al., 2019) and XLNet (Yang et al., 2019) regard historical encoded representations as memories to reuse hidden states. SparseTransformer (Child et al., 2019), in the opposite direction, stacks hundreds of layers by narrow the attention scope by sparse factorization. Information can still spread after propagations in several layers. Inspired by these observations, we design DISCOC (Discourse Context Oriented Classifier) to capture contextualized features by localized attentions and imitate recurrent models to reduce noises from long distance context. As shown in Figure 4, DISCOC predicts the argument impact through three steps. Attention Representative Query Key & Value Full BERT Ci C0, · · · , Cl Memory XLNet Ci (C0, · · · , Ci−1) Context SparseTransformer Ci Ci−1 Table 3: Different attention mechanisms. The Memory attention freezes the historical representations so that gradients of Ci would not propagate to the memory (C0, · · · , Ci−1). 3.2.1 Adjacent Claim Pair Encoding A difficult problem in such an argument claim tree is the noise in irrelevant contexts. A claim is connected to its parent claim because of a supporting or opposing stance, but claims in long distances are not high-correlated. Based on this observation, DISCOC conduct word-level representations by encoding claim pairs instead of the whole contexts. Given a claim Cl and its context path (C0, C1, · · · , Cl−1), all adjacent pairs are coupled together, i.e., (C0, C1), · · · , (Cl−1, Cl). We can observe that each claim appears twice except the first and the last. Next, each pair (Ci−1, Ci) is fed into the RoBERTa encoder to get the contextualized word representations. C0 and Cl are also encoded separately so that each claim has been encoded twice. We use −→ Hi to denote the encoded word representations of Ci when this claim is encoded with its parent Ci−1, or when it is computed alone as C0. Similarly, ←− Hi is the representations when encoding (Ci, Ci+1), or when it is fed as Cl. The encoding runs in parallel but we still use the term phase to demonstrate for better understanding. In 0-th phase, RoBERTa outputs −→ H0. One particular relationship between a parent-child pair is the stance, and we insert the one special token [Pro] or [Con] between them. It makes the sentiment and viewpoint of the child claim more accurate. On the other hand, discourse relations can also influence impact prediction, as reported in Section 3.1. However, discourse relations are not mutually exclusive, let alone predictions from BMGF-RoBERTa are not precise. Thus, we use the relation distributions as weights to get sense-related embeddings over 14 relations. We add additional W 1di for the parent and W 2di for the child except position embeddings and segment embeddings, where di is predicted discourse relation distribution for (Ci−1, Ci), W 1 and W 2 are trainable transformations for parents and children. Hence, RoBERTa outputs ←− Hi−1 and −→ Hi with the concatenation of two claims, [CTX] Ci−1 [SEP] [CLS] si Ci [SEP] in the i-th phase 3962 𝐶! Gated Transformer Layer 𝒗" 𝒗# 𝒗$%# 𝒗$ Impact Prediction 𝑦 𝑯" Phase 0 𝐶! 𝑯" 𝑯# Phase 1 𝑯$%# 𝑯# Phase 2 Phase 𝑙−1 𝑯$ 𝑯$%# Phase 𝑙 𝑯$ Phase 𝑙+ 1 𝒖" 𝒖# 𝒖$%# 𝒖$ Context Claim … … … 𝑠! 𝑾!𝒅! + +𝑾"𝒅! 𝑾!𝒅" + 𝐶#$! 𝑠#$! +𝑾"𝒅#$! 𝑾!𝒅# + 𝐶# 𝑠# +𝑾"𝒅# 𝑯* 𝐶" 𝑠" +𝑾"𝒅" … 𝑾!𝒅% + RoBERTa Figure 4: The architecture of DISCOC. si refers to the stance between Ci−1 and Ci, di is the discourse relation distribution obtained from F(Ci−1, Ci). Gray boxes represent the RoBERTa encoder and the violet is a gated transformer layer. [CTX], [CLS], and [SEP] are omitted in this figure. (i ∈{1, 2, · · · , l}), where [CTX] is a special token to indicate the parent claim and distinguish from [CLS]. Its embedding is initialized as a copy embedding of [CLS] but able to update by itself. And ←− Hl is computed by self-attention with no context in the last phase. In the end, each claim Ci has two contextualized representations ←− Hi and −→ Hi with limited surrounding context information. 3.2.2 Bidirectional Representation Fusion As claim representations {←− Hi} and {−→ Hi} from RoBERTa are not bidirectional, we need to combine them and control which of them matters more. The gated fusion (Liu et al., 2020) has been shown of a better mixture than the combination of multihead attention and layer normalization. We use it to maintain the powerful representative features and carry useful historical context information: ˆ Hi = MultiHead(←− Hi, −→ Hi, −→ Hi) (1) Aj = Sigmoid(W a[←− Hi, ˆ Hi]j + ba) (2) U i = A ⊙ˆ Hi + (1 −A) ⊙←− Hi, (3) where MultHead is the multi-head attention operation (Vaswani et al., 2017) whose query is ←− Hi and key & value is −→ Hi, Aj is the fusion gate for the j-th word embedding, [· · · ] is the concatenation, ⊙is the element product operation, and W a and ba are trainable matrix and bias for fusion gating. There are two reasons why using ←− Hi as the key of the multi-head attention: 1) [CLS] exists in the ←− Hi while the replaced token [CTX] appears in −→ Hi when i ̸= 0; 2) The position ids start from 0 when computing ←− Hi. The fused [CLS] token embedding ui is selected to represent the whole claim. 3.2.3 Context Path Information Gathering After extracting sentence-level claim representations u0, u1, · · · , ul, a transformer layer is used to gather longer-range context representations. The transformer layer includes a position embedding layer to provide sinusoid positional embeddings, a gated multi-head attention layer, a feed-forward network, and a layer normalization. The position embedding layer in DISCOC is different from that in the vanilla Transformer because it generates position ids in a reversed order, i.e. l, l −1, · · · , 0. The reversed order is helpful to model the contexts of variable length because the claim to be classified has the same position embedding. We also choose a gate to maintain the scale instead of using a residual connection. The gated transformer can generate meaningful representations because each claim can attend any other claims and itself. On the other hand, it perfectly fits the pair-wise encoding that imitates the recurrent networks to reduce the noise in irrelevant contexts and enhance the nearest context’s correlations. For example, in Figure 1, S2 is predicted as a result of S1 (with a probability 3963 of 39.17%) and a restatement (with a probability of 19.81%), and S1 is also a result of thesis (with a probability of 70.57%). Consequently, S2 is highrelevant to the thesis as a potential result if “physical torture is acceptable”, which can be captured by DISCOC. Finally, a 2-layer MLP with batch normalization is applied to vl of the last claim to predict its impact. 4 Experiments 4.1 Baseline Models Majority. The baseline simply returns Impactful. SVM. Durmus et al. (2019) created linguistic features for a SVM classifier, such as named entity types, POS tags, special marks, tf-idf scores for n-grams, etc. We report the result from their paper. HAN. HAN (Yang et al., 2016) computes document vectors in a hierarchical way of encoding and aggregation. We replace its BiGRU with BiLSTM for the sake of comparison. And we also extend it with pretrained encoders and transformer layers. Flat-MLMs. Pretrained masked languages, e.g., RoBERTa, learn word representations and predict masked words by self-attention. We use these encoders to encode the flat context concatenation like [CTX] C0 [SEP] [CTX] · · · [CTX] Cl−1 [SEP] as Segment A and [CLS] Cl [SEP] as Segment B. After getting [CTX] and [CLS] representations, a gated transformer layer and a MLP predict impacts. As for XLNet, we follow its default setting so that [CTX] and [CLS] are located at the end of claims. Interval-MLMs. Flat-MLMs regard the context path as a whole segment and ignore the real discourse structures except the adjacency, e.g., distances between two claims are missing. We borrow the idea from BERT-SUM (Liu and Lapata, 2019): segment embeddings of Ci are assigned depending on whether the distance to Cl is odd or even. Context-MLMs. We also compare pretrained encoders with context masks. A context mask is to localize the attention scope from the previous to the next. That is, Ci can attends words in Ci−1 and Ci+1 except for itself if 1 ≤i < l; C0 can only attend C0, C1, and Cl can only attend Cl−1, Cl. Memory-MLMs. XLNet utilizes memory to extend the capability of self-attention to learn super long historical text information. We also extend Flat-MLMs under this setting. 4.2 Model Configuration and Settings We use pretrained base models 2 in DISCOC and baselines. We follow the same finetuning setting: classifiers are optimized by Adam (Kingma and Ba, 2015) with a scheduler and a maximum learning rate 2e-5. The learning rate scheduler consists of a linear warmup for the 6% steps and a linear decay for the remaining steps. As for BiLSTM and HAN, the maximum learning rate is 1e-3. The hidden state dimension of linear layers, the hidden units of LSTM layers, and projected dimensions for attention are 128. The number of the multi-head attention is set as 8. Dropout is applied after each layer and the probability is 0.1. We pick the best context path length l for each model by grid search from 0 to 5 on validation data with the batch size of 32 in 10 epochs. Each model runs five times. 4.3 Argument Impact Classification Table 4 shows experimental results of different models. It is not surprising that neural models can easily beat traditional feature engineering methods in overall performance. But linguistic features still bring the highest precision. We also observe a significant 3.49% improvement with context vectors aggregating in HAN-BiLSTM compared with the simple BiLSTM. This indicates that it is necessary to model contexts with higher-level sentence features. Models with pretrained encoders benefit from representative embeddings, and HANRoBERTa achieves a gain of 5.49%. Flat context paths contain useful information to help detect the argument impact, but they also involve some noise from unrelated standpoints. Interval segment embeddings do not reduce noise but make BERT confused. It is counterintuitive that the segment embeddings depend on whether the distance is odd or even because BERT uses these for next sentence prediction. Since XLNet uses relative segment encodings instead of segment embeddings, Interval-XNet is better than Flat-XLNet in all three metrics. On the other hand, context masks bring another side effect for BERT, RoBERTa, and XLNet. Although these masks limit the attention scope at first sight, distant word information is able to flow to words with the increment of transformer layers. As a result, the uncertainty and attention bias increase after adding context masks. The memory storing context representations is also not helpful. The main reason is 2BERT-base-uncased, RoBERTa-base, and XLNet-basecased are downloaded from huggingface.co 3964 Model Precision Recall F1 Majority 19.43 33.33 24.55 SVM (Durmus et al., 2019) 65.67 38.58 35.42 BiLSTM 46.94 ± 1.08** 46.64 ± 0.71** 46.51 ± 1.11** HAN-BiLSTM 51.93 ± 1.37** 49.08 ± 1.52** 50.00 ± 1.49** HAN-BERT 53.72 ± 0.80** 53.45 ± 0.51** 53.46 ± 0.47** HAN-RoBERTa 55.71 ± 1.12** 55.95 ± 0.90** 55.49 ± 0.62** HAN-XLNet 53.91 ± 0.96** 55.56 ± 1.59** 54.53 ± 1.22** BERT (Durmus et al., 2019) 57.19 ± 0.92 55.77 ± 1.05** 55.98 ± 0.70** Flat-BERT 57.34 ± 1.56 57.07 ± 0.74* 56.75 ± 0.82** Flat-RoBERTa 58.11 ± 1.34 56.40 ± 0.61** 56.69 ± 0.63** Flat-XLNet 55.86 ± 1.74* 56.20 ± 1.17** 55.57 ± 0.95** Interval-BERT 55.56 ± 2.03* 55.52 ± 1.44** 55.34 ± 1.50** Interval-RoBERTa 58.31 ± 0.89 56.46 ± 1.44* 56.61 ± 1.24* Interval-XLNet 57.54 ± 0.50 56.78 ± 1.63* 56.52 ± 1.00** Context-BERT 54.96 ± 0.93** 56.09 ± 0.83** 55.44 ± 0.83** Context-RoBERTa 57.28 ± 0.97 55.29 ± 0.26** 55.83 ± 0.54** Context-XLNet 54.56 ± 0.71** 56.28 ± 1.22** 55.10 ± 0.72** Memory-BERT 54.33 ± 0.83** 57.57 ± 0.67* 55.22 ± 0.61** Memory-RoBERTa 55.08 ± 0.89** 55.55 ± 1.59** 54.76 ± 1.38** Memory-XLNet 55.44 ± 1.15** 55.45 ± 1.25** 54.91 ± 0.96** DISCOC 57.90 ± 0.70 59.41 ± 1.41 58.36 ± 0.52 Table 4: The averages and standard deviations of different models on the argument impact classification. The marker * refers to p-value < 0.05 and the marker ** refers to p-value < 0.001 in t-test compared with DISCOC. that the last claim’s update signal can not be used to update previous context representations. That is, Memory-models degenerate to models with frozen path features or even worth. DISCOC that we proposed can capture useful contexts and fuse in a comprehensive manner. Finally, DISCOC outperforms the second best model Flat-BERT over 1.61% and its backbone Flat-RoBERTa over 1.67%, the previous best model BERT by 2.38%. 4.4 Ablation Study Influence of the Context Path Length Different claims have different contexts. We only report the best performance with a fixed maximum context path length in Table 4. Figure 5 shows F1 scores of models with different hyper-parameters. DISCOC always benefits from longer discourse contexts while other models get stuck in performance fluctuation. Most models can handle one context claim, which is consistent with our idea of pair-wise encoding. DISCOC has consistent performance gains; instead, other models cannot learn long-distance structures better. Each token in Flat-RoBERTa and Interval-RoBERTa can attend all other tokens, and the two are the most competitive baselines. However, Context-RoBERTa and Memory-RoBERTa limit the attention scope to the tokens of one previous claim, making models unable to make use of long-distance context information. Figure 5: F1 scores of different models on varying the maximum path length. Model Precision Recall F1 DISCOC 57.90 59.41 58.36 DISCOC (E-BERT) 57.84 59.46 58.04 DISCOC (w/o StanceE) 58.68 58.12 57.74 DISCOC (w/o DiscoE) 57.81 58.42 57.29 DISCOC (F-BiLSTM) 58.58 57.87 57.72 DISCOC (F-Conv) 58.20 58.53 57.82 DISCOC (w/o GTrans) 56.04 54.71 54.78 Table 5: Ablation Studies of DISCOC. RoBERTa vs. BERT As shown in Table 4, there is little difference between the performance of RoBERTa variants and that of BERT variants. We conduct the experiment for DISCOC (E-BERT) with BERT as the encoder reported in Table 5. Its performance has achieved a significant boost over 1.29% despite the small gap between itself and DISCOC. 3965 Impactful Medium Impact Not Impactful Reason-Contrast Conjunction-Reason Restatement-Reason Restatement Conjunction-Contrast Contrast-Restatement Reason Contrast-Conjunction Chosen Alternative Restatement-Conjunction Conjunction-Restatement Restatement-Restatement Restatement-Contrast Contrast-Contrast Reason-Restatement Contrast-Instantiation Contrast-Reason Chosen Alternative-Reason Conjunction-Instantiation Conjunction-Conjunction-Restatement Contrast Restatement-Restatement Conjunction-Restatement-Conjunction Chosen Alternative-Conjunction Reason-Conjunction Conjunction-Reason-Conjunction Result-Reason Restatement-Result Conjunction-Conjunction Chosen Alternative-Restatement Table 6: Discourse path patterns that corresponding to the largest top 10 coefficients of the binary LR. Are Stances and Discourse Senses Helpful? We also remove either the stance token embedding or the discourse sense embeddings from DISCOC. The results in Table 5 suggest that both sides of structures are essential for modelling the correlation between the parent claim and the child claim. By comparison, discourse sense embeddings are more vital. Are Gated Transformers Necessary? We add a gated transformer layer to gather sentencelevel vectors. Such gathering is necessary for the proposed framework because each claim can only attend limited contexts. BiLSTM and convolutions can also be used for this purpose, so we replace the gated transformer layer with a BiLSTM or a convolutional layer. Moreover, we also remove it to make predictions by ul directly. The results in Table 5 show that the gated transformer is the irreplaceable part of DISCOC because it retains the contextualized representations and remains their scales. Simple removing it hurts recall enormously. 4.5 What Makes Claims Impactful? High-coefficient Discourse Relation Patterns We use Logistic Regression to mine several interesting discourse relation patterns. Detailed settings are described in Appendix A, and results including the most high-coefficient patterns are listed in Table 6. We observe that some discourse relation path patterns are distinguishing for classifying individual impact labels. Instantiation is a typical relation that only occurs in the top patterns of Impactful. Also, Restatement is relatively frequent for Impactful (5 of top 10), but it is the relation between the grandparent and the parent. Providing additional resources (Restatement-Result) or objecting others’ repetitions (Restatement-Contrast) can increase the persuasive power. For the Medium Impact class, its top 10 significant patterns are the longest on averDiscourse Patterns DISCOC DISCOC (w/o DiscoE) Reason-Contrast 65.56 43.33 Restatement 56.63 57.59 Reason 58.91 54.96 Conjunction-Reason 78.97 72.14 Conjunction-Contrast 80.64 66.17 Contrast-Conjunction 55.15 42.38 Restatement-Reason 38.00 37.35 Contrast-Restatement 66.10 76.24 Chosen Alternative 73.33 42.86 All 59.04 58.06 Table 7: F1 score differences between two best models on top 9 discourse relation patterns and all patterns. age. That indicates some views are usually considered ordinary in complex structures. Conjunction is the dominant relation (8 of top 10) so that we are suggested to avoid to go along with others. The case of Not Impactful is a little clearer, in the sense that it has a unique relation Chosen Alternative as one of the most significant patterns. Restatement also appears frequently, showing neither generalization, nor specification, nor paraphrasing of others’ views can help make claims stand out. Case Study In Appendix A, we define Pr(r1, · · · , rl) as the joint probability to generate the discourse relation path (r1, · · · , rl) given the context (C0, C1, · · · , Cl−1) and the claim Cl. For example, the Pr(Reason, Contrast) is 56.59% which corresponds to an Impactful claim “There is no evidence for this” with its parent claim “Our bodies know how to recognise and process current foods; changing them through genetic modification will create health issues”. Furthermore, we find 5 of top 5 and 8 of top 10 are voted as Impactful claims after sorting based on Pr(Reason, Contrast). For a complex pattern Restatement-Restatement appearing in both top patterns of the Impactful and the Not Impactful, 3 cases with the maximum probabil3966 ities are Not Impactful while the following 7 cases are Impactful. It is interesting that the thesis of the top 3 claims is the same discussion about an American politician. There are 25 Impactful claims and 22 Not Impactful claims in this topic, 24 of which are restatements of their parent claims. As for Restatement-Reason, the most top pattern of the Not Impactful, we find 7 of the top 10 claims relevant to politics, 2 of them about globalization, and one food-related. Therefore, there is no perfect answer in these quite controversial topics, and that is why Restatement and Reason appear frequently. Empirical Results On the other hand, we check the performance of testing examples to verify the effectiveness of these discourse relation patterns. We choose the best model of DISCOC, whose F1 score is 59.04% as well as the best model of DISCOC (w/o DiscoE) whose F1 score is 58.06%. We select testing examples with specific discourse patterns, and performance differences are shown in Table 7. DISCOC benefits from 7 of the top 9 patterns and the performance margins are even more significant than the improvement of the overall results. Without giving discourse relation patterns, the model still has trouble capturing such implicit context influences. Empirical results support our idea that implicit discourse relations could affect the persuasiveness. 5 Related Work There is an increasing interest in computational argumentation to evaluate the qualitative impact of arguments based on corpus extracted from Web Argumentation sources such as CMV sub-forum of Reddit (Tan et al., 2016). Studies explored the importance and effectiveness of various factors on determining the persuasiveness and convincingness of arguments, such as surface texture, social interaction and argumentation related features (Wei et al., 2016), characteristics of the source and audience (Durmus and Cardie, 2019; ShmueliScheuer et al., 2019; Durmus and Cardie, 2018), sequence ordering of arguments (Hidey and McKeown, 2018), and argument structure features (Li et al., 2020). The style feature is also proved to be significant in evaluating the persuasiveness of news editorial argumentation (Baff et al., 2020). Habernal and Gurevych (2016) conducted experiments in an entirely empirical manner, constructing a corpus for argument quality label classification and proposing several neural network models. In addition to the features mentioned above, the role of pragmatic and discourse contexts has shown to be crucial by not yet fully explored. Zeng et al. (2020) examined how the contexts and the dynamic progress of argumentative conversations influence the comparative persuasiveness of an argumentation process. Durmus et al. (2019) created a new dataset based on argument claims and impact votes from a debate platform kialo.com, and experiments showed that incorporating contexts is useful to classify the argument impact. Understanding discourse relations is one of the fundamental tasks of natural language understanding, and it is beneficial for various downstream tasks such as sentiment analysis (Nejat et al., 2017; Bhatia et al., 2015), machine translation (Li et al., 2014) and text generation (Bosselut et al., 2018). Discourse information is also considered indicative for various tasks of computational argumentation. Eckle-Kohler et al. (2015) analyzed the role of discourse markers for discriminating claims and premises in argumentative discourse and found that particular semantic group of discourse markers are highly predictive features. Hidey and McKeown (2018) concatenated sentence vectors with discourse relation embeddings as sentence features for persuasiveness prediction and showed that discourse embeddings helped improve performance. 6 Conclusion In this paper, we explicitly investigate how discourse structures influence the impact and the persuasiveness of an argument claim. We present DISCOC to produce discourse-dependent contextualized representations. Experiments and ablation studies show that our model improves its backbone RoBERTa around 1.67%. Instead, HAN and other attention mechanisms bring side effects. We discover distinct discourse relation path patterns and analyze representatives. In the future, we plan to explore discourse structures in other NLU tasks. Acknowledgements This paper was supported by the NSFC Grant (No. U20B2053) from China, the Early Career Scheme (ECS, No. 26206717), the General Research Fund (GRF, No. 16211520), and the Research Impact Fund (RIF, No. R6020-19 and No. R6021-20) from the Research Grants Council (RGC) of Hong Kong, with special thanks to the Huawei Noah’s Ark Lab for their gift fund. 3967 References Roxanne El Baff, Henning Wachsmuth, Khalid Al Khatib, and Benno Stein. 2020. Analyzing the persuasive effect of style in news editorial argumentation. In ACL, pages 3154–3160. Parminder Bhatia, Yangfeng Ji, and Jacob Eisenstein. 2015. Better document-level sentiment analysis from RST discourse parsing. In EMNLP, pages 2212–2218. Antoine Bosselut, Asli Celikyilmaz, Xiaodong He, Jianfeng Gao, Po-Sen Huang, and Yejin Choi. 2018. Discourse-aware neural rewards for coherent text generation. In NAACL-HLT, pages 173–184. Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. 2019. Generating long sequences with sparse transformers. CoRR, abs/1904.10509. Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G. Carbonell, Quoc Viet Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. In ACL, pages 2978– 2988. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186. Esin Durmus and Claire Cardie. 2018. Exploring the role of prior beliefs for argument persuasion. In NAACL-HLT, pages 1035–1045. Esin Durmus and Claire Cardie. 2019. Modeling the factors of user success in online debate. In WWW, pages 2701–2707. Esin Durmus, Faisal Ladhak, and Claire Cardie. 2019. The role of pragmatic and discourse context in determining argument impact. In EMNLP-IJCNLP, pages 5667–5677. Judith Eckle-Kohler, Roland Kluge, and Iryna Gurevych. 2015. On the role of discourse markers for discriminating claims and premises in argumentative discourse. In EMNLP, pages 2236–2242. Ivan Habernal and Iryna Gurevych. 2016. What makes a convincing argument? empirical analysis and detecting attributes of convincingness in web argumentation. In EMNLP, pages 1214–1223. Christopher Hidey and Kathleen R. McKeown. 2018. Persuasive influence detection: The role of argument sequencing. In AAAI, pages 5173–5180. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR. Man Lan, Jianxiang Wang, Yuanbin Wu, Zheng-Yu Niu, and Haifeng Wang. 2017. Multi-task attentionbased neural networks for implicit discourse relationship representation and identification. In EMNLP, pages 1299–1308. Jialu Li, Esin Durmus, and Claire Cardie. 2020. Exploring the role of argument structure in online debate persuasion. In EMNLP, pages 8905–8912. Junyi Jessy Li, Marine Carpuat, and Ani Nenkova. 2014. Assessing the discourse factors that influence the quality of machine translation. In ACL, pages 283–288. Ziheng Lin, Min-Yen Kan, and Hwee Tou Ng. 2009. Recognizing implicit discourse relations in the penn discourse treebank. In ACL, pages 343–351. Xin Liu, Jiefu Ou, Yangqiu Song, and Xin Jiang. 2020. On the importance of word and sentence representation learning in implicit discourse relation classification. In IJCAI, pages 3830–3836. Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In EMNLP-IJCNLP, pages 3728–3738. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. Bita Nejat, Giuseppe Carenini, and Raymond Ng. 2017. Exploring joint neural model for sentence level discourse parsing and sentiment analysis. In SIGDIAL, pages 289–298. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove: Global vectors for word representation. In EMNLP, pages 1532–1543. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind K. Joshi, and Bonnie L. Webber. 2008. The penn discourse treebank 2.0. In LREC. Michal Shmueli-Scheuer, Jonathan Herzig, David Konopnicki, and Tommy Sandbank. 2019. Detecting persuasive arguments based on author-reader personality traits and their interaction. In ACMUMAP, pages 211–215. Chenhao Tan, Vlad Niculae, Cristian DanescuNiculescu-Mizil, and Lillian Lee. 2016. Winning arguments: Interaction dynamics and persuasion strategies in good-faith online discussions. In WWW, pages 613–624. Siddharth Varia, Christopher Hidey, and Tuhin Chakrabarty. 2019. Discourse relation prediction: Revisiting word pairs with convolutional networks. In SIGdial, pages 442–452. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS, pages 5998–6008. Zhongyu Wei, Yang Liu, and Yi Li. 2016. Is this post persuasive? ranking argumentative comments in online forum. In ACL, pages 195–200. 3968 Nianwen Xue, Hwee Tou Ng, Sameer Pradhan, Rashmi Prasad, Christopher Bryant, and Attapol Rutherford. 2015. The conll-2015 shared task on shallow discourse parsing. In CoNLL, pages 1–16. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In NeurIPS, pages 5754– 5764. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alexander J. Smola, and Eduard H. Hovy. 2016. Hierarchical attention networks for document classification. In NAACL, pages 1480–1489. Jichuan Zeng, Jing Li, Yulan He, Cuiyun Gao, Michael R. Lyu, and Irwin King. 2020. What changed your mind: The roles of dynamic topics and discourse in argumentation process. In WWW, pages 1502–1513. 3969 A Discourse Relation Path Patterns To explicitly explore important high-order discourse relation patterns, we model the process of yielding a concrete discourse relation path pdisco(Cl) = (r1, · · · , rl) as a generative process. For a given context path (C0, C1, · · · , Cl−1) and the claim Cl, we define the pattern set as all possible patterns connected to Cl. Mathematically, it is denoted as P = Pl i=1 "l j=i R, where " is the Cartesian product. We assume that every ri ∈pdisco(Cl) is independent and identically distributed (i.i.d). Under this assumption, the joint probability of a given path of discourse relations (r1, · · · , rl) is Pr(r1, · · · , rl) = Πl i=1di[ri], (4) where di is the discourse relation distribution between Ci−1 and Ci, di[ri] is the probability of a specific relation sense ri. Observing the consistently increased performance of BiLSTM on discourse relations in Figure 3 when l starts from 1 to 3 and no noticeable enhancement with longer contexts, we analyze path-generated distributions for up to three previous claims. We compute the joint probabilities Pr(rl), Pr(rl−1, rl), Pr(rl−2, rl−1, rl) respectively and then concatenate these probabilities to get path pattern features x ∈R(|R|+|R|2+|R|3) where each dimension of x corresponds to the probability of a pattern belonging to P. Next, the feature vector x is fed into a logistic regression (LR) model to train a one-vs-rest binary classifier for each of the three impact labels. We report the largest top 10 coefficients of converged LR models in Table 6. Some relation path patterns are shown distinguishing for classifying individual impact labels. Coefficients vary differently among different LRs except for RestatementRestatement, which occurs in both Impactful and Not Impactful. In general, Instantiation is a typical relation that only occurs in the top patterns of Impactful. Also, Restatement is relatively frequent for Impactful (5 of top 10), but it is the relation between the grandparent and the parent. Providing additional resources (Restatement-Result) or objecting others’ repetitions (Restatement-Contrast) can increase the persuasive power. For the Medium Impact class, its top 10 significant patterns are the longest on average. That indicates some views are usually considered ordinary in complex structures. Conjunction is the dominant relation (8 of top 10) so that we are suggested to avoid to go along with others. The case of Not Impactful is a little clearer, in the sense that it has a unique relation Chosen Alternative as one of the most significant patterns. Restatement also appears frequently, showing that neither generalization, nor specification, nor paraphrasing of others’ views can help make claims stand out. These interesting correlations between discourse relation path patterns and argument quality could be further analysis from the linguistic perspective in future works.
|
2021
|
306
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3970–3979 August 1–6, 2021. ©2021 Association for Computational Linguistics 3970 Point, Disambiguate and Copy: Incorporating Bilingual Dictionaries for Neural Machine Translation Tong Zhang1,2, Long Zhang1,2, Wei Ye1,†, Bo Li1,2, Jinan Sun1, Xiaoyu Zhu3, Wen Zhao1, Shikun Zhang1,† 1 National Engineering Research Center for Software Engineering, Peking University 2 School of Software and Microelectronics, Peking University 3 BIGO {zhangtong17,zhanglong418, wye, zhangsk}@pku.edu.cn Abstract This paper proposes a sophisticated neural architecture to incorporate bilingual dictionaries into Neural Machine Translation (NMT) models. By introducing three novel components: Pointer, Disambiguator, and Copier, our method PDC achieves the following merits inherently compared with previous efforts: (1) Pointer leverages the semantic information from bilingual dictionaries, for the first time, to better locate source words whose translation in dictionaries can potentially be used; (2) Disambiguator synthesizes contextual information from the source view and the target view, both of which contribute to distinguishing the proper translation of a specific source word from multiple candidates in dictionaries; (3) Copier systematically connects Pointer and Disambiguator based on a hierarchical copy mechanism seamlessly integrated with Transformer, thereby building an end-to-end architecture that could avoid error propagation problems in alternative pipeline methods. The experimental results on Chinese-English and English-Japanese benchmarks demonstrate the PDC’s overall superiority and effectiveness of each component. 1 Introduction The past several years have witnessed the remarkable success of Neural machine translation (NMT), due to the development of sequence-to-sequence methods (Sutskever et al., 2014; Bahdanau et al., 2015; Vaswani et al., 2017). Since bilingual dictionaries cover rich prior knowledge, especially of low-frequency words, many efforts have been dedicated to incorporating bilingual dictionaries into NMT systems. These explorations can be roughly categorized into two broad paradigms. The first one transforms the bilingual dictionaries into pseudo parallel sentence pairs for training (Zhang †Corresponding authors. These patterns increase brake friction between tires and ground. 这些花纹可以增强轮胎与 地面 之间的制动摩擦 these pattern can increase tire and ground between of brake friction Source Input : Decoder Output : Bilingual Dictionary : Disambiguate rub friction clash conflict Copy Reference : These patterns increase brake Point 1 2 3 摩擦 : ？ Figure 1: Three key steps to translate with a bilingual dictionary: pointing, disambiguating and copying. This concrete illustrative example is chosen to conveniently show the primary intuition behind our method. and Zong, 2016; Zhao et al., 2020). The second one utilizes the bilingual dictionaries as external resources fed into neural architectures (Luong et al., 2015; Gulcehre et al., 2016; Arthur et al., 2016; Zhang et al., 2017b; Zhao et al., 2018a,b, 2019b), which is more widely used and the focus of this paper. In practice, bilingual dictionaries usually contain more than one translation for a word. From a highlevel perspective, we believe there are three critical steps to incorporate bilingual dictionaries into NMT models as shown in Figure 1: (1) pointing to a source word whose translation in dictionaries will be used at a decoding step, (2) disambiguating multiple translation candidates of the source word from dictionaries, and (3) copying the selected translation into the target side if necessary. Note that some works assume that only one translation exists for each word in dictionaries (Luong et al., 2015; Gulcehre et al., 2016). In this simplified scenario, the disambiguating step is unnecessary, hence the pointing and copying step can be merged into a single step similar to the classic copying mechanism (Gu et al., 2016). In more practical scenarios, however, this process suffers from the following bottlenecks corresponding to each step. 3971 (1) In the pointing step, semantic information of translations in dictionaries is underutilized. To locate source words whose translation in dictionaries may be used, some works (Luong et al., 2015; Gulcehre et al., 2016) use a classic copy mechanism, but in an oversimplified scenario mentioned above. More recent efforts further leverage statistics-based pre-processing methods (Zhao et al., 2018b, 2019b) to help identify, e.g., rare or troublesome source words. Note that the goal of locating a source word is to further use its translation in dictionaries. Intuitively, by exploring rich information of a source word’s translations in dictionaries, we can better understand the semantic meaning of the source word and distinguish whether we can its translation candidate. Unfortunately, this information is underutilized by most works, which could have boosted NMT performance, as shown in Section 5.2. (2) In the disambiguating step, the distinguishing information is from static prior knowledge or coarse-grained context information. To select the proper translation of one source word from multiple candidates in dictionaries, in addition to works that merely use the first-rank one (Luong et al., 2015; Gulcehre et al., 2016), existing explorations mainly involve exploiting prior probabilities, e.g., to adjust the distribution over the decoding vocabulary (Arthur et al., 2016; Zhao et al., 2018a). As a representative context-based disambiguation method, Zhao et al. (2019b) distinguish candidates by matching their embeddings with a decoder-oriented context embedding. Intuitively, an optimal translation candidate should not only accurately reflect the content of the source sentence, but also be consistent with the context of the current partial target sentence. Our observation is that both source information and target information is critical and complementary to distinguish candidates. Taking the source word “摩擦” in Figure 1 for example, the source context of “花 纹/pattern”, “轮胎/tire” and “地面/ground” helps to identify the candidates of “rub” and “friction” in the dictionary, and the target context of “these patterns increase brake” further makes “friction” the best choice. This observation inspires us to synthesize source information and target information in a more fine-grained manner to improve previous straightforward disambiguation methods. (3) A copying step is required to facilitate the collaboration between the pointing step and disambiguating step. Existing models usually do not explicitly emphasize a separate copying step 1, since it is a trivial task in their simplified or pipeline scenario. However, to deliver a sophisticated endto-end architecture that avoids error propagation problems, the pointing and disambiguating step must be appropriately connected as well as integrated into mature NMT models. The proposed copying step is the right place to complete this job. To address the above problems, we propose a novel neural architecture consisting of three novel components: Pointer, Disambiguator, and Copier, to effectively incorporate bilingual dictionaries into NMT models in an end-to-end manner. Pointer is a pioneering research effort on exploiting the semantic information from bilingual dictionaries to better locate source words whose translation in dictionaries may be used. Disambiguator synthesizes complementary contextual information from the source and target via a bi-view disambiguation mechanism, accurately distinguishing the proper translation of a specific source word from multiple candidates in dictionaries. Copier couples Pointer and Disambiguator based on a hierarchical copy mechanism seamlessly integrated with Transformer, thereby building a sophisticated endto-end architecture. Last but not least, we design a simple and effective method to integrate byte-pair encoding (BPE) with bilingual dictionaries in our architecture. Extensive experiments are performed on Chinese-English and English-Japanese benchmarks, and the results verify the PDC’s overall performance and effectiveness of each component. 2 Background: Transformer Transformer (Vaswani et al., 2017) is the most popular NMT architecture, which adopts the standard encoder-decoder framework and relies solely on stacked attention mechanisms. Specifically, given a source sequence x = {x1, x2..., xn}, the model is supposed to generate the target sequence y = {y1, y2..., ym} in an auto-regressive paradigm. Transformer Encoder. A Transformer encoder is constituted by a stack of N identical layers, each of which contains two sub-layers. The first is a multihead self-attention mechanism (SelfAtt), and the second is a fully connected feed-forward network (FFN). Layer normalization (LN) (Ba et al., 2016) and residual connection (He et al., 2016) is em1Note that previous works involve copy mechanism mainly correspond to the Pointing step. 3972 Source Embedding Dictionary Embedding Target Embedding 𝑥# LDC 𝑥$ 𝑥% 𝑥& 𝑦# 𝑦$ 𝑦% 𝑦& Self-Att FFN N× Dec-Enc-Att FFN Self-Att N× ℎ# ℎ$ ℎ% ℎ& Linear & Softmax 𝑠* 𝑞 𝛾-./0 Source Encoder Candidate Encoder Decoder 𝑎𝑡𝑡3 𝑃567 𝑃-./0 𝛾-./0 1 −𝛾-./0 𝑃:;7<= Source Sentence Target Sentence Translation Candidates ℎ′# ℎ′$ ℎ′% ℎ′& 𝑠′* 𝑑′$ 𝑑′% 𝑑′& 𝑑′# 𝑎𝑡𝑡*,# A 𝑞 𝑐# ($) 𝑐# (%) 𝑐# (#) 𝑐$ ($) 𝑐% ($) 𝑐% (%) 𝑐% (&) 𝑐$ (#) 𝑐% (#) 𝑐& (#) 𝑑# ($) 𝑑# (%) 𝑑# (#) 𝑞 Dic-Enc-Att Self-Att FFN 𝑑# ($) 𝑑# (%) 𝑑# (#) 𝑎𝑡𝑡E,# A Source-view Target-view Pointer Disambiguator Vanilla Transformer Copier Bilingual Dictionary Figure 2: An overview of our methods. The left is our PDC module as a copy mechanism, and the right is the vanilla Transformer. For each source word xi, we obtain a set of translation candidates {c(1) i , ..., c(k) i } via a bilingual dictionary. To better capture their semantics, candidate embeddings are shared with target embeddings and refined with self-attention before interacting with Transformer’s encoder states. The state h′ enriched by candidate semantics is utilized by Pointer to locate source words whose dictionary translations may be used. Disambiguator generates two disambiguation distributions over translation candidates from the source view and target view, respectively. Finally, Copier connects the outputs of Pointer and Disambiguator via a hierarchical copy operation. ployed around the two sub-layers in both encoder and decoder. ˜hl = LN(SelfAtt(hl−1) + hl−1), hl = LN(FFN(˜hl) + ˜hl), (1) where hl = {hl 1, hl 2..., hl n} is the output of the l-th layer. The final output hN of the last encoder layer serves as the encoder state h. Transformer Decoder. Similarly, the decoder employs the stack structure with N layers. Besides the two sub-layers, an additional cross attention (CrossAtt) sub-layer is inserted to capture the information from the encoder. ˜sl = LN(SelfAtt(sl−1) + sl−1), bsl = LN(CrossAtt(˜sl, h, h) + ˜sl), sl = LN(FFN(bsl) + bsl), (2) where sl is the output of the l-th decoder layer and the final output sN is taken as the decoder state s. Then, the translation probability p(yt|y<t, x) of the t-th target word is produced with a softmax layer: p(yt|y<t, x) ∝exp(Wost), (3) where y<t is the proceeding tokens before yt. 3 Methodology In this section, we mathematically describe our model in detail. We follow the notations in Section 2. ci = {c(1) i , ..., c(k) i } denotes the translation candidates of a source word xi, derived from a bilingual dictionary D. 3.1 Overview An overview of the proposed PDC model is shown in Figure 2. PDC aims to copy the correct translation candidate of the correct source word at a decoding step. Following the classic CopyNet (Gu et al., 2016), our model consists of two parts, an 3973 encoder-decoder translator to produce the generating probability and a copy mechanism to produce the copying probability. The above two probabilities will collaborate to emit the final probability. The procedure of our copy mechanism involves three critical components: (1) a Pointer that selects a source word whose translation candidates will potentially be copied, (2) a Disambiguator which distinguishes multiple translation candidates of the source word to find the optimal candidate to copy, and (3) a Copier that generates copying probability by combining the outputs from the above two components hierarchically. We will describe the details of each component in the following subsection. 3.2 Pointer The pointer aims to point which source word should be translated at a decoding step. We utilize the carefully extracted semantic information of translation candidates to promote pointing accuracy. Specifically, pointer first extracts the semantic information of candidates with candidate-wise encoding. Then the candidate representations of each source word are fused and interacted with the source representations from transformer encoder. An attention mechanism is applied on the refined source representations to point which word to be translated. Candidate Encoding. We first construct the candidate representations di = {d(1) i , ..., d(k) i } for the candidates of xi, through an candidate embedding matrix and a single layer candidate encoder. ˜di = LN(SelfAtt(Emb(ci)) + Emb(ci)), di = LN(FFN(˜di) + ˜di). (4) Note that this candidate-wise encoder exploits the same structure as a source encoder layer. Pointing with candidate semantics. Previous dictionary-enhanced NMT systems usually directly utilize encoder state h and the decoder state st at tth decoding step to point whose translation should be copied in the source sentence. Intuitively, translation candidates’ information contributes to pointing the right source word, while it is underutilized previously. Accordingly, we propose to explore the semantic information of translation candidates in our pointer. First, we fuse multiple translation candidates’ representations of each word by an attention mechanism between hi and di. d′ i = k X j=1 αsrc i,j ·d(j) i ; αsrc i,j = exp(hiWd(j) i ) Pk j′=1 exp(hiWd(j′) i ) , (5) where d′ i ∈d′ is the fused representation for all candidates of the source word xi. Next, the encoder state h and d′ are interacted to refine the representations of source words with the carefully-extracted candidate information. The refined encoder state h′ can be formalized as: ˜h′ = LN(CrossAtt(h′, d′, d′) + h′), h′ = LN(FFN( ˜h′) + ˜h′). (6) Then, we calculate the attention score to point which source word to be translated: s′ t = n X i=1 βi · h′ i; βi = exp(stWh′ i) Pn i′=1 exp(stWh′ i′), (7) where βi is the pointing probability for xi. s′ t denotes the refined decoder state. 3.3 Disambiguator When translating a specific word, our model has the whole source sentence and the partial target sentence as inputs. An optimal translation candidate should not only accurately reflect the content of source sentence, but also be consistent with the context of the partial target sentence. Thus, we propose a bi-view disambiguation module to select the optimal translation candidate in both source view and target view. Source-view Disambiguation. Source-view disambiguation chooses the optimal candidate for each word with the context information stored in source sentence. The attention score αsrc i = {αsrc i,1, ..., αsrc i,k}, which has been calculated in Equation 5, is employed as the source-view disambiguating distribution for the k translation candidates of xi. This disambiguating distribution is decodingagnostic, which means it serve as global information during decoding. Target-view Diambiguation. As analyzed in Section 1, translation candidates that seem proper from the source view may not well fit in the target context. Thus, we also perform a target view disambiguation to narrow down which candidates fit the partial target sentence’s context. Specifically, we leverage the refined decoder state s′ t to disambiguate the multiple candidates: αtgt i,j = exp(s′ tWdtd(j) i ) Pk j′=1 exp(s′ tWdtd(j′) i ) , (8) where αtgt i,j is the target-view disambiguating probability for c(j) i . In contrast to the decoding-agnostic 3974 source-view disambiguating probability, this targetview disambiguating probability varies during decoding steps. 3.4 Copier Finally, we combine the pointing distribution and the bi-view disambiguating distributions in a hierarchical way to constitute the copying distribution as follows: αi,j = [ρ × αsrc i,j + (1 −ρ) × αtgt i,j ] × βi, (9) where ρ is a scaling factor to adjust the contribution from source-view and target-view disambiguating probabilities. αi,j indicates the probability to copy c(j) i , the j-th translation candidate of the i-th source word. We transform this positional probability into word-level copying probability pcopy: pcopy = p(yt|y<t, x, c), (10) where c is the entire translation candidates for all source word in an instance. The final probability pfinal is constituted by a linear interpolation of pgen and pcopy: pfinal(yt|y<t, x, c) = γt ×pcopy +(1−γt)×pgen, (11) where pgen denotes the the generating probability from Transformer, calculated in Equation 3. γt is the dynamic weight at step t, formalized by: γt = sigmoid(Ws′ t). (12) 3.5 Selective BPE BPE (Sennrich et al., 2016) is commonly used in NMT to deal with the rare words by separating them into frequent subwords. However, it is nontrivial to incorporate BPE into NMT systems with copy mechanism, because the split subwords may not match the original word appearing in dictionaries, either in source side or target side. Simply applying BPE on dictionary words will complicates the scenario to disambiguate and copy, since the model needs to aggregate the representations of these subwords for disambiguation and copy the subwords sequentially. As revealed in Section 5.4, the experimental results demonstrate that whether applying original BPE on dictionary words or not will not yield promising results. In this paper, we present a simple and effective strategy named selective BPE, which only performs BPE on all source words and a portion of target words. All of the translation candidates from the dictionary remain intact. Concretely, in the target side, we keep the target word from being separated into subwords if we can copy it from the translation candidate set c of the source sentence. Such case is formalized as: Itgt(i) = ( 1, if yi ∈c 0, if yi /∈c , (13) where Itgt(i) is the BPE indicator for yi. A target word yi will be split by selective BPE only if Itgt(i) = 0. Note that selective BPE is only used in training, since the reference of validation sets and testing sets do not need BPE. By applying selective BPE, our model can implicitly exploit the information of which dictionary candidates are likely to be copied. Thus, rare words will be more inclined to be copied directly as a whole from the dictionary. 4 Experimental Settings In this section, we elaborate on the experiment setup to evaluate our proposed model. 4.1 Datasets We test our model on Chinese-to-Engish (Zh-En) and English-Japanese (En-Ja) translation tasks. For Zh-En translation, we carry out experiments on two datesets. We use 1.25M sentence pairs from news corpora LDC as the training set 1. We adopt NIST 2006 (MT06) as validation set. NIST 2002, 2003, 2004, 2005, 2008 datasets are used for testing. Besides, we use the Ted talks corpus from IWSLT 2014 and 2015 (Cettolo et al., 2012) including 0.22M sentence pairs for training. We use dev2010 with 0.9K sentence pairs for development and tst2010-2013 with 5.5K sentence pairs for testing. For En-Ja translation, we adopt Wikipedia article dataset KFTT2, which contains 0.44M sentence pairs for training, 1.2K sentence pairs for validation and 1.2K sentence pairs for testing. The bilingual dictionary we used is constructed by the open-source cross-lingual word translate dataset word2word (Choe et al., 2020). We limit the maximum number of translation candidates to 5 for each source word. 1The training set includes LDC2002E18, LDC2003E07, LDC2003E14, Hansards portion of LDC2004T07, LDC2004T08 and LDC2005T06. 2http://www.phontron.com/kftt/ 3975 Systems MT06 MT02 MT03 MT04 MT05 MT08 ∆ Exsisting NMT systems (Cheng et al., 2019) 46.95 47.06 46.48 47.39 46.58 37.38 (Yang et al., 2020) 44.69 46.56 46.04 37.53 (Yan et al., 2020) 47.80 47.72 46.60 48.30 38.70 Baseline NMT systems Transformer 44.11 46.38 45.05 47.07 44.82 34.74 ref Single-Copy 45.04 47.21 46.47 47.48 45.45 36.08 +0.93 Flat-Copy 44.93 46.33 46.26 46.83 45.38 35.19 +0.39 Our NMT systems PDC 46.74 48.85 48.43 48.57 47.71 37.45 +2.59 PDC(w/o Dict-Pointer） 45.79 47.58 47.81 47.98 46.32 36.53 +1.63 PDC(w/o Tgt-View) 45.80 47.43 47.91 48.49 46.81 36.99 +1.91 PDC(w/o Src-View) 45.97 47.42 47.90 47.92 47.07 36.81 +1.81 Table 1: The main results of NIST Zh-En task. ∆shows the average BLEU improvements over the test sets compared with Transformer (ref). The results of our models significantly outperform Transformer (p < 0.01). 4.2 Details for Training and Evaluation We implement our model on top of THUMT (Zhang et al., 2017a) toolkit. The dropout rate is set to be 0.1. The size of a mini-batch is 4096. We share the parameters in target embeddings and the output matrix of the Transformer decoder. The other hyper-parameters are the same as the default settings in Vaswani et al. (2017). The optimal value scaling factor ρ in bi-view disambiguation is 0.4. All these hyper-parameters are tuned on the validation set. We apply BPE (Sennrich et al., 2016) with 32K merge operations. The best single model in validation is used for testing. We use multi−bleu.perl3 to calculate the case-insensitive 4-gram BLEU. 4.3 Baselines Our models and the baselines use BPE in experiments by default. We compare our PDC with the following baselines: • Transformer is the most widely-used NMT system with self-attention (Vaswani et al., 2017). • Single-Copy is a Transformer-based copy mechanism that select a source word’s firstrank translation candidate exactly following Luong et al. (2015); Gulcehre et al. (2016). • Flat-Copy is a novel copy mechanism to perform automatic post-editing (APE) proposed 3https://github.com/moses-smt/mosesdecoder/blob/ master/scripts/generic/multi-bleu.perl by Huang et al. (2019). Note that APE focuses on copying from a draft generated by a pre-trained NMT system. We first arrange candidates of all source words into a sequence as a draft and then copy this flattened “draft” following Huang et al. (2019). 5 Experiment Results 5.1 Main Results Table 1 shows the performance of the baseline models and our method variants. We also list several existing robust NMT systems reported in previous work to validate PDC’s effectiveness. By investigating the results in Table 1, we have the following four observations. First, compared with existing state-of-the-art NMT systems, PDC achieves very competitive results, e.g., the best BLEU scores in 4 of the 5 test sets. Second, Single-Copy outperforms Transformer, indicating that even incorporating only the firstrank translation candidate can improve NMT models. However, since Single-Copy disregards many translation candidates in dictionaries, which could have been copied, the improvement is relatively small (e.g., +0.93 of average BLEU score on the test sets). Third, the performance of Flat-Copy is even worse than Single-Copy, though it considers all translation candidates in dictionaries. The reason lies in that Flat-Copy ignores the hierarchy formed by a source sentence and the corresponding translation candidates of its each word, making it much 3976 0.1 0.2 0.3 0.4 0.5 0.6 45.25 45.50 45.75 46.00 46.25 46.50 46.75 47.00 BLEU 45.69 45.70 46.01 46.20 46.05 45.78 46.00 46.16 46.28 46.74 46.52 46.08 Dev Test-avg Figure 3: The effect of hyper-parameter ρ on NIST ZhEn translation task. more challenging to identify the proper candidate to be copied. Finally, PDC substantially outperforms SingleCopy and Flat-Copy, with improvements of 1.66 and 2.20 average BLEU points, due to our effective hierarchical copy mechanism that connects the Pointer and the Disambiguator, which will be further analyzed in the next sections. 5.2 Effectiveness of Pointer What distinguishes our Pointer from its counterparts of other NMT models is the utilization of semantic information of translation candidates in dictionaries. To verify the effectiveness of this technical design, we implement a PDC variant named PDC(w/o Dict-Pointer) whose Pointer locates source words based on the encoder state (h) of the vanilla Transformer instead of the dictionaryenhanced encoder state (h′). So the semantic information from dictionaries is not incorporated into the pointing step. As expected, the performance of PDC(w/o DictPointer) demonstrates a decrement of nearly 1.0 average BLEU score on the test sets compared with PDC, verifying the promising effect of Pointer. The results also justify our intuition that the rich information of source words’ translations in dictionaries helps to point the proper source word. 5.3 Effectiveness of Disambiguator To investigate the effectiveness of our bi-view Disambiguator, we implement another two model variants: PDC(w/o Src-View) that is removed sourceview disambiguation and PDC(w/o Tgt-View) that is removed target-view disambiguation. As Table 1 shows, the performance of both models significantly decrease. To further investigate the collaboration between Strategies BPE target Dev Test Dict Src Tgt Avg None 43.94 43.68 Standard 45.16 44.75 Dict 45.71 44.84 Selective S 46.74 46.20 Table 2: The BLEU scores of different BPE strategies. For a BPE target (Dict means dictionary words, Src means source words, and Tgt means target words). , and S denote applying BPE, not applying BPE, and applying selective BPE, respectively. the source-view and target-view disambiguation, we analyze the impact of the hyper-parameter ρ, which denotes how to weight the disambiguation distribution generated from source-view and targetview. In Figure 3, the orange polyline shows the BLEU scores on the development set (MT06), and the blue polyline shows average BLEU scores on another five test sets. By looking into these two polylines’ trends, we find that PDC is bestperformed when ρ is 0.4, indicating neither the source view nor the target view can be ignored or overly dependent. These findings prove that both views’ contextual information is critical and complementary to identify a specific source word’s proper translation, and our Disambiguator synthesizes them effectively. 5.4 Effectiveness of Selective BPE We demonstrate the effects of different BPE strategies in Table 2, where None does not use BPE at all, Standard adopts the same BPE strategy as dictionary-independent NMT models, Dict simply apply BPE to dictionary candidates in addition to standard BPE, and Selective is our Selective BPE. More detailed settings of each strategy can be found in Table 2, from which we can also clearly observe the superiority of our selective BPE strategy. We attribute this superiority to the fine-grained collaboration between selective BPE and dictionaries, which implicitly yet effectively leveraging the information of which dictionary candidate are likely to be copied. It is worth mentioning that selective BPE on the target side will not prevent overcoming morphological variance compared with standard BPE. A morphologically inflected target word can be generated in two ways in our system. Firstly, if the target word is not in the candidate set, we will perform standard BPE decomposition. In this scenario, se3977 0 (0,0.05] (0.05,0.1] (0.1,0.15] (0.15,1] Proportion 30 35 40 45 50 55 BLEU 50.52 47.59 44.92 36.67 32.94 52.27 50.54 46.30 40.52 40.12 Transformer PDC Figure 4: Performance of Transformer and PDC on each subset with different rare word proportions. The figure is plotted based on the MT02 test set results. lective BPE is the same as standard BPE, and the target word will be generated in a standard way. Otherwise, if the target word is in the candidate set, it will not be decomposed and our method will encourage the model to copy this word directly. Thus, the morphological variance problem can be simply solved by copying. 5.5 Alleviation of the Rare Words Problem We notice that most dictionary-based NMT works aim to address the rare words problem. Though our work focuses on improving the overall process of incorporating dictionary information as external knowledge, we also conduct a rough experiment to see how our method alleviates the rare words problem. Specifically, we treat a source word as a rare word if it appears less than ten times in the training set. Then we split the test set into subsets according to the rare word proportions of source sentences. The performance on the subsets is shown in Figure 4. We find that PDC outperforms Transformer by a larger gap on the test subsets with more rare words (e.g., 7.18 for the proportion greater than 0.15), demonstrating that PDC can well alleviate the rare words issue. This observation is also consistent with previous investigations (Luong et al., 2015). 5.6 Results on IWSLT and KFTT To verify PDC’s generalization capability, we further conduct experiments on the IWSLT Zh-En translation task and KFTT En-Ja translation task. Due to space limitations, here we only report the performance of PDC and Transformer. PDC’s superiority can be easily observed from the results in Table 3, indicating that PDC can be effectively applied in translation tasks of different language pairs and domains (e.g., news, speech and Wiki). Method IWSLT KFTT Transformer 19.26 30.12 PDC 20.71 32.18 Table 3: Results on the tasks of IWSLT Zh-En translation and KFTT En-Ja translation. 6 Related Work 6.1 Dictionary-enhanced NMT Due to the rich prior information of parallel word pairs in bilingual dictionaries, many researchers have dedicated efforts to incorporating bilingual dictionaries into NMT systems. They either generate pseudo parallel sentence pairs based on bilingual dictionaries to boost training (Zhang and Zong, 2016; Zhao et al., 2020), or exploit the bilingual dictionaries as external resources fed into neural networks (Luong et al., 2015; Gulcehre et al., 2016; Arthur et al., 2016; Zhang et al., 2017b; Zhao et al., 2018a,b, 2019b). Our work can be categorized into the second direction, and focus on improving the overall process of incorporating bilingual dictionaries as external knowledge into the latest NMT systems. In particular, Luong et al. (2015); Gulcehre et al. (2016) first employed copy mechanism (Gu et al., 2016) into NMT to address rare words problem with one-to-one external bilingual dictionaries. Arthur et al. (2016); Zhao et al. (2018a) exploited the prior probabilities from external resource to adjust the distribution over the decoding vocabulary. (Zhao et al., 2018b, 2019b) leverage statisticsbased pre-processing method to filter out troublesome words and perform disambiguation on multiple candidates. Our work extends the above ideas and reforms the overall process into a novel end-toend framework consisting of three steps: pointing, disambiguating, and copying. 6.2 CopyNet CopyNet is also widely used in text summarization (See et al., 2017; Zhu et al., 2020), automatic postediting (Huang et al., 2019), grammar correction (Zhao et al., 2019a) and so on. From a high-level perspective, our methods share a similar Transformer-based architecture with Huang et al. (2019) and Zhu et al. (2020). Huang et al. (2019) employed CopyNet to copy from a draft generated by a pre-trained NMT system. Zhu 3978 et al. (2020) proposed a method that integrates the operation of attending, translating, and summarizing to do cross-lingual summarization. What distinguishes our PDC from other copy-based architectures lies in that the three novel components (Pointer, Disambiguator and Copier) and the selective BPE strategy can make full and effective use of dictionary knowledge. 7 Conclusion We have presented PDC, a new method to incorporate bilingual dictionaries into NMT models, mainly involving four techniques. (1) By integrating semantic information of dictionaries, the enhanced context representations help to locate source words whose dictionary translations will potentially be used. (2) The source and target information is well synthesized and contribute to identifying the optimal translation of a source word among multiple dictionary candidates, in a complementary way. (3) The above two steps are then systematically integrated based on a hierarchical copy mechanism. (4) We finally equip the architecture with a novel selective BPE strategy carefullydesigned for dictionary-enhanced NMT. Experiments show that we achieve competitive results on the Chinese-English and EnglishJapanese translation tasks, verifying that our approach favorably incorporates prior knowledge of bilingual dictionaries. Acknowledgements We thank anonymous reviewers for valuable comments. This research was supported by the National Key Research And Development Program of China under Grant No.2019YFB1405802 and the central government guided local science and technology development fund projects (science and technology innovation base projects) under Grant No.206Z0302G. References Philip Arthur, Graham Neubig, and Satoshi Nakamura. 2016. Incorporating discrete translation lexicons into neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1557–1567. Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization. CoRR, abs/1607.06450. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015. Mauro Cettolo, Christian Girardi, and Marcello Federico. 2012. Wit3: Web inventory of transcribed and translated talks. In Proceedings of the 16th Annual conference of the European Association for Machine Translation, pages 261–268. Yong Cheng, Lu Jiang, and Wolfgang Macherey. 2019. Robust neural machine translation with doubly adversarial inputs. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4324–4333. Yo Joong Choe, Kyubyong Park, and Dongwoo Kim. 2020. word2word: A collection of bilingual lexicons for 3,564 language pairs. In Proceedings of The 12th Language Resources and Evaluation Conference, pages 3036–3045. Jiatao Gu, Zhengdong Lu, Hang Li, and Victor OK Li. 2016. Incorporating copying mechanism in sequence-to-sequence learning. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1631–1640. Caglar Gulcehre, Sungjin Ahn, Ramesh Nallapati, Bowen Zhou, and Yoshua Bengio. 2016. Pointing the unknown words. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 140– 149. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778. Xuancheng Huang, Yang Liu, Huanbo Luan, Jingfang Xu, and Maosong Sun. 2019. Learning to copy for automatic post-editing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6124–6134. Minh-Thang Luong, Ilya Sutskever, Quoc Le, Oriol Vinyals, and Wojciech Zaremba. 2015. Addressing the rare word problem in neural machine translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 11–19. Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational 3979 Linguistics (Volume 1: Long Papers), pages 1073– 1083. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27:3104–3112. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. Jianhao Yan, Fandong Meng, and Jie Zhou. 2020. Multi-unit transformers for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1047–1059. Jian Yang, Shuming Ma, Dongdong Zhang, Zhoujun Li, and Ming Zhou. 2020. Improving neural machine translation with soft template prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5979– 5989. Jiacheng Zhang, Yanzhuo Ding, Shiqi Shen, Yong Cheng, Maosong Sun, Huanbo Luan, and Yang Liu. 2017a. Thumt: An open source toolkit for neural machine translation. arXiv preprint arXiv:1706.06415. Jiacheng Zhang, Yang Liu, Huanbo Luan, Jingfang Xu, and Maosong Sun. 2017b. Prior knowledge integration for neural machine translation using posterior regularization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1514– 1523. Jiajun Zhang and Chengqing Zong. 2016. Bridging neural machine translation and bilingual dictionaries. CoRR, abs/1610.07272. Wei Zhao, Liang Wang, Kewei Shen, Ruoyu Jia, and Jingming Liu. 2019a. Improving grammatical error correction via pre-training a copy-augmented architecture with unlabeled data. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 156–165. Yang Zhao, Yining Wang, Jiajun Zhang, and Chengqing Zong. 2018a. Phrase table as recommendation memory for neural machine translation. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4609–4615. Yang Zhao, Jiajun Zhang, Zhongjun He, Chengqing Zong, and Hua Wu. 2018b. Addressing troublesome words in neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 391–400. Yang Zhao, Jiajun Zhang, Yu Zhou, and Chengqing Zong. 2020. Knowledge graphs enhanced neural machine translation. In Proceedings of the TwentyNinth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 4039–4045. ijcai.org. Yang Zhao, Jiajun Zhang, Chengqing Zong, Zhongjun He, and Hua Wu. 2019b. Addressing the undertranslation problem from the entropy perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 451–458. Junnan Zhu, Yu Zhou, Jiajun Zhang, and Chengqing Zong. 2020. Attend, translate and summarize: An efficient method for neural cross-lingual summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1309–1321.
|
2021
|
307
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3980–3994 August 1–6, 2021. ©2021 Association for Computational Linguistics 3980 VECO: Variable and Flexible Cross-lingual Pre-training for Language Understanding and Generation Fuli Luo∗, Wei Wang∗, Jiahao Liu, Yijia Liu, Bin Bi, Songfang Huang, Fei Huang, Luo Si Alibaba Group {lfl259702,hebian.ww,glacier.ljh,yanshan.lyj}@alibaba-inc.com {b.bi,songfang.hsf,f.huang,luo.si}@alibaba-inc.com Abstract Existing work in multilingual pretraining has demonstrated the potential of cross-lingual transferability by training a unified Transformer encoder for multiple languages. However, much of this work only relies on the shared vocabulary and bilingual contexts to encourage the correlation across languages, which is loose and implicit for aligning the contextual representations between languages. In this paper, we plug a cross-attention module into the Transformer encoder to explicitly build the interdependence between languages. It can effectively avoid the degeneration of predicting masked words only conditioned on the context in its own language. More importantly, when fine-tuning on downstream tasks, the cross-attention module can be plugged in or out on-demand, thus naturally benefiting a wider range of cross-lingual tasks, from language understanding to generation. As a result, the proposed cross-lingual model delivers new state-of-the-art results on various cross-lingual understanding tasks of the XTREME benchmark, covering text classification, sequence labeling, question answering, and sentence retrieval. For cross-lingual generation tasks, it also outperforms all existing cross-lingual models and state-of-theart Transformer variants on WMT14 Englishto-German and English-to-French translation datasets, with gains of up to 1∼2 BLEU. 1 1 Introduction Cross-lingual pre-trained models like mBERT (Devlin et al., 2019), XLM (Lample and Conneau, 2019) and XLM-R (Conneau et al., 2019) that target providing contextualized representations for the inputs across languages, have shown large poten*Equal contribution. 1Code and model are available at https://github. com/alibaba/AliceMind/tree/main/VECO 2021/2/1 Lightshot screenshot chrome-extension://mbniclmhobmnbdlbpiphghaielnnpgdp/screenshot.html?id=screenshot_0.0007918474100581108 1/1 (a) XLM (MLM + TLM) (b) XLM-R (MLM) Figure 1: The attention scores of XLM and XLM-R with the input of a pair of parallel sentences: Take a seat and have a rest in English and its translated Chinese sentence. The darker line denotes a higher score. We can found that there are only a few attention patterns across English and Chinese subwords. tial on a variety of cross-lingual understanding and generation tasks. Behind the great success, two major factors play the role of aligning the contextual representations between languages: 1) build the shared vocabulary across languages through subword tokenization, which supports the simple extension of masked language modeling (MLM) from English corpus to multilingual corpus; 2) capture the alignment in parallel data via concatenating two sentences as input, called translation language modeling (TLM). However, both of these two mechanisms rely on the self-attention module (query=key/value) of the Transformer encoder to implicitly enhance the interdependence between languages, which may lead to few attention patterns across languages. Taking Figure 1 as an example, even though inputting a pair of parallel sentences, both models only attend to the English context to build the representation of English tokens, while ignoring the se3981 b) Translation Language Modeling (TLM) a) Masked Language Modeling (MLM) 𝑥! − 𝑥# 𝑥$ Feed-Forward Self-Attention 𝑥" Feed-Forward Self-Attention − 𝑦! 𝑦$ 𝑦# 𝑦" c) Cross-Attention MLM (CA-MLM) Plug-in Module 𝑥! − − − − 𝑦! 𝑥# 𝑥" 𝑦# 𝑦" 𝑥$ 𝑦$ Feed-Forward Self-Attention Cross-Attention 𝑥! −𝑥" 𝑥# 𝑦! 𝑦$ −𝑦# 𝑥! − 𝑥" 𝑥# 𝑦! 𝑦$ − 𝑦# 𝑥! −𝑥" 𝑥# 𝑦! 𝑦$ −𝑦# 𝑥! − 𝑥" 𝑥# 𝑦! 𝑦$ − 𝑦# 𝑥! −−𝑥# 𝑦! −−𝑦# 𝑥! − − 𝑥# 𝑦! − − 𝑦# 𝑥! − − 𝑦! 𝑥# 𝑦" 𝑥$ 𝑦$ Feed-Forward 𝑥" 𝑦# Self-Attention Figure 2: A schematic comparison of cross-lingual pre-training tasks and their attention matrices. When predicting the masked words of different languages: a) MLM can only attend to the context in its own language; b) TLM implicitly attend to a part of words across languages (as shown in Figure 1). However, c) the proposed CA-MLM can: (1) not only attend to the context in its own language to predict words x2 and y3, (2) but also can firstly attend to its own context and then explicitly attend to all words across languages to predict words x3 and y2 via a plug-in cross-attention module. mantically related Chinese tokens. That is, the self-attention module captures little communication across languages, which is crucial for learning universal cross-lingual representations. Based on the above observation, we propose to plug a cross-attention module (query!=key/value) into the Transformer encoder and design a crossattention MLM task to explicitly capture the interdependence between languages. As illustrated in Figure 2 (c), the cross-attention module takes the representation of x as query and y as key/value (purple lines) to build the representations of x in the next layer, thus explicitly aligning the representations across languages (purple attention matrices). It can effectively avoid the degeneration of predicting masked words only conditioned on the context in its own language. Moreover, what distinguishes our work from pre-training an encoderdecoder model (Liu et al., 2020b) is that we also keep the good nature (i.e., bidirectional contextual modeling) of the original encoder by unplugging the cross-attention from the model to predicting the masked words (e.g., x2 and y3). Furthermore, when fine-tuning on various downstream tasks, we can choose either plug-in or plugout the cross-attention module on-demand, thus making it suitable for both cross-lingual language understanding (NLU) and generation tasks (NLG). For cross-lingual NLU tasks, if plugging the crossattention module out, we can adopt the same finetuning methods as an encoder-only model like XLM. However, we find that plugging the crossattention module in fine-tuning can better utilize the bilingual context to boost the performance. For cross-lingual NLG like machine translation (MT), the cross attention is already jointly pre-trained with the whole network. Therefore, the parameters of the decoder do not need to be re-adjusted substantially in the following tuning process, thus fundamentally solving the main drawback of utilizing pre-trained encoders like XLM for initializing encoder-decoder models. We call our approach VECO for “Variable and Flexible Cross-lingual Pre-training”. We validate VECO on a variety of representative cross-lingual understanding and generation benchmarks. Regrading cross-lingual understanding tasks, we conduct experiments on the XTREME benchmark consisting of 9 cross-lingual tasks, including text classification, sequence labeling, question answering, and sentence retrieval. VECO ranks first at the XTREME leaderboard 2 at the submission deadline. Regrading cross-lingual generation tasks, we validate VECO on the widely used WMT14 EnglishGerman and English-French machine translation benchmarks. VECO obtains 44.5 and 31.7 BLEU scores, consistently outperforming existing crosslingual pre-training approaches and state-of-the-art Transformer variants by around 1∼2 BLEU. 2https://sites.research.google/xtreme 3982 2 Pre-training of VECO 2.1 Overview of VECO VECO extends from a multi-layer Transformer encoder and plugs a cross-attention module in each layer. Given a pair of input (x, y) and its corrupted version (ˆx, ˆy) via randomly masking part of its tokens, the model builds two types of contextualized vector representation for each token: • One suit of contextual representations H, denoted as green blocks and yellow blocks in Figure 2 (c), are only build on self-attention module (i.e., unpluging the cross-attention module) in each layer. • Another suit of contextual representations S, denoted as mixed color blocks in Figure 2 (c), are build on both the self-attention and cross-attention modules 3. The model is trained to predict the masked tokens via two corresponding representations, conditioning on both its own context and paired context, respectively. Take predicting the masked words in sequence x as an example, the training objective is the cross-entropy of the gold distribution and predicted distribution P(x|ˆx) and P(x|ˆy, ˆx) computed via the above two suits of contextual representations. Thus, the training objective of crossattention masked language modeling (CA-MLM) can be formulated as L(x, y) = −logP(x|ˆx; θs) −logP(x|ˆy, ˆx; θs, θc) −logP(y|ˆy; θs) −logP(y|ˆx, ˆy; θs, θc) (1) where θs and θc are the parameters of self-attention and cross-attention modules. 2.2 Architecture The backbone network of VECO is composed of a stack of N Transformer layers. Each layer has three modules: a required self-attention module, a plugand-play cross-attention module, and a required feed-forward linear module. Both self-attention and cross-attention modules are based on the multihead attention (Vaswani et al., 2017). An attention function can be described as mapping a query (Q) and a set of key-value (K-V) pairs to an output. 3For simplicity of illustration, we only show the mixed representations S of x3 and y2 in Figure 2 (c). For the self-attention module, all the queries, keys and values are the same representations from the previous layer. Specifically, for the l-th Transformer layer, the output of a self-attention head As l is computed via: Q = Hl−1WQ l (2) K = Hl−1WK l (3) V = Hl−1WV l (4) As l = softmax(QKT √dk )V (5) where Hl−1 are the previous layer’s outputs, WQ l , WK l , WV l are the parameter matrices of selfattention modules. For the cross-attention module, the queries come from the previous layer, and the keys and values come from the last layer’s representations of paired input. Specifically, for the l-th layer, the output of a cross-attention head Ac l is computed via: Q = Sl−1UQ l (6) K = HLUK l (7) V = HLUV l (8) Ac l = softmax(QKT √dk )V (9) where Sl−1 are the previous layer’s outputs, UQ l , UK l , UV l are the parameter matrices of crossattention modules. Finally, the output HL of the last layer is used to recover the masked tokens of x, conditioning on its own context. P(x|ˆx) = softmax(f(HL x)) (10) P(y|ˆy) = softmax(f(HL y )) (11) where f is the feed-forward network that maps the output vectors into the dictionary. HL x and HL y are computed via Eq 2∼5 when H0 x and H0 y are the word embeddings of x and y, respectively. Meanwhile, SL, conditioning on the context of the paired sequence ˆx and ˆy, is used to predict the masked tokens of y. P(x|ˆy, ˆx) = softmax(f(SL x)) (12) P(y|ˆx, ˆy) = softmax(f(SL y )) (13) where SL x and SL y are computed via Eq 6∼9 with the corresponding word embeddings and HL. 3983 VECO Fine-tuning: Flexible for NLU and NLG tasks Self-attention Cross-attention Feed-Forward Self-attention Feed-Forward Pre-training: Train a plug-and-play cross-attention module NLU Fine-tuning: NLG Fine-tuning： Initialize a Encoder-decoder Transformer 12 Plug-Out Fine-tuning Self-attention Feed-Forward Plug-In Fine-tuning Self-attention Feed-Forward Self-attention Cross-attention Feed-Forward Cross-attention Figure 3: The overview of VECO. During pre-training, a plug-and-play cross-attention module is jointly pretrained along with the self-attention module. When fine-tuning on natural language understanding (NLU) tasks, the cross-attention module can be either plug-in or plug-out on demand. When fine-tuning on natural language generation (NLG) tasks, VECO can initialize an encoder-decoder module (the mainstream backbone model of generation tasks) since all those necessary modules in the encoder and decoder are already pre-trained. Note that when optimizing the objectives based on Eq 12 and Eq 13, we apply a stop-gradients operation (Chen and He, 2020) to HL (i.e., HL is treated as a constant in this term). This operation can largely speed up the training by avoiding the backpropagation on a 2L-layer network. Moreover, it even stabilizes the training of deep postlayernorm Transformer, which requires non-trivial efforts regarding carefully designing learning rate schedulers and cutting-edge optimizers (Liu et al., 2020a; Bachlechner et al., 2020). 3 Fine-tuning VECO for Downstream Cross-lingual Understanding and Generation Tasks As Figure 3 illustrated, when fine-tuning on various downstream tasks, one advantage of VECO is its flexibility for initializing both the encoder-only Transformer for understanding tasks and encoderdecoder Transformer for generation tasks. Beyond it, we also explore a fine-tuning approach combined with the characteristics of VECO . 3.1 VECO for Cross-lingual Understanding Due to the plug-and-play cross-attention module, we explore two fine-tuning approaches: • Plug-Out fine-tuning is to unplug the crossattention module from the pre-trained model. In other words, the architecture of the finetuned model is almost the same as mBERT or XLM. Specifically, the contextual representations from the last layer HL x is used to predict the label of input x. • Plug-In fine-tuning is to plug the crossattention module into the fine-tuned model, if the bilingual or automatically translated training data y is available in the downstream task. Specifically, we concatenated the two representations [HL x : SL x] to predict the label of x, [HL y : SL y ] to predict the label of y. 4. 3.2 VECO for Cross-lingual Generation For pre-trained encoders like XLM, it is not a trivial problem to incorporate them into the sequenceto-sequence architecture – the mainstream backbone model of generation tasks (Zhu et al., 2020). One of the drawbacks or challenges could be that the encoder-to-decoder attention is not pre-trained. Therefore, the parameters of the decoder need to be re-adjusted along with the encoder in the following fine-tuning process (Ren et al., 2019). However, under the framework of VECO , the cross-attention is jointly pre-trained along with the whole network, making it easy to provide full initialization for sequence-to-sequence models. Specifically, the self-attention module is used to initialize both the corresponding modules in the encoder and decoder for contextual modeling, while the cross-attention module is used to initialize the encoder-to-decoder attention. It’s okay whether you continue to tie the self-attention parameters during fine-tuning. Directly pre-training a sequenceto-sequence model like mBART (Liu et al., 2020b) could be another solution for NLG tasks, but we found mBART is not so effective in cross-lingual NLU tasks. We refer the reader to the Section 7 for detailed experiments and analysis. 4Plug-In fine-tuning is not suitable for the zero-shot setting (also called cross-lingual transfer) due to the lack of bilingual or translated pair (x, y) 3984 Model Architecture #Parameters Enc Layers Dec Layers #Languages #Vocab Training Data mBERT (Devlin et al., 2019) Encoder-only 110M 12 104 110k Wikipedia XLM (Lample and Conneau, 2019) Encoder-only 570M 24 100 200k Wikipedia XLM-R (Conneau et al., 2019) Encoder-only 550M 24 100 250k CommonCrawl mRASP (Lin et al., 2020) Encoder-decoder 375M 6 6 32 64k Translation MMTE (Siddhant et al., 2020) Encoder-decoder 375M 6 6 103 64k Translation mBART (Liu et al., 2020b) Encoder-decoder 680M 12 12 25 250k CommonCrawl VECO Flexible 662M 24* 50 250k CommonCrawl + Translation Table 1: Comparison of large cross-lingual models. * denotes VECO unifies the encoder and decoder. 4 Pre-training Setup Model Configuration We pre-train a 24-layer model with 1024 embedding/hidden size and 4096 feed-forward size. We do not use language embeddings to allow our model to better deal with downstream tasks of unseen languages. We adopt the same 250K vocabulary that is also used by XLM-R (Conneau et al., 2019). Table 1 shows the other details of baselines and VECO . Pre-Training Data We collect monolingual and bilingual corpus covering 50 languages. For monolingual training datasets, we reconstruct CommonCrawl Corpus used in XLM-R (Conneau et al., 2019). We extract 1.36TB data in 50 languages, which contains 6.5G sentences and 0.4G documents. We up/down-sample the monolingual text like XLM from each language with a smoothing parameter α = 0.5. For bilingual data, we collect from the OPUS website 5 like previous works (Lample and Conneau, 2019; Chi et al., 2020b). There are 6.4G parallel sentences, covering 879 language pairs across 50 languages. See more statistics of training data in Appendix A. Optimization Settings For each iteration, we alternately sample a batch of adjacent segments from the monolingual corpus and a batch of parallel sentences from bilingual datasets to conduct a pair of masked input (ˆx, ˆy). We adopt the translation language modeling (TLM) when the inputs are parallel bilingual sentences. Thus the overall training objective is the sum of TLM and the proposed CA-MLM objectives. During training, the model parameters except for cross-attention are initialized by XLM-R. We first freeze the parameters of XLM-R and only update the cross-attention parameters for faster convergence. Then, we jointly train the whole model. We pre-train our model with mixed-precision training using 64 Nvidia Telsa V100 32GB GPUs. Appendix A shows additional details. 5http://opus.nlpl.eu/ 5 Experiments on Cross-lingual Understanding Tasks 5.1 Experimental Setup Downstream Tasks We conduct cross-lingual NLU evaluations on XTREME (Hu et al., 2020), a representative massively multilingual benchmark that consists of 9 understanding tasks over 40 languages. XTREME tasks can be classified into four different categories: (1) sentence-pair classification: XNLI (Conneau et al., 2018), PAWS-X (Yang et al., 2019); (2) structured prediction: POS (Nivre et al., 2018), Wikiann NER (Pan et al., 2017); (3) question answering: XQuAD (Artetxe et al., 2020), MLQA (Lewis et al., 2020), TyDiQA (Clark et al., 2020); (4) sentence retrieval: BUCC 2018 (Zweigenbaum et al., 2017), Tatoeba (Artetxe and Schwenk, 2019). Tasks in the first three categories are provided: 1) golden training corpus in English, 2) translated training corpus in other languages, and 3) dev/test set in all languages. For sentence retrieval tasks, no training datasets are provided. We refer the reader to Hu et al. (2020) for additional details about the datasets. Fine-tuning Setting Following previous works (Conneau et al., 2019; Hu et al., 2020), we consider two typical fine-tuning settings: (1) Cross-lingual Transfer which fine-tunes the pre-trained model using English golden data only and directly performs inference on the test data of different target languages; (2) TranslateTrain-All fine-tunes a multilingual model on the concatenation of all data (golden training corpus in English and translated training corpus in other languages). Note that for two sequence-labeling tasks (POS, NER), the position of token labels in the translated text generally differs from that in the source text. Following FILTER (Fang et al., 2020), we use the model trained only on the English training dataset as a teacher, to label the translated text. To have a fair comparison with the strong baseline XLM-R (Conneau et al., 2019) 3985 Datasets XNLI PAWS-X POS NER XQuAD MLQA TyDiQA BUCC Tatoeba #Languages 15 7 33 40 11 7 9 5 33 Metrics Acc Acc F1 F1 F1/EM F1/EM F1/EM F1 Acc Avg. Cross-lingual Transfer: Fine-tune model on English training set and test on all languages MMTE† 67.4 81.3 73.5 58.3 64.4/46.2 60.3/41.4 58.1/43.8 59.8 37.9 59.5 mBERT† 65.4 81.9 70.3 62.2 64.5/49.4 61.4/44.2 59.7/43.0 56.7 38.7 59.6 XLM† 69.1 80.9 70.1 61.2 59.8/44.3 48.5/32.6 43.6/29.1 56.8 32.6 55.5 XLM-R† 79.2 86.4 72.6 65.4 76.6/60.8 71.6/53.2 65.1/45.0 66.0 57.3 68.1 VECOout 79.9 88.7 75.1 65.7 77.3/61.8 71.7/53.2 67.6/49.1 85.0 75.1 73.1 Translate-Train-All: Fine-tune model on English training data and translated data of the target language XLM-R‡ 82.6 90.4 80.2/65.9 72.8/54.3 66.5/47.7 XLM-R∗ 82.8 90.2 72.6 65.4 80.0/65.8 73.0/54.3 74.5/58.3 80.2 75.2 74.4 FILTER 83.9 91.4 76.2 67.7 82.4/68.0 76.2/57.7 68.3/50.9 84.5 84.5 77.0 VECOout 83.0 91.1 75.1 65.7 79.9/66.3 73.1/54.9 75.0/58.9 89.3 86.9 77.2 VECOin 84.3 92.8 79.8 71.0 83.9/70.9 77.5/59.3 79.4/63.7 92.6 91.1 81.0 Table 2: XTREME results on each dataset (as of ACL submission deadline). Averaged results on the four categories can be found at leaderboard: https://sites.research.google/xtreme. “†” and “‡” indicates results from Hu et al. (2020) and Fang et al. (2020), respectively. “*” indicates the results obtained by our implementation. The detailed results for each language are in Appendix D. under the translate-train-all setting, we also show the results of XLM-R using the same fine-tuning hyperparameters as VECO . 5.2 Experimental Results The detailed test results of nine tasks on the XTREME benchmark are shown in Table 2. It demonstrates that the proposed VECO outperforms previous cross-lingual models on all datasets. Compared to XLM-R, it averagely scores 5.0 and 6.6 points higher under the cross-lingual transfer and translation-train-all settings, respectively. In the cross-lingual transfer setting, VECO delivers a large improvement compared to XLM-R, especially on zero-shot sentence retrieval tasks (BUCC, Tatoeba). This phenomenon reflects that our model can better build the interdependence between languages. Thus it can better mine parallel sentences in a multilingual corpus. Under the translation-train-all setting, it can be observed that VECO with Plug-In fine-tuning (VECOin) is better than Plug-Out fine-tuning (VECOout). We conclude the reasons as two-fold. On the input side, the Plug-Out fine-tuning individually takes multilingual instances as input, while the Plug-In fine-tuning considers the bilingual instances 6 at each run. On the model side, the Plug-In fine-tuning can encourage correspondence across language via the cross-attention module. Note that the Plug-In fine-tuning method also outperforms FILTER (Fang et al., 2020), an enhanced cross-lingual fine-tuning method that also takes the 6English instance with its translated one. bilingual instance as the input of XLM-R. It further demonstrates the effectiveness of VECO and its specialized fine-tuning method. We conclude the reasons for the above performance improvement as two-fold: 1) the introduction of bilingual data during pre-training, which is a direct way to enhance the cross-lingual ability of the model; 2) Stronger ability to enhance the interdependence and fusion among languages via the proposed CA-MLM pre-training tasks. To analyze which plays a leading role, we conduct a set of more fair experiments in Section 7. 6 Experiments on Cross-lingual Generation Tasks 6.1 Experimental Setup Datasets We choose the machine translation (MT) task, a typical cross-lingual generation scenario. In order to illustrate the generality of our approach and have a fair comparison with the most recent state-of-the-art Transformer work (Liu et al., 2020a), we choose two most widely used datasets: WMT14 English→German (En-De) and English→French (En-Fr) translation. WMT14 EnDe is a medium-resource dataset that provides 4.5M pairs for training and validation. We adopt standard newstest2014 as the test set. WMT14 En-Fr is a high-resource dataset that contains 36M pairs of parallel sentences. We use newstest2012+newstest2013 for validation and newstest2016 for test. We measure case-insensitive tokenized BLEU with multi-bleu.perl and de3986 Model WMT14 En-Fr WMT14 En-De BLEU SacreBLEU BLEU SacreBLEU Randomly Initialize Baseline 42.9 40.4 28.7 27.8 Liu et al. (2020a) 43.8 41.8 30.1 29.5 Randomly Initialize + More Bilingual Data* Baseline* 30.6 29.5 Cross-lingual Model Initialize mBART 43.2 41.0 30.0 29.1 mRASP 44.3 41.7 30.3 XLM-R 43.8 41.2 30.9 29.9 VECO 44.5 42.0 31.7 30.6 10 15 20 25 30 35 Epochs 25 26 27 28 29 30 sacreBLEU VECO Init. XLM-R Init. Random Init. Table 3: (left) Results on machine translation. (right) Learning curves of different initialization methods. tokenized SacreBLEU 7 to avoid the influence of different tokenization and normalization between models (Post, 2018). Fine-tuning Setting We fine-tune our model using fairseq 8 toolkit and adopt comparable training settings with baselines. We run WMT 14 EnDe and En-Fr MT experiments on 16 and 32 V100 GPUs, respectively. The batch size is 64k for EnDe and 256k for En-Fr. The total training updates are set to 100k. The learning rate is 1e-4/2e-4, with linear warm-up over the first 16k steps and linear decay. We average the last 10 checkpoints and use beam search with a beam size of 5. Baselines We consider two types of Transformer baselines: randomly initialized and cross-lingual models initialized. For random initialization, we reproduce a Transformer baseline that adopts the same architecture and fine-tuning hyperparameters as VECO but with random initialization. Besides, we compare to the state-of-the-art Deep Transformer (Liu et al., 2020a). For cross-lingual encoder-decoder models, we include mBART (Liu et al., 2020b) and mRASP (Lin et al., 2020), which show impressive results on MT. Note that since we tied the self-attention weights of each encoder layer with each decoder layer, the whole parameters of mBART and VECO are comparable. We also conduct the WMT experiments for XLM-R, following the totally same fine-tuning settings as VECO , but leaving the encoder-to-decoder attention un-initialized. 7Hash: BLEU+case.mixed+lang.en-{de,fr}+numrefs.1+ smooth.exp+test.wmt14/full+tok.13a+version.1.4.9 8https://github.com/pytorch/fairseq 6.2 Experimental Results Table 3 (left) shows the results on the machine translation. We can observe that VECO can largely outperform the randomly initialized same-sized Transformer baseline by 2.3 BLEU points. Moreover, it even beats the (randomly initialized) stateof-the-art Deep-Transformer (Liu et al., 2020a), which is three times deep as VECO . Among the cross-lingual models, VECO can consistently outperform the best models, averaged on two datasets, by 0.8 BLEU points. Table 3 (right) displays the BLEU scores of same-sized models during training. We find that VECO initialized model can get a surprising more than 28 SacreBLEU score just after 10 epochs, which is better than the final score of the randomly initialized model at 35 epochs. It reveals that VECO can provide a fairly good initialization for the machine translation model, which can converge quickly and further boost the results. One might suspect that the main reason for the performance improvement is leveraging parallel corpus during pre-training. To figure it out, we conduct a more comparable experiment. We first train an out-of-domain Transformer model using the whole En-De parallel data (∼68M) used in VECO pre-training, and then continue to train the model on the in-domain WMT14 En-De training dataset. Results are shown in Table 3 (left) marked with *. Under this set of a totally fair comparison, VECO still maintains a lead of 1.1 BLEU score. This directly confirms that the improvement in MT is not only due to the use of bilingual data. More importantly, CA-MLM ensures better use of bilingual and large-scale unlabeled multilingual corpus. 3987 Method #Layers WMT14 En-De BLEU SacreBLEU Randomly Initialize 3 28.5 27.6 6 28.6 27.7 VECO Initialize First-3 30.8 29.8 Last-3 31.2 30.3 First-6 31.1 30.1 Last-6 31.5 30.5 Full-24 31.7 30.6 Table 4: Results of utilizing VECO to initialize deep encoder and shallow decoder (3/6-layer) Transformers. 6.3 Potential of Initializing Shallow Decoder Online translation applications usually have a restriction of inference time. The most direct way is to reduce the decoder layers since previous MT works (Liu et al., 2020a) have shown that deeper encoders are more worthwhile than deeper decoders. Based on this, we also explore the potential of the VECO to initialize deep encoder and shallow decoder Transformers, which is a blank in the crosslingual pre-training works. Table 4 contrasts two ways of initializing a Transformer with n decoder layers (n < 24) via selecting: (1) the first n layers; (2) the last n layers from a 24-layer pre-trained VECO model. We consider n = {3, 6} to conduct experiments. We find that selecting the last n layers exhibits better performance than selecting the first n layers. It reveals that the last several layers play a more important role in making predictions over the whole vocabulary. Moreover, we can find that there is 0.2∼0.3 BLEU gain when increasing the decoder layers from 3 to 6. However, we observe that only marginal improvement can be gained when further increasing the decoder layers from 6 to 24, which is also in line with the findings in Liu et al. (2020a). Regardless of the initialization method, the VECO initialized model can gain consistent 1∼2 BLEU improvement over the randomly initialized model. 7 Analysis and Ablation Study We perform an ablation study to investigate where the improvement in cross-lingual NLU and NLG tasks mainly comes from. Specifically, there are three main aspects we have studied: 1. How much performance improvement comes from the parallel translation corpus used in pre-training? 2. How effective of the CA-MLM pre-training Data Models Tasks XNLI IWSLT Mono. XLM MLM 59.8 33.7 mBART MLM 57.3 32.9 VECO CA-MLM 60.6 34.0 Bili. XLM MLM+TLM 64.5 33.9 mBART MLM+TLM 60.8 34.5 VECO CA-MLM +TLM 67.7 36.0 Table 5: Ablation study of small-sized models on XNLI and IWSLT14 De-En translation dataset. task, especially compared to the MLM and TLM pre-training tasks? 3. How about pre-training a sequence-tosequence model like mBART for NLU and NLG tasks? To figure out these questions, we train XLM, mBART and VECO model from scratch using the same datasets and parameter settings (see Appendix A for more details). All of them is pre-trained via MLM and TLM tasks. Note that the MLM task generally refers to predict the masked words of source language, while the TLM task generally refers to predict the words of the target language. Specifically for mBART that is under the framework of encoder-decoder, the input of encoder is masked sequence ˆx, and the target of decoder is the masked words of source input x (for MLM task), or the parallel sentence y (for TLM task). Table 5 shows the results of two representative datasets of cross-lingual NLU and NLG. We can observe that, when using monolingual corpus only, VECO can outperform XLM by 0.8 points on the XNLI dataset and 0.3 BLEU scores on the IWSLT14 De-En translation dataset. It suggests that the CA-MLM can still benefit from adjacent sentences in monolingual corpus 9, to be equipped with a stronger ability of contextual modeling. Moreover, when pre-training both on the monolingual and bilingual corpus, VECO can even achieve a larger improvement compared to XLM, with 3.2 and 2.1 points improvement on two datasets, respectively. It reveals that CA-MLM objective of VECO can better utilize the bilingual corpus, compared to only optimized by TLM and MLM of XLM. Moreover, we find that pre-training a sequenceto-sequence model like mBART (Liu et al., 2020b) 9As noted in Section 4, we take two adjacent sentences in the monolingual corpus as (x, y). 3988 performs worst on NLU tasks like XNLI 10, almost 6 points worse than VECO and near 2 points worse than XLM. One possible explanation could be that the unidirectional language modeling in the decoder might be sub-optimal for NLU tasks. And even on the machine translation task, mBART still performs worse than VECO when pre-training on the same bilingual datasets. We conclude that it is because that VECO can do better in the contextual modeling of source input x via a explicit masked language modeling objective in Eq 10 applied to x2 in Figure 2 (c). 8 Related Work mBERT (Devlin et al., 2019) is a key step towards building a unified contextual language representation over multiple languages. It simply shares all languages’ vocabulary and trains a bidirectional Transformer encoder, achieving promising results in various cross-lingual NLU tasks. There have been several extensions that follow the same encoder-only backbone as mBERT. The main difference is the introduction of more training corpus (e.g., bilingual data) and pre-training tasks. XLM (Lample and Conneau, 2019) utilizes both monolingual and bilingual corpus to perform the masked language modeling. XLM-R (Conneau et al., 2019) extends to be built on RoBERTa (Liu et al., 2019) using larger monolingual training data. Other works (Huang et al., 2019; Yang et al., 2020; Chi et al., 2020b) propose new pre-training tasks to utilize the bilingual data better. However, there are two main drawbacks of these works. First, they mainly rely on the self-attention module in the Transformer encoder to implicitly build the interdependence between languages, leading to few attention patterns across languages due to the “lazy” network. Second, even though they show impressive performance improvement on cross-lingual understanding tasks like XNLI, only marginal improvement has been gained on cross-lingual generation tasks like machine translation, especially on high-resource languages. A feasible solution for cross-language generation is to pre-train a denoising auto-encoder like mBART (Liu et al., 2020b). It extends BART (Lewis et al., 2019) to the multilingual setting, demonstrating significant gains in low/medium-resource machine translation, but 10We follow BART (Lewis et al., 2019) by utilizing the final representation from the decoder for classification tasks. with a decrease in high resource languages. Unlike mBART, Chi et al. (2020a) first trains an encoder via MLM and then frozen the encoder to train the decoder only via two generative tasks. A similar approach is also proposed in Liang et al. (2020) and Lin et al. (2020), with the main difference in the joint training of encoder-decoder with code-switch tricks. However, all these cross-lingual models emphasize training a dedicated model for NLG. Thus they may hurt the NLU capabilities of the model. The ablation study in Section 7 also validates that it is sub-optimal to train an encoder-encoder network for NLU tasks. This paper endeavors to build a unified crosslingual model for NLU and NLG tasks via a plugand-play cross-attention module. More importantly, the cross-attention module plays a role in the explicit alignment of encoded representations of different languages, thus largely contributing to building a unified cross-lingual model. 9 Conclusion We present VECO, a variable and flexible crosslingual pre-training model, targets at explicitly capturing the interdependence between languages via a plug-and-play cross-attention module. Based on the flexible characteristics, VECO can initialize both NLU preferred encoder-only and NLG specialized encoder-decoder Transformer. Moreover, we also introduce a Plug-In fine-tuning approach to encourage the fusion between languages, combining the feature of VECO and cross-language downstream tasks. Taken together, VECO achieves consistent improvements on various language understanding and generation tasks, broadening the way of thinking about pre-trained backbone architecture and finetuning methods under the cross-lingual scenario. References Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 4623–4637. Association for Computational Linguistics. Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence embeddings for zeroshot cross-lingual transfer and beyond. Trans. Assoc. Comput. Linguistics, 7:597–610. 3989 Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Garrison W Cottrell, and Julian McAuley. 2020. Rezero is all you need: Fast convergence at large depth. arXiv preprint arXiv:2003.04887. Xinlei Chen and Kaiming He. 2020. Exploring simple siamese representation learning. CoRR, abs/2011.10566. Zewen Chi, Li Dong, Furu Wei, Wenhui Wang, XianLing Mao, and Heyan Huang. 2020a. Cross-lingual natural language generation via pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence. Zewen Chi, Li Dong, Furu Wei, Nan Yang, Saksham Singhal, Wenhui Wang, Xia Song, Xian-Ling Mao, Heyan Huang, and Ming Zhou. 2020b. InfoXLM: An information-theoretic framework for cross-lingual language model pre-training. arXiv preprint arXiv:2007.07834. Jonathan H. Clark, Eunsol Choi, Michael Collins, Dan Garrette, Tom Kwiatkowski, Vitaly Nikolaev, and Jennimaria Palomaki. 2020. TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages. In Transactions of the Association of Computational Linguistics. Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. Alexis Conneau, Ruty Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In Proceedings of EMNLP 2018, pages 2475–2485. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT. Yuwei Fang, Shuohang Wang, Zhe Gan, Siqi Sun, and Jingjing Liu. 2020. FILTER: An enhanced fusion method for cross-lingual language understanding. arXiv preprint arXiv:2009.05166. Junjie Hu, Sebastian Ruder, Aditya Siddhant, Graham Neubig, Orhan Firat, and Melvin Johnson. 2020. XTREME: A massively multilingual multitask benchmark for evaluating cross-lingual generalization. arXiv preprint arXiv:2003.11080. Haoyang Huang, Yaobo Liang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, and Ming Zhou. 2019. Unicoder: A universal language encoder by pretraining with multiple cross-lingual tasks. arXiv preprint arXiv:1909.00964. Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226. Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint arXiv:1901.07291. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461. Patrick Lewis, Barlas O˘guz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2020. MLQA: Evaluating Cross-lingual Extractive Question Answering. In Proceedings of ACL 2020. Yaobo Liang, Nan Duan, Yeyun Gong, Ning Wu, Fenfei Guo, Weizhen Qi, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, et al. 2020. XGLUE: A new benchmark dataset for cross-lingual pretraining, understanding and generation. arXiv preprint arXiv:2004.01401. Zehui Lin, Xiao Pan, Mingxuan Wang, Xipeng Qiu, Jiangtao Feng, Hao Zhou, and Lei Li. 2020. Pretraining multilingual neural machine translation by leveraging alignment information. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 2649–2663. Association for Computational Linguistics. Xiaodong Liu, Kevin Duh, Liyuan Liu, and Jianfeng Gao. 2020a. Very deep transformers for neural machine translation. arXiv preprint arXiv:2008.07772. Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020b. Multilingual denoising pre-training for neural machine translation. arXiv preprint arXiv:2001.08210. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692. Joakim Nivre, Mitchell Abrams, ˇZeljko Agi´c, Lars Ahrenberg, Lene Antonsen, Maria Jesus Aranzabe, Gashaw Arutie, Masayuki Asahara, Luma Ateyah, Mohammed Attia, et al. 2018. Universal dependencies 2.2. Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Crosslingual name tagging and linking for 282 languages. In Proceedings of ACL 2017, pages 1946–1958. 3990 Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation. Shuo Ren, Yu Wu, Shujie Liu, Ming Zhou, and Shuai Ma. 2019. Explicit cross-lingual pre-training for unsupervised machine translation. volume abs/1909.00180. Aditya Siddhant, Melvin Johnson, Henry Tsai, Naveen Ari, Jason Riesa, Ankur Bapna, Orhan Firat, and Karthik Raman. 2020. Evaluating the cross-lingual effectiveness of massively multilingual neural machine translation. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, pages 8854–8861. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems. Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vishrav Chaudhary, Francisco Guzman, Armand Joulin, and Edouard Grave. 2019. Ccnet: Extracting high quality monolingual datasets from web crawl data. arXiv preprint arXiv:1911.00359. Jian Yang, Shuming Ma, Dongdong Zhang, Shuangzhi Wu, Zhoujun Li, and Ming Zhou. 2020. Alternating language modeling for cross-lingual pre-training. In Proceedings of the AAAI Conference on Artificial Intelligence. Yinfei Yang, Yuan Zhang, Chris Tar, and Jason Baldridge. 2019. PAWS-X: A cross-lingual adversarial dataset for paraphrase identification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3685– 3690. Association for Computational Linguistics. Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. 2020. Incorporating BERT into neural machine translation. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Pierre Zweigenbaum, Serge Sharoff, and Reinhard Rapp. 2017. Overview of the second BUCC shared task: Spotting parallel sentences in comparable corpora. In Proceedings of the 10th Workshop on Building and Using Comparable Corpora, BUCC@ACL 2017, Vancouver, Canada, August 3, 2017, pages 60– 67. Association for Computational Linguistics. A Pre-Training Details For monolingual data, following XLM-R (Conneau et al., 2019), we build a clean CommonCrawl Corpus using an open-source tool CCNet (Wenzek et al., 2019). There are 1.36TB monolingual data in 50 languages before up/down-sampling. Table 6 reports the language codes and statistics of pretraining data. We collect bilingual corpus in 50 languages from the OPUS website11, including MultiUN, UNPC, Bombay, EU-bookshop, OpenSubtitles2018, Tanzil, GlobalVoices, ParaCrawl, MultiParaCrawl, DGT, Tilde, Europarl, Wikipedia, ECB, TED2013, News-Commentary, Ubuntu, Books, UN, infopankki-v1, EUconst, and Bianet. In total, there are 1TB bilingual training data before pre-processing, covering 879 language pairs. Table 7 lists the statistics for each language pair. We then apply subword tokenization directly on raw text data using Sentence Piece Model (Kudo and Richardson, 2018) without any additional preprocessing. We use the whole corpus to train VECO and a subset (∼1/4) that contains 33 languages to train small-sized XLM, mBART and VECO . The full set of pre-training hyperparameters for smallsized and large-sized VECO (default) are listed in Table 8. B More details about Illustrated Attention The models illustrated with attention patterns in Figure 1 of main paper (not appendix), are the base-sized XLM 12 and XLM-R 13. We show the attention scores averaged on all heads in the middle layer. C Fine-Tuning Details on XTERME We select the model with the best average result over all the languages on the dev sets, by searching the learning rate over [5e-6,8e-6,1e-5,2e-5,3e-5] for the Cross-lingual Transfer setting and [5e-6,6e6,7e-6,8e-6,9e-6] for Translate-Train-All setting, training epoch over [3,5,10], and batch size over [16,32,64]. D Detailed Results on XTREME The detailed results of each XTREME task under the cross-lingual transfer and translate-train-all settings on all languages are listed in the following tables. 11http://opus.nlpl.eu/ 12https://huggingface.co/ xlm-mlm-tlm-xnli15-1024 13https://huggingface.co/ xlm-roberta-base 3991 Language #Document(M) #Sentence(M) Size(GB) af 0.023 0.522 0.107 ar 2.823 42.659 11.786 bg 0.919 14.743 5.217 bn 0.750 9.217 4.264 cs 3.980 55.754 9.668 de 21.410 310.942 66.333 el 1.740 24.334 9.737 en 130.087 2,215.534 479.099 es 17.569 267.764 58.774 et 0.347 5.252 0.877 eu 0.342 5.216 0.613 fr 15.819 267.888 58.023 fa 2.506 43.570 13.831 fi 1.530 23.790 3.940 fy 0.027 0.537 0.054 gu 0.039 0.519 0.228 gd 0.009 0.126 0.020 he 0.755 12.338 3.073 hi 0.536 7.303 3.762 hu 1.816 29.962 6.421 id 3.417 60.908 11.528 it 9.336 133.006 30.854 ja 27.967 588.926 71.785 jv 0.002 0.138 0.030 ka 0.141 1.756 0.766 kk 0.061 1.545 0.448 ko 11.609 227.396 27.837 lt 0.552 7.996 1.480 lv 0.281 4.159 0.798 ms 0.334 3.762 0.455 ml 0.162 2.615 1.025 my 0.045 0.893 0.306 mr 0.059 0.708 0.365 pl 6.642 93.760 19.082 pt 8.623 128.107 25.612 ne 0.080 0.829 0.429 nl 6.513 85.997 16.648 ru 35.887 580.291 203.105 ro 1.944 31.929 7.056 si 0.132 2.927 0.902 sw 0.057 0.945 0.179 ta 0.876 20.376 6.422 te 0.288 4.995 1.721 tr 18.547 291.081 40.321 th 6.278 117.826 27.941 tl 0.166 5.611 0.679 vi 12.183 234.071 37.919 ur 0.460 7.509 2.003 yo 0.0002 0.003 0.0005 zh 27.067 497.408 87.005 Total 382.735 6,475.444 1,360.526 Table 6: The statistics of monolingual pre-training corpus. 3992 Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) Pair #Sent(K) af-ar 12.34 bg-my 0.08 de-he 12751.69 en-tr 46584.82 eu-zh 19.76 fy-vi 34.95 id-pt 6825.29 ko-sw 6.74 pl-es 46863.47 af-bg 18.19 bg-ne 0.01 de-hi 106.11 en-ur 781.60 fa-fi 4485.62 gd-es 21.62 id-ro 7944.59 ko-ta 13.74 pl-pt 72437.93 af-bn 1.19 bg-nl 30757.50 de-hu 24409.40 en-vi 3563.39 fa-fr 4507.06 gd-it 13.26 id-ru 5039.44 ko-te 0.93 pl-ru 19170.23 af-cs 17.93 bg-pl 33043.03 de-id 4786.89 en-yo 0.13 fa-he 4944.80 gd-pl 12.29 id-si 366.00 ko-th 230.84 pl-sw 1424.02 af-de 19.28 bg-pt 30058.54 de-it 35936.62 en-zh 28952.02 fa-hi 186.23 gd-pt 18.90 id-sw 30.56 ko-tl 1.21 pl-tl 1039.37 af-el 29.83 bg-ro 38925.52 de-ja 1472.72 es-et 18090.74 fa-hu 5201.51 gd-ru 10.39 id-ta 35.37 ko-tr 1246.58 pl-tr 32470.18 af-en 44.70 bg-ru 17423.43 de-ka 123.12 es-eu 793.59 fa-id 3220.00 gd-tr 14.12 id-te 13.30 ko-ur 57.21 pl-ur 391.99 af-es 34.31 bg-si 460.50 de-kk 3.72 es-fa 5696.70 fa-it 4243.56 he-hi 57.85 id-th 1562.94 ko-vi 345.79 pl-vi 3790.71 af-et 6.34 bg-sw 10.80 de-ko 776.89 es-fi 34222.07 fa-ja 1072.14 he-hu 23959.87 id-tl 7.80 ko-zh 56.43 pt-ro 33802.95 af-fa 3.07 bg-ta 27.14 de-lt 9134.99 es-fr 96233.21 fa-ka 96.32 he-id 6362.29 id-tr 8017.99 lt-lv 6546.76 pt-ru 14698.48 af-fi 10.25 bg-te 17.14 de-lv 8532.06 es-he 27060.49 fa-kk 1.01 he-it 19908.66 id-ur 172.71 lt-ml 66.40 pt-si 450.40 af-fr 18.56 bg-th 2733.84 de-ml 294.16 es-hi 85.35 fa-ko 627.97 he-ja 1683.29 id-vi 2081.70 lt-ms 393.89 pt-sw 13.06 af-fy 36.94 bg-tl 6.69 de-ms 1228.82 es-hu 43947.78 fa-lt 615.78 he-ka 149.06 id-zh 356.46 lt-nl 7497.18 pt-ta 26.37 af-he 14.53 bg-tr 31179.35 de-my 0.68 es-id 8015.69 fa-lv 228.40 he-kk 2.38 it-ja 1613.05 lt-pl 9965.36 pt-te 19.32 af-hi 1.15 bg-ur 71.60 de-ne 0.28 es-it 49423.51 fa-ml 308.49 he-ko 1094.72 it-ka 106.70 lt-pt 7663.84 pt-th 2561.09 af-hu 16.32 bg-vi 2855.13 de-nl 34909.49 es-ja 1929.41 fa-ms 1072.22 he-lt 1220.91 it-kk 2.54 lt-ro 5786.22 pt-tl 10.35 af-id 4.56 bg-zh 746.27 de-pt 32610.10 es-ka 181.19 fa-my 0.06 he-lv 461.81 it-ko 1125.97 lt-ru 950.02 pt-tr 27428.79 af-it 15.01 bn-cs 340.51 de-ro 24261.82 es-kk 2.48 fa-ne 0.01 he-ml 250.07 it-lt 7359.92 lt-si 106.53 pt-ur 73.57 af-ja 1.98 bn-de 346.51 de-ru 10904.25 es-ko 1229.50 fa-nl 5010.64 he-ms 1455.61 it-lv 6607.27 lt-sw 0.02 pt-vi 2963.83 af-lt 0.65 bn-el 340.94 de-si 324.86 es-lt 7702.99 fa-pt 4998.09 he-my 0.05 it-ml 235.96 lt-ta 13.04 pt-yo 0.05 af-lv 1.08 bn-en 752.08 de-sw 45.61 es-lv 6703.10 fa-ro 5714.73 he-nl 22186.61 it-ms 1269.97 lt-te 9.71 pt-zh 846.44 af-ml 2.18 bn-es 480.35 de-ta 42.32 es-ml 339.71 fa-ru 4205.20 he-pl 24962.23 it-my 0.36 lt-th 263.89 ro-ru 19568.56 af-ms 1.31 bn-et 252.68 de-te 12.81 es-ms 1731.36 fa-si 292.78 he-pt 21226.36 it-ne 1.02 lt-tl 1.36 ro-si 504.24 af-nl 22.61 bn-eu 42.42 de-th 1695.53 es-my 2.50 fa-sw 69.51 he-ro 26370.15 it-nl 37644.29 lt-tr 1377.40 ro-sw 10.72 af-pl 1096.89 bn-fa 391.89 de-tl 12.91 es-ne 2.87 fa-ta 83.30 he-ru 14873.77 it-pl 35037.31 lt-ur 4.47 ro-ta 33.50 af-pt 22.68 bn-fi 279.35 de-tr 17579.53 es-nl 46908.79 fa-te 10.11 he-si 435.87 it-pt 35301.98 lt-vi 486.84 ro-te 24.44 af-ro 32.19 bn-fr 373.13 de-ur 218.89 es-pt 47542.26 fa-th 1201.04 he-sw 0.06 it-ro 32153.38 lt-zh 40.65 ro-th 2874.73 af-ru 15.41 bn-he 302.62 de-vi 2284.70 es-ro 48229.60 fa-tl 7.02 he-ta 23.99 it-ru 17669.12 lv-ml 23.32 ro-tl 8.61 af-si 0.98 bn-hi 38.68 de-zh 587.96 es-ru 55569.05 fa-tr 6217.24 he-te 18.65 it-si 366.97 lv-ms 163.28 ro-tr 36549.61 af-ta 1.13 bn-hu 321.36 el-en 55078.46 es-si 512.22 fa-ur 568.00 he-th 2666.00 it-sw 15.77 lv-nl 6622.81 ro-ur 73.55 af-th 2.08 bn-id 360.65 el-es 46876.21 es-sw 41.33 fa-vi 1514.04 he-tl 6.58 it-ta 17.39 lv-pl 9460.93 ro-vi 3207.73 af-tr 24.22 bn-it 301.31 el-et 16463.57 es-ta 31.19 fa-zh 372.10 he-tr 25179.32 it-te 9.93 lv-pt 6672.14 ro-zh 947.91 af-vi 3.30 bn-ja 142.19 el-eu 673.93 es-te 21.76 fi-fr 28973.81 he-ur 20.57 it-th 2447.55 lv-ro 4833.77 ru-si 340.11 ar-bg 23090.32 bn-ka 8.68 el-fa 5137.52 es-th 2976.49 fi-he 17820.49 he-vi 2813.73 it-tl 13.30 lv-ru 435.73 ru-sw 84.77 ar-bn 378.28 bn-ko 93.92 el-fi 28885.65 es-tl 13.55 fi-hi 55.60 he-zh 563.24 it-tr 25770.29 lv-si 34.42 ru-ta 61.50 ar-cs 24147.25 bn-lt 96.24 el-fr 38560.84 es-tr 39805.02 fi-hu 27350.30 hi-hu 60.05 it-ur 69.89 lv-sw 0.01 ru-te 10.80 ar-de 12733.65 bn-lv 41.21 el-he 22042.85 es-ur 79.44 fi-id 5806.36 hi-id 85.85 it-vi 2542.41 lv-ta 4.10 ru-th 2194.91 ar-el 22486.60 bn-ml 93.14 el-hi 62.26 es-vi 3215.16 fi-it 26756.85 hi-it 60.12 it-yo 0.10 lv-te 4.01 ru-tl 13.43 ar-en 60392.55 bn-ms 203.84 el-hu 34559.75 es-yo 0.12 fi-ja 1599.82 hi-ja 46.14 it-zh 473.74 lv-th 108.92 ru-tr 19317.60 ar-es 57561.29 bn-my 0.78 el-id 7098.25 es-zh 28688.60 fi-ka 148.42 hi-ka 0.80 ja-ka 35.37 lv-tr 515.30 ru-ur 417.23 ar-et 9738.71 bn-ne 0.78 el-it 34337.63 et-eu 406.33 fi-kk 3.41 hi-ko 33.66 ja-kk 1.21 lv-ur 1.08 ru-vi 2289.72 ar-eu 578.30 bn-nl 331.34 el-ja 1740.08 et-fa 3085.41 fi-ko 859.31 hi-lt 23.67 ja-ko 308.30 lv-vi 209.40 ru-yo 0.10 ar-fa 5679.85 bn-pt 333.59 el-ka 167.39 et-fi 15969.08 fi-lt 7507.00 hi-lv 12.61 ja-lt 281.74 lv-zh 14.71 ru-zh 28138.59 ar-fi 17169.90 bn-ro 337.94 el-kk 2.33 et-fr 15697.59 fi-lv 6732.38 hi-ml 30.28 ja-lv 99.97 ml-ms 101.75 si-ta 6.33 ar-fr 50632.52 bn-ru 392.15 el-ko 1130.94 et-fy 51.63 fi-ml 232.48 hi-ms 40.38 ja-ml 79.78 ml-nl 268.10 si-te 1.85 ar-he 20577.16 bn-si 47.49 el-lt 7400.42 et-he 9814.49 fi-ms 1276.96 hi-my 0.01 ja-ms 489.33 ml-pt 280.62 si-th 109.38 ar-hi 96.26 bn-sw 23.91 el-lv 6549.40 et-hi 43.98 fi-nl 30693.72 hi-ne 0.04 ja-nl 1716.42 ml-ro 325.97 si-tl 3.02 ar-hu 23770.38 bn-ta 15.67 el-ml 302.85 et-hu 16819.43 fi-pl 29451.87 hi-nl 92.46 ja-pl 3295.60 ml-ru 310.59 si-tr 492.12 ar-id 6989.56 bn-th 129.60 el-ms 1547.63 et-id 4282.23 fi-pt 29269.50 hi-pl 681.08 ja-pt 1756.87 ml-si 28.01 si-ur 4.95 ar-it 20070.27 bn-tl 2.05 el-my 0.55 et-it 14462.11 fi-ro 27988.13 hi-pt 62.44 ja-ro 1843.14 ml-sw 12.47 si-vi 210.15 ar-ja 1847.98 bn-tr 441.74 el-ne 1.04 et-ja 1176.51 fi-ru 12403.26 hi-ro 82.89 ja-ru 1491.65 ml-ta 15.90 si-zh 14.28 ar-ka 161.65 bn-ur 108.74 el-nl 37188.78 et-ka 110.02 fi-si 391.99 hi-ru 142.53 ja-si 162.96 ml-th 81.03 sw-ta 6.24 ar-kk 1.28 bn-vi 219.57 el-pt 35491.54 et-kk 1.14 fi-sw 0.02 hi-si 11.41 ja-sw 6.24 ml-tl 3.30 sw-th 6.24 ar-ko 1262.60 bn-zh 85.24 el-ro 37986.26 et-ko 492.79 fi-ta 20.08 hi-sw 12.52 ja-ta 18.92 ml-tr 439.25 sw-tr 91.95 ar-lt 1177.67 cs-de 24049.84 el-ru 17052.36 et-lt 7431.17 fi-te 17.13 hi-ta 41.00 ja-te 5.68 ml-ur 100.52 sw-ur 50.29 ar-lv 433.66 cs-el 35372.28 el-si 466.44 et-lv 6728.85 fi-th 2288.65 hi-te 23.18 ja-th 632.26 ml-vi 124.30 sw-yo 0.03 ar-ml 348.33 cs-en 54470.47 el-sw 4.85 et-ml 179.99 fi-tl 5.91 hi-th 37.53 ja-tl 10.06 ml-zh 34.77 sw-zh 19.31 ar-ms 1555.33 cs-es 44962.42 el-ta 20.44 et-ms 1135.84 fi-tr 22551.99 hi-tl 0.51 ja-tr 1896.56 ms-nl 1409.07 ta-te 21.16 ar-my 0.18 cs-et 17819.46 el-te 18.10 et-nl 16560.63 fi-ur 19.43 hi-tr 176.39 ja-ur 61.41 ms-pt 1523.57 ta-th 14.15 ar-ne 0.41 cs-eu 686.53 el-th 2505.71 et-pl 19633.08 fi-vi 2517.08 hi-ur 101.10 ja-vi 679.31 ms-ro 1732.68 ta-tr 77.76 ar-nl 21273.78 cs-fa 5417.48 el-tl 10.13 et-pt 16768.45 fi-zh 630.12 hi-vi 32.99 ja-zh 104.37 ms-ru 1210.56 ta-ur 49.89 ar-pl 24819.83 cs-fi 28031.47 el-tr 31048.88 et-ro 15880.62 fr-he 21218.88 hi-zh 25.57 ka-ko 17.13 ms-si 204.06 ta-vi 12.65 ar-pt 20379.56 cs-fr 34876.02 el-ur 24.36 et-ru 6630.25 fr-hi 68.31 hu-id 7253.81 ka-lt 30.49 ms-sw 8.99 ta-zh 13.02 ar-ro 26187.15 cs-he 24503.29 el-vi 2966.14 et-si 331.22 fr-hu 37027.57 hu-it 33513.06 ka-lv 10.71 ms-ta 15.24 te-th 0.96 ar-ru 45992.72 cs-hi 86.86 el-yo 0.11 et-sw 0.01 fr-id 6235.29 hu-ja 1767.63 ka-ml 6.56 ms-te 4.70 te-tr 18.84 ar-si 483.96 cs-hu 39272.92 el-zh 649.81 et-ta 14.34 fr-it 41162.37 hu-ka 165.84 ka-ms 31.86 ms-th 413.17 te-vi 9.34 ar-sw 16.52 cs-id 7310.27 en-es 156560.00 et-te 14.44 fr-ja 1608.52 hu-kk 2.58 ka-nl 155.10 ms-tl 7.26 th-tl 7.28 ar-ta 37.15 cs-it 33935.96 en-et 22284.30 et-th 1746.50 fr-ka 139.63 hu-ko 1168.66 ka-pt 165.00 ms-tr 1754.22 th-tr 3054.07 ar-te 19.33 cs-ja 1806.97 en-eu 805.78 et-tl 3.09 fr-kk 1.34 hu-lt 7623.58 ka-ro 182.79 ms-ur 68.94 th-ur 58.65 ar-th 2959.96 cs-ka 163.35 en-fa 7462.52 et-tr 11408.82 fr-ko 991.60 hu-lv 6776.32 ka-ru 104.82 ms-vi 851.69 th-vi 672.82 ar-tl 7.58 cs-kk 1.26 en-fi 42783.36 et-ur 19.52 fr-lt 9440.34 hu-ml 279.13 ka-si 7.96 ms-zh 85.86 th-zh 133.45 ar-tr 26683.62 cs-ko 1199.62 en-fr 161519.91 et-vi 2048.37 fr-lv 8569.67 hu-ms 1581.43 ka-th 43.37 my-nl 0.10 tl-tr 14.51 ar-ur 126.33 cs-lt 7694.12 en-fy 126.19 et-zh 405.30 fr-ml 278.47 hu-my 0.06 ka-tl 1.27 my-pt 0.10 tl-vi 5.86 ar-vi 2875.00 cs-lv 6745.84 en-gd 47.02 eu-fa 245.78 fr-ms 1423.08 hu-nl 33904.34 ka-tr 178.79 my-ro 0.03 tr-ur 473.08 ar-yo 0.01 cs-ml 319.93 en-he 30028.28 eu-fi 581.61 fr-my 1.47 hu-pl 39869.14 ka-ur 1.98 my-ru 0.81 tr-vi 3178.03 ar-zh 28120.22 cs-ms 1592.17 en-hi 1844.38 eu-fr 636.16 fr-ne 1.45 hu-pt 31715.19 ka-vi 53.58 my-sw 0.15 tr-zh 1029.21 bg-bn 310.12 cs-my 0.08 en-hu 55233.87 eu-he 566.71 fr-nl 47363.70 hu-ro 38807.61 ka-zh 6.52 my-tr 0.03 ur-vi 12.52 bg-cs 34502.46 cs-ne 0.07 en-id 9677.33 eu-hi 9.98 fr-pt 42850.13 hu-ru 19172.99 kk-lt 0.83 my-ur 0.02 ur-zh 99.78 bg-de 19852.81 cs-nl 34427.07 en-it 76257.21 eu-hu 663.68 fr-ro 37249.80 hu-si 460.99 kk-lv 1.13 my-zh 0.13 vi-zh 148.22 bg-el 32130.86 cs-pt 32469.01 en-ja 2177.89 eu-id 307.85 fr-ru 54231.81 hu-sw 0.68 kk-ms 1.12 ne-nl 0.09 bg-en 47247.04 cs-ro 39226.31 en-ka 199.98 eu-it 568.66 fr-si 393.48 hu-ta 20.63 kk-nl 1.85 ne-pt 0.38 bg-es 39728.55 cs-ru 19703.43 en-kk 3.71 eu-ja 139.14 fr-sw 29.32 hu-te 17.57 kk-pl 77.88 ne-ro 0.04 bg-et 15188.54 cs-si 454.26 en-ko 1493.95 eu-ka 9.42 fr-ta 24.03 hu-th 2867.23 kk-pt 3.35 ne-ru 1.30 bg-eu 605.10 cs-sw 17.34 en-lt 10992.89 eu-ko 72.17 fr-te 11.93 hu-tl 10.79 kk-ro 2.35 ne-sw 0.05 bg-fa 4927.53 cs-ta 32.81 en-lv 9883.08 eu-lt 108.12 fr-th 2325.22 hu-tr 32494.90 kk-ru 2.22 ne-tr 0.03 bg-fi 25191.01 cs-te 18.72 en-ml 573.95 eu-lv 36.81 fr-tl 13.18 hu-ur 23.32 kk-th 0.93 ne-ur 0.06 bg-fr 30185.98 cs-th 2858.53 en-ms 2050.83 eu-ml 42.72 fr-tr 29245.91 hu-vi 2974.61 kk-tr 2.59 ne-zh 0.01 bg-he 22887.40 cs-tl 7.44 en-my 2.43 eu-ms 129.20 fr-ur 73.99 hu-zh 730.70 kk-vi 1.18 nl-pt 37775.73 bg-hi 71.38 cs-tr 32797.28 en-ne 2.89 eu-nl 619.88 fr-vi 2752.32 id-it 5831.16 ko-lt 148.54 nl-ro 36051.60 bg-hu 34293.44 cs-ur 122.87 en-nl 65918.54 eu-pt 641.30 fr-yo 0.12 id-ja 1271.31 ko-lv 57.10 nl-ru 16582.78 bg-id 7047.21 cs-vi 3040.14 en-pl 59729.77 eu-ro 715.99 fr-zh 28008.77 id-ka 85.07 ko-ml 42.92 nl-si 410.92 bg-it 27649.85 cs-zh 894.87 en-pt 61861.36 eu-ru 435.12 fy-es 49.12 id-kk 1.03 ko-ms 291.25 nl-sw 31.38 bg-ja 1658.40 de-el 30170.64 en-ro 60415.46 eu-si 34.56 fy-he 44.06 id-ko 605.78 ko-my 0.12 nl-ta 39.21 bg-ka 193.27 de-en 83872.47 en-ru 65105.13 eu-ta 3.35 fy-it 47.88 id-lt 855.43 ko-ne 0.01 nl-te 16.07 bg-kk 3.40 de-es 41634.80 en-si 601.16 eu-te 0.73 fy-ja 37.61 id-lv 342.36 ko-nl 1120.75 nl-th 2548.14 bg-ko 1056.96 de-et 15186.40 en-sw 171.65 eu-th 80.75 fy-pl 49.37 id-ml 230.67 ko-pl 2722.47 nl-tl 8.18 bg-lt 5604.11 de-eu 534.93 en-ta 125.96 eu-tl 2.60 fy-pt 95.81 id-ms 1614.63 ko-pt 1119.49 nl-tr 28822.22 bg-lv 4748.15 de-fa 3948.14 en-te 27.22 eu-tr 722.77 fy-ru 45.83 id-my 0.11 ko-ro 1242.76 nl-ur 171.71 bg-ml 283.77 de-fi 25753.06 en-th 3375.07 eu-ur 2.01 fy-sw 0.37 id-ne 0.07 ko-ru 959.46 nl-vi 2748.28 bg-ms 1506.56 de-fr 44392.06 en-tl 16.03 eu-vi 201.28 fy-tr 45.40 id-nl 6493.33 ko-si 58.66 nl-zh 866.75 Total 6,421,152.04 Table 7: The statistics of bilingual (parallel) pre-training corpus. 3993 Pre-training Hyperparameters Large Small Number of layers 24 6 Hidden Size 1024 768 FFN inner hidden size 4096 3072 Attention heads 16 12 Attention head size 64 64 Embedding Size 1024 768 Mask percent (monolingual/ bilingual) 15%/25% 15%/25% Learning Rate Decay Linear Linear Warmup steps 12k 12k Learning Rate 2e-4 3e-4 Adam ϵ 1e-6 1e-6 Adam β1 0.9 0.9 Adam β2 0.98 0.999 Attention Dropout 0.1 0.1 Dropout 0.1 0.1 Weight Decay 0.01 0.01 Max Sequence Length (monolingual/bilingual) 512/128 512/128 Batch Size (monolingual/bilingual) 1024/4096 1024/4096 Train Steps 240k 240k Total Parameters 662M 247M Table 8: The pre-training hyperparameters. Model en ar bg de el es fr hi ru sw th tr ur vi zh Avg. Cross-lingual Transfer XLM-R 88.7 77.2 83.0 82.5 80.8 83.7 82.2 75.6 79.1 71.2 77.4 78.0 71.7 79.3 78.2 79.2 VECOout 88.2 79.2 83.1 82.9 81.2 84.2 82.8 76.2 80.3 74.3 77.0 78.4 71.3 80.4 79.1 79.9 Translate-Train-All XLM-R 88.6 82.2 85.2 84.5 84.5 85.7 84.2 80.8 81.8 77.0 80.2 82.1 77.7 82.6 82.7 82.6 VECOout 88.9 82.4 86.0 84.7 85.3 86.2 85.8 80.1 83.0 77.2 80.9 82.8 75.3 83.1 83.0 83.0 VECOin 89.3 83.7 87.0 85.9 85.8 87.3 86.7 81.8 83.6 79.9 82.5 84.3 77.7 84.4 84.0 84.3 Table 9: XNLI accuracy scores for each language. Model en de es fr ja ko zh Avg. Cross-lingual Transfer XLM-R 94.7 89.7 90.1 90.4 78.7 79.0 82.3 86.4 VECOout 96.2 91.3 91.4 92.0 81.8 82.9 85.1 88.7 Translate-Train-All VECOout 96.4 93.0 93.0 93.5 87.2 86.8 87.9 91.1 VECOin 96.5 94.4 94.3 94.0 89.0 90.3 91.0 92.8 Table 10: PAWS-X accuracy scores. Model de fr ru zh Avg. Cross-lingual Transfer XLM-R 67.5 66.5 73.5 56.7 66.0 VECOout 89.6 84.6 87.4 78.5 85.0 Translate-Train-All VECOout 93.0 88.7 89.9 85.7 89.3 VECOin 95.4 91.9 93.1 89.9 92.6 Table 11: BUCC F1 results. Model af ar bg de el en es et eu fa fi fr he hi hu id it Cross-lingual Transfer XLM-R 89.8 67.5 88.1 88.5 86.3 96.1 88.3 86.5 72.5 70.6 85.8 87.2 68.3 76.4 82.6 72.4 89.4 VECOout 88.3 67.4 87.4 88.5 86.7 95.9 89.0 87.8 75.1 70.9 86.2 88.9 67.5 76.2 82.9 72.9 89.9 Translate-Train-All VECOin 92.5 73.7 93.4 91.8 90.4 95.2 91.3 90.6 79.1 79.8 89.5 91.4 79.1 80.6 88.4 74.8 91.8 ja kk ko mr nl pt ru ta te th tl tr ur vi yo zh Avg. Cross-lingual Transfer XLM-R 15.9 78.1 53.9 80.8 89.5 87.6 89.5 65.2 86.6 47.2 92.2 76.3 70.3 56.8 24.6 25.7 73.8 VECOout 31.4 79.3 53.1 84.3 89.8 88.3 90.2 64.3 85.8 48.0 93.7 77.2 69.2 58.1 26.2 39.4 75.1 Translate-Train-All VECOin 45.1 78.0 63.7 84.5 92.7 90.1 92.6 72.6 88.5 55.2 88.8 76.8 75.0 70.5 24.3 63.0 79.8 Table 12: POS results (Accuracy) for each language. 3994 Model en af ar bg bn de el es et eu fa fi fr he hi hu id it ja jv Cross-lingual Transfer XLM-R 84.7 78.9 53.0 81.4 78.8 78.8 79.5 79.6 79.1 60.9 61.9 79.2 80.5 56.8 73.0 79.8 53.0 81.3 23.2 62.5 VECOout 83.8 77.5 48.2 83.9 77.2 79.4 79.3 75.4 80.4 68.3 68.2 80.6 80.1 55.0 71.0 80.9 52.9 81.7 19.4 63.2 Translate-Train-All VECOin 80.7 82.5 66.4 84.1 78.4 82.2 82.4 79.7 84.7 78.2 68.8 84.9 79.1 69.7 76.6 85.1 77.3 83.8 21.3 70.3 ka kk ko ml mr ms my nl pt ru sw ta te th tl tr ur vi yo zh Cross-lingual Transfer XLM-R 71.6 56.2 60.0 67.8 68.1 57.1 54.3 84.0 81.9 69.1 70.5 59.5 55.8 1.3 73.2 76.1 56.4 79.4 33.6 33.1 VECOout 67.1 51.2 59.9 63.4 65.0 70.0 56.1 83.4 83.1 71.3 70.5 60.5 56.2 1.4 71.3 80.4 69.3 76.0 37.4 29.1 Translate-Train-All VECOin 77.0 67.2 71.0 73.3 74.1 71.8 63.8 85.5 80.8 72.8 77.0 69.1 67.5 2.6 74.0 85.2 71.5 76.4 32.8 31.0 Table 13: NER results (F1) for each language. Model en ar de el es hi ru th tr vi zh Avg. Cross-lingual Transfer XLM-R 86.5 / 75.7 68.6 / 49.0 80.4 / 63.4 79.8 / 61.7 82.0 / 63.9 76.7 / 59.7 80.1 / 64.3 74.2 / 62.8 75.9 / 59.3 79.1 / 59.0 59.3 / 50.0 76.6 / 60.8 VECOout 87.6 / 76.5 73.6 / 56.1 79.8 / 62.2 79.6 / 61.6 81.2 / 61.6 74.7 / 57.6 78.7 / 62.1 72.8 / 60.6 75.1 / 58.3 79.0 / 59.8 69.2 / 59.2 77.3 / 61.8 Translate-Train-All VECOout 88.3/77.9 76.9/61.1 80.5/64.6 81.5/64.1 84.2/66.8 78.8/62.5 80.2/66.1 77.0/70.4 77.8/62.2 82.5/63.7 71.6/69.4 79.9/66.3 VECOin 90.2/79.5 81.8/66.4 85.4/69.8 85.3/69.0 87.2/70.8 83.7/67.9 85.6/71.6 80.0/74.7 82.4/68.6 85.8/68.3 74.9/73.1 83.9/70.9 Table 14: XQuAD results (F1 / EM) for each language. Model en ar de es hi vi zh Avg. Cross-lingual Transfer XLM-R 83.5 / 70.6 66.6 / 47.1 70.1 / 54.9 74.1 / 56.6 70.6 / 53.1 74.0 / 52.9 62.1 / 37.0 71.6 / 53.2 VECOout 83.6 / 70.5 65.0 / 44.6 69.8 / 54.6 73.8 / 55.6 69.1 / 51.4 73.1 / 51.8 67.3 / 43.6 71.7 / 53.2 Translate-Train-All VECOout 84.1/71.3 67.8/47.1 70.7/55.8 74.6/56.6 71.1/53.4 74.8/54.4 68.8/45.8 73.1/54.9 VECOin 87.5/75.5 72.3/52.1 75.7/61.1 78.8/61.6 76.6/58.6 79.3/59.1 72.1/46.8 77.5/59.3 Table 15: MLQA results (F1 / EM) for each language. Model en ar bn fi id ko ru sw te Avg. Cross-lingual Transfer XLM-R 71.5 / 56.8 67.6 / 40.4 64.0 / 47.8 70.5 / 53.2 77.4 / 61.9 31.9 / 10.9 67.0 / 42.1 66.1 / 48.1 70.1 / 43.6 65.1 / 45.0 VECOout 71.3 / 58.2 73.1 / 52.8 58.9 / 42.5 70.9 / 55.1 77.2 / 60.0 54.2 / 39.9 66.1 / 37.6 65.8 / 45.7 70.6 / 50.7 67.6 / 49.1 Translate-Train-All VECOout 77.2/64.8 77.0/57.5 72.2/56.6 76.6/59.3 80.0/64.4 63.4/52.2 72.8/50.5 79.4/67.1 76.0/58.0 75.0/58.9 VECOin 79.4/65.2 80.1/60.9 80.8/68.1 81.6/65.5 84.3/69.7 65.4/50.4 77.8/55.8 83.7/74.1 81.0/63.4 79.4/63.7 Table 16: TyDiQA-GolP results (F1 / EM) for each language. Model af ar bg bn de el es et eu fa fi fr he hi hu id it ja Cross-lingual Transfer XLM-R 58.2 47.5 71.6 43 88.8 61.8 75.7 52.2 35.8 70.5 71.6 73.7 66.4 72.2 65.4 77 68.3 60.6 VECOout 48.2 70.9 86.7 57.7 97.5 81.5 94.8 89.7 62.9 82.1 87.9 88.8 74.7 80.7 87.6 89.6 89.2 83.2 Translate-Train-All VECOout 80.9 85.1 91.3 78.1 98.5 89.5 97.4 94.8 79.8 93.1 95.4 93.7 85.8 94.2 93.8 93.0 92.2 92.8 VECOin 88.5 88.7 91.5 84.2 98.9 91.5 97.9 96.4 85.8 95.3 95.9 95.6 89.6 97.0 95.1 94.2 94.1 94.0 jv ka kk ko ml mr nl pt ru sw ta te th tl tr ur vi zh Cross-lingual Transfer XLM-R 14.1 52.1 48.5 61.4 65.4 56.8 80.8 82.2 74.1 20.3 26.4 35.9 29.4 36.7 65.7 24.3 74.7 68.3 VECOout 17.6 58.5 53.9 75.3 80.1 64.2 94.4 92.8 88.6 37.4 61.9 65.8 84.5 52.5 89.3 64.3 85.8 82.7 Translate-Train-All VECOout 35.1 83.0 74.1 88.7 94.8 82.5 95.9 94.6 92.2 69.7 82.4 91.0 94.7 73.0 95.2 63.8 95.1 93.9 VECOin 49.3 86.6 83.7 91.2 97.1 87.9 97.6 96.1 93.8 82.6 88.9 95.3 95.1 79.8 97.6 91.4 97.2 95.2 Table 17: Tatoeba results (Accuracy) for each language
|
2021
|
308
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 3995–4007 August 1–6, 2021. ©2021 Association for Computational Linguistics 3995 A unified approach to sentence segmentation of punctuated text in many languages Rachel Wicks1 and Matt Post1,2 1Center for Language and Speech Processing 2Human Language Technology Center of Excellence Johns Hopkins University [email protected], [email protected] Abstract The sentence is a fundamental unit of text processing. Yet sentences in the wild are commonly encountered not in isolation, but unsegmented within larger paragraphs and documents. Therefore, the first step in many NLP pipelines is sentence segmentation. Despite its importance, this step is the subject of relatively little research. There are no standard test sets or even methods for evaluation, leaving researchers and engineers without a clear footing for evaluating and selecting models for the task. Existing tools have relatively small language coverage, and efforts to extend them to other languages are often ad hoc. We introduce a modern context-based modeling approach that provides a solution to the problem of segmenting punctuated text in many languages, and show how it can be trained on noisily-annotated data. We also establish a new 23-language multilingual evaluation set. Our approach exceeds high baselines set by existing methods on prior English corpora (WSJ and Brown corpora), and also performs well on average on our new evaluation set. We release our tool, ERSATZ, as open source. 1 Introduction In many ways, the sentence is the fundamental unit of text in natural language processing (NLP). From the user perspective, tasks such as sentiment analysis, POS tagging, or machine translation consume sentences and emit classifications, annotations, or transductions of those inputs. Even tasks that operate at the paragraph or document level, such as coreference resolution or summarization, often make use of sentences internally. Yet at the same time, sentences in the wild rarely exist with marked sentence boundaries. For many languages, punctuation serves as a cue for these Examples of Ambiguity in Punctuated Contexts en ... in the U.S. ⊗House of Representatives ... ...in the U.S. ✓Most Mexican Spanish ... cs ... podnikanie s.r.o. ⊗a hlavním investorem ... a Systémy s.r.o. ✓V roce 2017 ... ro ... W. Pauli s,.a. ⊗constituie direc¸tii ... ... de Robles s,.a. ✓A jucat în ... Table 1: Examples of ambiguous FULL STOP punctuation in English, Czech, and Romanian from Wikipedia. ✓denotes a sentence boundary while ⊗denotes an ambiguous sentence-internal position. boundaries, but this punctuation is ambiguous— as we might see with acronyms or abbreviations in English. When segmented sentences are required, they must be split using a sentence segmentation technique that can resolve these ambiguities. Despite its importance and early position in the NLP pipeline, sentence segmentation is the subject of relatively little research. Widely-used tools such as that in Moses (Koehn et al., 2007) are implemented with ad-hoc, manually-designed, language-specific rules, leaving them vulnerable to the long tail of languages and language phenomena. The little comparative work that does exist generally focuses on techniques that work in English or other Indo-European languages (Palmer and Hearst, 1997; Gillick, 2009). Secondly, there is not a well-understood methodology for training segmenters that do not make narrow assumptions about the features or characteristics of the languages they support. At the heart of this is the lack of labeled training data. Manually-split datasets that accompany annotation projects tend to be small, and larger datasets are typically (imperfectly) segmented by the very tools whose performance is under question. Tools such 3996 as NLTK (Bird and Loper, 2004), which packages Punkt (Kiss and Strunk, 2006), provide an unsupervised method to train a model, but it is unclear what the effect is when switching to non-Latin-script languages, or how a more supervised approach would handle such noisy data. Finally, and perhaps most importantly, there are no standard test sets or even metrics for evaluating segmenter performance, leaving researchers and engineers with no objective way to determine which one is best. The work described in this paper is aimed at these problems. We propose a simple window-based model and semi-supervised training paradigm for the segmentation of punctuated text (§3). We frame the task as binary classification applied to a set of candidate punctuation locations defined by a regular expression. Leveraging the similarity of the task across languages (Table 1), we show that our model is able to successfully bootstrap from multilingual data that has been imperfectly segmented. We define a common metric that works across different tools (§4), and assemble a multilingual test suite by semi-automatically splitting existing (undersegmented) test sets (§5), providing a basis for proper comparison. We release these data splits along with our tool, ERSATZ, as open source.1 2 Background A sentence is a sequence of grammatically linked words that conveys a complete thought. The term can be difficult to define in a precise manner that will not admit any exceptions, and in applications like machine translation, there are many times where the basic input unit is not a sentence, but a sentence fragment, such as a headline or an item from a list. In this work, we skirt these complexities, choosing instead to focus on the most common scenario, in which we are dealing with standard written language. For this, we adopt a functional definition: a sentence is a group of words that ends with a sentence-ending punctuation mark, such as (for many languages) a period, question mark, or exclamation point. Since punctuation is often used for non-sentence-ending purposes as well, the primary challenge for sentence segmentation is resolving this ambiguity for each segmentation candidate. 1https://github.com/rewicks/ersatz or pip install ersatz. Research in sentence segmentation2 has been limited in scope. Prior work either introduces methods that work under a set of assumptions unique to Latin-script languages (the existence and importance of casing, word length, or whitespace), or tackles new languages ad hoc, making adaptation to new languages and domains difficult. Statistical methods use text-based features such as casing, punctuation, or length of surrounding words to make decisions around punctuation. The earliest work we found (Riley, 1989) considered all sentence boundaries and used decision trees based on these features. Gillick (2009) trained two statistical models in the form of an SVM and Naive Bayes classifier. Palmer and Hearst (1997) introduced Satz and shifted the approach by only focusing potential sentence boundaries being near sentence-ending punctuation, using part-of-speech distribution vectors as input to a feed-forward neural network and additionally applied their technique to German and French. In order to work without labeled data, Kiss and Strunk (2006) used heuristics to craft scores based on likelihood values of occurrences of tokens, punctuation, casing and token length, and then manually tune a threshold of score to indicate a sentence boundary. This work expanded the most multilingually, considering 10 Indo-European languages as well as Estonian and Turkish. Other work has focused on specific non-English languages. Xue and Yang (2011) study Chinese and dissect the theoretical reasons behind segmenting Chinese sentences to match their English equivalents. To segment Thai, which lacks punctuation, Zhou et al. (2016) use POS-taggers. Some work has tackled the problem of domains. Sanchez (2019) approaches the problem of legal text, which has a set structure without punctuation; other approaches (Wang et al., 2019; Rehbein et al., 2020) have investigated speech, which lacks both punctuation and written textual structure. A popular splitter is packaged in the Moses toolkit (Koehn et al., 2007),3 which works by splitting on all sentence-final punctuation unless the preceding context is a “non-breaking prefix”—a hand-built, language-specific list of acronyms and abbreviations. This approach cannot resolve the ambiguity where punctuation legitimately exists at the end of a sentence and is indifferent to novel 2Alternately called sentence boundary detection. 3We use the repackaged Python module at https:// pypi.org/project/sentence-splitter/. 3997 abbreviations at inference time. It produces a conservative segmenter that is high precision (unlikely to oversegment) but low recall (prone to undersegmenting). This raises the question of what effect reliance on this tool has had on construction of recent massive bitexts, such as CCMatrix (Schwenk et al., 2019b, §4.3). Gillick (2009) credit a 0.75% increase in accuracy to reduction of summarization error by a factor of four. Errors in segmentation may therefore affect the top matches for a sentence when doing bitext construction. Another popular splitter is SpaCy, which has not been described or evaluated anywhere, as far as we could tell. With sentence splitting being a crucial piece of modern corpus creation for machine translation and other tasks, the lack of approaches and rigorous comparisons between tools limits the field. Additionally, the research field moving towards (often massively) multilingual settings, the need to build multilingual tools compare them in a proper scientific framework is both important and evident. 3 Approach Our general approach is to treat sentence segmentation as a binary classification problem, predicting sentence-internal (⊗) or sentence-ending (✓) positions. The input to the model (§3.1), shown in Figure 1, is the concatenated left and right token contexts, as depicted in Table 1. Predictions for both training and inference are done only at predefined candidate sites, which are determined by a regular expression (§3.2). We then train in a semisupervised setting where many of the labels may be missing (§3.3). 3.1 Models Our basic model is depicted in Figure 1. The encoder is a two-layer Transformer (Vaswani et al., 2017). Our hyperparameter search incorporates vocabulary size (V ), embedding size (e), and left and right context sizes (l and r). We also experiment with simpler architectures (§8.4), including single blocks of fully-connected linear layers with a TanH activation.4 These simpler models typically traded increased throughput for slight degradations in F1. Our training objective is binary cross-entropy loss. 4We initially experimented with various functions and layers (Sigmoid, ReLU, Pooling layers, etc) but found that TanH performs best. ENCODER !in !the !U.S. !Mr. !Rog ers … … linear + softmax ⊗ ✓ 0.64 0.46 embeddings (1) (2c, e) (2c, e) (V) (V) (V) (V) (V) (V) Figure 1: Model architecture. A binary predictor is constructed from token embeddings from the left and right context. Arrows denote output dimensions: V is the vocabulary, l and r the left and right context window sizes, and e the model/embedding size. 3.2 Candidate sites Our model works with segmentation candidate sites for both training and inference. This can be done in a fairly general, language-agnostic way. Let P be the set of all punctuation, and Pe ⊂P be the set of sentence-ending punctuation. For a given input, we examine every character boundary and match based on two regular expressions for the left and right context, respectively: • (.*PeP*) : The left context ends with sentence-final punctuation, optionally followed by any amount of punctuation; and • ([^0-9].*) : The right context does not start with a number. Raw text examples can be found in Table 1 and tokenized examples with fixed context sizes are shown in Table 2. Input to the model is in the form of documents. A linear pass over the data identifies all candidates sites and assembles them into a batch, with their associated left and right contexts. At training time, instances are extracted with their labels: ⊗for lineinternal sites, and ✓for sites that occur between input lines. At inference time, the trained classifier is applied, and newlines are inserted where ✓is predicted. This general definition carries benefits and risks. On the positive side, it allows us to work with many languages without having to develop language3998 Label Left context Right context ⊗ _the _ P . K . _ S h t ⊗ s o on er ? " _ h e _ ✓ B . A . T . _ " W e ✓ n er s . " ) _ I _ st Table 2: Candidate site examples with their labels. Left context-size (6) and right context-size (4) occurs after subword tokenization. In text, ‘_’ is subword beginning-of-word character. specific rules. It also speeds up training and inference, boosting both training speed and inference performance. On the downside, this loose definition can permit oversegmentation, since it permits, for example, word-internal segmentation in English and other languages. The criteria for identifying candidate sites can be easily altered to be more constrained or more general depending upon use case, and the list of punctuation to support more languages, if necessary. Our default list covers many languages.5 3.3 Training data As noted in our motivation, sentences in the wild are often not segmented but are part of paragraphs and documents. It is therefore unsurprising to find many segmentation errors in existing corpora. A particular problem one can observe is that of undersegmentation, perhaps resulting from application of conservative segmentation tools. This means the raw training data may contain many false negatives (✓sites mistakenly labeled as ⊗). Training a sentence segmentation model therefore presents a chicken-and-egg problem. We aim to train directly on existing data created for MT purposes, despite its having been either segmented by imperfect segmenters, or never segmented. While some data is undersegmented, the vast majority of the end-of-line contexts should be correct, since they are either (a) natural existing boundaries at the end of a paragraph or document or (b) the result of applying a conservative segmenter. We therefore hope to train classifiers even despite this noise. Because we are considering a binary classification problem (and using the associated binary cross entropy loss), we additionally consider 5Our punctuation set (by unicode name): Full Stop, Question Mark, Exclamation Mark, Ellipsis, Ideographic Full Stop, Devanagari Danda, Arabic Question Mark, Arabic Full Stop, Khmer Sign Khan adding a weighted λ value to the ✓class in order to give more credence to these contexts.6 For punctuation at the end of a line, the rightcontext is taken from the tokens at the beginning of the next sentence. In Section §7.3, we look into whether it matters if this right context is the true document context, or whether a random sentence will serve. 4 Evaluation: Metric For evaluation, we begin by removing sentences that do not end in punctuation, since none of the tools are able to segment these. We then concatenate the test set into a single line, joining sentences with a space. Evaluation among different tools contains subtle complexities. First, some tools normalize or tokenize the input text, complicating alignment between the input and the output. Second, different tools may attempt to segment at different subsets of input string locations, which might unfairly bias the results in favor of conservative tools. Finally, if we permit segmentation at any point of the input, there is a large class imbalance between ⊗and ✓. The class imbalance advocates for F1 as a natural metric. The use of F1 also addresses the second issue, since only the gold positive class (✓) factors into the score. The first two issues also require that we align a segmenter’s output with the gold standard segmented text. Since the texts are largely similar, we can do this efficiently using a modified Levenshtein distance7 that only considers a fixed maximum distance between any two characters. Once the text is aligned, we compute F1 against the set of ✓symbols in the gold text. An example is depicted in Figure 2. 5 Evaluation: Data We have noted the difficulty with making use of imperfect training data, and how we hope to work around it (§3.3). Unfortunately, this workaround cannot be used for evaluation, where we need goldstandard data. We construct test sets from the WMT News Translation test sets (Barrault et al., 2020), which 6Generally, we find no weight (λ = 1.0) is sufficient in punctuated English, but increasing the weight (λ = 20) improved performance in some languages and the multilingual setting where the data is noisier. 7While the distance itself can also be considered in comparing tools, we do not report these distances, and instead use the technique to align text within the window. 3999 … him. He added: “Mr. Rogers” … h i m . ✓H e a d d e d : “ M r . R o g e r s h i m . ✓H e a d d e d : " M r . ✓R o g e r s h i m . ✓H e a d d e d : ✓' ' M r . ✓R o g e r s text gold sys1 sys2 P R F1 – – – 0.5 1.0 0.67 0.3 1.0 0.5 Figure 2: Input text formatted as gold-standard data with two system outputs. Gold positive labels are marked with ✓. For scoring, system outputs are independently aligned to the gold text, which accounts for text transformations made by some tools and allows precision and recall to be computed. provides for decent-size test sets in many languages. We manually corrected all sentence segmentations. While some sets were already wellsegmented, some more recent years were extremely under-segmented. In Table 5, we show the test sets’ line counts before and after manual correction.8 Additionally, we report the % of candidate sites with a true ✓label, which provides a measure of the ambiguity of the punctuation. Many ⊗positions occur in acronyms, such as “U.S.A.", embedded quotes, ellipsis, or in company names such as “Yahoo!". 6 Experimental Setup We consider three language settings: (i) monolingual English, (ii) a multilingual setting that includes the set of recent WMT languages plus Arabic, and (iii) a much larger multilingual setting that includes the previous languages plus all languages with at least 10k lines in the WikiMatrix (Schwenk et al., 2019a) dataset. Starting with the English setting, we investigate the performance of a basic model and vary parameters such as context size, embedding size, and vocabulary size. After finding an optimal setting, we expand to the first multilingual setting and repeat. We train a single multilingual model that is agnostic of language and does not need language specification as input. Similar to the monolingual setting, we vary the aforementioned parameters, and compare the best model to baselines (§6.3). In order to test expandability, we then train with the same parameters on the largest set of languages (using the additional WikiMatrix data), and compare to the previous model’s performance. While we do not widely experiment with additional monolingual settings, we train monolingual models in each language to compare against the multilingual models’ performance. We report the 8iu was left uncorrected due to the fact that available bitext often aligned “sentences" with singular or compound sentences in English and a lack of automatic translation corresponding to sentences. comparison of these three settings to baselines in Table 5. 6.1 Datasets We train our English model on a subset of the WSJ9 and the English News Commentary datasets provided by WMT.10 To expand to a multilingual setting, we consider the set of all WMT Task languages and Arabic (23 in total) allowing us to leverage the various monolingual datasets (Joanis et al., 2020) released as part of the WMT workshops—often using News Commentary datasets, as well as WikiMatrix (Schwenk et al., 2019a), CCMatrix (Schwenk et al., 2019b), and Global Voices (Nguyen and Daumé III, 2019). For validation data, we use WMT test sets when available, and IWSLT (Cettolo et al., 2017) for Arabic. We experimented with (i) balancing the data so each language has equal amounts of data, (ii) normalizing the amount of data per language based on the relative ambiguity (measured by percent of candidate sites labeled as true ✓), and (iii) using all available data. We find that the third method performs the best and thus report under this setting. In the larger multilingual setting, we consider all WikiMatrix languages with more than 10k unique lines (64 additional languages) and do not expand the validation set. For a complete list of datasets, please see Table 7 in Appendix A. 6.2 Training For each vocabulary size, we train a SentencePiece (Kudo and Richardson, 2018) model over the training data. We use a binary cross-entropy loss over the labels, Adam optimizer with a learning rate of 0.0001, and a λ of 1.0 (English) and 20.0 (multilingual) 9Sections 1-2, 7-23 for training; section 24 for validation, and sections 03-06 for test in order to mirror the splits in Bird and Loper (2004) 10http://data.statmt.org/ news-commentary/v15/ 4000 on the ✓class (with the exception of the experiments in §7.4). We use a batch size of 25k instances, and compute F1 over the validation data every 500 batches, saving the model with the highest inference-time F1 score. This is the collective F1 score across all languages in the multilingual settings. If the model has not improved in 15 validations, training terminates. The models were trained on a Tesla V100 GPU. The monolingual models took approximately 2 hours to train while the multilingual models took approximately 10-15 hours. 6.3 Baselines We use the following existing tools as baselines: Always split on every candidate site. This serves as a lower-bound for our precision metric. Splitta (Gillick, 2009) ships with both SVM and Naive Bayes models. It targets English texts. We found similar performance and only report the Naive Bayes scores. NLTK Punkt Kiss and Strunk (2006) introduce an unsupervised training method for this task which uses frequency of occurences of input features such as casing, punctuation, and length in order to segment. Pretrained models for 18 languages (labeled as PUNKT in Table 5) are packaged with NLTK. NLTK additionally provides the framework to train a new model. We use this to train an additional model on all data (to simulate a multilingual model) and report the results in Table 5 as PUNKTML. PUNKT (and thus PUNKTML) does not segment around non-Latin punctuation. Moses Sentence Splitter uses a list of predefined acronyms and abbreviations for each language. If left token is in this list, it does not split. This circumvents the whole point behind the ambiguity "in the U.S." SpaCy Sentencizer is a “rule-based system" without specific details and varies from language to language. 7 Monolingual Experiments We first explore common questions and concerns while focusing on English data and results. We have three main parameters to study: context size, embedding size, and vocabulary size. We additionally consider how the training data affects results– both in relative noise in class labels in addition to 1 2 3 4 5 Right Context Size 1 2 3 4 5 6 7 8 Left Context Size 92.08 92.45 92.78 92.75 92.69 98.04 98.79 99.02 99.00 99.15 98.79 99.42 99.34 99.28 99.44 98.99 99.42 99.47 99.44 99.52 99.13 99.42 99.50 99.36 99.50 99.13 99.50 99.47 99.55 99.44 99.07 99.44 99.39 99.47 99.55 99.20 99.47 99.50 99.47 99.50 F1 on English DevTest 98.6 98.8 99.0 99.2 99.4 Figure 3: Heat map showing the change in F1 with respect to context size in a linear model. Embedding size and vocabulary size are kept constant at 32 and 125 respectively. training on shuffled sentences instead of documents. In general, we find our technique creates a monolingual English model (Table 3) that outperforms the baselines. F1 Precision Recall Always 86.9 76.9 100.0 Splitta 99.3 99.6 99.1 Punkt 98.6 98.8 98.4 Moses 98.8 99.7 98.0 SpaCy 88.0 86.3 89.7 Our Tool 99.8 99.8 99.8 Table 3: Scores on English WSJ 03-06 Test Data. The candidate set is determined the original English punctuation contexts as described in 3.2 7.1 Exploring context size Starting with a minimal model with an embedding size of 32, and a vocabulary size of 125, we investigate whether such a small model can solve this problem. Our method is rooted in a contextual encoding of the subword tokens inside its context windows, and may benefit from increasing the size of these windows. At the operating point with a very small embedding and vocabulary size, the window size is the determining factor on performance. The results on English in Figure 3 show that a minimal amount of left and right context is necessary; however, left context is more beneficial than right 4001 context. 7.2 How large of a model is necessary? We consider whether increasing the size of the model by doubling the embedding size and quadrupling the vocabulary size can produce better results. While varying the context windows (as seen in Figure 3) can result in increasingly higher scores, varying embedding size and vocabulary size did not produce the same effect. Keeping a fixed context window, we find that any given change in embedding size or vocabulary size increases F1 score by no more than 0.6%. While necessary to find the optimal model, it is clear that the context size is more important to experimentation. We note that a vocabulary size of 2000 tends to perform worse than smaller sizes while vocabulary sizes of 125 and 500 perform equally well when paired with any embedding size. Each of our monolingual models reported in Table 5 is the result of a grid search over various vocab sizes, and lambda weight (§3.3). We keep context sizes of left (6) and right (4) and embedding size (128) constant. 7.3 Is document context necessary? Because released monolingual data is often cleaned with sentences being removed and shuffled, it is unreasonable to assume that a set of consecutive sentences will always be available for training. In order to justify using this data, we repeat a subset of the previous English experiment—testing context and embedding sizes by training the model on the same data that has been shuffled. We test on the same validation data that has not been shuffled and retain its document order. In Table 4, we show that shuffling the training data has little impact on performance and document context is unnecessary in this punctuated setting. F1 Precision Recall Original 99.8 99.7 99.9 Shuffled 99.6 99.6 99.6 Undersegmented 97.5 95.2 99.8 Table 4: Scores on English WSJ 03-06 Test Data. Original is the best model trained on original English monolingual News Commentary data. Shuffled is trained on shuffled data, described in §7.3. Undersegmented is trained on raw Wikipedia, described in §7.4. 7.4 Can we train on undersegmented data? Uncleaned, unfiltered Wikipedia dumps do not have sentence boundaries in them. The smallest unit is the paragraph. Data scraped from internet sites is likely to have a similar form and much of our monolingual data is not guaranteed to be segmented. In order to justify that this approach works without already having segmented data, we show that we can achieve similar results as our previous English results in this setting. We train on one million randomly-selected paragraphs from an English Wikipedia dump. While many ⊗labels are now incorrect due to paragraphs being unsegmented, we assume the ✓class is relatively noise-free. Because we already established that shuffling the data does not affect performance in this setting, the random selection is sufficient. While maintaining previously chosen hyper-parameters—such as context sizes, learning rate, and dropout—we search among potential λ values to use as a weight for the ✓label. We find that increasing the λ value to 200.0 achieves the highest F1 of 97.5. An unweighted model performs poorly. While still distant from the cleanly-trained models, it performs significantly better than the poorer baselines. Comparison to our other English models can be seen in Table 4. 8 Multilingual Experiments After outperforming current baselines in a monolingual English setting, we generalized our approach to work multilingually. The multilingual model can segment text irrespective of input language. In parallel to the monolingual conditions, we train two-layer transformer models with 6 tokens of left context, and 4 tokens of right context with 128 embedding size. While we did experiment with scaling these for the multilingual model, we found little effect. We additionally scale the vocabulary size to 12,000 to accommodate the larger character sets in Chinese and Japanese. Because more of the additional languages have undersegmented data, we searched over potential lambda weights for the ✓class and report the best configuration (λ = 20.0) in Table 5. 8.1 Discussion Results of ERSATZ and baselines can be found in Table 5. In all cases, ERSATZ is at least competitive with baselines, if not outperforming them. Although most differences are small it outperforms 4002 # Orig. # Corr. % ✓ PUNKT PUNKTML ALWAYS SPACY MOSES ERSATZM ERSATZ ERSATZWM ar 1460 1504 84.9 92.7 93.5 90.3 98.2 98.0 98.0 cs 664 1726 80.1 99.8 99.6 96.3 85.3 99.7 99.8 99.8 99.8 de 785 1965 90.2 99.7 99.5 97.9 91.4 99.9 99.9 99.8 99.8 en 7706 7706 48.6 98.6 87.7 77.0 88.0 98.8 99.8 98.7 99.1 es 3000 3064 86.5 99.1 98.9 96.5 83.6 98.7 98.8 98.6 98.6 et 2000 2017 78.2 99.3 99.4 90.6 84.0 99.5 99.8 99.7 99.8 fi 1996 1996 95.0 99.7 99.7 98.9 97.9 99.8 99.9 99.9 99.9 fr 1619 1655 95.0 99.5 99.6 98.2 90.4 99.7 99.7 99.6 99.4 gu 1016 1018 92.3 100.0 97.9 3.8 99.7 99.8 100.0 100.0 hi 2507 2521 68.6 14.4 83.7 90.6 15.1 98.5 99.1 98.6 iu 2971 2971 59.1 91.3 63.9 86.1 93.7 93.6 ja 993 1072 89.4 0.2 98.1 93.7 99.9 99.9 99.9 kk 1000 1002 92.2 99.6 97.1 99.7 99.8 99.9 km 2320 2361 96.3 2.0 99.1 99.7 99.7 99.7 lt 1000 1000 59.2 94.7 85.5 76.6 98.6 98.8 98.8 98.9 lv 2001 2017 76.4 99.4 90.3 88.6 99.6 99.7 99.5 99.6 pl 1001 1005 70.7 98.3 94.8 90.1 78.9 92.8 93.4 99.1 99.2 ps 2719 2726 96.4 99.4 99.1 99.3 99.3 99.3 ro 1999 2000 89.1 98.7 97.0 90.9 98.5 99.3 99.3 99.2 ru 991 991 88.4 98.8 98.1 96.4 91.3 99.4 99.3 99.4 99.5 ta 997 1005 66.1 92.3 89.6 89.3 93.8 98.1 96.9 96.6 tr 3000 3009 67.5 95.8 85.2 85.1 99.5 99.6 99.5 99.5 zh 2000 2003 85.1 99.2 96.6 100.0 100.0 100.0 all 45k 48k 73.3 87.6 89.0 98.9 98.9 Table 5: Test set statistics (left block) and F1 scores (right block) on our test data. % ✓denotes the number of candidate sites with a true ✓label. PUNKTML denotes PUNKT model trained on our data. Lack of a score means the model was not available for that language. ERSATZM denotes monolingual models, ERSATZ the WMT-languages multilingual model, and ERSATZWM the model trained with additional WikiMatrix languages. SpaCy in all languages and often outperforms both Punkt and Moses. The Moses splitter is an interesting case. It identifies split points via a mix of general and languagespecific regular expressions, which are then filtered against a curated list of “non-breaking prefixes”. This results in a conservative segmenter that will not (for example) allow a sentence to end with the token U.S.. As such, its high performance is notable. However, the comparison is likely unfair, since it was likely built and refined against the news datasets that constitute our WMT test sets. This approach is therefore effective in this domain, but may not generalize. Our single multilingual model, trained on noisy data, performs nearly identically. 8.2 Performance across languages Sentence segmentation is not equally difficult in all languages or with respect to all punctuation. The ‘.’ is by far the most ambiguous form of punctuation and is frequently used as an abbreviation marker. Other scripts using their own punctuation, such as Hindi, have specified a particular marker (the Devanagari Danda) as a sentence-ending punctuation that is rarely used sentence-internally. In these cases, ambiguity is introduced when alternative punctuation (such as ‘.’ or ‘...’) is used. Additionally, even languages with the same scripts may not have the same level of ambiguity. French has the smallest number of punctuated contexts occurring sentence-internally within our test set, while English has the most. We note that the multilinguality of our model hurts the near-perfect performance that we see in the monolingual English models. We additionally note that some monolingual models perform worse than the multilingual model (see pl in Table 5). We hypothesize that this may be due to a lack of 4003 data, and the additional languages contain similar contexts, so the model may learn more about casing, punctuation, and length with additional data. 8.3 Scaling to more languages While we note that it is difficult to evaluate many of the world’s languages due to a lack of gold standard test data, we test for scalability by including additional languages (as described in §6) during training and noting any changes in performance on the evaluable languages. We include 64 additional languages (see Table 7 in the Appendix for comprehensive list) to bring us to a total of 87 languages. Table 5 also includes scores from a larger multilingual model (ERSATZWM) that was built with these 64 additional languages. Overall, we find very little change between these two settings. With en, we actually see some improvement in performance from the smaller multilingual model. Generally, there is not significant degradation of scores, implying this technique can generalize to additional languages. 8.4 How does size affect the speed? With our context construction method, we benefit from batching to decrease runtime, since the decision at each candidate point is dependent only on its immediate window. We benchmark our models as well as the baselines (Table 6). While our models are slower than some baselines, we find that increasing the size of the model does not dramatically increase the runtime. Additionally, the rate (in tokens per second) is roughly constant. Layer (# layers) # params Time (s) F1 Linear (x1) 1.7M 33 97.5 Linear (x2) 1.7M 35 98.0 Transformer (x1) 2.3M 74 98.7 Transformer (x2) 2.9M 172 98.7 Spacy 13.8 88.0 Moses 1164 98.8 Punkt 3.2 98.6 Table 6: Time in seconds for 1 million English tokens in input file. F1 is score on English Test Set. We show various size encoders for our method. Linear is a linear layer with TanH activation. 9 Summary As one of the earliest steps in NLP pipelines, sentence segmentation is an important task. However, it has not to this date received proper experimental attention, relying instead on ad hoc methods. It is a good time to correct this oversight, as NLP moves to the use of larger and larger corpora covering more and more languages. Even as the field moves towards processing text at the paragraph or document level directly, it is likely that sentence processing will be with us for some time. We show here that a simple context-based model can produce state-of-the-art results with a modest hyperparameter search, trained on noisy annotations from imperfectly-segmented data. Together with a straightforward multilingual approach to identifying candidate split points and training on noisy segmented data, our single model performs well across a range of languages. More fundamentally, we have defined an experimental framework for benchmarking and future comparative work. Missing from our paper is an evaluation of the effect of these tools on downstream tasks. An obvious candidate for future work is to conduct this evaluation. It is possible that some tasks will not be affected by small differences among the best performing models, but this work at least sheds light on those differences. Another obvious direction is to look at approaches that would work for unpunctuated text (e.g., Wang et al. (2019)). This would expand the functionality of segmenters into other important areas, such as speech translation, and to languages, like Thai, that do not mark ends of sentences. Acknowledgments The authors wish to thank Elizabeth Salesky, Carlos Aguirre, Jacob Bremerman and the anonymous reviewers for helpful technical discussions and feedback. References Loïc Barrault, Magdalena Biesialska, Ondˇrej Bojar, Marta R. Costa-jussà, Christian Federmann, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Matthias Huck, Eric Joanis, Tom Kocmi, Philipp Koehn, Chi-kiu Lo, Nikola Ljubeši´c, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Santanu Pal, Matt Post, and Marcos Zampieri. 2020. Findings of the 2020 conference on machine translation (WMT20). In Proceedings of 4004 the Fifth Conference on Machine Translation, pages 1–55, Online. Association for Computational Linguistics. Steven Bird and Edward Loper. 2004. NLTK: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, pages 214–217, Barcelona, Spain. Association for Computational Linguistics. Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuitho Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the iwslt 2017 evaluation campaign. In 14th International Workshop on Spoken Language Translation, pages 2–14, Tokyo, Japan. Dan Gillick. 2009. Sentence boundary detection and the problem with the U.S. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers, pages 241–244, Boulder, Colorado. Association for Computational Linguistics. Eric Joanis, Rebecca Knowles, Roland Kuhn, Samuel Larkin, Patrick Littell, Chi-kiu Lo, Darlene Stewart, and Jeffrey Micher. 2020. The Nunavut hansard Inuktitut–English parallel corpus 3.0 with preliminary machine translation results. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 2562–2572, Marseille, France. European Language Resources Association. Tibor Kiss and Jan Strunk. 2006. Unsupervised multilingual sentence boundary detection. Computational Linguistics, 32(4):485–525. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague, Czech Republic. Association for Computational Linguistics. Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, Brussels, Belgium. Association for Computational Linguistics. Khanh Nguyen and Hal Daumé III. 2019. Global Voices: Crossing borders in automatic news summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 90–97, Hong Kong, China. Association for Computational Linguistics. David D. Palmer and Marti A. Hearst. 1997. Adaptive multilingual sentence boundary disambiguation. Computational Linguistics, 23(2):241–267. Ines Rehbein, Josef Ruppenhofer, and Thomas Schmidt. 2020. Improving sentence boundary detection for spoken language transcripts. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 7102–7111, Marseille, France. European Language Resources Association. Michael D. Riley. 1989. Some applications of treebased modelling to speech and language. In Speech and Natural Language: Proceedings of a Workshop Held at Cape Cod, Massachusetts, October 15-18, 1989. George Sanchez. 2019. Sentence boundary detection in legal text. In Proceedings of the Natural Legal Language Processing Workshop 2019, pages 31–38, Minneapolis, Minnesota. Association for Computational Linguistics. Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2019a. Wikimatrix: Mining 135m parallel sentences in 1620 language pairs from wikipedia. CoRR, abs/1907.05791. Holger Schwenk, Guillaume Wenzek, Sergey Edunov, Edouard Grave, and Armand Joulin. 2019b. Ccmatrix: Mining billions of high-quality parallel sentences on the WEB. CoRR, abs/1911.04944. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc. Xiaolin Wang, Masao Utiyama, and Eiichiro Sumita. 2019. Online sentence segmentation for simultaneous interpretation using multi-shifted recurrent neural network. In Proceedings of Machine Translation Summit XVII Volume 1: Research Track, pages 1–11, Dublin, Ireland. European Association for Machine Translation. Nianwen Xue and Yaqin Yang. 2011. Chinese sentence segmentation as comma classification. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 631–635, Portland, Oregon, USA. Association for Computational Linguistics. Nina Zhou, AiTi Aw, Nattadaporn Lertcheva, and Xuancong Wang. 2016. A word labeling approach to Thai sentence boundary detection and POS tagging. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 319–327, Osaka, Japan. The COLING 2016 Organizing Committee. A Appendix 4005 Dataset # Lines # Tokens Dataset # Lines # Tokens ar News Comm. v15 181k 10M kk News Comm. v15 16.4k 280k WikiMatrix 774k 16M News Crawl 1.1M 14M cs News Comm. v15 277k 5.2M km JW Corpus 107k 4.6M WikiMatrix 429k 7.3M Common Crawl 343k 2.0M de News Comm. v15 422k 8.9M lt News Crawl 2.5M 37M WikiMatrix 1M 19M WikiMatrix 84.8k 1.1M en News Comm. v15 609k 13M lv News Crawl 1.8M 29M WSJ (sec 00-02;07-23) 40k 819k es News Comm. v15 465k 12M pl Global Voices 58k 890k Wikipedia 405k 6.6M News Crawl 3.0M 44M et News Crawl 1.6M 22M ps News Crawl 64.0k 1.8M WikiMatrix 152k 5.1M SADA 132k 4.1M SYSTRAN 196k 5.1M TRANSTAC 75k 1.2M fi News Crawl 4.7M 50M ro Global Voices 4043 76k WikiMatrix 207k 2.6M WikiMatrix 223k 4.7M News Crawl 6.9M 140M fr News Comm. v15 415k 10M ru News Comm. v15 377k 7.3M WikiMatrix 2.2M 50M WikiMatrix 2.2M 37M gu News Crawl 283k 3.8M ta News Crawl 501k 5.3M Common Crawl 164k 1.3M WikiMatrix 61.0k 532k hi News Comm. v15 7815 213k tr Global Voices 6529 80k WikiMatrix 1.1M 20M WikiMatrix 304k 4.5M News Crawl 135k 3.0M News Crawl 7.9M 108M iu N.H.I 3.0 1.3M 8.0M zh News Comm. v15 445k 772k WikiMatrix 492k 890k ja News Comm. 2983 4390 News Crawl 3.4M 6.9M Table 7: Multilingual Datasets Line Count and Token Count. 4006 Dataset # Lines # Tokens Dataset # Lines # Tokens ar News Comm v15 1637 38k kk News Comm v15 3000 38k IWSLT 2017 1504 20k WMT19 Test 1002 30k cs WMT18 Test 3008 47k km WMT WikiDev 2609 14k WMT20 Test 1726 26k WMT20 Test 2361 15k de WMT19 Test 2009 31k lt News Comm v15 3000 44k WMT20 Test 1965 31k WMT19 Test 1000 17k en News Commentary 3000 56k lv News Commentary 3000 49k WMT20 Test (en-cs) WMT17 Test 2017 33k WSJ 03-06; 24 10k 277k es WMT11 Test 3013 69k pl News Commentary v15 3000 16k WMT13 Test Set 3064 62k WMT20 Test Set 1005 16k et News Commentary 3000 41k ps WMT Wiki Dev 3166 64k WMT18 Test Set 2017 30k WMT20 Test Set 2726 55k fi WMT18 Test Set 3031 38k ro News Commentary v15 3000 60k WMT19 Test Set 1996 21k WMT16 Test Set 2000 43k fr WMT15 Test Set 1502 25k ru WMT18 Test Set 3000 52k WMT20 Test Set (fr-de) 1655 33k WMT20 Test Set 991 15k gu News Commentary 3000 40k ta News Commentary 3000 32k WMT19 Test Set 1018 14k WMT20 Test Set 1005 13k hi News Commentary 3000 56k tr WMT16 Test Set 3011 44k WMT14 Test Set 2521 57k WMT18 Test Set 3009 46k iu N.H.I 3.0 Dev 3028 27k zh WMT18 Test Set 4097 5.8k N.H.I 3.0 Dev 3028 27k WMT20 Test Set 2003 3.7k ja News Commentary 3000 5.9k WMT20 Test Set 1072 1888 Table 8: Dev and Test Data. Test Data is bolded. All News and Wikipedia sets come from WMT news translation tasks except ar. All test sets are lang-en unless otherwise noted. NHI is the Nunavut Hansard Inuktitut English Parallel Corpus-3.0. When News Commentary was used, the bottom N lines were taken. 4007 # lines # tokens # lines # tokens an 52k 93k la 45k 50k arz 35k 57k lb 43k 59k as 16k 15k lmo 10k 16k az 164k 170k mg 12k 18k bar 40k 49k mk 452k 672k ba 101k 112k ml 150k 130k be 164k 223k mr 216k 225k bg 454k 523k mwl 32k 78k bn 360k 452k nds-nl 14k 22k br 43k 53k nds 95k 145k bs 502k 831k ne 70k 81k ca 459k 417k nl 456k 348k ceb 80k 188k no 457k 472k da 453k 494k oc 171k 389k el 454k 555k pt 460k 367k eo 454k 554k sh 454k 582k eu 305k 369k simple 465k 666k fa 427k 818k si 182k 281k fo 38k 46k sk 453k 539k fy 56k 84k sl 451k 624k gl 453k 512k sq 262k 523k gom 22k 19k sr 452k 520k he 458k 387k sv 452k 388k hr 455k 551k sw 70k 118k hu 456k 353k te 213k 170k hy 23k 52k tg 17k 20k id 456k 468k tl 122k 237k is 124k 160k tt 78k 80k it 469k 386k uk 466k 313k jv 27k 40k vi 456k 646k ka 42k 81k wuu 46k 7k ko 454k 395k Table 9: Additional WikiMatrix languages with line and token counts for training data. Language code based on Wikipedia codes.
|
2021
|
309
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 366–376 August 1–6, 2021. ©2021 Association for Computational Linguistics 366 Deep Differential Amplifier for Extractive Summarization Ruipeng Jia1,2, Yanan Cao1,2, Fang Fang1,2∗, Yuchen Zhou1, Zheng Fang1, Yanbing Liu1,2 and Shi Wang3∗ 1Institute of Information Engineering, Chinese Academy of Sciences 2School of Cyber Security, University of Chinese Academy of Sciences 3Institute of Computing Technology, Chinese Academy of Sciences 1,2{jiaruipeng, caoyanan, fangfang0703, zhouyuchen, fangzheng, liuyanbing}@iie.ac.cn [email protected] Abstract For sentence-level extractive summarization, there is a disproportionate ratio of selected and unselected sentences, leading to flatting the summary features when optimizing the classification. The imbalanced sentence classification in extractive summarization is inherent, which can’t be addressed by data sampling or data augmentation algorithms easily. In order to address this problem, we innovatively consider the single-document extractive summarization as a rebalance problem and present a deep differential amplifier framework to enhance the features of summary sentences. Specifically, we calculate and amplify the semantic difference between each sentence and other sentences, and apply the residual unit to deepen the differential amplifier architecture. Furthermore, the corresponding objective loss of the minority class is boosted by a weighted cross-entropy. In this way, our model pays more attention to the pivotal information of one sentence, that is different from previous approaches which model all informative context in the source document. Experimental results on two benchmark datasets show that our summarizer performs competitively against state-of-the-art methods. Our source code will be available on Github. 1 Introduction Single-document extractive summarization forms summary by copying and concatenating the most important spans (usually sentences) in a document. Sentence-level summarization is a very challenging task, because it arguably requires an in-depth understanding of the source document sentences, and current automatic solutions are still far from human performance. Recent approaches frame the task as a sequence labeling problem, taking advantage of the success of neural network architectures. ∗Corresponding authors: Fang Fang and Shi Wang 10 20 30 40 50 Number of Sentences 20 25 30 35 40 ROUGE-1 ROUGE-2 ROUGE-L Figure 1: ROUGE score for documents with different length. The result is calculated on the test set of CNN/DM and the trained model is based on BERT. However, there are still two inherent obstacles for sentence-level extractive summarization: 1) It should be detrimental to keep tangential information (West et al., 2019). The intuitive limitation of those approaches is that they always prefer to model and retain all informative content from the source document. This goes against the fundamental goal of summarization, which crucially needs to forget all but the “pivotal” information. Recently, the Information Bottleneck principle (Tishby et al., 2000; West et al., 2019) is introduced to incorporate a tradeoff between information selection and pruning. Length penalty and the topic loss (Baziotis et al., 2019) are used in the autoencoding system to augment the reconstruction loss. However, these methods require external variables or augmentative terms, without enhancing the representation of pivotal information. 2) Imbalanced classes inherently result in models that have poor predictive performance, specifically for the minority class. The distribution of examples across the known classes can vary from a slight bias to a severe imbalance, where there is one example in the minority class for dozens of examples in the majority class. For instance, according to the statistics on the popular summarization dataset, only 7.33% sentences of 367 CNN/DM (Hermann et al., 2015) are labeled as “1” and others are “0”, indicating whether this sentence should be selected as summary or not. Conversely, most machine learning algorithms for classification predictive models are designed and demonstrated on problems that assume an equal distribution of classes. This means that a naive application of a model may only focus on learning the characteristics of the abundant observations, neglecting the examples from the minority class. Furthermore, as shown in Figure 1, the ROUGE score gradually declines along with the number of sentences accumulating, since the valuable summary sentences is generally a tiny minority (with the quantity of 1-4), while more and more majority sentences will swamp the minority ones. Unfortunately, the imbalance in summarization is inherent, which can’t be addressed by common data augmentation (He and Ma, 2013; Asai and Hajishirzi, 2020; Min et al., 2020; Zoph et al., 2019; Xie et al., 2020), for there is a rare influence on the 0/1 distribution by adding or deleting the entire document. These two obstacles are interrelated and interact with each other. Highlighting the pivotal information will strengthen the unique semantic and weaken the common informative content. Additionally, a more balanced distribution would make minority class more attractive. If we can’t resolve the category imbalance problem in extractive summarization by data augmentation, how to make the minority class more attractive? Inspired by the differential amplifier of analog electronics1, we propose a heuristic model, DifferSum, as shorthand for Differential Amplifier for Extractive Summarization to enhance the representation of the summary sentences. Specifically, we calculate and amplify the semantic difference between each sentence and other sentences, by the subtraction operation. The original differential amplifier consists of two terms and the second term is used to avoid making the final output zero. In our model, we use the residual unit instead of the second term to make the architecture deeper. We further design a more appropriate objective function to avoid biasing the data, by making the loss of a minority much greater than the majority. DifferSum shows superiority over other extractive methods in two aspects: 1) enhancing the representation of the pivotal information and 2) compensating the minority class and penalizing the majority ones. 1https://en.wikipedia.org/wiki/Differential amplifier Experimental results validate the effectiveness of DifferSum. The human evaluation also shows that our model is better in relevance compared with others. Our contributions in this work are concluded as follows: • We propose a novel conceptualization of extractive summarization as rebalance problem. • We introduce a heuristic approach, calculating and amplifying the semantic representation of pivotal information by integrating both the differential amplifier and residual learning. • Our proposed framework has achieved superior performance compared with strong baselines. 2 Related Work 2.1 Extractive Summarization Recent research work on extractive summarization spans a large range of approaches. These works usually instantiate their encoder-decoder architecture by choosing RNN (Nallapati et al., 2017; Zhou et al., 2018), Transformer (Wang et al., 2019; Zhong et al., 2019b; Liu and Lapata, 2019; Zhang et al., 2019b) or GNN (Wang et al., 2020; Jia et al., 2020b) as encoder, autoregressive (Jadhav and Rajan, 2018; Liu and Lapata, 2019) or RL-based (Narayan et al., 2018; Arumae and Liu, 2018; Luo et al., 2019) decoders. For two-stage summarization, Chen and Bansal (2018) and Bae et al. (2019) follow a hybrid extract-then-rewrite architecture, with policy-based RL to bridge the two networks together. Lebanoff et al. (2019), Xu and Durrett (2019) and Mendes et al. (2019) focus on the extract-then-compress learning paradigm, which will first train an extractor for content selection. Zhong et al. (2020) introduces extract-thenmatch framework, which employs BERTSUMEXT (Liu and Lapata, 2019) as first-stage to prune unnecessary information. However, these above extractive approaches prefer to model all source informative context and they pay little attention to the imbalance problem. 2.2 Deep Residual Learning The original deep residual learning is introduced in image recognition (He et al., 2016a) for the notorious degradation problem. Then, residual is introduced to the natural language process by Transformer (Vaswani et al., 2017). Essentially, we cannot determine the depth of the network very well 368 ݒ1 ݒ2 ݒ3 ݒ݅ ݒܯ ݒܯ ݒ݅ ݒ3 ݒ2 ݒ1 ݏ1 ݏ2 ݏ3 ݏܯ ݏ݅ ݓ11 ݓ12 ݓ13 ݓܯ1 ݓ݅1 ݑ11 ݑ12 ݑ13 ݑܯ1 ݑ݅1 Weighted Pooling ݒ1 ݒ2 ݒ3 ݒܯ ݒ݅ ݒ1 ݒ2 ݒ3 ݒܯ ݒ݅ FF & Sigmoid ݎ1 ݎ2 ݎ3 ݎ݅ ݎܯ Figure 2: Overview of DifferSum. when building a deep network. There will be optimal layers in the network, and outside the optimal layer is the redundant layer. We expect the redundant layer to correspond to the input and output, namely identity mapping (He et al., 2016a,b; Veit et al., 2016; Balduzzi et al., 2018). Resnet (He et al., 2016a) addresses the degradation problem by introducing a deep residual learning framework. If an identity mapping were optimal, it would be easier to push the residual to zero than to fit an identity mapping by a stack of nonlinear layers (Huang and Wang, 2017). In this paper, the residual unit serves as the second item of the differential amplifier to keep our architecture deep enough and capture pivotal information. 3 Methodology 3.1 Problem Definition We model the sentence extraction task as a sequence tagging problem (Kedzie et al., 2018). Given a document D consisting of a sequence of M sentences [s1, s2, ..., sM] and a sentence si consisting of a sequence of N words [wi1, wi2, ..., wiN]. We denote by hi and hij the embedding of sentences and words in a continuous space. The extractive summarizer aims to produce a summary S by selecting m sentences from D (where m ≤M). For each sentence si ∈D, there is ground-truth yi ∈{0, 1} and we will predict a label ˆyi ∈{0, 1}, where 1 means that si should be included in the summary. We assign a score p(ˆyi|si, D, θ) to quantify si’s relevance to the summary, where θ is the parameters of neural network model. Finally, we assemble a summary S by selecting m sentences, according to the probability of p(1|si, D, θ). 3.2 Sentence Encoder The sentence encoder in extractive summarization models is usually a recurrent neural network with Long-Short Term Memory (Hochreiter and Schmidhuber, 1997) or Gated Recurrent Units (Cho et al., 2014). In this paper, our sentence encoder builds on the BERT architecture (Devlin et al., 2019), a recently proposed highly efficient model which is based on the deep bidirectional Transformer (Vaswani et al., 2017) and has achieved state-ofthe-art performance in many NLP tasks. The Transformer aims at reducing the fundamental constraint of sequential computation which underlies most architecture (Liu et al., 2019). It eliminates recurrence in favor of applying a self-attention mechanism which directly models relationships between all words in a sentence. Our extractive model is composed of a sentencelevel Transformer (TS) and a document-level Transformer (TD) (Liu et al., 2019). For each sentence si in the input document, TS is applied to obtain a contextual representation for each word: [u11, u12, ..., uMN] = TS([w11, w12, ..., wMN]) (1) And the representation of a sentence is acquired by applying weighted-pooling: aij = W0uT ij si = 1 N N X j=1 aijuij (2) Document-level transformer TD takes si as input and yields a contextual representation for each sentence: [v1, v2, ..., vM] = TD([s1, s2, ..., sM]) (3) 3.3 Deep Differential Amplifier In the Transformer model sketched above, intersentence relations are modeled by multi-head attention based on softmax functions, which only capture shallow structural information (Liu et al., 2019). A differential amplifier is a type of electronic amplifier that amplifies the difference between two input voltages but suppresses any voltage common 369 to the two inputs. The output of an ideal differential amplifier is given by: Vout = Ad(V + in −V − in) (4) where V + in and V − in are the input voltage; Ad is the differential-mode gain. In practice, the gain should not be quite equal for the two inputs, V + in and V − in. For instance, even if V + in and V − in are equal, the output Vout should not be zero. So, modern differential amplifiers are usually implemented with a more realistic expression, which includes a second term: Vout = Ad(V + in −V − in) + Ac V + in + V − in 2 (5) where Ac is called the common-mode gain of the amplifier. Inspired by the differential amplifier above, we calculate and amplify the semantic difference between each sentence and other sentences by the subtraction operation of the sentence representations [v1, v2, ..., vM]. Particularly, for sentence si, V + in and V − in are calculated as follows: V + in = vi V − in = P j∈{1,2,...,M}\{i} vj M −1 (6) The original differential amplifier consists of two terms and the second one avoids making the final output zero. While for the deep neural network: 1) inputs of the differential amplifier are vector instances in the high dimensional space, which is practically impossible for the zero output, compared with scalar; 2) the second term of the differential amplifier is not suitable for the deep iterative architecture, since it is exposed to the degradation problem. Notably, residual learning is introduced in deep learning as shortcut connections to skip one or more layers, which is naturally an alternative to the second item of the differential amplifier. The advantages of this method are: 1) the residual architecture will highlight the pivotal information as well as reserving the original sentence representation; 2) it is easier to optimize the residual mapping than to optimize the original (He et al., 2016a). Hence, the residual unit is employed as the second item, along with an iterative refinement algorithm to enhance the final representation of sentences. 3.4 Residual Representation for Sentence The differential amplifier in our architecture consists of a few stacked layers to iteratively refine the pivotal representation. Let us consider H(x) as an underlying mapping to be fit, with x denoting the inputs to the first of these layers. Since multiple nonlinear layers can asymptotically approximate complicated functions (He et al., 2016a; Mont´ufar et al., 2014), the differential amplifier mapping H(x) is recast into a residual mapping F(x) and an identity mapping x: H(x) = F(x) + x (7) Obviously, residual learning is just a variant of the differential amplifier: H(x) := Vout F(x) := Ad(V + in −V − in) (8) where the output voltage Vout thus becomes the original mapping H(x) and the first item of amplifier Ad(V + in −V − in) equals to residual mapping F(x), In our model, the second item of the differential amplifier is replaced by the identity mapping x, which is the shortcut connection and the output is added to the outputs of F(x). Furthermore, 1) the identity shortcut connections advance the architecture without extra parameter; 2) the identity shortcut doesn’t add the computational complexity (He et al., 2016a); Thus, for sentence respresentation vi, the deep differential amplifier is: H(vi) = Ad(vi − P j∈{1,2,...,M}\{i} vj M −1 )+vi (9) 3.5 Iterative Structure Refinement The differential amplifier and residual unit specialize in modeling the pivotal information, while deeper neural networks with more parameters are able to infer semantic more accurately. So, an iterative refinement algorithm is introduced to enhance the final representation of pivotal information. For sentence vi, the fundamental iterative unit is: H(vi) = F(vi) + vi vi = H(vi) (10) 370 where we iteratively refine the representation vi for K times; and thanks to the built-in residual mechanism, most shorter paths are needed during training, as longer paths do not contribute any gradient. Along with the supervision, each iteration will pay more attention to the key semantic difference F(vi) of sentences with label 1, while trying to zero other F(vj). Conversely, previous extractive approaches without differential amplifier can only classify those sentences by compensating or penalizing vi / vj, which is more difficult to model. Following previous work (Nallapati et al., 2017; Liu et al., 2019), we use a sigmoid function after a linear transformation to calculate the probability ri of selecting si as a summary sentence: ri = sigmoid(W1vT i ) (11) 3.6 Weighted Objective Function To rebalance the bias of minority 1-class and majority 0-class, we have built a deep differential amplifier to amplify and capture the unique information for summary sentences. Besides, another heuristic method is to make our model pay more attention to 1-class: a weighted cross-entropy function. Particularly, we further design a more appropriate objective function to avoid biasing the data, by making the loss of a minority much greater than the majority. The weight we employed is to rebalance the observations for each class, so the sum of observations for each class are equal. Finally, we define the model’s loss function as the summation of the losses of all iterations: L = K X k=1 ( 1 M M X i=1 "P sj∈D I(sj /∈S) P sj∈D I(sj ∈S)y log(rk i ) +(1 −y) log(1 −rk i ) #) (12) where I(·) is an indicator function and K is the number of iterations. 4 Experiments 4.1 Datasets As shown in Table 1, we employ two datasets widely-used with multiple sentences summary: CNN and Dailymail (CNN/DM) (Hermann et al., 2015) and New York Times (NYT) (Sandhaus, 2008). Table 1: Data Statistics: CNN/Daily Mail and NYT. Datasets avg.doc length avg.summary length words sentences words sentences CNN 760.50 33.98 45.70 3.59 DailyMail 653.33 29.33 54.65 3.86 NYT 800.04 35.55 45.54 2.44 CNN/DM We used the standard split (Hermann et al., 2015) for training, validation, and test (90,266/1,220/1,093 for CNN and 196,96/12.148/10,397 for Daily Mail), with splitting sentences by Stanford CoreNLP (Manning et al., 2014) toolkit and pre-processing the dataset following (See et al., 2017) and (Zhong et al., 2020). This dataset contains news articles and several associated abstractive highlights. We use the unanonymized version as in previous summarization work and each document is truncated to 800 BPE tokens. NYT Following previous work (Zhang et al., 2019b; Xu and Durrett, 2019), we use 137,778, 17,222 and 17,223 samples for training, validation, and test, respectively. We also followed their filtering procedure, documents with summaries less than 50 words were removed from the dataset. Sentences were split with the Stanford CoreNLP toolkit (Manning et al., 2014). Input documents were truncated to 800 BPE tokens too. 4.2 Parameters Our code is based on Pytorch (Paszke et al., 2019) and the pre-trained model employed in DifferSum is ‘albert-xxlarge-v2’, which is based on the huggingface/transformers2. We train DifferSum two days for 100,000 steps on 2GPUs(Nvidia Tesla V100, 32GB) with gradient accumulation every two steps. Adam with β1 = 0.9, β2 = 0.999 is used as optimizer. Learning rate schedule follows the strategy with warming-up on first 10,000 steps. We have tried the iteration steps of 2/4/6/8 for iterative refinement, and K = 4 is the best choice based on the validation set. We select the top-3 checkpoints based on the evaluation loss on the validation set, and report the averaged results on the test set. Following Jia et al. (2020a) and Jia et al. (2021), we employ the greedy algorithm for the sentencelevel soft labels, which falls under the umbrella 2https://github.com/huggingface/transformers 371 Table 2: ROUGE F1 on CNN/DM. Models CNN/DM R-1 R-2 R-L Abstractive ABS (2015) 35.46 13.30 32.65 PGC (2017) 39.53 17.28 36.38 TransformerABS (2017) 40.21 17.76 37.09 T5Large (2020) 43.52 21.55 40.69 BARTLarge (2019a) 44.16 21.28 40.90 PEGASUSLarge (2019a) 44.17 21.47 41.11 ProphetNetLarge (2020) 44.20 21.17 41.30 Extractive Lead-3 40.42 17.62 36.67 Oracle (Sentence) 55.61 32.84 51.88 SummaRuNNer (2017) 39.60 16.20 35.30 Exconsumm (2019) 41.70 18.60 37.80 PNBERTBase (2019a) 42.69 19.60 38.85 HIBERTLarge (2019b) 42.37 19.95 38.83 BERT-ext+RLBase (2019) 42.76 19.87 39.11 BERTSUMEXTBase (2019) 43.25 20.24 39.63 BERTSUMEXTLarge (2019) 43.85 20.34 39.90 DiscoBERTBase (2020) 43.77 20.85 40.67 HSGBase (2020) 42.95 19.76 39.23 ETCSumBase (2020) 43.84 20.80 39.77 ARedSumBase (2020) 43.43 20.44 39.83 MATCHSUMBase (2020) 44.41 20.86 40.55 DifferSumLarge 44.70 21.36 40.83 of subset selection. Besides, we employ the Trigram Blocking strategy for decoding, which is a simple but powerful version of Maximal Marginal Relevance (Carbonell and Goldstein, 1998). Specifically, when predicting summaries for a new document, we first use the model to obtain the probability score p(1|si, D, θ) for each sentence, and then we rank sentences by their scores and discard those which have trigram overlappings with their predecessors. 4.3 Metric ROUGE (Lin, 2004) is the standard metric for evaluating the quality of summaries. We report the ROUGE-1, ROUGE-2, and ROUGE-L of DifferSum by ROUGE-1.5.5.pl, which calculates the overlap lexical units of extracted sentences and ground-truth. 5 Results and Analysis 5.1 Results on CNN/DM Table 2 shows the results on CNN/DailyMail. All of these scores are in accordance with original papers. Following Nallapati et al. (2017); Liu and Lapata (2019), we compare extractive summarizaTable 3: ROUGE F1 on NYT. Models NYT R-1 R-2 R-L Abstractive ABS (2015) 42.78 25.61 35.26 PGC (2017) 43.93 26.85 38.67 TransformerABS (2017) 45.36 27.34 39.53 BARTLarge (2019a) 48.73 29.25 44.48 Extractive Lead-3 41.80 22.60 35.00 Oracle (Sentence) 64.22 44.57 57.27 SummaRuNNer (2017) 42.37 23.89 38.74 Exconsumm (2019) 43.18 24.43 38.92 JECS (2019) 45.50 25.30 38.20 BERTSUMEXTBase (2019) 46.66 26.35 42.62 HIBERTLarge (2019b) 49.47 30.11 41.63 DifferSumLarge 49.52 29.78 43.86 tion models against abstractive models, and it is certainly that the abstractive paradigm is still on the frontier of summarization. The first part of extractive approaches is the Lead-3 baseline and Oracle upper bound, while the second part includes other extractive summarization models. We present our models finally at the bottom. It is obvious that our DifferSum outperforms all extractive baseline models. Compared with large version BERTSUMEXT, our DifferSum achieves 0.85/1.02/0.93 improvements on R-1, R-2, and R-L, which indicates the pivotal information captured by the differential amplifier is more powerful than the other structures. Compared with early approaches, especially for BERTSUMEXT, we observe that BERT outperforms all previous non-BERT-based summarization systems, and Trigram-Blocking leads to a great improvement on all ROUGE metrics. MATCHSUM is a comparable competitor to our DifferSum, which formulates the extractive summarization task as a two-step problem and extract-thenmatch summary based on a well-trained BERTSUMEXT. Therefore, we only train a large version DifferSum for a fair comparison. 5.2 Results on NYT Results on NYT are summarized in Table 3. Note that we use limited-length ROUGE recall as Durrett et al. (2016), where the selected sentences are truncated to the length of the human-written summaries. The parts of Table 3 is similar to Table 2. The first four lines are abstractive models, and the next two lines are our golden baselines for extrac372 Table 4: Ablation Study on CNN/DM. Models R-1 R-2 R-L DifferSum 44.70 21.36 40.83 DifferSum w/o ALBERT 44.41 20.80 40.57 DifferSum w/o Amplifier 44.17 20.74 40.42 DifferSum w/o Iteration 44.32 21.02 40.48 tive summarization. The third part reports the performance of other extractive works and our model respectively. Again, we observe that our differential amplifier modeling performs better than both LSTM and BERT. Meanwhile, we find that extractive approaches show superiority over abstractive models, and the ROUGE scores are higher than CNN/DailyMail. 5.3 Ablation Studies We propose several strategies to improve the performance of extractive summarization, including differential amplifier (vs. normal residual network), pre-trained ALBERT(vs. BERT), and iterative refinement (vs. None). To investigate the influence of these factors, we conduct experiments and list the results in Table 4. Significantly, 1) differential amplifier is more critical than ALBERT, for the reason that the pivotal information is essential and difficult for ALBERT to model; 2) iterative refinement mechanism enlarges the advantage of the differential amplifier, demonstrating the superiority of deep architecture. 5.4 Human Evaluation for Summarization It is not enough to only rely on the ROUGE evaluation for a summarization system, although the ROUGE correlates well with human judgments (Owczarzak et al., 2012). Therefore, we design an experiment based on a ranking method to evaluate the performance of DifferSum by humans. Following Cheng and Lapata (2016), Narayan et al. (2018) and Zhang et al. (2019b), firstly, we randomly select 40 samples from CNN/DM test set. Then the human participants are presented with one original document and a list of corresponding summaries produced by different model systems. Participants are requested to rank these summaries (ties allowed) by taking informativeness (Can the summary capture the important information from the document) and fluency (Is the summary grammatical) into account. Each document is annotated by three different participants separately. The input article and ground truth summaries are Table 5: Human Evaluation on CNN/DM. Models 1st 2nd 3rd 4th MeanR SummaRuNNer 0.20 0.27 0.30 0.23 2.56 BERTSUMEXT 0.25 0.30 0.28 0.17 2.37 DifferSum 0.48 0.27 0.20 0.05 1.82 Ground-Truth 0.68 0.22 0.07 0.03 1.45 also shown to the human participants in addition to the three model summaries (SummaRuNNer, BERTSUMEXT, and DifferSum). From the results shown in Table 5, it is obvious that DifferSum is better in relevance compared with others. 5.5 Trigram Blocking Strategy Trigram Blocking leads to a great improvement on all ROUGE metrics for many extractive approaches (Liu and Lapata, 2019; Wang et al., 2020). It is has become a fundamental module in extractive summarization. In this paper, DifferSum extracts summary sentences with the Trigram-Blocking algorithm, but whether there is a great improvement along with it, like in SummaRuNNer or BERTSUMEXT? It has been explained by Nallapati et al. (2017); Liu and Lapata (2019), that picking all sentences by comparing the predicted probability with a threshold may not be an optimal strategy since the training data is very imbalanced in terms of summarymembership of sentences. Therefore, the TrigramBlocking algorithm is introduced to select top-k sentences and reduce the redundancy. Coincidentally, our DifferSum is designed to 1) rebalance the distribution of majority and minority and 2) filter the tangential and redundant information. Thus, the Trigram-Blocking algorithm may be useless for our DifferSum. Table 6 further summarizes the performance gain of Trigram-Blocking strategy. It is obvious that this strategy is essential for BERTSUMEXT or SummaRuNNer, achieving more than 2.68 / 0.98 improvements on R-1 separately, for that there is no enough redundancy modeling for both of them. While on the other hand, the efficiency of the Trigram-Blocking strategy is weak for DifferSum. 5.6 Documents with a Different Number of Sentences In this paper, we emphasize the inherent imbalance problem of the majority 0-class and the minority 1-class. In fact, in CNN/DailyMail dataset, there are plenty of documents with a different num373 Table 6: ROUGE Scores about Trigram-Blocking on CNN/DM Test Set. Models R-1 R-L DifferSum (with Trigram-Blocking) 44.70 40.83 DifferSum 44.36 40.43 BERTSUMEXT (with Trigram-Blocking) 43.85 39.90 BERTSUMEXT 41.17 36.52 SummaRuNNer (with Trigram-Blocking) 40.58 36.61 SummaRuNNer 39.60 35.30 10 20 30 40 50 Number of Sentences 20 25 30 35 40 ROUGE-1 ROUGE-2 ROUGE-L (a) BERTSUMEXT 10 20 30 40 50 Number of Sentences 20 25 30 35 40 ROUGE-1 ROUGE-2 ROUGE-L (b) DifferSum Figure 3: Comparison Between the ROUGE Scores Tendencies of BERTSUMEXT and DifferSum ber of sentences, ranging from 3-sentences to 100sentences. While the number of summary sentences, labeled with 1, is from 1-sentences to 5sentences, and the average number of sentences labeled 1 in CNN/DailyMail is only 7.33%. What is worse is that the distribution of the number of sentences for documents is a uniform distribution, thus we could not avoid the imbalance by cleaning the data. In this paper, we design another experiment to analysis the harmful effect of imbalance classes. We train the BERTSUMEXT (12-layers) from scratch on CNN/DailyMail, and evaluate the model on the test set to check the tendency of ROUGE scores, along with the number of sentences accumulating. The result is shown in the line chart of Figure 1 and Figure 3a, and obviously we only pay attention to the document in which the number of sentences less than 55. Specifically, each document is truncated to 2000 BPE tokens to involve more sentences, but this can not cover those whole documents with more than 55-sentences. Therefore, we choose to calculate the ROUGE scores for documents with sentences from 3 to 55. For comparison, we train our DifferSum (12layers) from scratch, and each document is truncated to 2000 BPE tokens too. The tendency of our DifferSum is as Figure 3b. Compared with the tendency of BERTSUMEXT, there is no obvious ROUGE decrease, demonstrating that our approach has strengthened the representation of pivotal and rebalanced the disproportionate ratio of summary sentences and other sentences. Note that more truncated BPE tokens will increase the final average ROUGE slightly, for it may lose some summary sentences when truncating too many tokens. Unfortunately, our 24-layers DifferSum can only be trained with 800 BPE tokens for the limitation of GPU source. 5.7 Map Words Representation into Sentence Representation A key issue motivating the sentence-level Transformer (TS) and the document-level Transformer (TD) is that the features for words after the TS might be at different scales or magnitudes. This can be due to some words having very sharp or very distributed attention weights when summing over the features of the other words. In this paper, we apply two ways to map the words representation into its sentence representation: weighted-pooling at Equation 2 and picking [CLS] token as sentence (Liu and Lapata, 2019). Table 7 shows that [CLS] is not enough to convey enough informative information of words for both our DifferSum and BERTSUMEXT. Especially, DifferSum is more sensitive to the word features since our differential amplifier may amplify the semantic features effectively. Table 7: ROUGE Scores about Sentence Representation on CNN/DM Test Set. Models R-1 R-L DifferSum (Weighted-Pooling) 44.70 40.83 DifferSum ([CLS]) 44.41 40.43 BERTSUMEXT (Weighted-Pooling) 43.92 40.08 BERTSUMEXT ([CLS]) 43.85 39.90 6 Conclusion In this paper, we introduce a heuristic model, DifferSum, 1) to calculate and amplifier the pivotal information and 2) to rebalance the distribution of minority 1-class and majority 0-class. Besides, we employ another weighted cross-entropy function to compensate for the imbalance. Experimental results show that our method significantly outperforms previous models. In the future, we would like to generalize DifferSum to other fields. Acknowledgements This research is supported by the National Key Research and Development Program of China 374 (NO.2017YFC0820700) and National Natural Science Foundation of China (No.61902394). We thank all authors for their contributions and all anonymous reviewers for their constructive comments. References Kristjan Arumae and Fei Liu. 2018. Reinforced extractive summarization with question-focused rewards. In ACL, pages 105–111. Akari Asai and Hannaneh Hajishirzi. 2020. Logicguided data augmentation and regularization for consistent question answering. In ACL, pages 5642– 5650. Sanghwan Bae, Taeuk Kim, Jihoon Kim, and Sang goo Lee. 2019. Summary level training of sentence rewriting for abstractive summarization. In arXiv preprint arXiv:1909.08752. David Balduzzi, Marcus Frean, Lennox Leary, JP Lewis, Kurt Wan-Duo Ma, and Brian McWilliams. 2018. The shattered gradients problem: If resnets are the answer, then what is the question? In ICML, pages 342–350. Christos Baziotis, Ion Androutsopoulos, Ioannis Konstas, and Alexandros Potamianos. 2019. Seq3: Differentiable sequence-to-sequence-to-sequence autoencoder for unsupervised abstractive sentence compression. In NAACL-HLT, pages 673–681. Keping Bi, Rahul Jha, W. Bruce Croft, and Asli Celikyilmaz. 2020. Aredsum: Adaptive redundancy-aware iterative sentence ranking for extractive document summarization. Jaime Carbonell and Jade Goldstein. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In SIGIR, pages 209–210. Yen-Chun Chen and Mohit Bansal. 2018. Fast abstractive summarization with reinforce-selected sentence rewriting. In ACL, pages 675–686. Jianpeng Cheng and Mirella Lapata. 2016. Neural summarization by extracting sentences and words. In ACL. Kyunghyun Cho, Bart Van Merri¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. EMNLP, pages 1724–1734. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, pages 4171–4186. Greg Durrett, Taylor Berg-Kirkpatrick, and Dan Klein. 2016. Learning-based single-document summarization with compression and anaphoricity constraints. In arXiv preprint arXiv:1603.08887. Haibo He and Yunqian Ma. 2013. Imbalanced learning: foundations, algorithms, and applications. John Wiley & Sons. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016a. Deep residual learning for image recognition. In CVPR, pages 770–778. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016b. Identity mappings in deep residual networks. In ECCV, pages 630–645. Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In NIPS, pages 1693–1701. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural computation, pages 1735–1780. Yi Yao Huang and William Yang Wang. 2017. Deep residual learning for weakly-supervised relation extraction. In EMNLP, pages 1803–1807. Aishwarya Jadhav and Vaibhav Rajan. 2018. Extractive summarization with swap-net: Sentences and words from alternating pointer networks. In ACL, pages 142–151. Ruipeng Jia, Yanan Cao, Haichao Shi, Fang Fang, Cong Cao, and Shi Wang. 2021. Flexible nonautoregressive extractive summarization with threshold: How to extract a non-fixed number of summary sentences. In AAAI. Ruipeng Jia, Yanan Cao, Haichao Shi, Fang Fang, Yanbing Liu, and Jianlong Tan. 2020a. Distilsum: Distilling the knowledge for extractive summarization. In CIKM, pages 2069–2072. Ruipeng Jia, Yanan Cao, Hengzhu Tang, Fang Fang, Cong Cao, and Shi Wang. 2020b. Neural extractive summarization with hierarchical attentive heterogeneous graph network. In EMNLP, pages 3622– 3631. Chris Kedzie, Kathleen McKeown, and Hal Daum´e III. 2018. Content selection in deep learning models of summarization. In EMNLP, pages 1818–1828. Logan Lebanoff, Kaiqiang Song, Franck Dernoncourt, Doo Soon Kim, Seokhwan Kim, Walter Chang, and Fei Liu. 2019. Scoring sentence singletons and pairs for abstractive summarization. ACL, pages 2175– 2189. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81. 375 Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In EMNLP, pages 3728–3738. Yang Liu, Ivan Titov, and Mirella Lapata. 2019. Single document summarization as tree induction. In NAACL-HLT, pages 1745–1755. Ling Luo, Xiang Ao, Yan Song, Feiyang Pan, Min Yang, and Qing He. 2019. Reading like HER: Human reading inspired extractive summarization. In EMNLP, pages 3033–3043. Christopher Manning, Mihai Surdeanu, John Bauer, Jenny Finkel, Steven Bethard, and David McClosky. 2014. The stanford corenlp natural language processing toolkit. In ACL, pages 55–60. Afonso Mendes, Shashi Narayan, Sebasti˜ao Miranda, Zita Marinho, Andr´e FT Martins, and Shay B Cohen. 2019. Jointly extracting and compressing documents with summary state representations. In NAACL-HLT, pages 3955–3966. Junghyun Min, R. Thomas McCoy, Dipanjan Das, Emily Pitler, and Tal Linzen. 2020. Syntactic data augmentation increases robustness to inference heuristics. In ACL, pages 2339–2352. Guido Mont´ufar, Razvan Pascanu, Kyunghyun Cho, and Yoshua Bengio. 2014. On the number of linear regions of deep neural networks. In NIPS, pages 2924–2932. Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. Summarunner: A recurrent neural network based sequence model for extractive summarization of documents. In AAAI, pages 3075–3081. Shashi Narayan, Shay B Cohen, and Mirella Lapata. 2018. Ranking sentences for extractive summarization with reinforcement learning. In NAACL-HLT, pages 1747–1759. Shashi Narayan, Joshua Maynez, Jakub Adamek, Daniele Pighin, Blaˇz Brataniˇc, and Ryan McDonald. 2020. Stepwise extractive summarization and planning with structured transformers. In EMNLP, pages 4143–4159. Karolina Owczarzak, John M Conroy, Hoa Trang Dang, and Ani Nenkova. 2012. An assessment of the accuracy of automatic evaluation in summarization. In Proceedings of Workshop on Evaluation Metrics and System Comparison for Automatic Summarization, pages 1–9. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K¨opf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In NIPS, pages 8024–8035. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., pages 140:1–140:67. Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In EMNLP, pages 379–389. Evan Sandhaus. 2008. The new york times annotated corpus. In Linguistic Data Consortium, Philadelphia. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In ACL, pages 1073–1083. Naftali Tishby, Fernando C Pereira, and William Bialek. 2000. The information bottleneck method. arXiv preprint physics/0004057. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS, pages 5998–6008. Andreas Veit, Michael Wilber, and Serge Belongie. 2016. Residual networks behave like ensembles of relatively shallow networks. In NIPS, pages 550– 558. Danqing Wang, Pengfei Liu, Yining Zheng, Xipeng Qiu, and Xuanjing Huang. 2020. Heterogeneous graph neural networks for extractive document summarization. In ACL, pages 6209–6219. Danqing Wang, Pengfei Liu, Ming Zhong, Jie Fu, Xipeng Qiu, and Xuanjing Huang. 2019. Exploring domain shift in extractive text summarization. In arXiv preprint arXiv:1908.11664. Peter West, Ari Holtzman, Jan Buys, and Yejin Choi. 2019. Bottlesum: Unsupervised and self-supervised sentence summarization using the information bottleneck principle. In EMNLP, pages 3750–3759. Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. 2020. Unsupervised data augmentation for consistency training. In NIPS. Jiacheng Xu and Greg Durrett. 2019. Neural extractive text summarization with syntactic compression. EMNLP, pages 3290–3301. Jiacheng Xu, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020. Discourse-aware neural extractive text summarization. In ACL, pages 5021–5031. Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang, and Ming Zhou. 2020. Prophetnet: Predicting future n-gram for sequence-to-sequence pre-training. In arXiv preprint arXiv:2001.04063, pages 2401–2410. 376 Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter J Liu. 2019a. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In arXiv preprint arXiv:1912.08777, pages 11328– 11339. Xingxing Zhang, Furu Wei, and Ming Zhou. 2019b. Hibert: Document level pre-training of hierarchical bidirectional transformers for document summarization. In ACL, pages 5059–5069. Ming Zhong, Pengfei Liu, Yiran Chen, Danqing Wang, Xipeng Qiu, and Xuanjing Huang. 2020. Extractive summarization as text matching. In ACL, pages 6197–6208. Ming Zhong, Pengfei Liu, Danqing Wang, Xipeng Qiu, and Xuanjing Huang. 2019a. Searching for effective neural extractive summarization: What works and what’s next. In ACL, pages 1049–1058. Ming Zhong, Danqing Wang, Pengfei Liu, Xipeng Qiu, and Xuanjing Huang. 2019b. A closer look at data bias in neural extractive summarization models. In arXiv preprint arXiv:1909.13705. Qingyu Zhou, Nan Yang, Furu Wei, Shaohan Huang, Ming Zhou, and Tiejun Zhao. 2018. Neural document summarization by jointly learning to score and select sentences. In ACL, pages 654–663. Barret Zoph, Ekin D. Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V. Le. 2019. Learning data augmentation strategies for object detection. In ECCV, pages 566–583.
|
2021
|
31
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4008–4018 August 1–6, 2021. ©2021 Association for Computational Linguistics 4008 Towards User-Driven Neural Machine Translation Huan Lin1,2 Liang Yao3 Baosong Yang3 Dayiheng Liu3 Haibo Zhang3 Weihua Luo3 Degen Huang4 Jinsong Su1,2,5∗ 1School of Informatics, Xiamen University 2Institute of Artificial Intelligence, Xiamen University 3Alibaba Group 4Dalian University of Technology 5Pengcheng Lab, Shenzhen [email protected] {yaoliang.yl,yangbaosong.ybs,liudayiheng.ldyh,zhanhui.zhb}@alibaba-inc.com [email protected] [email protected] [email protected] Abstract A good translation should not only translate the original content semantically, but also incarnate personal traits of the original text. For a real-world neural machine translation (NMT) system, these user traits (e.g., topic preference, stylistic characteristics and expression habits) can be preserved in user behavior (e.g., historical inputs). However, current NMT systems marginally consider the user behavior due to: 1) the difficulty of modeling user portraits in zero-shot scenarios, and 2) the lack of userbehavior annotated parallel dataset. To fill this gap, we introduce a novel framework called user-driven NMT. Specifically, a cache-based module and a user-driven contrastive learning method are proposed to offer NMT the ability to capture potential user traits from their historical inputs under a zero-shot learning fashion. Furthermore, we contribute the first ChineseEnglish parallel corpus annotated with user behavior called UDT-Corpus. Experimental results confirm that the proposed user-driven NMT can generate user-specific translations. 1 1 Introduction In recent years, neural machine translation (NMT) models (Sutskever et al., 2014; Luong et al., 2015; Vaswani et al., 2017) have shown promising quality and thus increasingly attracted users. When drawing on a translation system, every user has his own traits, including topic preference, stylistic characteristics, and expression habits, which can be implicitly embodied in their behavior, e.g., the historical inputs of these users. A good translation should implicitly mirror user traits rather than ∗Jinsong Su is the corresponding author. This work was done when Huan Lin was interning at DAMO Academy, Alibaba Group. 1We release our source code and the associated benchmark at https://github.com/DeepLearnXMU/ User-Driven-NMT. That is amazing ! Cool ! [cheerful, outgoing, active] [polite, formal, gentle] 太棒了！ 太棒了！ A B Figure 1: An example in which user traits leads to synonymous yet stylistically different translations. merely translate the original content, as the example shown in Figure 1. However, current NMT models are mainly designed for the semantic transformation between the source and target sentences regardless of subtle traits with respect to user behavior. It can be said that the effect of user behavior on translation modeling is still far from utilization, which, to some extent, limits the applicability of NMT models in real-world scenarios. More recently, several studies have shown that the prominent signals in terms of personal characteristics can be served as inductive biases and reflected in translation results using domain adaptation approaches, such as personality (Mirkin et al., 2015), gender (Rabinovich et al., 2017), and politeness (Sennrich et al., 2016a). However, previously explored signals characterize users from a single dimension, which insufficiently represent fine-grained user traits. Furthermore, Michel and Neubig (2018) pay their attention to personalized TED talk translation, in which they train a speakerspecific bias to revise the prediction distribution. In contrast with these studies, our work investigates a more realistic online scenario: a real-world MT system serves extensive users, where the user-behavior annotated data covering all users is unavailable. Previous methods (Mirkin et al., 2015; Michel and Neubig, 2018) require the users in the training set and the test set to be consistent, therefore can not 4009 deal with this zero-shot issue. Starting from this concern, we explore userdriven NMT that generates personalized translations for users unseen in the training dataset according to their behavior. Specifically, we choose the historical inputs to represent user behavior since they can not only be easily obtained in the real-world scenarios, but also reflect the topic preference, stylistic characteristic, and context of user. Moreover, compared with pre-defined or userspecific labels, historical inputs can be updated with current source sentences, which is also in line with realistic scenario. In this work, we propose a novel framework for this task, where the NMT model is equipped with a cache module to restore and update historical inputs. Besides, in order to further transfer the traits from the seen users to the unseen ones, we design a regularization framework based on contrastive learning (Bose et al., 2018; Yang et al., 2019), which forces our model to decrease the divergence between translations of similar users while increasing the diversity on dissimilar users. In order to further train and assess the proposed framework, we construct a new User-Driven Machine Translation dataset called UDT-Corpus. This corpus consists of 6,550 users with totally 57,639 Chinese sentences collected from a realworld online MT system. Among them, 17,099 Chinese sentences are annotated with their English translations by linguistic experts according to the user-specific historical inputs. Experimental results demonstrate that the proposed framework facilitates the translation quality, and exactly generates diverse translations for different users. To summarize, major contributions of our work are four-fold: • We introduce and explore user-driven NMT task that leverages user behavior to enhance translation model. We hope our study can attract more attention to explore techniques on this topic. • We propose a novel framework for user-driven NMT based on cache module and contrastive learning, which is able to model user traits in zero-shot scenarios. • We collect UDT-Corpus and make it publicly available, which may contribute to the subsequent researches in the communities of NMT and user-driven models. • Extensive analyses indicate the effectiveness of our work and verify that NMT can profit from user behavior to generate diverse translations conforming to user traits. 2 Related Work This section mainly includes the related studies of personalized machine translation, cache-based NMT and contrastive learning for NMT. Personalized Machine Translation Recently, some researchers have employed domain adaptation (Zhang et al., 2019; Gururangan et al., 2020; Yao et al., 2020) to generate personalized translations. For example, Mirkin et al. (2015) show that the translation generated by the SMT model has an adverse effect on the prediction of author personalities, demonstrating the necessity of personalized machine translation. Furthermore, Sennrich et al. (2016a) control the politeness in the translation by adding a politeness label on the source side. Rabinovich et al. (2017) explore a gender-personalized SMT system that retains the original gender traits. These domain labels represent users in single dimension separately, which are insufficient to distinguish large-scale users in a fine-grained way. The most correlated work to ours is Michel and Neubig (2018) which introduces a speaker-specific bias into the conventional NMT model. However, these methods are unable to deal with users unseen at the training time. Different from them, user-driven NMT can generate personalized translations for these unseen users in a zero-shot manner. Cache-Based Machine Translation Inspired by the great success of cache on language modeling (Kuhn and de Mori, 1990; Goodman, 2001; Federico et al., 2008), Nepveu et al. (2004) propose a cache-based adaptive SMT system. Tiedemann (2010) explore a cache-based translation model that fills the cache with bilingual phrase pairs extracted from previous sentence pairs in a document. Bertoldi et al. (2013) use a cache mechanism to achieve online learning in phrase-based SMT. Gong et al. (2011), Kuang et al. (2018), and Tu et al. (2018) further exploit cache-based approaches to leverage contextual information for document-level machine translation. Contrast with the documentlevel NMT that learns to capture contextual information, our study aims at modeling user traits, such as, topic preference, stylistic characteristics, and expression habits. Moreover, historical inputs of user has relatively fewer dependencies than the contexts 4010 used in document-level translation. Contrastive Learning for NMT Contrastive learning has been extensively applied in the communities of computer vision and natural language processing due to its effectiveness and generality on self-supervised learning (Vaswani et al., 2013; Mnih and Kavukcuoglu, 2013; Liu and Sun, 2015; Bose et al., 2018). Towards raising the ability of NMT in capturing global dependencies, Wiseman and Rush (2016) first introduce contrastive learning into NMT, where the ground-truth translation and the model output are considered as the positive and contrastive samples, respectively. Yang et al. (2019) construct contrastive examples by deleting words from ground-truth translation to reduce word omission errors in NMT. Contrast to these studies, we employ contrastive learning to create broader learning signals for our user-driven NMT model, where the prediction distribution of translations with respect to similar users and dissimilar users are considered as positive and contrastive samples, respectively. Thus, our model can better transfer the knowledge of the seen users to the unseen ones. 3 User-Driven Translation Dataset In order to build a user-driven NMT system, we construct a new dataset called UDT-Corpus containing 57,639 inputs of 6,550 users, 17,099 among them are Chinese-to-English translation examples. 3.1 Data Collection and Preprocessing We collect raw examples from Alibaba Translate2 which contain the user inputs and the translations given by the translation system. For data preprocessing, we first anonymize data and perform data deduplication within each user. Then, we utilize a pre-trained n-gram language model KenLM3 to filter out translation examples with low-quality source data. Moreover, we remove such pairs whose source sentence is shorter than 2 words or longer than 100 words. 3.2 Data Annotation In the corpus, we represent each translation example as a triplet ⟨X(u), Y (u), H(u)⟩, where H(u) is the historical inputs of the user u, X(u) is the current source sentence and Y (u) is the target translation sentence annotated with H(u). To obtain 2https://www.aliyun.com/product/ai/ base_alimt 3https://github.com/kpu/kenlm. such a triplet, we first sequentially sample up to 10 source sentences which are the historical inputs of each user. Then, for the given historical inputs, we collect their followed source input paired with the pseudo translation given by the translation system. Afterwards, we assign these historical inputs and the current input pairs to two professional annotators and ask them to revise the pseudo translation according to the source sentence and historical inputs. Specifically, we first ask one of them to annotate and the other to evaluate, and then resolve annotation disagreements by reviewing. During annotation, 91.8% of the original data are revised. Moreover, annotators are asked to record whether their revision is affected by user history. The result shows that 76.25% of the sentences are impacted. 4 User-Driven NMT Framework In this section, we first give a brief description about the problem formulation of user-driven NMT, and then introduce our proposed framework in detail. We choose Transformer (Vaswani et al., 2017) as the basic NMT model due to its competitive performance. In fact, our framework is transparent and applicable to other NMT models. Figure 2 illustrates the basic framework of the proposed user-driven NMT. Most typically, we equip the NMT model with two user-specific caches to exploit user behavior for better translation (See Section § 4.2). Besides, we augment the conventional NMT training objective with contrastive learning, which allows the model to learn translation diversity across users (See Section § 4.3). 4.1 Problem Formulation Given the source sentence X and the previously generated words Y<i = y1, ..., yi−1, the conventional NMT model with parameter θ predicts the current target word yi by P (yi|X, Y<i; θ). As a significant extension of conventional NMT, userdriven NMT with parameter θ aims to model P y(u) i |X(u), Y (u) <i , u; θ , that is, generates the translation that can reflect the traits of user u. Unlike previous studies (Mirkin et al., 2015; Michel and Neubig, 2018) only caring for generating translations for users seen at the training time, our userdriven NMT mainly focuses on a more realistic online MT scenario, where the users for testing are unseen in the training dataset. Moreover, the conventional domain adaptation methods can not be directly applied to this zero-shot scenario. 4011 Topic Cache 𝐜! (#) Historical Inputs ⊕ Context Cache 𝐜% (#) 𝐻(#) 𝑟(") 𝑟("!) 𝑟("") 𝑋(") 𝑌(") 𝑃(𝑦! " |𝑋" , 𝑌#! " , 𝐻" ) 𝑃(𝑦! " |𝑋" , 𝑌#! " , 𝐻"! ) 𝑃(𝑦! " |𝑋" , 𝑌#! " , 𝐻"" ) 𝐿$% + 𝐿&%' User-Driven NMT Model Figure 2: The architecture of our user-driven NMT model. We use the topic cache and context cache to capture the long-term and short-term user traits for user u from corresponding historical inputs H(u), respectively. Then, we combine the representations of two caches to get a user behavior representation r(u), which is fed into the NMT model for personalized translation. Furthermore, we use contrastive learning involving similar user u+ and dissimilar user u−to increase the translation diversity among different users. 4.2 Cache-based User Behavior Modeling Due to the advantages of cache mechanism on dynamic representations (Gong et al., 2011; Kuang et al., 2018; Tu et al., 2018), we equip the conventional Transformer-based NMT model with two user-specific caches to leverage user behavior for NMT: 1) topic cache c(u) t that aims at capturing the global and long-term traits of user u; and 2) context cache c(u) c , which is introduced to capture the short-term traits from the recent source inputs of user u. During this process, we focus on the following three operations on cache: Cache Representation In order to facilitate the efficient computation of the user behavior encoded by our caches, we define each cache as an embedding sequence of keywords. We first calculate TF-IDF values of input words, and then extract words with TF-IDF weights higher than a predefined threshold to represent user behavior. Note that the calculation of TF-IDF value of a word mainly depends on its frequency in the document and inverse document frequency in the corpus. Since two caches play different roles in the userdriven NMT model, we identify keywords for two caches based on different definitions of “document” and “corpus”. Specifically, when constructing topic cache c(u) t , we treat the historical inputs H(u) of the user u as the “document” and the historical inputs H(u) of all users U as the “corpus”, then define topic cache c(u) t as an embedding sequence of historical keywords. Unlike the topic cache, for context cache c(u) c , we individually consider the current source sentence X(u) and historical inputs H(u) as the TF-IDF “document” and “corpus”, defining c(u) c as an embedding sequence of current keywords. Besides, in the real-world MT scenario, there exists a large number of users without any historical input. For these users, we find the most similar user according to the cosine similarity based on their TF-IDF bag-of-word representations of topic keywords, and initialize the corresponding topic cache with that of the most similar user. Updating Caches When using an online MT system, users often continuously input multiple sentences. Thus, our caches should be dynamically updated to ensure the accurate encoding of user behavior. To update topic cache, we first recalcualte the TF-IDF values of all historical input words, so as to redetermine the keywords stored in this cache. As for context cache, we consider it as a filter window sliding across historical inputs, and apply first-infirst-out rule to replace its earliest keywords with the recently input ones. Reading from Caches During the translation of the NMT model, we perform a gating operation on c(u) t and c(u) c , producing a vector r(u) that reflects user behavior as follows: r(u) = αc(u) t + (1 −α)c(u) c (1) α = Sigmoid(Wtc(u) t + Wrc(u) c ), (2) c(u) t = MeanPooling h c(u) t i , (3) c(u) c = MeanPooling h c(u) c i , (4) where both Wt and Wr are learnable parameter matrices. Then, we directly add r(u) into the embedding sequence of original current source sentence X(u), forming a source embedding sequence with user behavior as follows: ˆX(u) = {x(u) i + r(u)}1≤i<|X(u)|. (5) Finally, the NMT model is fed with ˆX(u) to generate the translation for u. Due to the limitation 4012 of pages, we omit the detailed descriptions of the NMT model. Please refer to Vaswani et al. (2017) for the details. 4.3 Model Training with a Contrastive Loss Given training instances ⟨X(u), Y (u), H(u)⟩, we train the user-driven NMT model using the following objective function: L = Lmle + Lcl. (6) Here, Lmle is the maximum likelihood translation loss extended from the conventional NMT training objective. Formally, it is defined as: Lmle = X i −log P(y(u) i |X(u), Y (u) <i , H(u); θ). (7) Lcl is a triplet-margin-based constrastive loss, which allows the NMT model to learn the translation diversity across users. Specifically, for an input sentence, an ideal userdriven NMT model should be able to generate translations with non-divergent user traits for similar users, while producing translations with diverse user traits for dissimilar users. However, using only Lmle cannot guarantee this since it separately considers each training instance during the model training. To deal with this issue, for each training instance ⟨X(u), Y (u), H(u)⟩, we first determine the most similar user u+ according to the cosine similarity based on their bag-of-keyword representations, and randomly select a user without any same keyword as the dissimilar user u−of u. Finally, using historical inputs of u+ and u−, we construct several pseudo training instances to define Lcl as follows: Lcl = X u∈U max[d(X(u), Y (u), H(u), H(u+)) (8) −d(X(u), Y (u), H(u), H(u−)) + η, 0], where d X(u), Y (u), H(u), H(u+) = || 1 |Y (u)| X i log P y(u) i |X(u), Y (u) <i , H(u) − 1 |Y (u)| X i log P y(u) i |X(u), Y (u) <i , H(u+) ||2 (9) and η is a predefined threshold, which is set to 2 in our experiments. Here, we omit the definition of Train Dev Test #user 5,350 600 600 #historical input 33,441 3,629 3,470 #current sentence pairs 14,006 1,557 1,536 Table 1: Dataset for fine-tuning experiments. d X(u), Y (u), H(u), H(u−) , which is similar to d X(u), Y (u), H(u), H(u+) . Formally, Lcl will encourage the NMT model to minimize the prediction difference between the training instances ⟨X(u), Y (u), H(u)⟩and ⟨X(u), Y (u), H(u+)⟩, and maximize the difference between the training instances ⟨X(u), Y (u), H(u)⟩ and ⟨X(u), Y (u), H(u−)⟩. In this way, the NMT model can not only exploit pesudo training instances, but also produce more consistent translations with user traits. 5 Experiments In this section, we carry out several groups of experiments to investigate the effectiveness of our proposed framework on UDT-Corpus. 5.1 Setup We develop the user-driven NMT model based on Open-NMT Transformer (Klein et al., 2017), and adopt a two-stage strategy to train this model: we first pre-train a Transformer-based NMT model on the WMT2017 Chinese-to-English dataset, and then fine-tune this model to our user-driven NMT model using UDT-Corpus. Datasets The WMT2017 Chinese-to-English dataset is composed of the News Commentary v12, UN Parallel Corpus v1.0, and CWMT corpora, with totally 25M parallel sentences. To fine-tune our model, we split UDT-Corpus into training, validation and test set, respectively. Table 1 provides more detailed statistics of these datasets. To improve the efficiency of model training, we train the model using only parallel sentences with no more than 100 words. Following common practices, we employ byte pair encoding (Sennrich et al., 2016b) with 32K merge operations to deal with all sentences. Training Details Following Vaswani et al. (2017), we use the following hyper-parameters: the word embedding dimension is set to 512, the hidden layer dimension is 2048, the layer numbers of 4013 both encoder and decoder are set to 6, and the number of attention heads is set to 8. Besides, we use 4 GPUs for training. At the pre-training stage, we employ the Adam optimizer with β2 = 0.998. We use the batch size of 16,384 tokens and pre-train the model for 200,000 steps. Particularly, we adopt the dropout strategy (Srivastava et al., 2014) with rate 0.1 to enhance the robustness of our model. When fine-tuning the model, we keep the other settings consistent with the pre-training stage, but reduce the batch size to 2048 tokens and fine-tune the model with early-stopping strategy. Evaluation We assess the translation quality with two metrics: one is case-insensitive BLEU (mteval-v13a.pl, Papineni et al., 2002)4 and the other is METEOR5 (Denkowski and Lavie, 2011). 5.2 Baselines We represent our user-driven NMT model as UDNMT and compare it with the following baselines: • TF. It is a Transformer-based NMT model pretrained on the WMT2017 corpus. This model yields 24.61 BLEU score on WMT2017 Chinese-to-English translation task, which is comparable with reported results in (Wan et al., 2020; Zhou et al., 2020), which makes our subsequent experiments convincing. • TF-FT. This model is also a Transformerbased NMT model that is further fine-tuned on the parallel sentences of UDT-Corpus. • TF-FT + PesuData. This model is a variant of TF-FT. When constructing it, we pair historical inputs with their translations produced by our online translation system, forming additional data for fine-tuning TF-FT. • TF-FT + ConcHist (Tiedemann and Scherrer, 2017). In this model, we introduce user behavior into TF-FT by concatenating each input sentence with several historical inputs. We mark all tokens in historical inputs with a special prefix to indicate that they are additional information. • TF-FT + UserBias (Michel and Neubig, 2018). It introduces user-specific biases to refine softmax-based predictions of Transformer NMT model. We change it to a zeroshot method similar to (Farajian et al., 2017) 4https://github.com/moses-smt/ mosesdecoder/blob/master/scripts/ generic/multi-bleu.perl 5https://github.com/cmu-mtlab/meteor 5 15 25 35 45 BLEU 33.0 32.9 32.8 32.7 32.6 32.5 32.4 32.3 5 15 25 35 45 BLEU (a) Topic Cache Size 𝒔𝒕(𝒔𝒄=35) (b) Context Cache Size 𝒔𝒄(𝒔𝒕=25) 33.0 32.9 32.8 32.7 32.6 32.5 32.4 32.3 Figure 3: Effects of cache size on translation quality. Model BLEU METEOR w/o user behavior TF 27.52 44.05 TF-FT 28.61 45.35 TF-FT + PesuData 29.02 45.40 w/ user behavior TF-FT + ConcHist 30.85 46.08 TF-FT + UserBias 31.36 46.79 UD-NMT 32.35 48.20 Table 2: Main results on UDT-Corpus. “w/o”, “w/” denote “without” and “with”, respectively. since (Michel and Neubig, 2018) can not be directly applied to our scenario. In particular, we replace the user ID in the test set with that of the most similar user in the training set. Note that the first two baselines, e.g., TF and TFFT, are conventional NMT models without exploiting user behavior. 5.3 Effect of Cache Sizes Since cache size directly determines the utility of user behavior, we investigate its effect on the performance of UD-NMT. We denote the sizes of topic cache and context cache as st and sc for simplicity. Figure 3 lists the performance of our model with different st and sc on validation set. We observe that st larger than 25 and sc larger than 35 do not lead to significant improvements. For this result, we speculate that small cache sizes are unable to capture sufficient user behavior for NMT. However, since the number of keywords are limited, larger cache sizes only bring limited information gain. Therefore, we directly use st = 25 and sc = 35 in the subsequent experiments. 5.4 Main Results From Table 2, we observe that our UD-NMT model consistently outperforms all baselines in terms of two metrics. Moreover, we draw several interesting conclusions: 4014 Model BLEU↑ METEOR↑ s-BLEU↑ d-BLEU↑ s-Sim.↓ d-Sim.↓ UD-NMT 32.35 48.20 32.17 32.23 93.18 80.10 w/o topic cache 31.88† 48.00 – – – – w/o context cache 31.86† 47.84† 31.94† 31.58† 88.61 69.32 w/o similar user initialization 32.02 48.14 31.86† 31.13‡ 93.54† 80.16 w/o contrastive learning 32.00 48.09 31.88† 31.94 93.49† 81.59† Table 3: Ablation Study. ↑: higher is better, ↓: lower is better. Since the user similarity is calculated based on the topic keywords, the model can not find similar user and dissimilar user without it. Thus w/o topic cache does not have the s-BLEU, s-Sim., d-BLEU and d-Sim.. ‡/†: indicates the drop of translation quality is statistically significant comparing to “UD-NMT” (p<0.01/0.05). 1) All NMT models leveraging user behavior surpass vanilla models, including TF, TF-FT, showing that user behavior is useful for NMT. 2) UD-NMT exhibits better than TF-FT + PesuData, which uses the same training data as ours. The underlying reason is that UD-NMT can leverage user traits to generate better translations. 3) Although both TF-FT + UserBias and UDNMT exploit user behavior for NMT, UD-NMT achieves better performance than TF-FT + UserBias without introducing extra parameters. This result demonstrates the advantage of cache on modeling user behavior than introducing user-specific biases into model parameters. 5.5 Ablation Study To explore the effectiveness of different components in our model, we further compare UD-NMT with its several variants, as shown in Table 3. Particularly, we propose to evaluate translations using the following variant metrics: s-BLEU, sSim., d-BLEU and d-Sim.. When using s-BLEU, we replace the topic cache of current user with that of his most similar user. Keeping the same current input, we calculate the BLEU score with ground-truth as reference and the translation for this similar user as hypothesis. As for s-Sim., we adopt the same strategy as s-BLEU, but use the translation for original user as reference to evaluate the BLEU score. In other words, s-BLEU and dBLEU assesses the translation quality given unsuitable user. Therefore, higher s-BLEU and d-BLEU indicates better model robustness, while s-BLEU and d-BLEU measures how much the translation changes given different user. Thus lower s-Sim. and d-Sim. show larger translation diversity. Our conclusions are shown as follows: 1) w/o topic cache. To build this variant, we remove topic cache from our model. The result in Line 2 indicates that removing topic cache leads to a performance drop, suggesting that topic cache is useful for modeling user behavior. 2) w/o context cache. Unlike the above variant, we only use topic cache to represent user traits in this variant. According to the results shown in Line 3, we observe that this change results in a significant performance decline of our model, demonstrating that context cache also effectively captures user behavior for NMT. However, the translation diversity among users increases since the model will not be affected by the context cache in this variant, which is the same between different users when calculating s-Sim. and d-Sim.. 3) w/o similar user initialization. In this variant, we do not initialize topic caches of the users without historical inputs using that of the most similar users. From Line 4, we observe that the performance of our model degrades without similar user initialization. 4) w/o contrastive learning. In this variant, we remove the contrastive learning from the whole training objective to inspect the performance change of our model. As shown in Line 4, the performance of our model drops, proving that the contrastive learning is important for the training of our model. Moreover, we can infer from Column 6 and 7 that our model can generate diverse translations. Specifically, the translations of dissimilar users has larger diversity than that of similar ones. Furthermore, we conclude that our model is robust, since it still performs well when we replace the topic cache of current user with those of other users (See Column 4 and 5). 5.6 Analysis of Contrastive Margin Inspired by Yang et al. (2019), we argue that the contrastive learning may increase the prediction diversity of our model between users compared with using the MLE loss. To confirm this, we randomly 4015 Historical inputs 面料成分氨纶风格性感款式连体裤颜色白色, 黑色(Fabric Composition Spandex Style Sexy Type Jumpsuit Color White, Black) 2020 秋冬季新款港风复古落肩外套女宽松学生毛绒短款上衣(2020 Autumn and Winter New Hong Kong Style Retro Drop Sleeves Jacket Female Loose Student Plush Crop Top) Src 牛津纺面料防水耐磨, 15英寸大小 Ref Oxford Fabric Waterproof and Wear Resistant, 15 Inch in Size TF-FT + PesuData Oxford Textile Fabrics Waterproof and Waterproof, 15 Inches Size TF-FT + UserBias Oxford Woven Fabric is Waterproof and Resistant, 15 Inches in Size UD-FT Oxford Woven Fabric Waterproof and Wear - Resistant, 15 Inch Size (a) Translations of UD-NMT given different user traits. Words highlighted in green indicate the unnatural or awkward translations. (b) Translations of different models. Words highlighted in red are incorrect translations. Src 基因芯片分析发现, 在crf 多突变体中, 许多受b型arrs 调控的基因同时也受crfs的调制剂起负反馈调节作用。 Ref Gene chip analysis found out that in multiple CRF mutants, many genes regulated by type b arrs were also negatively regulated by the modulators of CRFs. User A Topic Cache 木马（trojan）| 渔夫（fisherman）| 百洁布（cleaning cloth）| 耳机（earphone）| 蓝牙（bluetooth） Translation Gene chip analysis found that in the CRF mutant , many genes regulated by b arrs are also negatively fed by CRFs toning . User B Topic Cache 乙烯（ethylene）| 伸长（extension）| 促进（promote）| 基因（gene）| 生长素（auxin）| 调控（regulate） Translation Gene chip analysis found that in the CRF mutant , many genes regulated by type b arrs are also subject to negative feedback adjustment by CRFs modulating agents . Figure 4: Two examples of user-driven machine translation. sample 300 examples from the training dataset, and compute the following margin: ∆= h d(u+) (·)−d(u−) (·) i − h d(u+) mle (·)−d(u−) mle (·) i , where d(u+)(·) is defined in Equation 9. The definition of d(u+) mle (·) is the same with d(·), the only difference lies in that the NMT model is only trained by the conventional MLE loss. We find that d(·) has a larger margin than dmle(·) on 88% of sampled sentence pairs, with an average margin of 0.19. The results indicate again that the contrastive learning increases the translation diversity. 5.7 Qualitative Analysis In order to intuitively understand how our cache module exactly affects the translations, we feed our model with the same current source sentence but different users, and display the 1-best translations generated by our model. As shown in the Figure 4 (a), our model is able to produce correct but diverse translations according to different topic caches. Moreover, it is interesting to observe that specific topic keywords such as “type b arr”, “negatively regulated” and “modulators” are translated to synonymous but “out-of-domain” phrases if the topic cache does not conform to input sentence. On the contrary, the model conversely generates “indomain” translation if the topic cache comes from the same topic of input sentence. Correlation Order Proportion UD-NMT > TF-FT + PesuData 86% UD-NMT > TF-FT + UserBias 74% Table 4: The proportion of translations more related to historical inputs assessed by human translators. A > B indicates the translations generated by A system is more correlated to history inputs. Besides, to further reveal the effect of user behavior, we provide an example in Figure 4 (b), which lists different translations by compared models for the same inputs. The historical inputs indicate that this user may be an apparel seller, since his historical inputs contain the product titles and descriptions of clothing. Thus, the keywords “Wear Resistant” in the source sentence are correlated with this user. However, two baselines translate it to “Waterproof” and “Resistant”, respectively. Moreover, TF-FT + UserBias generates a subject–verb–object structured sentence by adding the auxiliary verb “is”, which does not conform to the expression habit of the product title. By contrast, with the hint of the keywords in historical inputs, our UD-NMT is able to produce suitable translation consistent with the topic preference of this user. 5.8 Manual Evaluation To further find out weather the improvements of our model are contributed by user traits, we ran4016 domly sample 100 examples from the test dataset and ask the linguist experts to sort different systems according to the relevance between the generated translations and the historical input. The results in Table 4 show that our model can generate translations more in line with history inputs than baseline models in most cases, proving that our method can make better use of user traits. 6 Conclusion We propose user-driven NMT task, which aims to leverage user behavior to generate personalized translations. With the help of cache module and contrastive estimation, we successfully build an end-to-end NMT model that is able to capture potential user traits from their historical inputs and generate diverse translations under a zero-shot learning fashion. Furthermore, we contribute UDTCorpus, which is the first Chinese-English parallel corpus annotated with user behavior. We expect our study can attract more attention towards this topic. It is a promising direction to explore other behavior in future, such as clickthrough and editing operations. Moreover, following recent advancements in domain adaptation for NMT, we plan to further improve our model via adversial training based knowledge transfer (Zeng et al., 2018; Yao et al., 2020; Su et al., 2021) and dual knowledge transfer (Zeng et al., 2019). Acknowledgments The project was supported by National Key Research and Development Program of China (No. 2020AAA0108004 and No. 2018YFB1403202), National Natural Science Foundation of China (No. 61672440), Natural Science Foundation of Fujian Province of China (No. 2020J06001), Youth Innovation Fund of Xiamen (No. 3502Z20206059), and the Fundamental Research Funds for the Central Universities (No. ZK20720200077). We also thank the reviewers for their insightful comments. References Nicola Bertoldi, Mauro Cettolo, and Marcello Federico. 2013. Cache-based online adaptation for machine translation enhanced computer assisted translation. In Proceedings of the 2013 Machine Translation Summit, pages 35–42. Avishek Joey Bose, Huan Ling, and Yanshuai Cao. 2018. Adversarial contrastive estimation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 1021– 1032. Michael J. Denkowski and Alon Lavie. 2011. Meteor 1.3: Automatic metric for reliable optimization and evaluation of machine translation systems. In Proceedings of the 6th Workshop on Statistical Machine Translation, pages 85–91. M. Amin Farajian, Marco Turchi, Matteo Negri, and Marcello Federico. 2017. Multi-domain neural machine translation through unsupervised adaptation. In Proceedings of the Second Conference on Machine Translation, pages 127–137. Marcello Federico, Nicola Bertoldi, and Mauro Cettolo. 2008. IRSTLM: an open source toolkit for handling large scale language models. In Proceedings of the 9th Annual Conference of the International Speech Communication Association, pages 1618–1621. Zhengxian Gong, Min Zhang, and Guodong Zhou. 2011. Cache-based document-level statistical machine translation. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 909–919. Joshua T. Goodman. 2001. A bit of progress in language modeling. Computer Speech and Language, 15(4):403–434. Suchin Gururangan, Ana Marasovic, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8342–8360. Guillaume Klein, Yoon Kim, Yuntian Deng, Jean Senellart, and Alexander Rush. 2017. OpenNMT: Opensource toolkit for neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 67–72. Shaohui Kuang, Deyi Xiong, Weihua Luo, and Guodong Zhou. 2018. Modeling coherence for neural machine translation with dynamic and topic caches. In Proceedings of the 27th International Conference on Computational Linguistics, pages 596–606. Roland Kuhn and Renato de Mori. 1990. A cachebased natural language model for speech recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 12(6):570–583. Yang Liu and Maosong Sun. 2015. Contrastive unsupervised word alignment with non-local features. In Proceedings of the 29th Association for the Advancement of Artificial Intelligence, pages 2295–2301. Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412–1421. 4017 Paul Michel and Graham Neubig. 2018. Extreme adaptation for personalized neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 312– 318. Shachar Mirkin, Scott Nowson, Caroline Brun, and Julien Perez. 2015. Motivating personality-aware machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1102–1108. Andriy Mnih and Koray Kavukcuoglu. 2013. Learning word embeddings efficiently with noise-contrastive estimation. In Proceedings of the 27th Annual Conference on Neural Information Processing Systems, pages 2265–2273. Laurent Nepveu, Guy Lapalme, Philippe Langlais, and George F. Foster. 2004. Adaptive language and translation models for interactive machine translation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 190–197. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318. Ella Rabinovich, Raj Nath Patel, Shachar Mirkin, Lucia Specia, and Shuly Wintner. 2017. Personalized machine translation: Preserving original author traits. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 1074–1084. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Controlling politeness in neural machine translation via side constraints. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics, pages 35–40. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1715–1725. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. The Journal of Machine Learning Research, 15(1):1929–1958. Jinsong Su, Jiali Zeng, Jun Xie, Huating Wen, Yongjing Yin, and Yang Liu. 2021. Exploring discriminative word-level domain contexts for multidomain neural machine translation. IEEE Trans. Pattern Anal. Mach. Intell., 43:1530–1545. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Proceedings of the 28th Annual Conference on Neural Information Processing Systems, pages 3104–3112. J¨org Tiedemann. 2010. Context adaptation in statistical machine translation using models with exponentially decaying cache. In Proceedings of the 2010 Workshop on Domain Adaptation for Natural Language Processing, pages 8–15. J¨org Tiedemann and Yves Scherrer. 2017. Neural machine translation with extended context. In Proceedings of the 3rd Workshop on Discourse in Machine Translation, pages 82–92. Zhaopeng Tu, Yang Liu, Shuming Shi, and Tong Zhang. 2018. Learning to remember translation history with a continuous cache. Transactions of the Association for Computational Linguistics, 6:407–420. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31th Annual Conference on Neural Information Processing Systems, pages 5998–6008. Ashish Vaswani, Yinggong Zhao, Victoria Fossum, and David Chiang. 2013. Decoding with large-scale neural language models improves translation. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1387– 1392. Yu Wan, Baosong Yang, Derek F. Wong, Yikai Zhou, Lidia S. Chao, Haibo Zhang, and Boxing Chen. 2020. Self-paced learning for neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1074–1080. Sam Wiseman and Alexander M. Rush. 2016. Sequence-to-sequence learning as beam-search optimization. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1296–1306. Zonghan Yang, Yong Cheng, Yang Liu, and Maosong Sun. 2019. Reducing word omission errors in neural machine translation: A contrastive learning approach. In Proceedings of the 57th Conference of the Association for Computational Linguistics, pages 6191–6196. Liang Yao, Baosong Yang, Haibo Zhang, Boxing Chen, and Weihua Luo. 2020. Domain transfer based data augmentation for neural query translation. In Proceedings of the 28th International Conference on Computational Linguistics, COLING 2020, Barcelona, Spain (Online), December 8-13, 2020, pages 4521–4533. International Committee on Computational Linguistics. 4018 Jiali Zeng, Yang Liu, Jinsong Su, Yubin Ge, Yaojie Lu, Yongjing Yin, and Jiebo Luo. 2019. Iterative dual domain adaptation for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 845–855. Jiali Zeng, Jinsong Su, Huating Wen, Yang Liu, Jun Xie, Yongjing Yin, and Jianqiang Zhao. 2018. Multidomain neural machine translation with word-level domain context discrimination. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 447–457. Xuan Zhang, Pamela Shapiro, Gaurav Kumar, Paul McNamee, Marine Carpuat, and Kevin Duh. 2019. Curriculum learning for domain adaptation in neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, pages 1903– 1915. Yikai Zhou, Baosong Yang, Derek F. Wong, Yu Wan, and Lidia S. Chao. 2020. Uncertainty-aware curriculum learning for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6934– 6944.
|
2021
|
310
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4019–4033 August 1–6, 2021. ©2021 Association for Computational Linguistics 4019 End-to-End Lexically Constrained Machine Translation for Morphologically Rich Languages Josef Jon and João Paulo Aires and Dušan Variš and Ondˇrej Bojar Charles University [email protected] Abstract Lexically constrained machine translation allows the user to manipulate the output sentence by enforcing the presence or absence of certain words and phrases. Although current approaches can enforce terms to appear in the translation, they often struggle to make the constraint word form agree with the rest of the generated output. Our manual analysis shows that 46% of the errors in the output of a baseline constrained model for English to Czech translation are related to agreement. We investigate mechanisms to allow neural machine translation to infer the correct word inflection given lemmatized constraints. In particular, we focus on methods based on training the model with constraints provided as part of the input sequence. Our experiments on the English-Czech language pair show that this approach improves the translation of constrained terms in both automatic and manual evaluation by reducing errors in agreement. Our approach thus eliminates inflection errors, without introducing new errors or decreasing the overall quality of the translation. 1 Introduction In Neural Machine Translation (NMT), lexical constraining (Song et al., 2019; Hokamp and Liu, 2017; Post and Vilar, 2018) involves changing the translation process in a way that desired terms appear in the model’s output. Translation constraints are useful in domain adaptation, interactive machine translation or named entities translation. Current approaches focus either on manipulating beam search decoding (Hokamp and Liu, 2017; Post and Vilar, 2018; Hu et al., 2019) or training an NMT model using constraints alongside the input (Dinu et al., 2019; Song et al., 2019; Chen et al., 2020). In inflected languages, constraints from both source and target sides may appear in numerous surface forms, which may result in errors during Likud party has merged with an even more hawkish lot under Avigdor Lieberman. Input (EN) No constraint translation (CS) Strana Likud se spojila s ještě jestřábím losem pod Avigdorem Liebermanem. Surface form model output (CS) Strana Likud se spojila s ještě radikální partou pod vedením Avigdora Liebermana Lemmatized model output (CS) Strana Likud se spojila s ještě radikálnější partií pod vedením Avigdora Liebermana. radikální radikální Figure 1: Comparison between constrained translations from English to Czech. translation. By enforcing the presence of a certain exact term on the target side, existing approaches fail to deal with word inflections. As we show, they preserve the surface form of the word provided as constraint regardless of the context. Morphologically rich languages have multiple forms of each word, e.g. inflections to nouns. For satisfactory results in these languages, the constraint processing method needs to be capable of detecting any surface form on the source side and generating the correct surface form on the target side. To illustrate the problem, Figure 1 shows a sentence translation from English to Czech with outputs from three methods. The first one is a noconstraint translation where “hawkish” is translated as “jestˇrábím” (literally “hawkish”, no figurative meaning; followed by a further mis-translation of “lot”). The second is a constrained model requested to use the word form “radikální” (“radical”) in the output. The constraint was satisfied but the adjective should have taken the comparative degree to match the rest of the translation. The third output is the result of a model that processes the input along with the canonical form constraint (“radikální”) and modifies the constraint inflection in the final translation (“radikálnˇejší”) to correctly express the comparative form (although the translation of “lot” is worse than in previous case). 4020 We evaluate different methods of lexically constrained machine translation on the Czech language. We propose an approach to deal with word inflection in lexically constrained translation. By training a model that receives lemmatized target constraints as the input alongside the source sentence, we improve the generation of constraints in forms matching the output context. We run experiments on both synthetic and real-world test scenarios. 2 Related work In MT, there are scenarios where words that should or should not appear in the output are known upfront. Common use cases include integration of domain-specific terminology and translation of named entities or rare words using a dictionary. Such functionality was previously implemented in phrase-based systems (Okuma et al., 2008), like Moses (Koehn et al., 2007). In NMT, this task is not yet definitely solved, since the translation process is hard to interpret and influence. 2.1 Output post-processing In order to enforce the presence of specific terms, some approaches post-process the output. Prior to subword handling (Sennrich et al., 2016; Kudo and Richardson, 2018), unknown words were corrected by replacing them with word translation pairs from a bilingual dictionary (Luong et al., 2015). Crego et al. (2016) use placeholders to translate numbers and named entities. Placeholders have also been found useful for translation of text with formal mark-up and its interaction with content (Hanneman and Dinu, 2020). 2.2 Constrained decoding An alternative way of adding constraints to the final translation is by manipulating the beam search decoding process. Anderson et al. (2017) use a finite state machine (FSM) that recognizes target sentence with constraint patterns. Each state of the FSM has its own beam and only hypotheses in beams that are in accepting states can be finished. Hasler et al. (2018) improve upon this work by utilizing encoder-decoder attention weights to guide the placement of a constraint. Chatterjee et al. (2017) also use attention weights and beam search look-ahead to choose constraint positions. Hokamp and Liu (2017) present Grid Beam Search, which extends the usual beam search (Och and Ney, 2004) with a mechanism to ensure the coverage of all constrains. Post and Vilar (2018) propose a similar but more efficient algorithm. By dynamically reallocating the beam capacity, an arbitrary number of constraints can be processed within a constant width of the beam. One shortcoming of the above methods is the slower inference compared to unmodified beam search models. This issue is in large part solved by effective vectorized beam allocation (Hu et al., 2019). Another drawback of constrained decoding is a less fluent output, especially in morphologically rich languages, since we force the output to contain a phrase that may not be in agreement with the rest of the output. 2.3 Learned constraining One way of integrating constraints into NMT is to provide them alongside the input sentence and train the model to be biased towards utilizing them. This gives the user less direct control over the output translation and requires specially trained models. On the other hand, these approaches are simple to implement, do not incur inference slowdown, and make the translation more robust in case of wrongly chosen constraints. NMT models are often able to produce very fluent output (Popel et al., 2020a), making them capable to cope with inflections properly. Thus, using this capability may yield better results than constrained decoding with heuristics for inflections in inflected languages. Dinu et al. (2019) use input factors to annotate source sentences with desired translations and train the model to copy these translations into the output sequence. Chen et al. (2020) append constraints to the end of the source sentence. Their goal is to train the model to place constraints in the output translation without the need of a bilingual dictionary or a specified word alignment. Song et al. (2019) also propose a data augmentation approach that uses constraints along the source as input during the model training. Concurrently to our work, Bergmanis and Pinnis (2021) modify Dinu et al. (2019) approach by providing lemmatized word factors associated to random tokens in the source sentence. With the lemmatized factors, they force the model to learn the correct inflection of the word in the translation. The main difference between our work and most of the existing approaches is the use of lemmatized constraints to allow the model to correctly inflect them to agree with the output context. The 4021 concurrent work by Bergmanis and Pinnis (2021) presents a very similar idea. They also use lemmatized forms of the constraints and let the model itself to generate correct surface form. While their choice of languages (English to Latvian) and their experimental setup was slightly different, the overall conclusions of their work agree with ours. The main difference is the approach to integration of the constraints. Bergmanis and Pinnis (2021) use factors to directly annotate to the source tokens with lemmas of their desired translations. We experimented with this approach (see B.5), but in most of the experiments, we opted for a simpler integration method, by concatenating desired target lemmas to the source sentence. This simplifies preparation of the training data by removing the need for source to target word alignment and as we show, hurts the performance only by a very slight margin. 3 Proposed methods Building upon the described techniques, we focus on allowing the model to choose the correct word form. Our approaches are based on learned constraining, where the constraints are lemmatized during both training and test time. 3.1 Learned constraining In our approach, we append the target constraints as a suffix of the input sentences, same as Chen et al. (2020). We use <sep> token to separate constraints from the input sentence, and <c> token to separate constraints from each other. Inspired by Chen et al. (2020), we shift the positional embeddings by 1024 for the constraint tokens. However, while Chen et al. (2020) start each constraint on the same position, we shift the start of the constraint string and continue monotonically from there. We do not use any other techniques described in their work. The following example illustrates an input to our baseline constrained model, passing two constraints (“plánováno” and “obcích”) along with the source text. In this case, both constraints are in correct target surface forms, which are obtained from the reference translation. Without knowledge of the reference, it is necessary to solve the problem of agreement of the constraint with the rest of the translation, which is the main goal of our work. Source: Price increase is planned mainly in larger municipalities. <sep> plánováno <c> obcích Reference: Zvýšení cen je plánováno pˇredevším ve vˇetších obcích. We also experimented with the factored translation approach introduced by Dinu et al. (2019) as a second constraint integration method. In Appendix B, we present the description of the method and a comparison with appending the constraints as a suffix. 3.2 Preparing synthetic constraints To our current knowledge, there is no EnglishCzech dataset with provided constraints. Thus, we generate constraints from the existing parallel data. We consider two approaches to generate constraints for the training and test data. Training The simplest method of obtaining target-side constraints is sampling random token subsequences from the reference sentence. In our experiments, every token in the sentence can become a start of a constraint with probability of 0.3. An open constraint finishes on each subsequent token with probability of 0.85 and multiple constraints for a single sentence are permitted (without overlapping). We did not optimize these probabilities, further gains may be obtained by a search for better values. The constraint order is randomly permuted, since during the test time, order of constraints in the target is not known beforehand. The second approach makes use of either a bilingual dictionary or a terminology database. If a translation pair from the dictionary is found in the source and target sentences, its target side can serve as the constraint. By this method, we also obtain alignment of the source and target expressions, which is useful for the factored translation approach (see Appendix B.5). Test time Given an input sentence and no reference translation, we can synthesize constraints by searching for source expressions in a dictionary or a terminology database. Dictionaries generally map one expression to many target ones and we or the model have to decide which of them to use. Terminology databases are usually unambiguous and the target translation serves as the constraint. We experiment with terminology in Section 4.3. 4022 Lemmatization Our methods use lemmatized1 constraints. For the random target subsequence method, we lemmatize the selected words. For the dictionary search method, we lemmatize both the dictionary and training data and we search for matching expression pairs using the lemmas. During the actual training, we use the original, non-lemmatized sentence with lemmatized constraints. This scenario is more similar to real-life use cases, since target word form which should be produced is not known beforehand. With constraint lemmatization, the above example would be: Input: Price increase is planned mainly in larger municipalities. <sep> obec <c> plánovat 4 Experiments In this section, methods presented above are compared on various tasks and datasets. First, we use an oracle test set, which is created with previous knowledge of the reference. We use it to assess the ability of the models to integrate the constraints themselves without additional noise caused by problems of the real world. In the subsequent experiments, we present a more realistic scenario – we use official terminology for EU-related expressions to translate parts of Europarl corpus. Finally, we evaluate the approaches on translation of general, open-domain rare words using dictionary. 4.1 Data We train English-Czech NMT models for our experiments. Czech has a high degree of inflection with seven cases and three genders for nouns and adjectives. We train our models on CzEng 2.0 (Kocmi et al., 2020) using all authentic parallel sentences (61M), as well as back-translated Czech monolingual sentences (51M). Newstest-2019 (Barrault et al., 2019) is used as a validation set and newstest2020 (Barrault et al., 2020) as a test set. We break the text into subwords using SentencePiece (Kudo and Richardson, 2018) and lemmatize using UDPipe (Straka and Straková, 2017). BLEU scores are computed using SacreBLEU (Post, 2018).2 For experiments mentioning dictionaries, we extracted pairs of terms from English and Czech Wik1In Appendix B, we show that simple stemming heuristic performs at least as well as proper lemmatization in automated metrics described further. 2SacreBLEU signature: BLEU+case.mixed+lang.encs+numrefs.1+smooth.exp+test.wmt20+tok.13a+version.1.4.14 tionary3 and a large commercial dictionary. In appendix B.2 we show that using Wiktionary also improves performance upon baseline, but the commercial dictionary offers better coverage of the expressions and thus provides better overall results. For this reason, all the experimets shown further are based on the commercial dictionary data. We use the Czech government database for EU terminology4 to evaluate integration of domainspecific terminology through constraints. We select all Czech terms and their translation to English, which corresponds to 14203 expressions per language. Then, we search the Europarl5 corpus (Koehn, 2005) for sentence pairs containing English terms in the source side and lemmas of the Czech translation in a lemmatized version of the target side, ignoring trivial terms. Keeping at most the first ten sentence pairs containing specific source term, the final dataset consists of 6585 examples, covering 1433 terms. We remove these sentences from the training data, since Europarl is part of the CzEng corpus. 4.1.1 Model We use MarianNMT (Junczys-Dowmunt et al., 2018) to train Transformer-base models with standard parameters (Vaswani et al., 2017). Inspired by Popel et al. (2020b), we alternate between authentic and backtranslated data every 25 million training sentences, while using exponential smoothing of the parameters. Four NVIDIA V100 GPUs were used for the training and one training run (400500k steps) takes approximately 40 hours with this configuration. A large portion of the computation time can be saved by finetuning an existing NMT model on the proposed dataset. By finetuning the baseline model we reached the same performance after 30-50k steps. However, all the results provided in this paper are obtained by training from scratch. Since we integrate constraints in the target language into the source sequence, we share source and target vocabularies (and embeddings), consisting of 32000 subwords, to allow easier copying of the subwords from source to target sequence. 4.2 Oracle constraints To assess the ability of a model to produce the provided constraints in the output, we use newstest3www.wiktionary.org 4sap.vlada.cz/dul/zavaznet.nsf/ca? OpenView 5www.statmt.org/europarl/ 4023 Train const. Train form Test form BLEU Cvg BLEUL CvgL baseline 32.0 68.84 38.2 78.14 random 31.2 69.59 37.1 78.47 random surface surface 34.5 94.00 39.9 94.55 random surface lemma 27.1 61.31 36.8 94.26 random lemma lemma 33.3 82.37 39.7 93.61 dict surface surface 16.5 57.34 20.4 68.69 dict surface surface 37.7 93.46 42.2 93.23 dict surface lemma 30.6 64.11 39.6 91.55 dict lemma lemma 34.2 78.61 40.5 89.02 dict, skip half surface 31.7 68.88 38.2 78.06 dict, skip half surface surface 36.9 91.37 42.3 93.00 dict, skip half surface lemma 31.4 68.0 40.0 90.79 dict, skip half lemma lemma 33.1 75.36 39.3 85.30 Table 1: Results on newstest-2020 with oracle constraints. The first column shows the methods used for obtaining constraints for training. ‘random’ means sampling random subsequences of target tokens, ‘dict’ stands for terms matched by dictionary. In the ‘skip half’ variant, a half of the training examples is presented with no constraint. For test sets, only constraints from the dictionary are used, still chosen so that the reference sentence contains the requested words. The second and third column indicate if the appended constraints are lemmatized or not, at training and test time, respectively. 2020 test set with oracle constraints. These constraints are obtained via dictionary search on the test set as described above, i.e., the constraints are terms from a English-Czech dictionary, where both source and target sides are present in the sentence pair. Note that we know the reference beforehand, thus, this evaluation may not reflect improvement in translation in a real world setting. We only use it to measure the ability of constraint integration. We trained two sets of constrained models. The first one, baseline constrained models, use original target side forms of the constraint expressions. The second set consists of models trained using lemmatized forms of the constraints. Our goal with the lemmatized models was to harness the language modeling capacity of the model to generate a surface form of lemmatized constraint that agrees with the rest of the translation. Table 1 presents the results. We used two forms of the test set constraints – original reference forms and lemmatized constraints (column Test form). The lemmatized constraints are closer to real world scenario, where we do not know the output form of the constraint expression beforehand. As a sanity check, we compute standard BLEU and BLEU calculated on lemmatized hypothesis against lemmatized reference (BLEUL). More importantly, we assess target constraint coverage (Cvg and CvgL) on original and lemmatized test set by comparing the constraints in the output with the reference. Note that in theory, Cvg value should always be lower or equal to CvgL, since surface form coverage is equal to lemma coverage minus proportion of incorrectly generated surface forms. This is not always the case, since the lemmatizer takes the sentence context into consideration and lemmatized versions of stand-alone terms in the terminology database may not match lemmatized versions of the same terms inside a reference sentence. This causes a slight underestimation of CvgL. The Cvg and CvgL columns document that both methods of constraint synthesis for training (random target subsequences and dictionary terms) lead to models capable of producing more than 93% of the constraints when constraints are not lemmatized. Surface coverage of surface form trained models drops to 61–68% when using lemmatized form of the test set constraints, but lemma coverage is only slightly lower – this is expected, as these models are trained to reproduce exact form of the given constraints. The results of models trained on lemmatized constraints with lemmatized test constraints show that the surface form coverage increases compared to surface form trained models with lemmatized test constraints (rows lemma/lemma vs. surface/lemma). While the coverage is lower than when using surface form test set for the surface 4024 Train Test BLEU Cvg Baseline No constraints 37.9 75.02 All No constraints 19.1 61.40 Terms 37.3 91.73 Dict 43.3 84.14 Terms + Dict 44.0 93.75 Skip half No constraints 38.2 75.32 Terms 38.4 90.52 Dict 43.5 83.49 Terms + Dict 43.1 91.22 Table 2: Performance of models trained using surface forms of dictionary constraints on the same Europarl test set split. Train column documents whether all of the training sentences were accompanied by constraints, or we left 50% of them without constraints (Skip half). Term constraints come from a terminology database, Dict constraints are expressions from a general dictionary. Note that for applying Dict constraints at test time, we used test reference for dictionary target term disambiguation, which makes this combined approach not feasible in realistic conditions. All test set constraints are used in reference surface forms. form model, we show in Section 5 that this is mainly an artifact of reference-based evaluation and that the model inflects the constraints correctly. The model trained with constraints based on dictionary reaches the best performance on the oracle constraint test set, for which the constraints are generated in the same way. However, when constraints are not supplied, BLEU and coverage drops sharply (the row dict/surface/-). This may be caused by the fact that sentences containing expressions present in the dictionary are almost always accompanied by the constraint during the training. Therefore, the model is not presented with many examples where the translation appears without the corresponding constraint and generates constraint expression with much lower probability when this happens during the test time. We experimented with skipping half of the sentences during the constraint generation, leaving them without any constraints (“skip half” in the table). As shown in Table 1, this largely reduces the problem – without any test time constraints, the model reaches baseline results (the row dict, skip half/surface/-). However, when the constraints are supplied, the coverage is slightly lower than for a model trained with constraints for all the sentences (e.g. 91.4% instead of 93.5% for surface form models). Fine-tuning the ratio or choosing the sentences to leave without the constraints dynamically during the training might help to solve this problem. Train c. Test c. BLEU Cvg CvgL 38.2 69.90 84.37 SF 38.8 70.27 85.0 canon. 36.6 44.0 96.56 Ref SF 40.6 96.97 95.08 lemma 35.1 30.88 96.74 Lemma 38.6 69.87 84.05 canon. 38.9 77.1 95.44 Ref SF 39.1 81.44 94.15 lemma 38.9 77.22 95.55 Table 3: Results on whole Europarl test set. None of the BLEU scores for constrained models (except Ref SF) is significantly better than the best unconstrained score. 4.3 Terminology Integration Since the studied methods proved to work well with oracle surface form of constraints, we moved to a realistic use-case with the Europarl test set described in Section 4.1. We split the test set into two parts: • same contains examples where the form of the constraint in the reference is the same as in the terminology database (and as provided to the baseline constrained model), • diff contains examples where the form of the constraint in the target sentence is different from the database form. The target lemmas of the constraint should match in both cases. This split allows us to better assess the translation in inflected languages, since the problems we focus on are more pronounced in the diff test set. Table 2 shows that the model trained with dictionary constraints underperforms in terms of BLEU when only the constraints from terminology database are supplied (BLEU of 19.1). This is caused by the issue described earlier – during the training, the model does not encounter the words which are present in the dictionary enough times without the constraint. When the dictionary constraints are used alongside the terminology database constraints (rows denoted by “Terms + Dict”), the BLEU score increases. This approach requires either prior knowledge of the reference, or a mechanism for the target dictionary term disambiguation. To mitigate this issue, we skip half of the sentences when generating the constraints, i.e., half of the training corpus is seen without any constraints. This alleviates the problem to a large extent, see the “Skip half” results. 4025 Train c. Test c. BLEU Cvg CvgL 38.3 67.1 84.12 SF 38.8 67.14 84.68 canon. 35.0 15.20 96.20 Ref SF 40.8 96.32 93.92 lemma 34.3 15.38 96.41 Lemma 38.7 66.61 83.42 canon. 38.9 72.31 94.76 Ref SF 39.2 79.16 92.78 lemma 39.0 72.62 94.88 Table 4: Results on diff Europarl test set split, where we only consider cases where the constraint is provided in different form than in the reference, i.e. reference contains different form than the canonical one present in the terminology database. None of the BLEU scores for constrained models (except Ref SF) is significantly better than the best unconstrained score. We present the results on the whole test set in Table 3. The first and second columns show word form of the constraints during the training and test time, respectively. Canon. constraint is in its canonical, original form from the the terminology database. Ref SF rows show results with constraints in the same form as in the reference translation (this requires prior knowledge of the reference). First, let us focus on results of models trained with surface form constraints. Three trends in the results hint that generating the correct constraint form is challenging for the model, if the correct form is different from the one supplied in the input. First, the difference between surface form and lemma coverage (44% vs 96.6%) shows the model generates the correct constraint words, but in a form not matching the reference. Second, the difference is more pronounced in the diff split (Table 4), while in the same split (Table 5), surface form coverage is almost the same as the lemma coverage. This is because in the same split, target constraints are already in the canonical form, same as in the terminology database, so there is no need for further inflection. Third, using constraints in the same surface form as in the reference (Ref SF) improves the observed coverage compared to using the canonical form from the terminology database (e.g., 97% vs 44% on the whole test set, see Table 3). This “oracle” setting, using the reference to determine the correct surface form, shows the upper limits of the constraint integration approach, if the inflection issue is solved optimally. As stated earlier, we trained the models again using lemmatized versions of the constraints. When we supply lemmatized constraints to these modTrain c. Test c. BLEU Cvg CvgL 37.9 75.02 84.72 SF 38.8 75.94 85.50 canon. 39.9 97.69 97.03 lemma 36.6 59.56 97.38 Lemma 38.4 75.89 85.15 canon. 38.8 85.81 96.55 lemma 38.8 85.58 96.55 Table 5: Results on same Europarl test set split. In this subset, the constraints from terminology database are already in the same form as in reference, i.e. canon. is the same as Ref SF. BLEU score that is significantly better than the best BLEU without constraints is in bold (bootstrap resampling, p ≤0.05). els during the test time, the coverage rises from 44% (surface form trained model with canonical constraint forms) to 77%, but this is still far from the oracle 97%. This suggests that a large room for improvement exists, but as we show in Section 5, most of these discrepancies are caused by reference-based evaluation and are not real errors. In majority (92%) of the cases marked as not covered when using lemmatized model, the form of the constraint is different from the reference, but correct given the context, as the model translates the sentences differently (but correctly). 4.4 Comparison with constrained decoding Our work is based on training the NMT model to include provided constraints in the output translation. Another popular way of constraint integration is modifying the decoding process. We hypothesize that this approach will not be useful in our scenario, since the constraints are enforced in their surface forms, which is the issue we are trying to solve. To verify this, we evaluated lexically constrained decoding by Hu et al. (2019) as implemented in fairseq (Ott et al., 2019) on the Europarl test sets described in Section 4.3. Split Con. BLEU Cvg BLEUL CvgL Pos ρ Same no 36.4 69.3 42.8 79.7 0.95 Same yes 35.7 97.1 41.5 97.3 0.83 Diff no 36.4 63.1 43.1 81.3 0.95 Diff yes 30.6 26.0 39.3 94.7 0.80 Whole no 36.4 65.2 43.0 80.8 0.95 Whole yes 32.3 50.7 40.0 95.6 0.81 Table 6: Lexically constrained decoding The results in Table 6 show that while the constrained decoding indeed produces the target constraints in the output, they stay in the same form as in the terminology database. This is shown by the low surface form constraint coverage (column 4026 Constraint src BLEU % as ref % correct No constraint 21.6 35.4 64.6 Reference term 23.1 91.7 91.7 Random term 22.6 54.2 83.3 Table 7: Translation of sentences containing rare words. For source expressions with multiple possible translations according to the dictionary, we compare choosing a translation variant randomly (Random term) against choosing the same translation variant as in the reference. All constraints are lemmatized. Column % as ref shows the percentage of examples with the constraint translated with the same term as in the reference. Column % correct shows human evaluation of rare word translation. Cvg) for the diff and whole dataset splits, while for the same split, where the constraints are in the same form in the translation as in the terminology database, the coverage is high. On lemma level (CvgL), coverage on all splits remains high, again showing that the system produces exactly the surface form provided, instead of correct target sentence form. Note that the results are not directly comparable with the results in previous subsection, since here we use only a part of the training data (first 25M sentence pairs from parallel part of CzEng) for the preliminary experiments. We also observed that the Pearson correlation of constraint placement in respect to reference translation (see Appendix A.1 for details) is lower (0.81) when using constrained decoding than when using the training approach as in the main experiments (0.94). 4.5 Semi-parametric rare words translation We define rare words as terms from a dictionary that occur in the source side of the training corpus at most 50 times. We create a subset of our general dictionary by only using expression pairs with rare words on source side. We search WMT 2007-2020 English-Czech news test sets (Barrault et al., 2020) for sentence pairs containing term pairs from this rare word dictionary, resulting in 48 examples. A dictionary generally provides 1-to-many mappings of source terms to a target language, so the correct target expression needs to be disambiguated. Table 7 presents results with no constraints, with constraints where the lemmatized target constraint is chosen based on the lemmatized reference, and with constraints where the target expression is chosen randomly from all the possible translations. We used a model trained on lemmatized random target token subsequences for the translation. On average, each rare word in the test set has 3.3 possible dictionary translations. Aside from BLEU score, we show the percentage of rare words translated correctly, meaning that either they are the same expression as in the reference, or that they are synonymous expressions that are correct in the given context. This is different from the terminology use case, since we do not strictly enforce single possible translation. The results show that even with the random choice of the dictionary constraint translation, our model improves the translation of rare words. 5 Manual analysis In this section, we analyse examples marked as errors by automatic evaluation. In Appendix A.1, we analyse the position of constraints in translation outputs, showing that they are placed correctly. In Appendix A.2, we look closely at the constrained translation of an out-of-domain document. 5.1 Error analysis We manually analysed outputs marked as not having the desired constraint in the reference surface form by the automatic coverage evaluation introduced in the previous section. Table 9 presents the results. We compare three models. First, the baseline without any constraints (column B). Second, the best model trained with non-lemmatized constraints (SF), and, finally, the best model trained on lemmatized constraints (column L). The baseline model outputs have constraint surface form coverage of 69.9% on the whole Europarl test set, which results in 1982 out of 6585 examples being marked as different from the reference by the automatic evaluation. The SF model reached 44% coverage (4346 differences). The lemmatized model agreed with the reference in 77.1% (1508 differences). For each model, we randomly sample 100 supposedly erroneous translations to be analysed. The first row of Table 9 shows the number of examples with constraints incorrectly inflected in the context of the generated output. Rows 2 and 3 show cases where the constraint form agrees with rest of the translation: Correct in correct context (CCC) indicates that the target sentence is a valid translation, whereas Correct in incorrect context (CIC) indicates that the constraint was inflected correctly given its context but as a whole, the translation is wrong. Thus, CCC cases are not in fact errors, but were wrongly classified as such by the automatic 4027 Source Canon Ref Translation Error They are seeking to weaken the Commission’s proposal to benefit the industry. návrh návrhu Snaží se oslabit návrh Komise ve prospˇech pr˚umyslu. CCC Snaží se oslabit návrh Komise na prospˇech pr˚umyslu. CIC Snaží se oslabit návrhu Komise ve prospˇech pr˚umyslu. Inflection Table 8: Example of three error types given canonical and reference target form constraints. Error type B SF L Incorrect inflection 2 46 0 Correct in correct context 65 44 92 Correct in incorrect context 0 3 2 Different correct word choice 28 2 4 Different incorrect word choice 0 0 1 Invalid translation 5 5 1 Table 9: Analysis of 100 outputs marked as errors by the automatic evaluation, which means that either they do not contain the constraint or they contain it in a different surface form compared to the reference. We analysed three models – baseline (B), a model trained with surface form constraints using canonical forms of the constraints at test time (SF), and a model trained with lemmatized constraints using lemmatized terminology entries at test time (L). evaluation, based on a direct comparison with the reference. The cases where the model ignores the constraint and generates a different word are in the categories Different correct/incorrect word choice (fourth and fifth rows), based on whether the generated word is a plausible translation of the source constraint. Examples where the translation generally goes wrong and the issue does not fit into the previous categories are under Invalid translation. Our analysis shows that for the lemmatized model (L), the vast majority of the examples classified as errors are actually correctly translated and contain the requested constraint in the correct surface form. The presumed error is an artifact of the reference-based evaluation. Only 8% of these examples are real errors, compared to 66% for the surface form model. In Table 8, we show three examples of errors found by the automatic evaluation. Given the canonical and reference source form of a constraint (návrh and návrhu, respectively, meaning “proposal”), some errors may arise in the translation. In the first row, although different from the reference source form, the constraint is correctly inflected given the context generated and in a correct translation, which configures a “correct in correct context” error (CCC). Similarly, in the second row, the same constraint with the same source form is correctly inflected given the context but in a wrong translation, which describes a “correct in incorrect context” (CIC) error. Finally, the third translation has a wrong inflection given the context generated (Inflection error). 6 Conclusion We described the problem of word inflection in lexically constrained machine translation. Our solution capitalizes on the ability of NMT models to generate correct word forms in the output translation. We train a Transformer model using lemmatized constraints supplied alongside the input sentences, and correct surface forms of the constraints in the reference. This training leads to a model producing the constraints in the output with high coverage, correct placement, and in a correct surface form. We compare several methods of obtaining constraints and integrating them into the input. In the realistic use case of terminology integration, we evaluated our methods and show that without lemmatizing the training constraints, the chosen approach of integrating constraints into NMT does not work well for Czech. We effectively solve the issue of inflection errors by lemmatizing constraints, taking advantage of the Transformer’s language modelling capacity with no additional inference costs. This has been proven by both automatic and manual evaluation. We show our method is also effective in translating general domain rare words using a bilingual dictionary and we plan future work in solving the problem of choosing correct translation term from number of variants. Acknowledgements Our work is supported by the Bergamot project (European Union’s Horizon 2020 research and innovation programme under grant agreement No 825303) aiming for fast and private user-side browser translation, GA ˇCR NEUREM3 grant (Neural Representations in Multi-modal and Multi-lingual Modelling, 19-26934X (RIV: GX19-26934X)) and by SVV 260 453 grant. We also want to thank Michal Novák for his useful feedback and discussions. 4028 References Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2017. Guided open vocabulary image captioning with constrained beam search. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 936–945, Copenhagen, Denmark. Association for Computational Linguistics. Loïc Barrault, Magdalena Biesialska, Ondˇrej Bojar, Marta R. Costa-jussà, Christian Federmann, Yvette Graham, Roman Grundkiewicz, Barry Haddow, Matthias Huck, Eric Joanis, Tom Kocmi, Philipp Koehn, Chi-kiu Lo, Nikola Ljubeši´c, Christof Monz, Makoto Morishita, Masaaki Nagata, Toshiaki Nakazawa, Santanu Pal, Matt Post, and Marcos Zampieri. 2020. Findings of the 2020 conference on machine translation (WMT20). In Proceedings of the Fifth Conference on Machine Translation, pages 1–55, Online. Association for Computational Linguistics. Loïc Barrault, Ondˇrej Bojar, Marta R. Costa-jussà, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Müller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 1–61, Florence, Italy. Association for Computational Linguistics. Toms Bergmanis and M¯arcis Pinnis. 2021. Facilitating terminology translation with target lemma annotations. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3105–3111, Online. Association for Computational Linguistics. Rajen Chatterjee, Matteo Negri, Marco Turchi, Marcello Federico, Lucia Specia, and Frédéric Blain. 2017. Guiding neural machine translation decoding with external knowledge. In Proceedings of the Second Conference on Machine Translation, pages 157– 168, Copenhagen, Denmark. Association for Computational Linguistics. Guanhua Chen, Yun Chen, Yong Wang, and Victor O.K. Li. 2020. Lexical-constraint-aware neural machine translation via data augmentation. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20, pages 3587–3593. International Joint Conferences on Artificial Intelligence Organization. Main track. Josep Crego, Jungi Kim, Guillaume Klein, Anabel Rebollo, Kathy Yang, Jean Senellart, Egor Akhanov, Patrice Brunelle, Aurelien Coquard, Yongchao Deng, Satoshi Enoue, Chiyo Geiss, Joshua Johanson, Ardas Khalsa, Raoum Khiari, Byeongil Ko, Catherine Kobus, Jean Lorieux, Leidiana Martins, Dang-Chuan Nguyen, Alexandra Priori, Thomas Riccardi, Natalia Segal, Christophe Servan, Cyril Tiquet, Bo Wang, Jin Yang, Dakun Zhang, Jing Zhou, and Peter Zoldan. 2016. Systran’s pure neural machine translation systems. Georgiana Dinu, Prashant Mathur, Marcello Federico, and Yaser Al-Onaizan. 2019. Training neural machine translation to apply terminology constraints. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3063–3068, Florence, Italy. Association for Computational Linguistics. Greg Hanneman and Georgiana Dinu. 2020. How should markup tags be translated? In Proceedings of the Fifth Conference on Machine Translation, pages 1160–1173, Online. Association for Computational Linguistics. Eva Hasler, Adrià de Gispert, Gonzalo Iglesias, and Bill Byrne. 2018. Neural machine translation decoding with terminology constraints. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 506–512, New Orleans, Louisiana. Association for Computational Linguistics. Chris Hokamp and Qun Liu. 2017. Lexically constrained decoding for sequence generation using grid beam search. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1535–1546, Vancouver, Canada. Association for Computational Linguistics. J. Edward Hu, Huda Khayrallah, Ryan Culkin, Patrick Xia, Tongfei Chen, Matt Post, and Benjamin Van Durme. 2019. Improved lexically constrained decoding for translation and monolingual rewriting. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 839–850, Minneapolis, Minnesota. Association for Computational Linguistics. Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu Hoang, Kenneth Heafield, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, André F. T. Martins, and Alexandra Birch. 2018. Marian: Fast neural machine translation in C++. In Proceedings of ACL 2018, System Demonstrations, pages 116– 121, Melbourne, Australia. Association for Computational Linguistics. Tom Kocmi, Martin Popel, and Ondrej Bojar. 2020. Announcing czeng 2.0 parallel corpus with over 2 gigawords. arXiv preprint arXiv:2007.03006. Philipp Koehn. 2005. Europarl: A parallel corpus for statistical machine translation. In MT summit, volume 5, pages 79–86. Citeseer. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, 4029 Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondˇrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague, Czech Republic. Association for Computational Linguistics. Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, Brussels, Belgium. Association for Computational Linguistics. Thang Luong, Ilya Sutskever, Quoc Le, Oriol Vinyals, and Wojciech Zaremba. 2015. Addressing the rare word problem in neural machine translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 11–19, Beijing, China. Association for Computational Linguistics. Franz Josef Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417–449. Hideo Okuma, Hirofumi Yamamoto, and Eiichiro Sumita. 2008. Introducing a translation dictionary into phrase-based smt. IEICE - Trans. Inf. Syst., E91-D(7):2051–2057. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 48–53, Minneapolis, Minnesota. Association for Computational Linguistics. Martin Popel, Marketa Tomkova, Jakub Tomek, Łukasz Kaiser, Jakob Uszkoreit, Ondˇrej Bojar, and Zdenˇek Žabokrtský. 2020a. Transforming machine translation: a deep learning system reaches news translation quality comparable to human professionals. Nature Communications, 11(4381):1–15. Martin Popel, Marketa Tomkova, Jakub Tomek, Łukasz Kaiser, Jakob Uszkoreit, Ondˇrej Bojar, and Zdenˇek Žabokrtský. 2020b. Transforming machine translation: a deep learning system reaches news translation quality comparable to human professionals. Nature Communications, 11(4381):1–15. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186– 191, Brussels, Belgium. Association for Computational Linguistics. Matt Post and David Vilar. 2018. Fast lexically constrained decoding with dynamic beam allocation for neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1314–1324, New Orleans, Louisiana. Association for Computational Linguistics. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715– 1725, Berlin, Germany. Association for Computational Linguistics. Kai Song, Yue Zhang, Heng Yu, Weihua Luo, Kun Wang, and Min Zhang. 2019. Code-switching for enhancing NMT with pre-specified translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 449–459, Minneapolis, Minnesota. Association for Computational Linguistics. Milan Straka and Jana Straková. 2017. Tokenizing, POS tagging, lemmatizing and parsing UD 2.0 with UDPipe. In Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies, pages 88–99, Vancouver, Canada. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, undefinedukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6000–6010, Red Hook, NY, USA. Curran Associates Inc. Vilém Zouhar, Tereza Vojtˇechová, and Ondˇrej Bojar. 2020. WMT20 document-level markable error exploration. In Proceedings of the Fifth Conference on Machine Translation, pages 371–380, Online. Association for Computational Linguistics. 4030 A Further analysis A.1 Constraint placement Increased BLEU and constraint coverage show that the evaluated methods are able to generate correct constraint string in the output. However, these metrics do not tell much about placement of constraints. If all the constraints are appended at the end of the output, we would get perfect coverage and, in some cases, possible increase in BLEU score – but this is not a desired behavior of the system. To evaluate the correctness of constraint placement, we record starting indices of each satisfied constraint in both MT output and reference, and we compute Pearson’s correlation between these two variables. As a sanity check of the correlation measure, we also modify the output of the constrained system and move the constraints it correctly produced to random positions. Both BLEU and the Pearson correlation drop considerably, see the line marked with “*” in Table 10. The second row shows the case of supplying constraints as a suffix for the baseline model, which was not trained to utilize them. Coverage of the constraints has increased – but, as expected, the model only generates some of the constraints at the end of the translation. Lower correlation with the reference placement shows that the placement is incorrect. In the fourth row, we randomly change positions of the constraints as described above. Again, the correlation decreases. These experiments indicate that the evaluated systems can generate constraints at correct positions in the output. A.2 Lease agreement case study Our method proved to work well on the Europarl terminology test set. Since Europarl is included in the training data (only the actual test sentences are filtered out), we used an out-of-domain test document to assess the results using unknown terminology. For this purpose, we used a sublease agreement translated from Czech into English, which is included in WMT20 Markables test suite6 (Zouhar et al., 2020). There are minor translation errors in the reference of the test set version used at WMT20, which we fixed.7 The difficulty of translating this agreement accurately lies in the translation of some of the legal terms, e.g. tenant, lessee, lease or sub6https://github.com/ELITR/ wmt20-elitr-testsuite/ 7We will provide the link to the fixed test set in the cameraready version. Model Constr. BLEU Cvg Pos ρ Baseline 30.9 70.70 0.9362 Baseline suffix 27.6 76.93 0.8407 Suffix suffix 35.3 95.05 0.9382 Suffix * suffix 16.8 95.05 0.3486 Table 10: Correlation between start character indices of the satisfied constraints in system’s output and reference. The table shows that the evaluated methods place constraints at the correct positions in the output. When we move the constraints (marked with an asterisk), the correlation between the positions drops. lease. These terms are often used interchangeably in common language. In this case, tenant (nájemník in Czech) is a person who has an apartment in a lease from the owner and lessee (podnájemník) is a person that is using the apartment based on the agreement with the tenant. We manually created a database of 11 legal terms and their translations used in the document. Note that we know that the sublease agreement is between two women, so we used feminine forms of the translations for lessee and tenant. Table 11 compares translations produced by our systems against existing approaches on one problematic sentence. We used following term pairs as our constraints for this sentence: Source Target Term of the Lease Doba podnájmu lessee podnájemkynˇe tenant nájemkynˇe apartment in question pˇredmˇetný byt Sublease agreement Smlouva o podnájmu bytu Our three systems are: (1) the model based on suffixed surface form constraints, (2) the same model using lemmatized constraints, and (3) the two-factored model using surface form factors as described in Appendix B.5. They are compared with the outputs of CUBBITT8, the state-of-the-art English-Czech system by Popel et al. (2020b), and two commercial translation engines (Google Translate9 and Lingea Translator10). Constraint terms typeset in green are translated correctly according to the terminology, orange terms are very similar 8https://lindat.mff.cuni.cz/services/ translation/ 9https://translate.google.com/ 10https://translator.lingea.com/ 4031 Model Translation Source In Art. III of the Sublease agreement1, entitled “ Term of the Lease2 ,” the tenant3, and the lessee4 agreed that the apartment in question5 would be rented to the lessee6 for a fixed period from 13th May 2016 to 31st December 2018. Google Translate V ˇcl. III smlouvy o podnájmu1 s názvem „Doba nájmu“2 se nájemce3 a nájemce4 dohodli, že pˇredmˇetný byt5 bude nájemci6 pronajat na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018. Lingea Translator V ˇcl. III podnájemní smlouvy1, nadepsané „Lh˚uta nájmu2 ,„ se nájemce3 a nájemce4 dohodli, že dotˇcený byt5 bude nájemci6 pronajat na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018. CUBBITT V umˇení. III podnájemní smlouvy1 nazvané „ Podmínky pronájmu2 “ se nájemce3 a nájemce4 dohodli, že pˇredmˇetný byt5 bude nájemci6 pronajímán na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018 Suffix surface form V ˇcl. III podnájemní smlouva o podnájmu1, nadepsaném „ Lh˚uta nájmu2“, se nájemkynˇe3 a podnájemkynˇe4 dohodly, že pˇredmˇetný byt5 bude nájemci6 pronajat na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018. Suffix lemmatized V ˇclánku III smlouvy o podnájmu1, nazvaném „doba podnájmu2 ,“ se nájemkynˇe3 a podnájemkynˇe dohodli4, že pˇredmˇetný byt5 bude nájemci6 pronajat na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018. Factored SF V ˇcl. III smlouvy o podnájmu bytu1, nadepsaný „podnájmu2 ,“ nájemkynˇe3 a podnájemkynˇe4souhlasily s tím, že pˇredmˇetný byt5 bude pronajat podnájemkyni6 na dobu urˇcitou od 13. kvˇetna 2016 do 31. prosince 2018. Ref V ˇcl. III Smlouvy o podnájmu1 bytu, nazvaném „Doba podnájmu2“, se nájemkynˇe3 a podnájemkynˇe4 dohodly, že pˇredmˇetný byt5 bude podnájemkyni6 pˇrenechán k užívání na dobu urˇcitou od 13. 5. 2016 do 31. 12. 2018. Table 11: Translations of one of the difficult sentences from WMT20 ELITR test set. in meaning to the terminology database translation, and red ones are clear translation errors. We note that especially the word podnájemkynˇe (lessee in feminine form) poses some difficulties for the model to produce, since it does not appear in the training data. Its masculine forms, podnájemce or podnájemník appear 182 times in different inflections. Another difficulty is added by the fact that the word lessee appears two times in the sentence. All of the systems produce the correct constrained translation at most for the first occurrence, with exception of factored model, which is supplied explicit alignment between source and target part of the constraints. We hypothesize that other models consider the constraint as covered after it is generated for the first time. Overall, the constrained models provide more accurate translations compared to the unconstrained SOTA models, effectively integrating the constraints even in a difficult out-of-domain example. B Other related experiments We present experiments that influenced our architectural choices in the paper, but are not discussed in the main text. Note that the results are not directly comparable, since a slightly different preprocessing was used. B.1 Constraints as prefix or suffix In Table 12, we compare passing the constraints as a prefix of the source sentence, as a suffix and as a suffix with all positional embeddings of the constraint part starting with 1024. Using prefix resulted in the best coverage, but, as visible in column Pos ρ, correlation of constraint positions is lower than for other models. Upon manual inspection, we saw that the constraints were in some cases generated also as a prefix of the target sentence. For the main experiments, we decided to use suffix integration with positional embedding shifting, since it provided slightly better coverage than the basic suffix variant. B.2 Wiktionary vs. proprietary dictionary Dictionary is necessary for one of the training methods we explore. For our main results, we used a proprietary dictionary, which provides better coverage of the possible term pairs, but harms the reproducibility of this part of our experiments. Thus, we also evaluated our method using Wiktionary11 to 11https://www.wiktionary.org/ 4032 Train const. Integration BLEU Cvg BLEUL CvgL Pos ρ baseline 30.9 70.51 37.0 77.46 0.9322 random prefix 34.7 96.15 39.4 95.51 0.8468 random suffix 34.9 93.02 40.0 92.99 0.9336 random, shift suffix+shift 34.9 93.12 40.1 93.25 0.9349 Table 12: Comparison of integrating the constraints as a prefix, suffix and suffix with positional embedding shifting. Note the results are not directly comparable to main paper results, as the train and test set preprocessing is different. Words Price increase is planned plánováno mainly in larger municipalities obcích . Factor 0 0 0 SRC TGT 0 0 0 SRC TGT 0 Table 13: Example of the constrained translation process using factors. Train c. Test c. BLEU Cvg CvgL 38.2 69.90 84.37 SF 38.8 70.27 85.0 canon. 36.6 44.0 96.56 Ref SF 40.6 96.97 95.08 lemma 35.1 30.88 96.74 Lemma 38.6 69.87 84.05 canon. 38.9 77.1 95.44 Ref SF 39.1 81.44 94.15 lemma 38.9 77.22 95.55 Mixed 37.7 69.37 83.51 canon. 37.5 69.68 95.08 Ref SF 39 91.65 94.72 lemma 38 76.57 95.25 Table 14: Performance of the model mixing half of the training examples with surface form constraints and half of them lemmatized on the whole Europarl test set. Compared with the models, where either lemmatization was never applied on constraints during training (SF), or it was applied on all data examples (Lemma). obtain constraints in the same way as described in the main experiments section (see Section 4). We present the results in Table 15. Looking up term pairs from the commercial dictionary in the test set, we found 7201 term pairs that were used as a constraint. On the other hand, we found only 2529 term pairs using Wiktionary. We see that both models are able to incorporate constraints from the dictionary used during the training with similar success – about 94% of the constraints are covered. However, Large dictionary provides better BLEU scores, since more constraint pairs are found overall in the test set. B.3 Mixed lemma and surface form training As we noticed in Section 4, lemmatized models have lower surface form coverage than nonlemmatized models when supplied with constraints in the reference surface form. As we show in our manual analysis, this is mostly an issue of Training dict. Test dict. BLEU Cvg Wiki 29.2 79.5 Large 29.2 69.2 Wiki Wiki 30.1 93.7 Wiki Large 29.6 81.8 Large Wiki 24.6 91.7 Large Large 34.3 94.3 Table 15: Comparison between using large, commercial dictionary (Large) as opposed to Wiktionary (Wiki) to obtain both training and test constrains. The results are computed on the oracle newstest-2020 test set, see 4.2 for details. automated evaluation based on comparison with reference, as the constraints are produced in correct form given the context of the output sentence produced by the model. Nevertheless, we experimented with a way to improve results of this automatic evaluation. We trained another batch of models with 50% of the constraints lemmatized and 50% left in the surface forms. Table 14 shows that this type of training improves integration of reference surface form constraints over the training where all constraints are lemmatized, while performance on lemmatized constraints does not decrease by a large margin. B.4 Stemming In Table 17, we compare stemming12 and lemmatization as the contraint preprocessing methods. The models in the table are trained with suffix constraints. The results are very similar, with stemming obtaining better results in terms of surface form coverage whereas lemmatization is better in lemma coverage. Since in Section 5 we have shown that the difference between surface and lemma coverage for lemmatized model is caused 12https://research.variancia.com/czech_ stemmer/ 4033 Train const. Integration Const. form BLEU Cvg BLEUL CvgeL baseline 30.9 70.70 37.1 77.73 random suffix surface 35.3 95.05 40.4 94.67 dict suffix surface 37.7 93.46 42.2 93.23 dict factors surface 37.5 95.72 42.0 95.11 Table 16: Comparison of constraint integration methods on the oracle test set. All the models were trained on non-lemmatized, surface form constraints. Train prep. Test prep. BLEU Cvg BLEUL CvgL Pos ρ baseline no constraints 32 68.84 38.2 78.14 0.9404 lemma no constraints 31.8 69.76 37.9 79.01 0.9367 lemma lemma 33.3 82.15 39.6 93.42 0.9341 stemming no constraints 31.3 69.51 37.4 78.48 0.9338 stemming stemming 33.2 84.10 39.5 92.86 0.9235 Table 17: Comparison of stemming and lematization as a preprocessing for training and test constraints. by the automatic reference-based evaluation and not by real errors in the translation, we opted for lemmatization in the paper. B.5 Constraint integration using factors We also present preliminary experiments with the factored model for constraining inspired by Dinu et al. (2019). We use a two-factor model, where the first factor comprises of the input sequence of words. For each source constraint in the input sequence, the translation tokens are inserted immediately after. In the second factor, one of the following three label values is assigned to a corresponding input token: • 0: ordinary source token without constraint • SRC: source side of a constraint • TGT: translation of the constraint For instance, consider the example in Table 13. Each word in the first factor has an associated label in the second factor according to its role in the translation. The words plánováno and obcích are Czech constraints that must appear in the translation of the English sentence. As a part of the target constraint, both words are labeled with the value TGT in the second factor. The words planned and municipalities are English words representing the source part of the constraints, thus receiving the value SRC. Words that are not constrained are labeled by 0 in the second factor. The values of the second factor are copied over all subwords of the constraint sequence. Embeddings of the values in both factors have the same dimensionality demb and they are summed to obtain a complete embedding, which is used by the model. For example, function Esub produces an embedding of an input subword and function Ef produces an embedding of its label. Final embedding of the word planned in the above example is computed by the following formula: E(plannedSRC) = Esub(planned) + Ef(SRC) Table 16 shows the comparison with the other integration methods on the oracle test set, similar to Section 4.2. We see factors provide the best coverage of the constraints. The factored approach makes use of alignment between source and target. This additional information probably helps with generating the correct constraints, but also complicates the preprocessing. Since the differences are only minor and the goal of our paper is not to reach state-of-the-art results in constrained translation, we opted for the suffix-based approaches for simplicity. Nevertheless, we note that factored approach is promising and we plan further research in this direction.
|
2021
|
311
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4034–4045 August 1–6, 2021. ©2021 Association for Computational Linguistics 4034 Handling Extreme Class Imbalance in Technical Logbook Datasets Farhad Akhbardeh, Cecilia Ovesdotter Alm, Marcos Zampieri, Travis Desell Rochester Institute of Technology Rochester, NY, USA {fa3019, coagla, mazgla, tjdvse}@rit.edu Abstract Technical logbooks are a challenging and under-explored text type in automated event identification. These texts are typically short and written in non-standard yet technical language, posing challenges to off-the-shelf NLP pipelines. The granularity of issue types described in these datasets additionally leads to class imbalance, making it challenging for models to accurately predict which issue each logbook entry describes. In this paper we focus on the problem of technical issue classification by considering logbook datasets from the automotive, aviation, and facilities maintenance domains. We adapt a feedback strategy from computer vision for handling extreme class imbalance, which resamples the training data based on its error in the prediction process. Our experiments show that with statistical significance this feedback strategy provides the best results for four different neural network models trained across a suite of seven different technical logbook datasets from distinct technical domains. The feedback strategy is also generic and could be applied to any learning problem with substantial class imbalances. 1 Introduction Predictive maintenance techniques are applied to engineering systems to estimate when maintenance should be performed to reduce costs and improve operational efficiency (Carvalho et al., 2019), as well as mitigate risk and increase safety. Maintenance records are an important source of information for predictive maintenance (McArthur et al., 2018). These records are often stored in the form of technical logbooks in which each entry contains fields that identify and describe a maintenance issue (Akhbardeh et al., 2020a). Being able to classify these technical events is an important step in the development of predictive maintenance systems. In most technical logbooks, issues are manually labeled by domain experts (e.g., mechanics) in free text fields. This text can then be used to classify or cluster events by semantic similarity. Classifying events in technical logbooks is a challenging problem for the NLP community for several reasons: (a) the technical logbooks are written by various domain experts and contain short text entries with nonstandard language including domain-specific abbreviated words (see Table 1 for examples), which makes them distinct from other short non-standard text corpora (e.g., social media); (b) off-the-shelf NLP tools struggle to perform well on this type of data as they tend to be trained on standard contemporary corpora such as newspaper texts; (c) outside of the clinical and biomedical sciences, there is a lack of domain-specific, expert-based datasets for studying expert-based event classification, and in particular few resources are available for technical problem domains; and (d) technical logbooks tend to be characterized by a large number of event classes that are highly imbalanced. Original Entry Pre-processed Entry fwd eng baff seeal needs resecured. forward engine baffle seal needs resecured. r/h eng #3 intake gsk leaking. right engine number 3 intake gasket leaking. bird struck on p/w at twy. bird rmvd. bird struck on pilot window at taxiway. bird removed location rptd as nm from rwy aprch end. location reported as new mexico from runway approach end. Table 1: Original and text-normalized example data instances illustrating that domain-specific terms (baffle), abbreviations (gsk - gasket, eng - engine), and misspellings (seeal - seal) are abundant in logbook data. We address the aforementioned challenges with a special focus on exploring strategies to address class imbalance. There is wide variation in the number of instances among the technical event classes examined in this work, as shown in Figure 1 and Ta4035 Figure 1: Number of instances in 39 unbalanced classes of the aviation maintenance (Avi-Main) dataset. ble 3. This extreme class imbalance is an obstacle when processing logbooks as it causes most learning algorithms to become biased and mainly predict the large classes (Kim et al., 2019). To overcome this issue, we introduce a feedback loop strategy, which is a repurposing of a method used to address extreme class imbalance in computer vision (Bowley et al., 2019), and examine it for classification of textual technical event descriptions. This technique is applied in the training of a suite of common classification models on seven predictive maintenance datasets representing the aviation, automotive, and facility maintenance domains. This paper addresses these research questions: RQ1: To which extent does the class granularity and class imbalance present in technical logbooks impact technical event classification performance, and can a feedback loop for training data selection effectively address this issue? RQ2: Which classification models are better suited to classify technical events for predictive maintenance across logbook datasets representing different technical domains? The main contributions of this work include: 1. Experimental results showing strong performance of the feedback loop in addressing the class imbalance problem in technical event classification across all datasets and models; 2. A thorough empirical evaluation of the performance of the technical event classifier considering multiple models and seven logbook datasets from three different domains. 2 Related Work Most expert-domain datasets containing events have focused on healthcare. For instance, Altuncu et al. (2019) analyzed patient incidents in unstructured electronic health records provided by the U.K. National Health Service. They evaluated a deep artificial neural network model on the expertannotated textual dataset of a safety incident to identify similar events that occurred. Del´eger et al. (2010) proposed a method to deal with unstructured clinical records, using rule-based techniques to extract names of medicines and related information such as prescribed dosage. Savova et al. (2010) considered free-text electronic medical records for information extraction purposes and developed a system to obtain clinical domain knowledge. Patrick and Li (2009) proposed the cascade methods of extracting the medication records such as treatment duration or reason, obtained from patient’s historical records. Their approach for event extraction includes text normalization, tokenization, and context identification. A system using multiple features outperformed a baseline method using a bag of words model. Yetisgen-Yildiz et al. (2013) proposed the lung disease phenotypes identification method to prevent the use of a handoperated identification strategy. They employed NLP pipelines including text pre-processing and further text classification on the textual reports to identify the patients with a positive diagnosis for the disease. Based on the outcome, they achieve 4036 Tech. Event or Issue Label Example Instance of Technical Logbook Entry Abbr., Misspelling, Terminology SUBSTANTIAL DAMAGE (1) AFT ON TAXI, WING STRUECK FUEL TRUCK, CHANDLER, AZ AFT, WING, STRUECK, FUEL BAFFLE DAMAGE (2) R/H FWD UPPER BAFF SEAL NEEDS TO BE RESECURED R/H, FWD, BAFL MINOR DAMAGE (1) SAW SML FLOCK FLYING UPON LDG FLARE, ACROSS RWY SML, LDG, RWY UNKNOWN (1) NO DMG. BIRD REMAINS ON F/O WINDSCREEN DMG, F/O, WINDSCREEN PM SERVICE (3) PM SERVICES CHECK TIRES FOR LEAKS CHECK PLOW BATT PM,TIRES, PLOW, BATT DRIVING ISSUE (4) FAILURE TO YIELD RIGHT, OVE CORRECTING OVER STEERING OVE, STEERING STOP SIGN RUNNING (4) MOTORISTS REGULARLY ILLEGAL U-TURNS IN R/HOUR U-TURNS, R/HOUR BUILDING PM (5) THE A/C UNIT IN THE KITCHEN ON 3TH FLOOR DMG/LEAK A/C, DMG ENG NEED REPAIR (3) CHANGE OIL & FILTER: L/H ENG, CHECK COMP & PLUGS OIL, ENG, L/H, COMP, PLUGS PREVENTIVE MAINT (5) RESET BOILER #2 TMER, CHECKED BLDG. THROUGHOUT BOILER, BLDG Table 2: Example instances of technical logbook entries spanning the aviation accident (1), aviation maintenance (2), automotive maintenance (3), automotive safety (4), and facility maintenance (5). Each instance shows how domain-specific terminology, abbreviations (Abbr.), and misspelled words (in bold font) are used by the domain expert, and also illustrates some of the event types covered. More details are provided in Section 3. notable performance by using the n-gram features with the Maximum Entropy (MaxEnt) classifier. There is also relevant research on event classification in social media. For example, Ritter et al. (2012) proposed an open-source event extraction and supervised tagger for noisy microblogs. Cherry and Guo (2015) applied word embedding-based modeling for information extraction on news-wire and tweets, comparing named entity taggers to improve their method. Hammar et al. (2018) performed experimental work on Instagram text using weakly supervised text classification to extracted clothing brand based on user descriptions in posts. The problem of class imbalance has been studied in recent years for numerous natural language processing tasks. Tayyar Madabushi et al. (2019) studied automatic propaganda event detection from a news dataset using a pre-trained BERT model. They recognized that the BERT model had issues in generalizing. To overcome this issue, they proposed a cost-weighting method. Al-Azani and ElAlfy (2017) analyzed polarity measurement in imbalanced tweet datasets utilizing features learned with word embeddings. Li and Nenkova (2014) studied the class imbalance problem in the task of discourse relation identification by comparing the accuracy of multiple classifiers. They showed that utilizing a unified method and further downsampling the negative instances can significantly enhance the performance of the prediction model on unbalanced binary and multi-classes. Dealing with unbalance classes is also studied well in the sentiment classification task. Li et al. (2012) introduced an active learning method that overcomes the problem of data class unbalance by choosing the significant sample of minority class for manual annotation and majority class for automatic annotation to lower the amount of human annotation required. Furthermore, Damaschk et al. (2019) examined techniques to overcome the problem of dealing with high-class imbalance in classifying a collection of song lyrics. They employed neural network models including a multi-layer perceptron and a Doc2Vec model in their experiments where the finding was that undersampling the majority class can be a reasonable approach to remove the data sparsity and further improve the classification performance. Li et al. (2020) also explored the problem of high data imbalance using cross-entropy criteria as well as standard performance metrics. They proposed a loss function called Dice loss that assigns equal importance to the false negatives and the false positives. In computer vision, Bowley et al. (2019) developed an automated feedback loop method to identify and classify wildlife species from Unmanned Aerial Systems imagery, for training CNNs to overcome the unbalanced class issue. On their expert imagery dataset, the error rate decreased substantially from 0.88 to 0.05. This work adapts this feedback loop strategy to the NLP problem of classifying technical events. 3 Technical Event Datasets In this work, we used a set of 7 logbook datasets from the aviation, automotive, and facility domains available at MaintNet (Akhbardeh et al., 2020a). MaintNet is a collaborative open-source platform for predictive maintenance language resources featuring multiple technical logbook datasets and tools. These datasets include: 1) Avi-Main contains seven years of maintenance logbook reports collected by 4037 Code Inst Avg N Class Size Toks Cls Min Med Avg Max Avi-Main 6,169 13.85 39 21 56 158 1,674 Avi-Acc 4,130 14.31 5 179 966 826 1,595 Avi-Safe 17,718 19.52 2 2,134 8,859 8,859 15,584 Auto-Main 617 7.34 5 23 48 123 268 Auto-Acc 52,707 4.59 3 1,085 11,060 17,569 40,562 Auto-Safe 4,824 25.11 17 86 213 284 678 Faci-Main 74,360 31.50 70 25 303 1,062 10,748 Table 3: Number of instances (Inst), average number of tokens per instance (Avg Toks), number of classes (N Cls), and class size statistics: minimum, average, median, and maximum (Min, Med, Avg, Max) for each dataset. the University of North Dakota aviation program on aircraft maintenance that were reported by the mechanic or pilot. 2) Avi-Acc contains four years of aviation accident and reported damages. 3) AviSafe contains eleven years of aviation safety and incident reports. Accidents were caused by foreign objects/birds during the flights which led to safety inspection and maintenance, where safety crews indicated the damage (safety) level for further analysis. 4) Auto-Main is a single year report with maintenance records for cars. 5) Auto-Acc contains twelve years of car accidents and crash reports describing the related car maintenance issue and property damaged in the accident. 6) Auto-Safe contains four years of noted hazards and incidents on the roadway from the driver. 7) Faci-Main contains six years of logbook reports collected for building maintenance. These technical logbooks include short, compact, and descriptive domain-specific English texts single instances usually contain between 2 and 20 tokens on average including abbreviations and domain-specific words. An example instance from Table 2, r/h fwd upper baff seal needs to be resecured, shows how the instances for a specific issue class are comprised from specific vocabulary (less ambiguity), and therefore contain a high level of granularity (level of description for an event from multiple words) (Mulkar-Mehta et al., 2011). Table 3 presents statistics for each dataset, in terms of the number of instances, average instance length, number of classes, and the minimum, average, median and maximum class size to represent how imbalanced the datasets are. An instance in the logbook can be formed as a complete description of the technical event (such as a safety or maintenance inspection) like: #2 & #4 cyl rocker cover gsk are leaking, or it might contain an incomplete description by solely referring to the damaged part/section of machinery (hyd cap chck eng light on) using few domain words. In either form of the problem description, the given annotation (label) is at the issue type-level, e.g., baffle damage. Table 2 shows multiple examples with associated instances. Further characteristics of these log entries include compound words (antifreeze, engine-holder, driftangle, dashboard). Many of these words (e.g., a compound word: dashboard) essentially represent the items, or domain-specific parts used in the descriptions. Additionally, function words (e.g., prepositions) are important and removing them could alter the meaning of the entry. The logbook datasets also have both the following shared and distinct characteristics: Shared Characteristics: Each instance contains a descriptive observation of the issue and/or the suggested action that should be taken (eng inspection panel missing screw). Each instance also refers to maintaining a single event, which means the recognized problem applies to the only single-issue type. As an example, the instance cyl #1 baff cracked at screw support & forward baff below #1 includes a combination of sequences that refers to the location and/or specific part of the machinery. Distinct Characteristics: In each domain, terminologies, a list of terms, and abbreviations are distinct, and an abbreviation can have different expansion depending on the domain context (Sproat et al., 2001), e.g., a/c can mean aircraft in aviation and in the automotive domain air conditioner. However, the abbreviations and acronyms of the domain words (e.g. atc - air traffic control) in these technical datasets should not be approached as a word sense disambiguation problem as they require character level expansion. 4 Methods and Models 4.1 Handling Class Imbalance Collecting additional data to augment datasets is a common approach for tackling the problem of skewed class distributions. However, as discussed earlier, technical logbooks are proprietary and very hard to obtain. In addition, each domain captures domain-specific lexical semantics, preventing the use of techniques such as domain adaption (Ma 4038 Algorithm 1 Feedback Loop Pseudocode ▷Gets MCS random instances from each class function SAMPLERANDOM(C, MCS) Array A for i ←1 to SIZE(C) do SHUFFLE(Ci) A ←A ∪GETFIRSTN(MCS, Ci) return A ▷Gets MCS instances from each class with the worst error function RESAMPLE(C, M, MCS) Array A for i ←1 to SIZE(C) do CALCULATEERROR(Ci) SORTBYERROR(Ci) A ←A ∪GETFIRSTN(MCS, Ci) return A Input: Training Data D = Instance(1, 2, . . . , N) Input: Feedback Loop Iterations FLI Input: Epochs Per Loop Iteration FLE Input: Minimum Class Size MCS ▷Divide training data by class Array C ←SPLITBYCLASS(D) ▷Get initial active training data A randomly Array A ←SAMPLERANDOM(C, MCS) Model M for l ←1 to FLI do ▷Train the model for the number of epochs per iteration M ←TRAIN(M, FLE, A) ▷Update the active training data A ←RESAMPLE(D, M, MCS) Output: M et al., 2019) to apply a large class data from one technical domain to another. For example, instances that describe an engine failure in the aviation domain are distinct from engine failure instances reported in the automotive domain. In this paper we apply five different methods for selecting training data for the models to analyze their effects on classification performance: (1) under(down)and (2) over-sampling, (3) random down-sampling, (4) a feedback loop strategy, and (5) a baseline strategy which simply uses all available data. Re-sampling Under- and over-sampling are resampling techniques (Maragoudakis et al., 2006) that were used to create balanced class sizes for model training. For over-sampling, instances of the minority classes are randomly copied so that all classes would have the same number of instances as the largest class. For under-sampling, observations are randomly removed from the majority classes, so that all classes have the same number of instances as the smallest class. For both approaches, we first divided our datasets into test and training sets before performing over-sampling to prevent contamination of the test set by having the same observations in both the training and test data. Feedback Loop To address class imbalances in text classification, this work adapts the approach in Bowley et al. (2019) from the computer vision domain. The goal of this approach is not only to alleviate the bias towards majority classes but also to adjust the training data instances such that the models are always being trained on the instances that was performing the worst on. It should be noted that this approach is very similar to adaptive learning strategies which have been shown to aid in human learning (Kerr, 2015; Midgley, 2014). Algorithm 1 presents pseudocode for the feedback loop. In this process, the active training data (the data used to actually train the models in each iteration of the loop) is continually resampled from the training data. The model is first initially trained with an undersampled number of random instances from each class, which becomes the initial active training data. The model M then performs inference over the entire training set, and then selects MCS instances from each class Ci which had the worst error during inference, where MCS is the minority (smallest) class size. The model is then retrained with this new active training data and the process of training, inference and selection of the MCS worst instances repeats for a fixed number of feedback loop iterations, FLI. In this way the model is always being trained on the instances it has classified the worst. To measure the effect of resampling the worst performing instances, the feedback loop approach was also compared to a random downsampling (DS) loop, where instead of evaluating the model over each instance and selecting the worst performing instances, MCS instances from each class are instead randomly sampled. As performing inference over the entire training set adds overhead, a comparison to the random DS loop method would show if performing this inference is worth the performance cost over simple random resampling. This approach is the same as Algorithm 1 except that SampleRandom is used instead of Resample in the feedback loop. Section 4.3 describes how the number of training epochs and loop iterations were determined such that all the training data selection methods are given a fair evaluation with the same amount of computational time. 4039 Evaluation Metrics For imbalanced datasets, simply using precision, recall or F1 score metrics for the entire datasets would not accurately reflect how well a model or method performs, as they emphasize the majority classes. To overcome this, alternative evaluation metrics to handle the class imbalance problem were used, as recommended by Banerjee et al. (2019). Specifically, we report the models performance based on precision, recall, and F1 score by utilizing a macro-average over all classes, as this gives every class equal weight, and hence reveals how well the models and training data selection strategies perform. 4.2 Model Architecture and Training Different machine learning methods were considered for technical event/issue classification (e.g. engine failure, turbine failure). Each instance is an individual short logbook entry and contains approximately 2 to 20 tokens (12 words on average per instance including function words), as shown in Table 3.The methods used in this study were a Deep Neural Network (DNN) (Dernoncourt et al., 2017), a Long Short-Term Memory (LSTM) (Suzgun et al., 2019), recurrent neural network (RNN) (Pascanu et al., 2013), a Convolutional Neural Network (CNN) (Lin et al., 2018), and BERT (Devlin et al., 2019). Deep Neural Network A deep artificial neural network (DNN), as described by Dernoncourt et al. (2017), can learn abstract representation and features of the input instances that would help to achieve better performance on predicting the issue type in the logbook dataset. The DNN used was a 3 layer, fully connected feed forward neural network with an input embedding layer of dimension 300 and equal size number of words followed by 2 dense layers with 512 hidden units with ReLU activation functions followed by a dropout layer. Finally, we added a fully connected dense layer with size equal to the number of classes, with a SoftMax activation function. Long Short-Term Memory An LSTM RNN was also used to perform a sequence-to-label classification. As described by Suzgun et al. (2019) LSTM RNNs utilize several vector gates at each state to regulate the passing of data by the sequence which enhances the modeling of the long-term dependencies. We used a 3 layer LSTM model with a word embedding layer of dimension 300 and the equal size number of words followed by an LSTM layer with setting the number of hidden units equal to the embedding dimension, followed by a dropout layer. Finally, we added a fully connected layer with size equal to the number of classes, with a SoftMax activation function. Convolutional Neural Network Convolutional neural networks (CNNs) have demonstrated exceptional success in NLP tasks such as document classification, language modeling, or machine translation (Lin et al., 2018). As Xu et al. (2020) described, CNN models can produce consistent performance when applied to the various text types such as short sequences. We evaluated a CNN architecture (Shen et al., 2018) with a convolutional layer, followed by batch normalization, ReLU, and a dropout layer, which was followed by a maxpooling layer. The model contained 300 convolutional filters with the size of 1 by n-gram length pooling with the size of 1 by the length of the input sequence, followed by concatenation layer, then finally connected to a fully connected dense layer, and an output layer equal to the size of the dataset class using a SoftMax activation function. Bidirectional Encoder Representations We also evaluated using the pre-trained uncased Bidirectional Encoder Representations (BERT) for English (Devlin et al., 2019). We fine-tuned the model, and used a word piece based BERT tokenizer for the tokenization process and the RandomSampler and SequentialSampler for training and testing respectively. To better optimize this model, a schedule was created for the learning rate that decayed linearly from the initial learning rate we set in the optimizer to 0. 4.3 Experimental Settings Datasets and Baselines First, the technical text pre-processing pipeline developed by Akhbardeh et al. (2020b) was applied, which comprises domain-specific noise entity removal, dictionarybased standardization, lexical normalization, part of speech tagging, and domain-specific lemmatization. We divided the datasets selecting randomly from each class independently to maintain a similar class size distribution, using 80% of the instances for training and 20% of the instances for testing data. For feature extraction, two methods were considered: a bag-of-word model (n-grams:1) (Pedregosa et al., 2011) and pre-trained 300 dimensional GloVe word embeddings (Pennington et al., 2014). 4040 Hyperparameter and Tuning The coarse to fine learning (CFL) approach (Lee et al., 2018) was used to set parameters and hyperparameters for the DNN, LSTM, and CNN models. Experiments considered batch sizes of 32, 64, and 128, an initial learning rate ranging from 0.01 to 0.001 with a learning decay rate of 0.9, and dropout regularization in the range from 0.2 to 0.5 in all models, as well as ReLU and SoftMax activation functions (Nair and Hinton, 2010), categorical cross-entropy (Zhang and Sabuncu, 2018) as the loss function, and the Adam optimizer (Kingma and Ba, 2015) in the DNN, LSTM, CNN and BERT models. Based on experiments and network training accuracy, a batch size of 64 and drop out regularization of 0.3 was selected for model training. Each model with each training data selection strategy was trained 20 times to generate results for each dataset. To ensure each training data selection strategy was fairly compared with a similar computational budget, the number of training epochs and loop iterations (if the strategy had a feedback or random downsampling loop) were adjusted so that the total number of training instances evaluations each model performed was the same. For each dataset, the number of forward and backward passes, ‘T’ for 100 epochs of the baseline strategy was used as the standard. As an example, Table 4 shows how many loop iterations, epochs per loop, and inference passes were done for each training data selection strategy on the Auto-Safe dataset. Given the differences between the min and max class sizes it was not possible to get exact matches but the strategies came as close as possible. We counted each inference pass for the feedback loop the same as a forward and backward training pass, which actually was a slight computational disadvantage for the feedback loop, as a forward and backward pass in training takes approximately 1x to 2x the time as an inference pass. 5 Results Table 5 shows a comparison between the baseline and the four different class balancing methods (over-sampling, under-sampling, the random downsampling (DS) loop and the feedback loop). Based on these outcomes, the feedback loop strategy almost entirely outperforms the other methods over all datasets and models, showing that performing inference over the training set and reselecting the training data from the worst performing instances Dataset L EPL LTI INM T Baseline 1 100 3,859 0 385,900 Downsampling 1 329 1,173 0 385,917 Oversampling 1 42 9,214 0 386,988 Random DS Loop 33 10 1,173 0 387,090 Feedback Loop 25 10 1,173 3,859 389,725 Table 4: Details regarding different training process using the various methods for handling the unbalanced class in automotive safety (Auto-Safe) dataset with 17 total classes. Loop (L), Epochs Per Loop (EPL), Active Training instance Size (LTI), Inference for New Misclassified (INM) and Total Instances Evaluated (T). does provide a benefit to the learning process. A plausible explanation is that this strategy does not introduce bias into the larger class and also does not effect the minority class size distribution. It also does not waste training time on instances the model has already well learned. Table 5 also shows the empirical analysis of the four classification models, with the model and training data selection strategy providing the overall best results shown in bold and italics. Using technical text pre-processing techniques described in Section 4.3, and the feedback loop strategy described in Section 4.1, the precision, recall, and F1 score improved compared to the baseline performance. The CNN model outperformed the other algorithms with improved precision, recall, and F1 score for almost all datasets except for Avi-Main, where BERT had the similar results, and Auto-Main where CNN and BERT tied. This is interesting, given the current popularity of the BERT model, however it may be due to the substantial lexical, topical, and structural linguistic differences between the technical logbook data and the English corpus that BERT was pre-trained on. Furthermore, we conducted the Mann-Whitney U-test of statistical significance by using the F1 scores of each of the 20 repeated experiments of the classification models, using the baseline and the feedback loop approach as the two different populations. The outcomes are shown in Table 6, with the differences being highly statistically significant. 6 Discussion To investigate the optimal strategies for dealing with these imbalanced technical datasets, we studied various methods on how to process the data, extract features, and classify the type of event. Re4041 Down Over Random Feedback Dataset Model Baseline (%) Sampling (%) Sampling (%) DS Loop (%) Loop (%) Pre Rec F1 Pre Rec F1 Pre Rec F1 Pre Rec F1 Pre Rec F1 Avi-Main DNN 0.90 0.89 0.89 0.67 0.78 0.70 0.90 0.90 0.90 0.90 0.90 0.90 0.93 0.91 0.91 LSTM 0.84 0.85 0.84 0.81 0.83 0.81 0.85 0.84 0.84 0.84 0.84 0.84 0.86 0.88 0.87 CNN 0.93 0.92 0.92 0.89 0.88 0.88 0.94 0.92 0.92 0.93 0.91 0.91 0.95 0.94 0.94 BERT 0.93 0.93 0.93 0.85 0.86 0.85 0.94 0.94 0.94 0.94 0.93 0.93 0.95 0.96 0.95 Avi-Acc DNN 0.47 0.44 0.43 0.35 0.45 0.35 0.48 0.47 0.47 0.50 0.44 0.46 0.52 0.45 0.48 LSTM 0.38 0.37 0.37 0.35 0.35 0.35 0.39 0.39 0.39 0.38 0.39 0.38 0.40 0.39 0.39 CNN 0.50 0.49 0.49 0.43 0.42 0.42 0.52 0.44 0.47 0.51 0.44 0.47 0.52 0.46 0.48 BERT 0.48 0.42 0.44 0.41 0.40 0.40 0.50 0.44 0.46 0.50 0.44 0.46 0.51 0.45 0.47 Avi-Safe DNN 0.43 0.41 0.41 0.36 0.36 0.36 0.50 0.50 0.50 0.50 0.49 0.49 0.53 0.51 0.51 LSTM 0.47 0.46 0.46 0.43 0.42 0.42 0.49 0.50 0.49 0.48 0.46 0.47 0.49 0.50 0.49 CNN 0.59 0.57 0.57 0.50 0.50 0.50 0.60 0.59 0.59 0.59 0.59 0.59 0.62 0.61 0.61 BERT 0.50 0.50 0.50 0.44 0.46 0.44 0.54 0.54 0.54 0.53 0.53 0.53 0.56 0.57 0.56 Auto-Main DNN 0.58 0.45 0.49 0.33 0.49 0.39 0.60 0.55 0.56 0.58 0.54 0.55 0.61 0.55 0.57 LSTM 0.49 0.55 0.51 0.41 0.42 0.41 0.50 0.60 0.54 0.51 0.58 0.54 0.53 0.61 0.55 CNN 0.61 0.61 0.61 0.53 0.53 0.53 0.64 0.64 0.64 0.63 0.64 0.63 0.65 0.64 0.64 BERT 0.60 0.60 0.60 0.54 0.53 0.53 0.63 0.64 0.63 0.63 0.63 0.63 0.64 0.64 0.64 Auto-Acc DNN 0.43 0.34 0.30 0.35 0.42 0.27 0.39 0.42 0.31 0.40 0.39 0.39 0.48 0.40 0.40 LSTM 0.45 0.39 0.41 0.40 0.40 0.40 0.42 0.41 0.41 0.42 0.40 0.40 0.48 0.41 0.44 CNN 0.46 0.43 0.44 0.44 0.41 0.42 0.49 0.50 0.49 0.50 0.51 0.50 0.51 0.53 0.52 BERT 0.50 0.49 0.49 0.47 0.47 0.47 0.50 0.50 0.50 0.51 0.49 0.50 0.52 0.51 0.51 Auto-Safe DNN 0.52 0.46 0.48 0.40 0.47 0.41 0.54 0.51 0.51 0.54 0.51 0.51 0.55 0.52 0.53 LSTM 0.40 0.40 0.40 0.38 0.39 0.38 0.41 0.42 0.41 0.41 0.41 0.41 0.43 0.42 0.42 CNN 0.59 0.58 0.58 0.52 0.51 0.51 0.59 0.60 0.59 0.59 0.59 0.59 0.62 0.60 0.61 BERT 0.57 0.56 0.56 0.52 0.50 0.50 0.58 0.56 0.56 0.57 0.57 0.57 0.58 0.59 0.59 Faci-Main DNN 0.57 0.48 0.50 0.33 0.40 0.34 0.56 0.48 0.50 0.57 0.50 0.53 0.59 0.51 0.54 LSTM 0.56 0.56 0.56 0.53 0.52 0.52 0.59 0.55 0.56 0.59 0.56 0.57 0.63 0.56 0.60 CNN 0.64 0.64 0.64 0.61 0.60 0.60 0.66 0.66 0.66 0.65 0.65 0.65 0.69 0.67 0.68 BERT 0.63 0.64 0.63 0.60 0.60 0.60 0.65 0.64 0.64 0.64 0.65 0.64 0.68 0.67 0.67 Table 5: Comparison of results for the 7 datasets, for the baseline and four methods to address class imbalance for the four evaluated models (DNN, LSTM, CNN and BERT). Each model’s macro average performance is shown as precision (Pre), recall (Rec) and F1 score. The best results over the training data selection strategies are shown in bold, and the best results over all models are additionally in italics. garding the discussion provided in Section 3 about the nature of such a dataset, there are key challenges that effect the performance of employed algorithms. As discussed in Section 1, the extreme class imbalance observed in these technical datasets substantially affects learning algorithms’ performance. To overcome this issue, we first explored oversampling and undersampling, which both result in balanced class sizes. Undersampling removed portions of dataset that could be important for certain technical events or issues, which resulted in underfitting and weak generalization for important classes. On the other hand, oversampling may introduce overfitting in the minority class, as some of the event types are very short tokens containing domain-specific words. Following this, to minimize the possibility of overfitting and underfitting, a random downsampling loop and a feedback loop were investigated to minimize bias in the training process. It was found that the added computational cost of the feedback loop inference was worth the reduction in training time it caused over the random downsampling loop. The scarce data available in a dataset such as Auto-Main is certainly an issue for deep learning methods. Examining the accuracy improvement by using the proposed feedback loop strategy, requires incorporating more instances to the event classes. 4042 Similar to any supervised learning models, we noticed some limitations that could be addressed in future work. As shown in the previous sections (such as Table 2), logbook instances contain short text (ranging from 2 to 20 tokens per instance), and utilizing recurrent deep learning algorithms such as LSTM RNNs which are heavily based on the context leads to weak performance compared to other algorithms. One possible explanation is that logbooks with short instances (sequences) are not providing sufficient context for the algorithm to make better predictions. Another could be that RNNs are notoriously difficult to train (Pascanu et al., 2013), and the LSTM models may simply require more training time to achieve similar results. There is some evidence for this, as the dataset with the most instances, which also had the second largest number of tokens per instance on average was FaciMain, which is the dataset which the LSTM model had the closest performance to the CNN and BERT models, and was also the only one which the LSTM model outperformed the DNN model. The pre-trained BERT model provided a reasonable classification performance compared to the other deep learning models, however as BERT is pre-trained on standard language, the performance when applying to logbook data was not optimal. Training or fine-tunning BERT to technical logbook data is likely to improve performance as observed in the legal and scientific domains (Chalkidis et al., 2020; Beltagy et al., 2019). As training or finetuning BERT requires large amounts of data, a limitation for fine-tuning a domain-specific BERT is the amount of logbook data available. 7 Conclusion and Future Work This work focused on predictive maintenance and technical event/issue classification, with a special focus on addressing class imbalance. We acquired seven logbook datasets from three technical domains containing short instances with non-standard grammar and spelling, and many abbreviations. To address RQ1, we evaluated multiple strategies to address the extreme class imbalance in these datasets and we showed that the feedback loop strategy performs best, almost entirely providing the best results for the 7 different datasets and 4 different models investigated. To address RQ2, we empirically compared different classification algorithms (DNN, LSTM, CNN, and pre-tuned BERT). Results show that the CNN model outperforms the Dataset DNN LSTM CNN BERT Avi-Main 0.0020 0.0043 0.0002 0.0004 Avi-Acc 0.0011 0.0399 0.0103 0.0015 Avi-Safe 0.0000 0.0023 0.0059 0.0012 Auto-Main 0.0001 0.0181 0.0009 0.0004 Auto-Acc 0.0000 0.0055 0.0001 0.0161 Auto-Safe 0.0003 0.0106 0.0011 0.0083 Faci-Main 0.0002 0.0001 0.0003 0.0005 Table 6: Statistical significance of the various classification models between the Baseline approach and Feedback Loop approach F1 scores using the MannWhitney U test. Experiments indicate statistical significance with a p value of 0.05. other classifiers. The methodology presented in this paper could be applied to other maintenance corpora from a variety of technical domains. The feedback loop approach for selecting training data is generic and could easily be applied to any learning problem with substantial class imbalances. This is useful as extreme class imbalance is a challenge at the heart of a number of natural language tasks. In future work, we would like to fine-tune BERT using logbook data, as described in Section 6, and extend this work to datasets in other languages. The biggest challenge for these two research directions is the limited availability of logbook datasets. Furthermore, we are exploring various methods of domain adaptation and transfer learning on these datasets to further improve the performance of classification models. Acknowledgments We would like to thanks the University of North Dakota aviation program for providing the valuable aviation maintenance logbook datasets to the MaintNet research. We further thank the aviation domain expert Zechariah Morgan for evaluating the outcomes of the various algorithms and providing valuable feedback for the aviation domain dataset. We also would like to thank the anonymous ACL reviewers for providing us with helpful comments and feedback. 4043 References Farhad Akhbardeh, Travis Desell, and Marcos Zampieri. 2020a. MaintNet: A collaborative opensource library for predictive maintenance language resources. In Proceedings of the 28th International Conference on Computational Linguistics, pages 7– 11, Barcelona, Spain. International Committee on Computational Linguistics. Farhad Akhbardeh, Travis Desell, and Marcos Zampieri. 2020b. NLP tools for predictive maintenance records in MaintNet. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 26–32, Suzhou, China. Association for Computational Linguistics. Sadam Al-Azani and El-Sayed El-Alfy. 2017. Using word embedding and ensemble learning for highly imbalanced data sentiment analysis in short arabic text. Procedia Computer Science, Vol 109:359–366. M. Altuncu, Erik Mayer, Sophia Yaliraki, and Mauricio Barahona. 2019. From free text to clusters of content in health records: an unsupervised graph partitioning approach. Applied Network Science, Vol 4. Siddhartha Banerjee, Cem Akkaya, Francisco PerezSorrosal, and Kostas Tsioutsiouliklis. 2019. Hierarchical transfer learning for multi-label text classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics ACL, pages 6295–6300, Florence, Italy. Association for Computational Linguistics. Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3615– 3620, Hong Kong, China. Association for Computational Linguistics. Connor Bowley, Marshall Mattingly, Andrew Barnas, Susan Ellis-Felege, and Travis Desell. 2019. An analysis of altitude, citizen science and a convolutional neural network feedback loop on object detection in unmanned aerial systems. Journal of Computational Science, Vol 34:102 – 116. Thyago P. Carvalho, Fabr´ızzio A. A. M. N. Soares, Roberto Vita, Roberto da P. Francisco, Jo˜ao P. Basto, and Symone G. S. Alcal´a. 2019. A systematic literature review of machine learning methods applied to predictive maintenance. Computers and Industrial Engineering, 137:106024. Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. 2020. LEGAL-BERT: The muppets straight out of law school. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2898– 2904, Online. Association for Computational Linguistics. Colin Cherry and Hongyu Guo. 2015. The unreasonable effectiveness of word representations for Twitter named entity recognition. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies HLT-NAACL, pages 735–745, Denver, Colorado. Association for Computational Linguistics. Matthias Damaschk, Tillmann D¨onicke, and Florian Lux. 2019. Multiclass text classification on unbalanced, sparse and noisy data. In Proceedings of the First NLPL Workshop on Deep Learning for Natural Language Processing, pages 58–65, Turku, Finland. Link¨oping University Electronic Press. Louise Del´eger, Cyril Grouin, and Pierre Zweigenbaum. 2010. Extracting medical information from narrative patient records: The case of medicationrelated information. Journal of the American Medical Informatics Association, Vol 17:555 – 558. Franck Dernoncourt, Ji Young Lee, and Peter Szolovits. 2017. Neural networks for joint sentence classification in medical paper abstracts. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, pages 694–700, Valencia, Spain. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Kim Hammar, Shatha Jaradat, Nima Dokoohaki, and Mihhail Matskin. 2018. Deep text mining of instagram data without strong supervision. In International Conference on Web Intelligence (WI), pages 158 – 165, Santiago, Chile. Philip Kerr. 2015. Adaptive learning. English Language Teaching (ELT) Journal, 70(1):88–93. Donghwa Kim, Deokseong Seo, Suhyoun Cho, and Pilsung Kang. 2019. Multi-co-training for document classification using various document representations: TF-IDF, LDA, and Doc2Vec. Information Sciences, Vol 477:15 – 29. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-tofine inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2, pages 687–692, New 4044 Orleans, Louisiana. Association for Computational Linguistics. Junyi Jessy Li and Ani Nenkova. 2014. Addressing class imbalance for improved recognition of implicit discourse relations. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), pages 142–150, Philadelphia, PA, U.S.A. Association for Computational Linguistics. Shoushan Li, Shengfeng Ju, Guodong Zhou, and Xiaojun Li. 2012. Active learning for imbalanced sentiment classification. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 139–148, Jeju Island, Korea. Association for Computational Linguistics. Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. 2020. Dice loss for dataimbalanced NLP tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 465–476, Online. Association for Computational Linguistics. Junyang Lin, Qi Su, Pengcheng Yang, Shuming Ma, and Xu Sun. 2018. Semantic-unit-based dilated convolution for multi-label text classification. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4554– 4564, Brussels, Belgium. Association for Computational Linguistics. Xiaofei Ma, Peng Xu, Zhiguo Wang, Ramesh Nallapati, and Bing Xiang. 2019. Domain adaptation with BERT-based domain classification and data selection. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP (DeepLo 2019), pages 76–83, Hong Kong, China. Association for Computational Linguistics. Manolis Maragoudakis, Katia Kermanidis, Aristogiannis Garbis, and Nikos Fakotakis. 2006. Dealing with imbalanced data using Bayesian techniques. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06), Genoa, Italy. European Language Resources Association (ELRA). J.J. McArthur, Nima Shahbazi, Ricky Fok, Christopher Raghubar, Brandon Bortoluzzi, and Aijun An. 2018. Machine learning and bim visualization for maintenance issue classification and enhanced data collection. Advanced Engineering Informatics, 38:101 – 112. Carol Midgley. 2014. Goals, goal structures, and patterns of adaptive learning. Routledge. Rutu Mulkar-Mehta, Jerry Hobbs, and Eduard Hovy. 2011. Granularity in natural language discourse. In Proceedings of the Ninth International Conference on Computational Semantics, IWCS ’11, page 360–364, USA. Association for Computational Linguistics. Vinod Nair and Geoffrey E. Hinton. 2010. Rectified linear units improve restricted Boltzmann Machines. In Proceedings of the 27 th International Conference on Machine Learning, Haifa, Israel. ICML. Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013. On the difficulty of training recurrent neural networks. In International conference on machine learning, pages 1310–1318. PMLR. Jon Patrick and Min Li. 2009. A cascade approach to extracting medication events. In Proceedings of the Australasian Language Technology Association Workshop 2009, pages 99–103, Sydney, Australia. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825–2830. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543, Doha, Qatar. Association for Computational Linguistics. Alan Ritter, Mausam Mausam, Oren Etzioni, and Sam Clark. 2012. Open domain event extraction from twitter. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1104-1112:1104 – 1112. Guergana K. Savova, James J. Masanz, Philip V. Ogren, Jiaping Zheng, Sunghwan Sohn, Karin Kipper Schuler, and Christopher G. Chute. 2010. Mayo clinical text analysis and knowledge extraction system (ctakes): architecture, component evaluation and applications. Journal of the American Medical Informatics Association : JAMIA, 17 5:507–13. Dinghan Shen, Martin Renqiang Min, Yitong Li, and Lawrence Carin. 2018. Learning context-sensitive convolutional filters for text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1839–1848, Brussels, Belgium. Association for Computational Linguistics. Richard Sproat, Alan W. Black, Stanley Chen, Shankar Kumar, Mari Ostendorf, and Christopher Richards. 2001. Normalization of non-standard words. Computer Speech & Language, 15(3):287 – 333. Mirac Suzgun, Yonatan Belinkov, and Stuart M. Shieber. 2019. On evaluating the generalization of LSTM models in formal languages. In Proceedings of the Society for Computation in Linguistics (SCiL) 2019, pages 277–286. Harish Tayyar Madabushi, Elena Kochkina, and Michael Castelle. 2019. Cost-sensitive BERT for generalisable sentence classification on imbalanced 4045 data. In Proceedings of the Second Workshop on Natural Language Processing for Internet Freedom: Censorship, Disinformation, and Propaganda, pages 125–134, Hong Kong, China. Association for Computational Linguistics. Jingyun Xu, Yi Cai, Xin Wu, Xue Lei, Qingbao Huang, Ho fung Leung, and Qing Li. 2020. Incorporating context-relevant concepts into convolutional neural networks for short text classification. Neurocomputing, 386:42 – 53. Meliha Yetisgen-Yildiz, Cosmin Bejan, and Mark Wurfel. 2013. Identification of patients with acute lung injury from free-text chest X-ray reports. In Proceedings of the 2013 Workshop on Biomedical Natural Language Processing, pages 10–17, Sofia, Bulgaria. Association for Computational Linguistics. Zhilu Zhang and Mert R. Sabuncu. 2018. Generalized cross entropy loss for training deep neural networks with noisy labels. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS’18, page 8792–8802, Red Hook, NY, USA. Curran Associates Inc.
|
2021
|
312
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4046–4062 August 1–6, 2021. ©2021 Association for Computational Linguistics 4046 ILDC for CJPE: Indian Legal Documents Corpus for Court Judgment Prediction and Explanation Vijit Malik1 Rishabh Sanjay1 Shubham Kumar Nigam1 Kripa Ghosh2 Shouvik Kumar Guha3 Arnab Bhattacharya1 Ashutosh Modi1 1Indian Institute of Technology Kanpur (IIT-K) 2Indian Institute of Science Education and Research Kolkata (IISER-K) 3West Bengal National University of Juridical Sciences (WBNUJS) {vijitvm,rsan,sknigam}@iitk.ac.in [email protected] [email protected] {arnabb,ashutoshm}@cse.iitk.ac.in Abstract An automated system that could assist a judge in predicting the outcome of a case would help expedite the judicial process. For such a system to be practically useful, predictions by the system should be explainable. To promote research in developing such a system, we introduce ILDC (Indian Legal Documents Corpus). ILDC is a large corpus of 35k Indian Supreme Court cases annotated with original court decisions. A portion of the corpus (a separate test set) is annotated with gold standard explanations by legal experts. Based on ILDC, we propose the task of Court Judgment Prediction and Explanation (CJPE). The task requires an automated system to predict an explainable outcome of a case. We experiment with a battery of baseline models for case predictions and propose a hierarchical occlusion based model for explainability. Our best prediction model has an accuracy of 78% versus 94% for human legal experts, pointing towards the complexity of the prediction task. The analysis of explanations by the proposed algorithm reveals a significant difference in the point of view of the algorithm and legal experts for explaining the judgments, pointing towards scope for future research. 1 Introduction In many of the highly populated countries like India, there is a vast number of pending backlog of legal cases that impede the judicial process (Katju, 2019). The backlog is due to multiple factors, including the unavailability of competent judges. Therefore, a system capable of assisting a judge by suggesting the outcome of an ongoing court case is likely to be useful for expediting the judicial process. However, an automated decision system is not tenable in law unless it is well explained in terms of how humans understand the legal process. Hence, it is necessary to explain the suggestion. In other words, we would like such a system to predict not only what should be the final decision of a court case but also how one arrives at that decision. In this paper, we introduce INDIAN LEGAL DOCUMENTS CORPUS (ILDC) intending to promote research in developing a system that could assist in legal case judgment prediction in an explainable way. ILDC is a corpus of case proceedings from the Supreme Court of India (SCI) that are annotated with original court decisions. A portion of ILDC (i.e., a separate test set) is additionally annotated with gold standard judgment decision explanations by legal experts to evaluate how well the judgment prediction algorithms explain themselves. Based on ILDC, we propose a new task: COURT JUDGMENT PREDICTION AND EXPLANATION (CJPE). This task aims to predict the final decision given all the facts and arguments of the case and provide an explanation for the predicted decision. The decision can be either allowed, which indicates ruling in favor of the appellant/petitioner, or dismissed, which indicates a ruling in favor of the respondent. The explanations in the CJPE task refer to sentences/phrases in the case description that best justify the final decision. Since, we are addressing mainly the SCI cases, one might argue that the usefulness of the task may be limited since, the legislative provisions can always change with time. However, the legal principles of how to apply a given law to a given set of facts remain constant for prolonged periods. Judgment prediction and explanation in the CJPE task are far more challenging than a standard text-classification task for multiple reasons. Firstly, the legal court case documents (especially 4047 in Indian context) are unstructured and are usually quite long, verbose, and noisy. There is no easy way of extracting and directly using the facts and arguments. Secondly, the domain-specific lexicon used in court cases makes models pre-trained on generally available texts ineffective on such documents. Consequently, the standard models need to be adapted to the legal domain for the proposed judgment prediction on court cases. Thirdly, explaining prediction in legal documents is considerably more challenging as it requires understanding the facts, following the arguments and applying legal rules, and principles to arrive at the final decision. Our main contributions can be summarized as: 1. We create a new corpus, INDIAN LEGAL DOCUMENTS CORPUS (ILDC), annotated with court decisions. A portion of the corpus (i.e. a separate test set) is additionally annotated with explanations corresponding to the court decisions. We perform detailed case studies on the corpus to understand differences in prediction and explanation annotations by legal experts, indicative of the computational challenges of modeling the data. 2. We introduce a new task, COURT JUDGMENT PREDICTION AND EXPLANATION (CJPE), with the two sub-tasks: (a) Court Judgment Prediction (CJP) and (b) Explanation of the Prediction. While CJP is not a novel task per se; however, in combination with the explanation part, the CJPE task is new. Moreover, the requirement for explanations also puts restrictions on the type of techniques that could be tried for CJP. In the CJPE task, gold explanations are not provided in the train set; the task expects that the trained algorithms should explain the predictions without requiring additional information in the form of annotations during training. 3. We develop a battery of baseline models for the CJPE task. We perform extensive experimentation with state-of-the-art machine learning algorithms for the judgment prediction task. We develop a new method for explaining machine predictions since none of the existing methods could be readily applied in our setting. We compare model explainability results with annotations by legal experts, showing significant differences between the point of view of algorithms and experts. ILDC is introduced to promote the development of a system/models that will augment humans and not replace them. We have covered the ethical considerations in the paper. Nevertheless, the community needs to pursue more research in this regard to fully understand the unforeseen social implications of such models. This paper takes initial steps by introducing the corpus and baseline models to the community. Moreover, we plan to continue to grow, revise and upgrade ILDC. We release the ILDC and code for the prediction and explanation models via GitHub1. 2 Related Work There has been extensive research on legal domain text, and various corpora and tasks have been proposed e.g., prior case retrieval (Jackson et al., 2003), summarization (Tran et al., 2019; Bhattacharya et al., 2019a), catchphrase extraction (Galgani et al., 2012), crime classification (Wang et al., 2019), and judgment prediction (Zhong et al., 2020). Why ILDC? The task of Legal Judgment Prediction (LJP) and its corresponding corpora (Chalkidis et al., 2019; Zhong et al., 2020; Yang et al., 2019a; Xiao et al., 2018) are related to our setting. In the LJP task, given the facts of a case, violations, charges (e.g., theft) and terms of penalty are predicted. However, the ILDC and the CJPE task introduced in this paper differ from the existing LJP corpora and task in multiple ways. Firstly, we require prediction algorithms to explain the decisions in the CJPE task, to evaluate the explanations we provide a separate test set annotated with gold explanations. Secondly, in the LJP task, typically, the facts of a case are explicitly provided. However, in our case, only unannotated unstructured documents are provided. ILDC addresses a more realistic/practical setting, and consequently, CJPE is a much more challenging task. Moreover, the bare facts do not form the judgment premise of a case since facts are subject to interpretations. A court case description, in practice, has other vital aspects like Ruling by Lower Court, Arguments, Statutes, Precedents, and Ratio of the decision (Bhattacharya et al., 2019b) that are instrumental in decision making by the judge(s). Unlike LJP, we consider (along with the facts) the entire case (except the judgment), and we predict the judgment only. Work by Strickson and de la Iglesia (2020) comes close to our setting, where the authors prepared the test set on UK court cases by removing the final decision from rulings and employed classical machine learning models. Thirdly, to the best of our knowledge, 1https://github.com/Exploration-Lab/ CJPE 4048 we are the first to create the largest legal corpus (34, 816 documents) for the Indian setting. It is important because India has roots in the common law system and case decisions are not strictly as per the statute law, with the judiciary having the discretion to interpret their version of the legal provisions as applicable to the case at hand; this can sometimes make the decision process subjective. Fourth, we do not focus on any particular class of cases (e.g., criminal, civil) but address publicly available generic SCI case documents. Xiao et al. (2018) released the Chinese AI and Law challenge dataset (CAIL2018) in Chinese for judgment prediction, that contains more than 2.68 million criminal cases published by the Supreme People’s Court of China. Chalkidis et al. (2019) released an English legal judgment prediction dataset, containing 11, 478 cases from the European Court of Human Rights (ECHR). It contains facts, articles violated (if any), and an importance score for each case. ILDC contrasts with the existing LJP corpora, where mainly the civil law system and cases are considered. Though the proposed corpus focuses on Indian cases, our analysis reveals (§ 4.2) that the language used in the cases is quite challenging to process computationally and provides a good playground for developing realistic legal text understanding systems. Several different approaches and corpora have been proposed for the LJP task. Chalkidis et al. (2019) proposed a hierarchical version of BERT (Devlin et al., 2019) to alleviate BERT’s input token count limitation for the LJP task. Yang et al. (2019a) applied Multi-Perspective Bi-Feedback Network for predicting the relevant law articles, charges, and terms of penalty on Chinese AI and Law challenge (CAIL2018) datasets. Xu et al. (2020) proposed a system for distinguishing confusing law articles in the LJP task. Zhong et al. (2018) applied topological multi-task learning on a directed acyclic graph to predict charges like theft, traffic violation, intentional homicide on three Chinese datasets (CJO, PKU, and CAIL). Luo et al. (2017) proposed an attention-based model to predict the charges given the facts of the case along with the relevant articles on a dataset of Criminal Law of the People’s Republic of China. Hu et al. (2018) used an attribute-attentive model in a fewshot setup for charge prediction from facts of the case. Long et al. (2019) predicts the decision of the case using a Legal Reading Comprehension techCorpus (Avg. tokens) Number of docs (Accepted Class %) Train Validation Test ILDCmulti (3231) 32305 (41.43%) 994 (50%) 1517 (50.23%) ILDCsingle (3884) 5082 (38.08%) ILDCexpert (2894) 56 (51.78%) Table 1: ILDC Statistics nique on a Chinese dataset. Chen et al. (2019) used a deep gating network for prison term prediction, given the facts and charges on a dataset constructed from documents of the Supreme People’s Court of China. Aletras et al. (2016) used linear SVM to predict violations from facts on European Court of Human Rights cases. S¸ulea et al. (2017) used SVM in the LJP task on French Supreme Court cases. Katz et al. (2017) presented a random forest model to predict the “Reverse”, “Affirm”, and “Other” decisions of US Supreme Court judges. We also experiment with some of these models as baselines for the CJPE task (§ 5). Explainability in a system is of paramount importance in the legal domain. Zhong et al. (2020) presented a QA based model using reinforcement learning for explainable LJP task on three Chinese datasets (CJO, PKU, and CAIL). The model aims to predict the appropriate crime by asking relevant questions related to the facts of the case. Jiang et al. (2018) used a rationale augmented classification model for the charge prediction task. The model selects as rationale the relevant textual portions in the fact description. Ye et al. (2018) used labelconditioned Seq2Seq model for charge prediction on Chinese legal documents, and the interpretation comprise the selection of the relevant rationales in the text for the charge. We develop an explainability model based on the occlusion method (§ 5.2). 3 Indian Legal Document Corpus In this paper, we introduce the INDIAN LEGAL DOCUMENTS CORPUS (ILDC), a collection of case proceedings (in the English language) from the Supreme Court of India (SCI). For a case filed at the SCI, a decision (“accepted” v/s “rejected”) is taken between the appellant/petitioner versus the respondent by a judge while taking into account the facts of the case, ruling by lower Court(s), if any, arguments, statutes, and precedents. For every case filed in the Supreme Court of India (SCI), the judge 4049 (or a bench) decides on whether the claim(s) filed by the appellant/petitioner against the respondent should be “accepted” or “rejected”. The decision is relative to the appellant. In ILDC, each of the case proceeding document is labeled with the original decision made by the judge(s) of the SCI, which serve as the gold labels. In addition to the ground truth decision, a separate test set documents are annotated (by legal experts) with explanations that led to the decision. The explanations annotations are ranked in the order of importance. ILDC Creation. We extracted all the publicly available SCI2 case proceedings from the year 1947 to April 2020 from the website: https: //indiankanoon.org. Case proceedings are unstructured documents and have different formats and sizes, have spelling mistakes (since these are typed during the court hearing), making it challenging to (pre-)process. We used regular expressions to remove the noisy text and meta-information (e.g., initial portions of the document containing case number, judge name, dates, and other meta information) from the proceedings. In practice, as pointed by the legal experts, the judge deciding the case and other meta information influence the final decision. In SCI case proceedings, the decisions are written towards the end of the document. These end section(s) directly stating the decision have been deleted from the documents in ILDC since that is what we aim to predict. Each case’s actual decision label has been extracted from the deleted end sections of the proceeding using regular expressions. Another challenge with SCI case proceedings is the presence of cases with multiple petitions where, in a single case, multiple petitions have been filed by the appellant leading to multiple decisions. Consequently, we divided ILDC documents into two sets. The first set, called ILDCsingle, either have documents where there is a single petition (and, thus, a single decision) or multiple petitions, but the decisions are the same across all those petitions. The second set, called ILDCmulti, is a superset of ILDCsingle and has multiple appeals leading to different decisions. Predicting multiple different decisions for cases with multiple appeals is significantly challenging. In this paper, we do not develop any baseline computational models for this setting; we plan to address this in future work. For the com2Although IndianKanoon includes lower court cases as well, they do not have a common structural format and many of the case documents in lower courts may be in a regional Indian language. Hence, for now we only use SCI documents. putational models for the CJPE task, in the case of ILDCmulti, even if a single appeal was accepted in the case having multiple appeals/petitions, we assigned it the label as accepted. Table 1 shows the corpus statistics for ILDC. Note that the validation and test sets are the same for both ILDCmulti and ILDCsingle. Temporal Aspect. The corpus is randomly divided into train, validation, and test sets, with the restriction that validation and test sets should be balanced w.r.t. the decisions. The division into train, development, and test set was not based on any temporal consideration or stratification because the system’s objective that may eventually emerge from the project is not meant to be limited to any particular law(s), nor focused on any particular period of time. On the contrary, the aim is to identify standard features of judgments pronounced in relation to various legislation by different judges and across different temporal phases, to be able to use the said features to decipher the judicial decision-making process and successfully predict the nature of the order finally pronounced by the court given a set of facts and legal arguments. While there would be a degree of subjectivity involved, given the difference in the thoughts and interpretations adopted by different judges, such differences are also found between two judges who are contemporaries of each other, as much as between two judges who have pronounced judgments on similar matters across a gap of decades. The focus is, therefore, to develop a system that would be equally successful in predicting the outcome of a judgment given the law that had been in vogue twenty years back, as it would in relation to the law that is currently in practice. The validity and efficacy of the system can therefore be equally tested by applying it to cases from years back, as to cases from a more recent period. In fact, if the system cannot be temporally independent, and remains limited to only successful prediction of contemporary judgments, then it is likely to fail any test of application because by the time the final version of the system can be ready for practical applications on a large scale, the laws might get amended or replaced, and therefore, the judgments that would subsequently be rendered by the court might be as different from one pronounced today, as the latter might differ from one pronounced in the twentieth century. Not acknowledging time as a factor during data sample choice, therefore, appears to be the prudent step in this case, especially 4050 given the exponential rate at which legislation is getting amended today, as well as the fast-paced growth of technological development. Legal Expert Annotations. In our case, the legal expert team consisted of a law professor and his students at a reputed national law school. We took a set of 56 documents (ILDCexpert) from the test set, and these were given to 5 legal experts. Experts were requested to (i) predict the judgment, and (ii) mark the sentences that they think are explanations for their judgment. Each document was annotated by all the 5 experts (in isolation) using the WebAnno framework (de Castilho et al., 2016). The annotators could assign ranks to the sentences selected as explanations; a higher rank indicates more importance for the final judgment. The rationale for rank assignment to the sentences is as follows. Rank 1 was given to sentences immediately leading to the decision. Rank 2 was assigned to sentences that contributed to the decision. Rank 3 was given to sentences indicative of the disagreement of the current court with a lower court/tribunal decision. Sentences containing the facts of the case, not immediately, leading to decision making, but are essential for the case were assigned Rank 4 (or lower). Note in practice, only a small set of sentences of a document were assigned a rank. Although documents were annotated with explanations in order of ranks, we did not have a similar mechanism in our automated explainability models. From the machine learning perspective, this is a very challenging task, and to the best of our knowledge, none of the state-of-the-art explainability models are capable of doing this. Annotation of explanations is a very specialized, time-consuming, and laborious effort. In the current version of ILDC we provide explanation annotations to only a small portion of the test set, this is for evaluating prediction algorithms for the explainability aspect. Even this small set of documents is enough to highlight the difference between the ML-based explainability methods and how a legal expert would explain a decision (§ 5.3). Nevertheless, we plan to continue to grow the corpus by adding more explainability annotations and other types of annotations. Moreover, we plan to include lower courts like Indian High Court cases and tribunal cases. The corpus provides new research avenues to be explored by the community. Fairness and Bias. While creating the corpus, we took all possible steps to mitigate any biases that might creep in. We have not made any specific choice with regard to any specific law or any category of cases, i.e., the sampling of cases was completely random. As explained earlier, we took care of the temporal aspect. Importantly, the names of the judge(s), appellants, petitioners, etc., were anonymized in the documents so that no inherent bias regarding these creeps in. The anonymization with respect to judge names is necessary as legal experts pointed out that a judge’s identity can sometimes be a strong indicator of the case outcome. It is noteworthy that according to the legal experts if we had not done the same, we could have had higher prediction accuracy. The subjectivity associated with judicial decision-making may also be controlled in this way since the system focuses on how consideration of the facts and applicable law are supposed to determine the outcome of the cases, instead of any individual bias on the judge’s part. We also address the ethical concerns in the end. 4 Annotation Analysis We performed a detailed analysis of case predictions and the explanations annotations. With assistance from a legal expert, we also performed detailed studies for some court cases to understand the task’s complexity and possible reasons for deviations between the annotators. 4.1 Case Judgment Accuracy We computed the case judgment accuracy of the annotators with respect to original decisions by judges of SCI. The results are shown in Table 2. Though the values are high, none of these are 100%. The accuracy indicates that no annotator agrees with the original judgment in all the cases. This possibly depicts the subjectivity in the legal domain with regard to decision making. The subjectivity aspect has also been observed in other tasks that involve human decision-making, e.g., sentiment and emotion analysis. We performed detailed case studies with the help of experts to further probe into this difference in judgment. Due to space limitations, we are not able to present the studies here; please refer to appendix A and GitHub repository for details. To summarize, the study indicated that the sources of confusion are mainly due to differences in linguistic interpretation (by the annotators) of the legal language given in the case document. 4.2 Inter-Annotator Agreements Agreement in the judgment prediction: For the quantitative evaluation, we calculate pair-wise 4051 Expert Accuracy (%) Expert 1 94.64 Expert 2 91.07 Expert 3 98.21 Expert 4 89.28 Expert 5 96.43 Table 2: Annotators’ accuracy. Agreement (%) Expert 1 Expert 2 Expert 3 Expert 4 Expert 5 Expert 1 100.0 87.5 94.6 85.7 89.3 Expert 2 87.5 100.0 92.9 87.5 91.1 Expert 3 94.6 92.9 100.0 91.1 94.6 Expert 4 85.7 87.5 91.1 100.0 89.3 Expert 5 89.3 91.1 94.6 89.3 100.0 Table 3: Pairwise inter-annotator agreement for judgment prediction. User 1 User 2 User 3 User 4 User 5 User 1 User 2 User 3 User 4 User 5 1.0 0.6313 0.7869 0.8048 0.6509 0.6309 1.0 0.6223 0.6241 0.6083 0.7869 0.6224 1.0 0.8694 0.6593 0.8048 0.624 0.8695 1.0 0.6765 0.6516 0.6097 0.6598 0.6763 1.0 ROUGE-L 0.64 0.72 0.80 0.88 0.96 Figure 1: Explanation agreement among the annotators agreement between the annotators as shown in Table 3. The highest agreement (94.6%) is between Experts 1-3 and 3-5. We also calculate Fleiss’ kappa (Fleiss, 1971) as 0.820, among all the five annotators, which indicates high agreement. Agreement in the explanation: There are no standard metrics for evaluating annotator agreements for textual annotations. For quantitative evaluation of agreements among the annotators for explanations, we took inspiration from machine translation community and used metrics like ROUGE-L, ROUGE-1, ROUGE-2 (Lin, 2004), BLEU (Papineni et al., 2002) (unigram and bigram averaging), METEOR (Lavie and Agarwal, 2007), Jaccard Similarity, Overlap Maximum and Overlap Minimum3. The result for ROUGE-L (averaged out over all documents)4 is shown in Figure 1. The highest overlap across all the metrics is observed between Expert 3 and Expert 4. The highest value (0.9129) is between Expert 2 and Expert 4 for Overlap-Min. We also performed a qualitative evaluation of the agreements in the explanations. We observed that Expert 1, Expert 3, and Expert 4 consider holis3Overlap Max: Size of the intersection divided by the maximum size out of the two sample sets that are being compared. Overlap Min: Size of the intersection divided by the minimum size out of the two sample sets that are being compared 4Due to space constraints we are not able to show heatmaps corresponding to other metrics but they showed similar trends. For the heatmaps for other metrics please refer to our GitHub repository. tic reasoning for the decision. They look at both Substantive (sections applicable) and Procedural (about the jurisdiction of a lower court) aspects of the case. The differences among them are largely due to consideration/non-consideration of the factual sentences. On the other hand, Expert 2 and Expert 5 often use bare-minimum reasoning leading to the final judgment instead of looking at the exhaustive set of reasons and did not always cover both Substantive and Procedural aspects of the case. Analysis of annotations gives insights into the inherent complexity and subjectivity of the task. Legal proceedings are long, verbose, often challenging to comprehend, and exhibit interesting (and computationally challenging) linguistic phenomena. For example, in a case numbered “1962 47” (appendix A), sentence 17 of the case appears to refer to the Supreme Court having accepted a previous appeal for which a review has been requested (i.e., the current appeal). This amounted to the fact that the court actually rejected the present appeal while accepting the previous one. Such intricacies can confuse even legal experts. 5 CJPE Task Given a case proceeding from the SCI, the task of COURT JUDGMENT PREDICTION AND EXPLANATION (CJPE) is to automatically predict the decision for the case (with respect to the appellant) and provide the explanation for the decision. We address the CJPE task via two sub-tasks in the following sequence: Prediction and Explanation. Prediction: Given a case proceeding D, the task is to predict the decision y ∈{0, 1}, where the label 1 corresponds to the acceptance of the appeal/petition of the appellant/petitioner. Explanation: Given the case proceeding and the predicted decision for the case, the task is to explain the decision by predicting important sentences that lead to the decision. Annotated explanations are not provided during training; the rationale is that a model learned for prediction should explain the decision without explicit training on explanations, since explanation annotations are difficult to obtain. 4052 5.1 Case Decision Prediction ILDC documents are long and have specialized vocabulary compared to typical corpora used for training text classification models and language models. We initially experimented with non-neural models based on text features (e.g., n-grams, tfidf, word based features, and syntactic features) and existing pre-trained models (e.g., pre-trained word embeddings based models, transformers), but none of them were better than a random classifier. Consequently, we retrained/fine-tuned/developed neural models for our setting. In particular, we ran a battery of experiments and came up with four different types of models: classical models, sequential models, transformer models, and hierarchical transformer models. Table 4 summarizes the performance of different models. Due to space constraints, we are not able to describe each of the models here. We give a very detailed description of model implementations in appendix B. Classical Models: We considered classical ML models like word/sentence embedding based Logistic Regression, SVM, and Random Forest. We also tried prediction with summarized legal (Bhattacharya et al., 2019a) documents; however, these resulted in a classifier no better than random classifier. As shown in Table 4, classical models did not perform so well. However, model based on Doc2vec embeddings had similar performance as sequential models. We extensively experimented with dividing documents into chunks and training the model using each of the chunks separately. We empirically determined that sequential and transformer-based models performed the best on the validation set using the last 512 tokens5 of the document. Intuitively, this makes sense since the last parts of case proceedings usually contain the main information about the case and the rationale behind the judgment. We also experimented with different sections of a document, and we observed last 512 tokens gave the best performance. Sequence Models: We experimented with standard BiGRU (2 layers) with attention model. We tried 3 different types of embeddings: (i) Word level trained GloVe embeddings (Pennington et al., 2014), with last 512 tokens as input, (ii) Sentence level embeddings (Sent2Vec), where last 150 sen5length of 512 was partly influenced by the maximum input token limit of BERT Model Macro Precision (%) Macro Recall (%) Macro F1 (%) Accuracy (%) Classical Models on ILDCmulti train set Doc2Vec + LR 63.03 61.00 62.00 60.91 Sent2vec + LR 57.19 55.55 56.36 55.44 Sequential Models on ILDCmulti train set Sent2vec + BiGRU + att. 60.98 58.40 59.66 58.31 Doc2vec + BiGRU + att. 57.18 56.03 56.60 57.44 GloVe + BiGRU + att. 68.26 60.87 64.35 60.75 HAN 59.96 59.57 59.77 59.53 Sequential Models on ILDCsingle train set Sent2Vec + BiGRU+ att. 60.05 55.8 57.85 55.67 Doc2vec + BiGRU + att. 58.07 57.44 57.75 59.23 GloVe + BiGRU + att. 66.92 62.30 64.53 62.2 HAN 57.64 55.56 56.58 55.44 Catchphrases + Sent2Vec + BiGRU + att. 61.90 60.13 61.00 60.06 Transformer Models on ILDCmulti train set BERT Base 60.56 57.64 59.06 57.65 BERT Base 67.54 62.22 64.77 62.10 BERT Base 67.24 63.85 65.50 63.74 BERT Base 66.12 60.58 63.23 60.45 BERT Base 69.33 67.31 68.31 67.24 DistillBERT 65.21 64.26 64.73 64.21 RoBERTa 72.25 71.31 71.77 71.26 XLNet 72.09 70.07 71.07 70.01 Hierarchical Models on ILDCmulti train set BERT + BiGRU 70.98 70.42 70.69 70.38 RoBERTa + BiGRU 75.13 74.30 74.71 (±0.01) 74.33 (±1.99) XLNet + BiGRU 77.80 77.78 77.79 77.78 BERT + CNN 71.68 70.17 70.92 70.12 RoBERTa + CNN 74.74 73.17 73.95 73.22 XLNet + CNN 77.84 77.21 77.53 77.24 Hierarchical Models on ILDCsingle train set BERT + BiGRU 65.28 63.95 64.27 (±0.0116) 63.89 (±1.10) RoBERTa + BiGRU 73.24 72.93 73.09 (±0.0022) 72.95 (±0.25) XLNet + BiGRU 75.11 75.06 75.09 (±0.0043) 75.06 (±0.42) Hierarchical Models with Attention on ILDCmulti train set BERT + BiGRU + att. 71.31 70.98 71.14 (±0.0011) 71.26 (±0.09) RoBERTa + BiGRU + att. 75.89 74.88 75.38 (±0.0004) 74.91 (±0.11) XLNet + BiGRU + att. 77.32 76.82 77.07 (±0.0077) 77.01 (±0.52) Hierarchical Models with Attention on ILDCsingle train set BERT + BiGRU + att. 68.30 62.05 65.03 (±0.0084) 61.93 (±0.68) RoBERTa + BiGRU + att. 73.39 72.66 73.02 (±0.0017) 72.69 (±0.29) XLNet + BiGRU + att. 75.26 75.22 75.25 (±0.0009) 75.22 (±0.13) Transformers Voting Ensemble RoBERTa 68.20 62.55 65.26 62.43 XLNet 67.84 60.07 63.72 59.92 Hierarchical concatenated model with attention on ILDCsingle train XLNet + BiGRU 76.85 76.31 76.55 (±0.0140) 76.32 (±2.43) Table 4: Prediction Results using different models. Some of the transformer and hierarchical models vary in performance across runs, we average out performance across 3 runs (variance in the parenthesis). tences were input6, and (iii) Chunk level embeddings (trained via Doc2Vec). We also trained Hierarchical Attention Network (HAN) (Yang et al., 2016) model. GloVe embeddings with BiGRU and 6last 150 sentences covered around 90% of the documents 4053 XLNet XLNet XLNet | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Dense . . . . . . . . . . . . . . . . . . . . . Chunk Input ids [CLS] embedding BiGRU BiGRU BiGRU BiGRU BiGRU BiGRU . . . . . . . . . . . . . . Concat BiGRU Half XLNet Half + Figure 2: Hierarchical XLNet architecture (XLNet + BiGRU) attention model gave the best performance (64% F1) among the sequential models. Sequential models trained on ILDCmulti and ILDCsingle have similar performances Transformer Models: We experimented with BERT (Devlin et al., 2019), DistilBERT (Sanh et al., 2019), RoBERTa (Liu et al., 2019), and XLNet (Yang et al., 2019b). Due to limitation on the number of input tokens to BERT and other transformer models, we experimented with different sections (begin tokens, middle tokens, end tokens, combinations of these) of the documents and as shown in Table 4, the last 512 tokens gave the best performance. In general, transformer models outperform classical and sequential models. RoBERTa gave the best performance (72% F1) and DistilBERT was the worst. We did not experiment with domain specific transformers like LEGALBERT (Chalkidis et al., 2020), since these have been trained upon US/EU legal texts, hence, they do not work well in the Indian setting as the legal systems are entirely different. Hierarchical Transformer Models: Taking inspiration from hierarchical topic prediction model (Chitkara et al., 2019), we developed Hierarchical Transformer model architecture (Chalkidis et al., 2019). We divided each document into chunks using a moving window approach where each chunk was of length 512 tokens, and there was an overlap of 100 tokens. We obtained the [CLS] representation of these chunks, which were then used as input to sequential models (BiGRU + attention) or feed-forward model (CNN (Kim, 2014)). We also tried an ensemble of individual transformer models on each of the chunks. In general, all the hierarchical models outperform transformer models. The best performing model (78% F1) for predicting the case decision is XLNet with BiGRU on the top (Figure 2). Comparing best model accuracy with average annotator accuracy (78% vs. 94%) indicates the task’s inherent complexity and motivates more research in this direction. 5.2 Case Decision Explanation We experimented with a variety explainability algorithms as a post-prediction step. We experimented with the best judgment prediction model (Hierarchical Transformer (XLNet + BiGRU)) for all the explainable algorithms. We explored three class of explainability methods (Xie et al., 2020): attribution based, model agnostic, and attention-based. In the class of attribution based methods, Layerwise Relevance Propagation (LRP) (Bach et al., 2015) and DeepLIFT (Shrikumar et al., 2017) methods did not work in our case. Due to the long length of documents, model agnostic explainability methods like LIME (Ribeiro et al., 2016) and Anchors (Ribeiro et al., 2018) were not applicable. We also experimented with attention-based methods, and Integrated Gradients (Sundararajan et al., 2017) method using the CAPTUM library (Kokhlikyan et al., 2019). However, these highlighted only a few tokens or short phrases. Moreover, attention-based scores are not necessarily indicative of explanations (Jain and Wallace, 2019). To extract explanations, we propose a method inspired from Li et al. (2016) and Zeiler and Fergus (2014). The idea is to use the occlusion method at both levels of the hierarchy. For each document, for the BiGRU part of the model, we mask each complete chunk embedding one at a time. The masked input is passed through the trained BiGRU, and the output probability (masked probability) of the label obtained by the original unmasked model is calculated. The masked probability is compared with unmasked probability to calculate the chunk explainability score. Formally, for a chunk c, if the sigmoid outputs (of the BiGRU) are σm (when the chunk was not masked) and σm′ (when the chunk was masked) and the predicted label is y then the probabilities and chunk score sc = pm −pm′ and pm′/m = ( σm′/m, y = 1 1 −σm′/m, y = 0 We obtain sentences that explain the decision from the transformer part of the model (XLNet) using the chunks that were assigned positive scores. Each chunk (length 512 tokens) is segmented into sentences using NLTK sentence splitter (Loper and Bird, 2002). Similar to BiGRU, each sentence is masked and the output of the transformer at the classification head (softmax logits) is compared 4054 Metric Explainability Model vs Experts Expert 1 2 3 4 5 Jaccard Similarity 0.333 0.317 0.328 0.324 0.318 Overlap-Min 0.744 0.589 0.81 0.834 0.617 Overlap-Max 0.39 0.414 0.36 0.35 0.401 ROUGE-1 0.444 0.517 0.401 0.391 0.501 ROUGE-2 0.303 0.295 0.296 0.297 0.294 ROUGE-L 0.439 0.407 0.423 0.444 0.407 BLEU 0.16 0.28 0.099 0.093 0.248 Meteor 0.22 0.3 0.18 0.177 0.279 Table 5: Machine explanations v/s Expert explanations with logits of the label corresponding to original hierarchical model. The difference between the logits normalized by the length of the sentence is the explanation score of the sentence. Finally, top-k sentences (∼40%) in each chunk are selected. To understand and analyze which parts of the documents were contributing towards prediction, we examined the attention weights (scores) in the case of the XLNet+BiGRU+Attention model and the occlusion scores of the XLNet+BiGRU model. Plots for some of the documents are shown in Figure 3. Plots for different chunk sizes are provided in Data/images folder in our GitHub repository. We also provide the t-SNE visualization on the test set using the BERT and Doc2Vec embeddings. Token visualization heatmap using Integrated Gradient for document name 1951 33.txt for BERT model is also provided in GitHub. Plots of scores averaged out over the entire test set for each chunk size can be visualized in appendix B.2. Two things can be noted: firstly, the largest attention and occlusion scores are assigned to chunks corresponding to the end of the document; this is in line with our hypothesis that most of the important information and rationale for judgment is mainly towards the end of the document. Secondly, although attention scores are optimized (via loss minimization or accuracy maximization) to concentrate on the last chunks, this is not the case with occlusion scores. There is no optimization of occlusion scores; yet they still focus on the chunks at the end, which affirms our hypothesis. 5.3 Model Explainability versus Annotators We compare the performance of occlusion method explanations with the expert annotators’ gold explanations by measuring the overlap between the two. We used the same measures (§ 4.2) ROUGE-L, ROUGE-1, ROUGE-2, Jaccard Similarity, BLEU, METEOR, Overlap Maximum, and Overlap Minimum Table 5 compares machine explanations with Figure 3: Averaged chunk scores for attention and occlusion the gold explanations. The highest overlap value (0.8337) is observed for the measure Overlap-Min with Expert 4. The values for Overlap-Min depict high agreements of the explainability model with all the experts. However, the values for the other evaluation measures, e.g., ROUGE-L, are in the low to medium range, the highest being 0.4445 for ROUGE-L and Expert 4. The results show the wide gap between how a machine would explain a judgment and the way a legal expert would explain it. The results motivate us for future research in this direction of developing an explainable model. 6 Conclusion This paper introduces the ILDC corpus and corresponding CJPE task. The corpus is annotated with case decisions and explanations for the decisions for a separate test set. Analysis of the corpus and modeling results shows the complexity of legal documents that pose challenges from a computational perspective. We hope that the corpus and the task would provide a challenging and interesting resource for the Legal NLP researchers. For future work, we would like to train a legal transformer similar to LEGAL-BERT (Chalkidis et al., 2020) on our Indian legal case documents. Moreover, we would also like to focus upon using rhetorical roles Bhattacharya et al. (2019b) of the sentences to include structural information of the documents for CJPE task as well. Acknowledgements We would like to thank anonymous reviewers for their insightful comments. We would like to thank student research assistants Abin Thomas Alex, Amrita Ghosh, Parmeet Singh, and Unnati Jhunjhunwala from West Bengal National University of Juridical Sciences (WBNUJS) for annotating the documents. This work would not have been possible without their help. 4055 Ethical Concerns The corpus is created from publicly available data: proceedings of Supreme Court of India (SCI). The data was scraped from the website: www. indiankanoon.org. The website allows scrapping of the data and no copyrights were infringed. Annotators were selected randomly and they participated voluntarily. The proposed corpus aims to promote the development of an explainable case judgment prediction system. The system intends to assist legal professionals in their research and decision-making and not replace them. Therefore, ethical considerations such as allowing legal rights and obligations of human beings to be decided and pronounced upon by non-human intelligence are not being breached by the system. The system proposes to provide valuable information that might be useful to a legal professional to make strategic decisions, but the actual decision-making process is still going to be carried out by the professional himself. Therefore, the system is not intended to produce a host of artificial lawyers and judges regulating human behavior. At the same time, the final expert human analysis of the systemic output should ensure that any existing flaw, absurdity, or overt or latent bias gets subjected to an additional layer of ethical scrutiny. In this way, the usual ethical concerns associated with the concept of case-law prediction also get addressed to a considerable extent since the system is not performing any judicial role herein nor deciding the legal rights or liabilities of human beings. Instead, the system is purported to be used primarily by legal professionals to make strategic decisions of their own, said decisions being still subjected to legal and judicial scrutiny performed by human experts. Nevertheless, the community needs to pursue more research in this regard to fully understand the unforeseen social implications of such system. This paper takes initial steps by introducing the corpus and baseline models to the community. Care has been taken to select cases in a completely random manner, without any particular focus on the type of law or the identities or sociopolitico-economic background of the parties or the judges involved. Specifically, the aforementioned identities have been deliberately anonymized so as to minimize or eliminate any possible bias in the course of prediction. The subjectivity that is associated with the judicial decision-making may also be controlled in this way, since the system is focusing on how consideration of the facts and applicable law are supposed to determine the outcome of the cases, instead of any individual bias on the judge’s part; another judge might not share such bias, and therefore the only common point of reference that the two judges would have would be the relevant facts of the case and the laws involved. This also gets reflected in the objective methodology used in the selection of annotators and by eliminating any interaction between the annotators themselves while at the same time paying attention to the factors or observations common to the output from the various annotators. The only specification with regard to the forum has been made by taking all the cases from the domain of the Supreme Court of India, owing to the propensity of the apex court of the land towards focusing on the legalities of the issues involved rather than rendering mere fact-specific judgments, as well as the binding nature of such decisions on the subordinate courts of the land. This would also allow the results to be further generalized and applied to a broader set of cases filed before other forums, too, since the subordinate courts are supposed to follow the reasoning of the Supreme Court’s judgments to the greatest possible extent. As a result, the impact of the training and testing opportunities provided to the system by a few Supreme Court cases is likely to be much greater than the mere absolute numbers would otherwise suggest. References Nikolaos Aletras, Dimitrios Tsarapatsanis, Daniel Preotiuc-Pietro, and Vasileios Lampos. 2016. Predicting judicial decisions of the European Court of Human Rights:A Natural Language Processing perspective. PeerJ Computer Science, 2:93. Sebastian Bach, Alexander Binder, Gr´egoire Montavon, Frederick Klauschen, Klaus-Robert M¨uller, and Wojciech Samek. 2015. On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one, 10(7):e0130140. Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. Paheli Bhattacharya, Kaustubh Hiware, Subham Rajgaria, Nilay Pochhi, Kripabandhu Ghosh, and Saptarshi Ghosh. 2019a. A comparative study of summarization algorithms applied to legal case judgments. In European Conference on Information Retrieval, pages 413–428. Springer. 4056 Paheli Bhattacharya, Shounak Paul, Kripabandhu Ghosh, Saptarshi Ghosh, and Adam Wyner. 2019b. Identification of Rhetorical Roles of Sentences in Indian Legal Judgments. In Legal Knowledge and Information Systems - JURIX 2019, volume 322 of Frontiers in Artificial Intelligence and Applications, pages 3–12. IOS Press. Richard Eckart de Castilho, Eva Mujdricza-Maydt, Seid Muhie Yimam, Silvana Hartmann, Iryna Gurevych, Anette Frank, and Chris Biemann. 2016. A web-based tool for the integrated annotation of semantic and syntactic structures. In Proceedings of the Workshop on Language Technology Resources and Tools for Digital Humanities (LT4DH), pages 76–84. Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. 2019. Neural Legal Judgment Prediction in English. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4317–4323, Florence, Italy. Association for Computational Linguistics. Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of Law School. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2898– 2904, Online. Association for Computational Linguistics. Huajie Chen, Deng Cai, Wei Dai, Zehui Dai, and Yadong Ding. 2019. Charge-Based Prison Term Prediction with Deep Gating Network. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6362–6367, Hong Kong, China. Association for Computational Linguistics. Pooja Chitkara, Ashutosh Modi, Pravalika Avvaru, Sepehr Janghorbani, and Mubbasir Kapadia. 2019. Topic Spotting using Hierarchical Networks with Self Attention. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3755–3761, Minneapolis, Minnesota. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. J. L. Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin, 75(5). Filippo Galgani, Paul Compton, and Achim Hoffmann. 2012. Towards automatic generation of catchphrases for legal case reports. In International Conference on Intelligent Text Processing and Computational Linguistics, pages 414–425. Springer. Zikun Hu, Xiang Li, Cunchao Tu, Zhiyuan Liu, and Maosong Sun. 2018. Few-Shot Charge Prediction with Discriminative Legal Attributes. In Proceedings of the 27th International Conference on Computational Linguistics, pages 487–498, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Peter Jackson, Khalid Al-Kofahi, Alex Tyrrell, and Arun Vachher. 2003. Information extraction from case law and retrieval of prior cases. Artificial Intelligence, 150(1-2):239–290. Sarthak Jain and Byron C. Wallace. 2019. Attention is not Explanation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3543–3556, Minneapolis, Minnesota. Association for Computational Linguistics. Xin Jiang, Hai Ye, Zhunchen Luo, WenHan Chao, and Wenjia Ma. 2018. Interpretable Rationale Augmented Charge Prediction System. In Proceedings of the 27th International Conference on Computational Linguistics: System Demonstrations, pages 146–151, Santa Fe, New Mexico. Association for Computational Linguistics. Justice Markandey Katju. 2019. Backlog of cases crippling judiciary. https://tinyurl.com/ v4xu6mvk. Daniel Martin Katz, Michael J. Bommarito, II, and Josh Blackman. 2017. A general approach for predicting the behavior of the Supreme Court of the United States. PLOS ONE, 12:1–18. Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751, Doha, Qatar. Association for Computational Linguistics. Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The Efficient Transformer. In International Conference on Learning Representations. Narine Kokhlikyan, Vivek Miglani, Miguel Martin, Edward Wang, Jonathan Reynolds, Alexander Melnikov, Natalia Lunova, and Orion ReblitzRichardson. 2019. Pytorch Captum. https:// github.com/pytorch/captum. Alon Lavie and Abhaya Agarwal. 2007. METEOR: An automatic metric for MT evaluation with high levels of correlation with human judgments. In Proceedings of the second workshop on statistical machine translation, pages 228–231. 4057 Quoc Le and Tomas Mikolov. 2014. Distributed representations of sentences and documents. In International conference on machine learning, pages 1188– 1196. Jiwei Li, Will Monroe, and Dan Jurafsky. 2016. Understanding neural networks through representation erasure. arXiv preprint arXiv:1612.08220. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692. Shangbang Long, Cunchao Tu, Zhiyuan Liu, and Maosong Sun. 2019. Automatic judgment prediction via legal reading comprehension. In China National Conference on Chinese Computational Linguistics, pages 558–572. Springer. Edward Loper and Steven Bird. 2002. NLTK: the natural language toolkit. arXiv preprint cs/0205028. Bingfeng Luo, Yansong Feng, Jianbo Xu, Xiang Zhang, and Dongyan Zhao. 2017. Learning to Predict Charges for Criminal Cases with Legal Basis. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2727–2736, Copenhagen, Denmark. Association for Computational Linguistics. Arpan Mandal, Kripabandhu Ghosh, Arindam Pal, and Saptarshi Ghosh. 2017. Automatic catchphrase identification from legal court case documents. In Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, pages 2187– 2190. Matteo Pagliardini, Prakhar Gupta, and Martin Jaggi. 2018. Unsupervised Learning of Sentence Embeddings Using Compositional n-Gram features. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 528–540, New Orleans, Louisiana. Association for Computational Linguistics. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12:2825–2830. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. ”Why should I trust you?” Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144. Marco T´ulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Anchors: High-Precision ModelAgnostic Explanations. In Proceedings of the ThirtySecond AAAI Conference on Artificial Intelligence, (AAAI-18), pages 1527–1535. AAAI Press. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. Avanti Shrikumar, Peyton Greenside, and Anshul Kundaje. 2017. Learning important features through propagating activation differences. In International Conference on Machine Learning, pages 3145–3153. PMLR. Benjamin Strickson and Beatriz de la Iglesia. 2020. Legal Judgement Prediction for UK Courts. ICISS 2020: The 3rd International Conference on Information Science and System, Cambridge, UK, March 19-22, 2020, pages 204–209. Octavia-Maria S¸ulea, Marcos Zampieri, Mihaela Vela, and Josef van Genabith. 2017. Predicting the Law Area and Decisions of French Supreme Court Cases. In Proceedings of the International Conference Recent Advances in Natural Language Processing, RANLP 2017, pages 716–722, Varna, Bulgaria. INCOMA Ltd. Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In International Conference on Machine Learning, pages 3319–3328. PMLR. Vu Tran, Minh Le Nguyen, and Ken Satoh. 2019. Building Legal Case Retrieval Systems with Lexical Matching and Summarization Using A Pre-Trained Phrase Scoring Model. In Proceedings of the Seventeenth International Conference on Artificial Intelligence and Law, ICAIL ’19, page 275–282, New York, NY, USA. Association for Computing Machinery. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. 4058 Pengfei Wang, Yu Fan, Shuzi Niu, Ze Yang, Yongfeng Zhang, and Jiafeng Guo. 2019. Hierarchical Matching Network for Crime Classification. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2019, Paris, France, July 21-25, 2019, pages 325–334. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Chaojun Xiao, Haoxi Zhong, Zhipeng Guo, Cunchao Tu, Zhiyuan Liu, Maosong Sun, Yansong Feng, Xianpei Han, Zhen Hu, Heng Wang, et al. 2018. Cail2018: A large-scale legal dataset for judgment prediction. arXiv preprint arXiv:1807.02478. Ning Xie, Gabrielle Ras, Marcel van Gerven, and Derek Doran. 2020. Explainable deep learning: A field guide for the uninitiated. arXiv preprint arXiv:2004.14545. Nuo Xu, Pinghui Wang, Long Chen, Li Pan, Xiaoyan Wang, and Junzhou Zhao. 2020. Distinguish Confusing Law Articles for Legal Judgment Prediction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3086–3095, Online. Association for Computational Linguistics. Wenmian Yang, Weijia Jia, Xiaojie Zhou, and Yutao Luo. 2019a. Legal Judgment Prediction via Multi-Perspective Bi-Feedback Network. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019, pages 4085– 4091. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019b. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5753– 5763. Zichao Yang, Diyi Yang, Chris Dyer, Xiaodong He, Alex Smola, and Eduard Hovy. 2016. Hierarchical attention networks for document classification. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 1480–1489. Hai Ye, Xin Jiang, Zhunchen Luo, and Wenhan Chao. 2018. Interpretable Charge Predictions for Criminal Cases: Learning to Generate Court Views from Fact Descriptions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1854–1864, New Orleans, Louisiana. Association for Computational Linguistics. Matthew D Zeiler and Rob Fergus. 2014. Visualizing and understanding convolutional networks. In European conference on computer vision, pages 818–833. Springer. Haoxi Zhong, Zhipeng Guo, Cunchao Tu, Chaojun Xiao, Zhiyuan Liu, and Maosong Sun. 2018. Legal Judgment Prediction via Topological Learning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3540–3549, Brussels, Belgium. Association for Computational Linguistics. Haoxi Zhong, Yuzhong Wang, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and Maosong Sun. 2020. Iteratively Questioning and Answering for Interpretable Legal Judgment Prediction. Proceedings of the AAAI Conference on Artificial Intelligence, 34(01):1250–1257. 4059 Appendix A Annotations and Case studies: Agreement in Judgment Prediction for Annotators Annotation Assignment 1954 13: In this case, although the original decision is that the appeal has been rejected, Experts 1-4 have reached the decision that it has been accepted, while Expert 5 has decided that it has been rejected. This discrepancy appears to owe its origin to the very nature of the case and the issues considered by the court. There had been more than one such issue and separate arguments had been made by appellant in favour of each of such issue and associated prayer. The court appears to have agreed to some of the arguments and disagreed with the rest. Annotation Assignment 1961 417: In this case, although the original decision is that the appeal has been rejected, Experts 2 and 4 have decided that it has been accepted. Expert 2 appears to have misconstrued certain positions of law and relied unduly upon one of the other cases being cited as precedent (but not considered relevant by the Supreme Court), which might account for the divergence. In case of Expert 4, however, the issue appears to be more of a linguistic matter. Expert 4 has referred to a particular statement made by the court, “The main question that arises in this appeal is whether an illegitimate son of a sudra vis-a-vis his self acquired property, after having succeeded to a half share of his putative fathers estate, will be entitled to succeed to the other half share got by the widow, after the succession opened out to his putative father on the death of the said widow.” From this sentence, Expert 4 has drawn the inference that the appellant was the one asking to establish such entitlement. Since the court in subsequent comments agreed that such entitlement does exist, Expert 4 inferred that the appeal had been accepted. However, in reality, the appellant had been contesting such entitlement. Annotation Assignment 1962 47: In this case, although the original decision is that the appeal has been rejected, Experts 2 and 5 have decided that it has been accepted. This discrepancy appears to owe its origin to both of them having been misled by Sentence 17 of the case, which appears to refer to the Supreme Court having accepted an appeal and merely giving reasons for such order in the present case. However, the case in point was actually arising from an application for review of the court’s earlier judgment (acceptance of the appeal), and therefore, when the court was affirming its earlier judgment and giving reasons behind it, it was in reality rejecting this present application for review, that had been made by the party (respondent in the original appeal) aggrieved by the acceptance of such appeal by the court earlier. Experts 2 and 5 could not apparently distinguish the appeal from the review petition and that appears to have led to such discrepancy. B Models Details Table 6 summarizes hyperparameter settings for all the models. All the experiments were run on Google Colab7 and used the default single GPU Tesla P100-PCIE-16GB, provided by Colab. B.1 Case Prediction Model Details Classical Models: We considered classical ML models like Logistic Regression, SVM, and Random Forest. We used sentence embeddings via Sent2Vec (Pagliardini et al., 2018) and document embeddings via Doc2Vec (Le and Mikolov, 2014) as input features. Both embeddings were trained on ILDCmulti as our data is domain-specific. Legal proceedings are typically long documents, we tried out extractive summarization methods (as described in Bhattacharya et al. (2019a)) for gleaning relevant information from the documents and passing these as input to neural models. However, this approach also resulted in classifiers that were no better than random classifier. We also experimented by using TF-IDF vectors with the classical models like Logistic Regression (LR), Random Forests (RF) and Support Vector Machines (SVM) from the scikit-learn library in python (Pedregosa et al., 2011). However, the results were no better than a random classifier, which, according to us, could be due to the huge length of the documents and they were not able to capture such long term dependencies well enough. Results: Classical models based on logistic regression and Sent2Vec embeddings performed much worse than the one based on Doc2vec embeddings. It is interesting to see that Doc2Vec+LR has performance competitive to Sequential models. The simple word embedding based model has 7https://colab.research.google.com/ 4060 similar performance as the more complicated hierarchical attention network model (HAN). The best results are recorded in the Table 4, each for Sent2Vec and Doc2Vec. Sequential Models: We experimented with standard BiGRU (2 layers) with attention model. We tried 3 different types of embeddings: (i) Word level trained GloVe embeddings (Pennington et al., 2014), with last 512 tokens as input, (ii) Sentence level embeddings (Sent2Vec), where last 150 sentences were input8, and (iii) Chunk level embeddings (trained via Doc2Vec). Both Sequential models and HAN were trained on both ILDCmulti and ILDCsingle. All the models from here on were trained on Colab9. We extracted catchphrases (Mandal et al., 2017) from the ILDCsingle (we could not use this method on ILDCmulti due to requirement of huge compute resources). After extracting these catchphrases we ranked the sentences from the documents accordingly and used upto 200 sentences only10. These top 200 sentences were then mapped to their Sent2Vec embeddings and passed through BiGRU as above. Results: Sequential models trained on ILDCmulti and ILDCsingle have similar performances. We also experimented with extracting key sentences from ILDCsingle documents with the help of catchphrases and using these sentences as input (via the Sent2Vec embeddings) to a sequence model. Extracting the key sentences performs better than the using all the sentences but the performance is worse (61% versus 64% F1) than using GloVe embeddings on last 512 words. GloVe embeddings with BiGRU and attention model gave the best performance (64% F1) among the sequential models. The GloVe embeddings (last 512 tokens) with BiGRU + Attention gave the best results among the models mentioned above. Transformer Models: Recently, SOTA language models have been developed using Transformer Architectures (Vaswani et al., 2017). A number of transformer architectures have been introduced recently. We experimented with BERT (Devlin et al., 2019), DistilBERT (Sanh et al., 2019), RoBERTa (Liu et al., 2019), and XLNet (Yang et al., 2019b). We used HuggingFace library (Wolf et al., 2020) to fine tune BASE models of above transformers 8last 150 sentences covered around 90% of the documents 9https://colab.research.google.com/ 10These covered more than 90% of the ILDCsingle. from HuggingFace (Wolf et al., 2020) on the last 512 tokens of ILDCmulti11. Due to high compute requirements we could not utilize Longformer (Beltagy et al., 2020) and Reformer (Kitaev et al., 2020) models developed especially for long documents. For the other transformer models we used only the last 512 tokens as input. Results: Among the combinations of input tokens, the best performance was obtained by using last 512 tokens as input to the BERT Base model. We can observe the trend that the more the tokens from the final parts of the document are taken as input, the better is the prediction performance. This observation agrees with the fact that there are more clues towards the correct prediction in the final parts of the document (since Arguments, Ratio of the decision etc. Bhattacharya et al. (2019b) most aligned to the judgment are expected to appear more towards the end, closer to the judgment). As for the comparison between different transformers, unsurprisingly, RoBERTa and XLNet perform better than BERT in the prediction sub-task. Similarly, among DistilBERT and BERT, the latter outperforms the other. Hierarchical Models: In order to use transformers hierarchically, it was first necessary to fine-tune these models on the downstream task of classification. We use two different strategies to fine-tune these: • On ILDCmulti: Using last 512 tokens only from the documents. • On ILDCsingle: We fine-tune the transformer by dividing each document into chunks of 512 with an overlap of 100 tokens, the label for each chunk is given as the whole document label. Then we extracted the 768 dimension, [CLS] token embeddings from the transformers for each chunk in all the documents. This was done on ILDCmulti corpus irrespective of whether it was fine-tuned on ILDCmulti or ILDCsingle. As mentioned in (Devlin et al., 2019) we also experimented with concatenating the last 4 hidden layers of the [CLS] token and taking that as the chunk embedding. After getting the chunk embeddings we used two types of neural networks: BiGRU and CNN. For some models, the results varied over multiple runs. For these we recorded their mean and variance on F1 and Accuracy in the table 4. 11As shown in Table 4, we also experimented with different sections of documents and we observed last 512 tokens gave the best performance 4061 Results: Information is lost in considering only the last portion of the case proceeding for prediction and this is reflected in the performance of hierarchical models. In general, all the hierarchical models outperform transformer models. Adding attention on top of BiGRU in the hierarchical model does not boost the performance significantly. However, adding a CNN (instead of BiGRU + Attention) on top gives a competitive performance. As for the comparison between the strategies of fine-tuning between ILDCmulti and ILDCsingle, the later seemed to perform worse on prediction. For the hierarchical concatenated model fine tuned on ILDCsingle, there was a slight boost in performance. B.2 Explanability Models and Results Details To extract explanations from our best model (XLNet + BiGRU), we propose a method inspired from Li et al. (2016) and Zeiler and Fergus (2014). The idea is to use occlusion method at both levels of the hierarchy. For the BiGRU part of the model, for each document we mask each complete chunk embedding one at a time. The masked input is passed through the trained BiGRU and output probability (masked probability) of the label obtained by original unmasked model is calculated. The masked probability is compared with unmasked probability to calculate chunk explainability score. Formally, for a chunk c, if the sigmoid outputs (of the BiGRU) are σm (when the chunk was not masked) and σm′ (when the chunk was masked) and the predicted label is y then the probabilities and chunk score sc = pm −pm′ and pm′/m = ( σm′/m, y = 1 1 −σm′/m, y = 0 We obtain sentences that explain the decision from the transformer part of the model (XLNet) using the chunks that were assigned positive scores. Each chunk (length 512 tokens) is segmented into Figure 4: Visualization of Occlusion scores accross full Test set. Figure 5: Visualization of Attention scores accross full Test set. 4062 sentences using NLTK sentence splitter (Loper and Bird, 2002). Similar to BiGRU, each sentence is masked and the output of the transformer at the classification head (softmax logits) is compared with logits of the label corresponding to original hierarchical model. The difference between the logits normalized by the length of the sentence is the explanation score of the sentence. Finally, top-k sentences (∼40%) in each chunk are selected. In Figure 4 and Figure 5 we visualize the mean chunk importance scores. Out of the 1517 test documents we average out chunk scores of the documents having same number of chunks. As shown in Figure 5, the attention weights are biased towards the last chunks, thus giving negligible attention to the chunks before. However, in Figure 4, in some of the graphs, the last chunk is given the secondhighest score and in 7 out of 10 graphs, it has the highest score. Due to space limitation, we are not providing the graphs for occlusion and attention scores for chunks 1 to 15. But we observed that for these chunks pattern matches for occlusion scores with attention scores. From these observations, we believe it is safe to say that both the methods of visualization affirm our hypothesis that the most relevant syntactic and semantic information lies towards the end of the case. Although attention scores are optimized (via loss minimization or accuracy maximization) to concentrate on last chunks, this is not the case with occlusion scores. There is no optimization of occlusion scores, yet they still focus on the chunks at the end which affirms our hypothesis. One might argue that this observation might be due to the transformer being trained on last 512 tokens only. To check this, we also visualized the hierarchical transformers trained on ILDCsingle, but the results were similar as to what we have observed in this case. Model Hyper-Parameters (E = Epochs), (Dim = Embedding Dimension), (L = Layers), (att. = attention), (default setting= 512 tokens with overlapping 100 tokens) Classical Models on ILDCmulti train set Doc2Vec + LR dim = 1000 , E = 20 Sent2vec + LR dim=500, E = 20, Avg Pool Sequential Models on ILDCmulti train set Sent2vec + BiGRU + att. dim = 200, E = 1, L = 2 Doc2vec + BiGRU + att. dim = 1000, E = 2, L = 2 GloVe + BiGRU + att. dim = 180, E = 3, L = 2 HAN word dim = 100, sent dim = 100, E = 10 Sequential Models on ILDCsingle train set Sent2Vec + BiGRU+ att. dim = 200, E = 1, L = 2 Doc2vec + BiGRU + att. dim = 1000, E = 2, L = 2 GloVe + BiGRU + att. dim = 180, E = 10, L = 2 HAN word dim = 100, sent dim = 100, E = 10 Catchphrases + Sent2Vec + BiGRU + att. dim =180, E =5, L = 2 Transformer Models on ILDCmulti train set BERT Base 512 begin tokens, E = 3 BERT Base 256 begin, 256 end tokens, E = 3 BERT Base 256 mid, 256 end tokens, E = 3 BERT Base 128 begin, 256 mid, 128 end, E = 3 BERT Base 512 end tokens, E = 3 DistillBERT 512 end tokens, E = 5 RoBERTa 512 end tokens, E = 5 XLNet 512 end tokens, E = 3 Hierarchical Models on ILDCmulti train set BERT + BiGRU default setting, E = 5, L = 3 RoBERTa + BiGRU default setting, E = 2, L = 3, runs = 3 XLNet + BiGRU default setting, E = 5, L = 2 BERT + CNN default setting, E = 3, L = 3 (Conv1D) RoBERTa + CNN default setting, E = 3, L = 3 (Conv1D) XLNet + CNN default setting, E = 3, L = 3 (Conv1D) Hierarchical Models on ILDCsingle train set BERT + BiGRU default setting, E = 1, L = 2, 3 runs RoBERTa + BiGRU default setting, E = 1, L = 2, 3 runs XLNet + BiGRU default setting, E = 2, L = 2, 3 runs Hierarchical Models with Attention on ILDCmulti train set BERT + BiGRU + att. default setting, E = 2, L = 2, 3 runs RoBERTa + BiGRU + att. default setting, E = 2, L = 3, 3 runs XLNet + BiGRU + att. default setting, E = 3, L = 2, 3 runs Hierarchical Models with Attention on ILDCsingle train set BERT + BiGRU + att. default setting, E = 1, L = 2, 3 runs RoBERTa + BiGRU + att. default setting, E = 1, L = 3, 3 runs XLNet + BiGRU + att. default setting, E = 1, L = 2, 3 runs Transformers Voting Ensemble RoBERTa fine tuned on last 512 tokens, voting XLNet fine tuned on last 512 tokens, voting Hierarchical concatenated model with att on ILDCsingle train XLNet + BiGRU last 4 layers concat, E = 1, L = 2, 3 runs Table 6: Hyper-parameters corresponding to every model.
|
2021
|
313
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4063–4077 August 1–6, 2021. ©2021 Association for Computational Linguistics 4063 Supporting Cognitive and Emotional Empathic Writing of Students Thiemo Wambsganss1,2, Christina Niklaus1, 3, Matthias S¨ollner4, Siegfried Handschuh1, 3 and Jan Marco Leimeister1, 4 1 University of St.Gallen {thiemo.wambsganss, christina.niklaus, siegfried.handschuh, janmarco.leimeister}@unisg.ch 2 Carnegie Mellon University [email protected] 3 University of Passau {christina.niklaus, siegfried.handschuh}@uni-passau.de 4 University of Kassel {soellner, leimeister}@uni-kassel.de Abstract We present an annotation approach to capturing emotional and cognitive empathy in student-written peer reviews on business models in German. We propose an annotation scheme that allows us to model emotional and cognitive empathy scores based on three types of review components. Also, we conducted an annotation study with three annotators based on 92 student essays to evaluate our annotation scheme. The obtained inter-rater agreement of α=0.79 for the components and the multi-π=0.41 for the empathy scores indicate that the proposed annotation scheme successfully guides annotators to a substantial to moderate agreement. Moreover, we trained predictive models to detect the annotated empathy structures and embedded them in an adaptive writing support system for students to receive individual empathy feedback independent of an instructor, time, and location. We evaluated our tool in a peer learning exercise with 58 students and found promising results for perceived empathy skill learning, perceived feedback accuracy, and intention to use. Finally, we present our freely available corpus of 500 empathy-annotated, student-written peer reviews on business models and our annotation guidelines to encourage future research on the design and development of empathy support systems. 1 Introduction Empathy is an elementary skill in society for daily interaction and professional communication and is therefore elementary for educational curricula (e.g., Learning Framework 2030 (OECD, 2018)). It is the “ability to simply understand the other person’s perspective [. . .] and to react to the obFigure 1: Empathy annotation scheme. First, a text paragraph is classified into a peer review component (strengths, weakness, improvement suggestions). Second, the same annotator is then scoring the cognitive and emotional empathy level of the components based on our annotation guideline on a 1-to-5 scale. served experiences of another,” (Davis, 1983, p.1)1. Empathy skills not only pave the foundation for successful interactions in digital companies, e.g., in agile work environments (Luca and Tarricone, 2001), but they are also one of the key abilities in the future that will distinguish the human workforce and artificial intelligence agents from one another (Poser and Bittner, 2020). However, besides the growing importance of empathy, research has shown that empathy skills of US college students decreased from 1979 to 2009 by more than thirty percent and even more rapidly between 2000 to 2009 (Konrath et al., 2011). On these grounds, the Organization for Economic Cooperation and Development (OECD) claims that the training for empathy skills should receive a more prominent role in today’s higher education (OECD, 2018). 1Being aware that empathy is a multidimensional construct, in this study, we focus on emotional and cognitive empathy (Spreng et al., 2009; Davis, 1983). 4064 To train students with regard to empathy, educational institutions traditionally rely on experiential learning scenarios, such as shadowing, communication skills training, or role playing (Lok and Foster, 2019; van Berkhout and Malouff, 2016). Individual empathy training is only available for a limited number of students since individual feedback through a student’s learning journey is often hindered due to large-scale lectures or the growing field of distance learning scenarios such as Massive Open Online Classes (MOOCs) (Seaman et al., 2018; Hattie and Timperley, 2007). One possible path for providing individual learning conditions is to leverage recent developments in computational linguistics. Language-based models enable the development of writing support systems that provide tailored feedback and recommendations (Santos et al., 2018), e.g., like those already used for argumentation skill learning (Wambsganss et al., 2020a, 2021b). Recently, studies have started investigating elaborated models of human emotions (e.g., Wang et al. (2016), Abdul-Mageed and Ungar (2017), Buechel and Hahn (2018), or Sharma et al. (2020)), but available corpora for empathy detection are still rare. Only a few studies address the detection and prediction of empathy in natural texts (Khanpour et al., 2017; Xiao et al., 2012), and, to the best of our knowledge, only one corpus is publicly available for empathy modelling based on news story reactions (Buechel et al., 2018). Past literature therefore lacks 1) publicly available empathy annotated data sets, 2) empathy annotation models based on rigorous annotation guidelines combined with annotation studies to assess the quality of the data, 3) the alignment of empathy in literature on psychological constructs and theories, and 4) an embedding and real-world evaluation of novel modelling approaches in collaborative learning scenarios (Ros´e et al., 2008). We introduce an empathy annotation scheme and a corpus of 500 student-written reviews that are annotated for the three types of review components, strengths, weaknesses, and suggestions for improvements, and their embedded emotional and cognitive empathy level based on psychological theory (Davis, 1983; Spreng et al., 2009). We trained different models and embedded them as feedback algorithms in a novel writing support tool, which provided students with individual empathy feedback and recommendations in peer learning scenarios. The measured empathy skill learning (Spreng et al., 2009), the perceived feedback accuracy (Podsakoff and Farh, 1989), and the intention to use (Venkatesh and Bala, 2008) in a controlled evaluation with 58 students provided promising results for using our approach in different peer learning scenarios to offer quality education independent of an instructor, time, and location. Our contribution is fourfold: 1) we derive a novel annotation scheme for empathy modeling based on psychological theory and previous work on empathy annotation (Buechel et al., 2018); 2) we present an annotation study based on 92 student peer reviews and three annotators to show that the annotation of empathy in student peer reviews is reliably possible; 3) to the best of our knowledge, we present the second freely available corpus for empathy detection in general and the first corpus for empathy detection in the educational domain based on 500 student peer reviews collected in our lecture about business innovation in German; 4) we embedded our annotation approach as predictive models in a writing support system and evaluated it with 58 students in a controlled peer learning scenario. We hope to encourage research on student-written empathetic texts and writing support systems to train students’ empathy skills based on NLP towards a quality education independent of a student’s location or instructors. 2 Background The Construct of Empathy The ability to perceive the feelings of another person and react to their emotions in the right way requires empathy – the ability “of one individual to react to the observed experiences of another” (Davis (1983), p.1). Empathy plays an essential role in daily life in many practical situations, such as client communication, leadership, or agile teamwork. Despite the interdisciplinary research interest, the term empathy is defined from multiple perspectives in terms of its dimensions or components (Decety and Jackson, 2004). Aware of the multiple perspectives on empathy, in this annotation study, we focused on the cognitive and emotional components of empathy as defined by Davis (1983) and Lawrence et al. (2004). Therefore, we follow the ‘Toronto Empathy Scale’ (Spreng et al., 2009) as a synthesis of instruments for measuring and validating empathy. Hence, empathy consists of both emotional and cognitive components (Spreng et al., 2009). While emotional empathy lets us perceive what 4065 other people feel, cognitive empathy is the human ability to recognize and understand other individuals (Lawrence et al., 2004). Emotion and Empathy Detection In NLP, the detection of empathy in texts is usually regarded as a subset of emotion detection, which in turn is often referred to as part of sentiment analysis. The detection of emotions in texts has made major progress, with sentiment analysis being one of the most prominent areas in recent years (Liu, 2015). However, most scientific studies have been focusing on the prediction of the polarity of words for assessing negative and positive notions (e.g., in online forums (Abbasi et al., 2008) or twitter postings (Rosenthal et al., 2018)). Moreover, researchers have also started investigating more elaborated models of human emotions (e.g., Wang et al. (2016), Abdul-Mageed and Ungar (2017), and Mohammad and Bravo-Marquez (2017)). Several corpora exist where researchers have annotated and assessed the emotional level of texts. For example, Scherer and Wallbott (1994) published an emotionlabelled corpus based on seven different emotional states. Strapparava and Mihalcea (2007) classified news headlines based on the basic emotions scale of Ekman (1992) (i.e., anger, disgust, fear, happiness, sadness and surprise). More recently, Chen et al. (2018) published EmotionLines, an emotion corpus of multi-party conversations, as the first data set with emotion labels for all utterances was only based on their textual content. Bostan and Klinger (2018) presented a novel unified domainindependent corpus based on eleven emotions as the common label set. However, besides the multiple corpora available for emotion detection in texts, corpora for empathy detection are rather rare. As Buechel et al. (2018) also outline, the construction of corpora for empathy detection and empathy modelling might be less investigated due to various psychological perspectives on the construct of empathy. Most of the works for empathy detection focus, therefore, on spoken dialogue, addressing conversational agents, psychological interventions, or call center applications (e.g., McQuiggan and Lester (2007), P´erez-Rosas et al. (2017), Alam et al. (2018), Sharma et al. (2020)) rather than written texts. Consequently, there are hardly any corpora available in different domains and languages that enable researchers in training models to detect the empathy level in texts, e.g., by providing students with individual empathy feedback (Buechel et al., 2018). Empathy Annotated Corpora and Annotation Schemes Only a few studies address the detection and prediction of empathy in natural language texts (e.g., Khanpour et al. (2017) and Xiao et al. (2012)). Presenting the first and only available gold standard data set for empathy detection, Buechel et al. (2018) constructed a corpus in which crowdworkers were asked to write emphatic reactions to news stories. Before the writing tasks, the crowdworkers were asked to conduct a short survey with self-reported items to measure their empathy level and their personal distress based on Batson et al. (1987). The scores from the survey were then taken as the annotation score for the overall news reaction message. The final corpus consisted of 1,860 annotated messages (Buechel et al., 2018). Nevertheless, previous empathy annotations on natural texts merely focused on intuition-based labels instead of rigorous annotation guidelines combined with annotation studies by researchers to assess the quality of the corpora (i.e., as is done for corpora of other writing support tasks, e.g., argumentative student essays by Stab and Gurevych (2017)). Moreover, previous annotations have mostly been conducted at the overall document level, resulting in one generic score for the whole document, which makes the corpus harder to apply to writing support systems. Consequently, there is a lack of linguistic corpora for empathy detection in general and, more specifically, for training models that provide students with adaptive support and feedback about their empathy in common pedagogical scenarios like large-scale lectures or the growing field of MOOCs (Wambsganss et al., 2021c, 2020b). In fact, in the literature about computer-supported collaborative learning (Dillenbourg et al., 2009), we found only one approach by Santos et al. (2018) that used a dictionary-based approach to provide students with feedback on the empathy level of their texts. We aim to address this literature gap by presenting and evaluating an annotation scheme and an annotated empathy corpus built on studentwritten texts with the objective to develop intelligent and accurate empathy writing support systems for students. 3 Corpus Construction We compiled a corpus of 500 student-generated peer reviews in which students provided each other 4066 with feedback on previously developed business models. Peer reviews are a modern learning scenario in large-scale lectures, enabling students to reflect on their content, receive individual feedback from peers, and thus deepen their understanding of the content (Rietsche and S¨ollner, 2019). Moreover, they are easy to set up in traditional largescale learning scenarios or the growing field of distance-learning scenarios such as MOOCs. This can be leveraged to train skills such as the ability to appropriately react to other students’ perspectives (e.g., Santos et al. (2018)). Therefore, we aim to create an annotated corpus to provide empathy feedback based on a data set that A) is based on real-world student peer reviews, B) consists of a sufficient corpus size to be able to train models in a real-world scenario and C) follows a novel annotation guideline for guiding the annotators towards an adequate agreement. Hence, we propose a new annotation scheme to model peer review components and their emotional and cognitive empathy levels that reflect the feedback discourse in peer review texts. We base our empathy annotation scheme on emotional and cognitive empathy following Davis (1983) and Spreng et al. (2009) guided by the study of Buechel et al. (2018). To build a reliable corpus, we followed a 4-step methodology: 1) we examined scientific literature and theory on the construct of empathy and on how to model empathy structures in texts from different domains; 2) we randomly sampled 92 student-generated peer reviews and, on the basis of our findings from literature and theory, developed a set of annotation guidelines consisting of rules and limitations on how to annotate emphatic review discourse structures; 3) we applied, evaluated, and improved our guidelines with three native speakers of German in a total of eight consecutive workshops to resolve annotation ambiguities; 4) we followed the final annotation scheme based on our 14-page guidelines to annotate a corpus of 500 student-generated peer reviews.2 3.1 Data Source We gathered a corpus of 500 student-generated peer reviews written in German. The data was collected in a business innovation lecture in a master’s program at a Western European university. In this lecture, around 200 students develop and present a 2The annotation guidelines as well as the entire corpus can be accessed at https://github.com/thiemowa/ empathy_annotated_peer_reviews. new business model for which they receive three peer reviews each. Here, a fellow student from the same course elaborates on the strengths and weaknesses of the business model and gives recommendations on what could be improved. We collected a random subset of 500 of these reviews from around 7,000 documents collected from the years 2014 to 2018 in line with the ethical guidelines of our university and with approval from the students to utilize the writings for scientific purposes. An average peer review consists of 200 to 300 tokens (in our corpus we counted a mean of 19 sentences and 254 tokens per document). A peer review example is displayed in Figure 2. 3.2 Annotation Scheme Our objective is to model the empathy structures of student-generated peer reviews by annotating the review components and their emotional and cognitive empathy levels. Most of the peer reviews in our corpus followed a similar structure. They described several strengths or weaknesses of the business model under consideration, backing them up by examples or further elaboration. Moreover, the students formulated certain suggestions for improvements of the business model. These review components (i.e., strengths, weaknesses, and suggestions for improvement) were written with different empathetic levels, sometimes directly criticizing the content harshly, sometimes empathetically referring to weaknesses as further potentials for improvement with examples and explanation. We aim to capture these empathic differences between the peer reviews with two empathy level scores, the cognitive empathy level of a certain review component and the emotional empathy level of a certain component. Our basic annotation scheme is illustrated in Figure 1. 3.2.1 Review Components For the review components, we follow established models of feedback structures suggested by feedback theory (e.g., Hattie and Timperley (2007) or Black and Wiliam (2009)). A typical peer review, therefore, consists of three parts: 1) elaboration of strengths, 2) elaboration of weaknesses, and 3) suggestions for improvements (to answer “Where am I going and how am I going?” and “Where do I go next?”, i.e., Hattie and Timperley (2007)). Accordingly, the content of a review consists of multiple components, including several controversial statements (e.g., a claim about a strength or 4067 weakness of a business model) that are usually supported by elaborations or examples (i.e., a premise) (Toulmin, 1984). Also, in the domain of studentwritten peer reviews, we found that a standpoint and its elaboration are the central element of a review component. Accordingly, we summarized all the claims and premises which described positive aspects of a business model as strengths. All content (claims and premises) describing negative aspects were modelled as weaknesses, while claims and premises with certain content for improvement were modelled as suggestions for improvement, following the structure of a typical review. Besides the content, syntactical elements and key words were used as characteristics for the compound classification, e.g., most students introduced a review component by starting with structural indications such as ”Strengths:” or ”Weaknesses:” in their peer review texts. 3.2.2 Empathy Level To capture the differences in the empathy levels of the peer reviews (i.e., the way the writer was conveying their feedback (Hattie and Timperley, 2007)), we followed the approach of Davis (1983) and Spreng et al. (2009) for cognitive and emotional empathy. Cognitive empathy (perspective taking) is the writer’s ability to use cognitive processes, such as role taking, perspective taking, or “decentering,” while evaluating the peers’ submitted tasks. The student sets aside their own perspective and “steps into the shoes of the other.” Cognitive empathy can happen purely cognitively, in that there is no reference to any affective state, (BaronCohen and Wheelwright, 2004) but it mostly includes understanding the other’s emotional state as well. The following example displays high cognitive empathy: “You could then say, for example, ‘Since market services are not differentiated according to customer segments and locations, the following business areas result... And that due to the given scope of this task you will focus on the Concierge-Service business segment.’ After that, you have correctly only dealt with this business segment.” Emotional empathy (emphatic concern) is the writer’s emotional response to the peers’ affective state. The students can either show the same emotions as read in the review or simply state an appropriate feeling towards the peer. Typical examples include sharing excitement with the peer about the business model submitted or showing concern over the peer’s opinion. The following example depicts high emotional empathy: “I think your idea is brilliant!”. Both constructs are measured on a scale from 1-5 following the empathy scale range of Moyers and Martin (2010), with every level being precisely defined in our annotation guidelines. A summary of the definitions for both empathy level scores are displayed in Table 1 and Table 2. A more detailed description of both scores can be found in the appendix in Table 7 and Table 8.3 Figure 2 illustrates an example of an entire peer review that is annotated for strength, weakness and suggestion for improvement and the cognitive and emotional empathy scores.4 Figure 2: Fully annotated example of a peer review. 3.3 Annotation Process Three native German speakers annotated the peer reviews independently from each other for the components strengths, weaknesses and suggestions for improvement, as well as their cognitive and emotional empathy levels according to the annotation guidelines we specified. The annotators were master’s students in business innovation from a European university with bachelor’s degrees in business administration and were, therefore, domain experts in the field of business models. Inspired by Stab 3More elaborated definitions, examples, and key word lists for both empathy scales can be found in our annotation guidelines. 4Since the original texts are written in German, we translated the examples to English for the sake of this paper. 4068 ScoreDescription 5 The student fully understands the peer’s thoughts. She completely stepped outside her own perspective and thinks from the peer’s perspective. She does that by carefully evaluating the peer’s idea with rich explanations. Questions, personal pronouns, or direct addressing of the author could be used in order to better understand and elaborate on the peer’s perspective. 4 The student thinks from the perspective of the peer. She elaborates in a way that serves the peer best to further establish the idea or activity. Each component is affirmed with further explanations. 3 The student tries to understand the perspective of the peer and adds further elaborations to her statements. However, her elaborations are not completely thought through, and her feedback is missing some essential explanations, examples, or questions to make sure she understood everything correctly. 2 The student did not try to understand the peer’s perspective. The student rather just tried to accomplish the task of giving feedback. 1 The student’s feedback is very short and does not include the peer’s perspective. She does not add any further elaboration in her thoughts. Table 1: Description of the cognitive empathy scores. ScoreDescription 5 The student was able to respond very emotionally to the peer’s work and fully represents the affectional state in her entire review. She illustrates this by writing in a very emotional and personal manner and expressing her feelings (positive or negative) throughout the review. Strong expressions include exclamation marks (!). 4 The student was able to respond emotionally to the peer’s submitted activity with suitable emotions (positive or negative). She returns emotions in her feedback on various locations and expresses her feelings by using the personal pronouns (“I”, “You”). Some sentences might include exclamations marks (!). 3 The student occasionally includes emotions or personal emotional statements in the peer review. They could be quite strong. However, the student’s review is missing personal pronouns (“I”, “You”) and is mostly written in third person. Emotions can both be positive or negative. Negative emotions can be demonstrated with concern, missing understanding or insecurity (e.g., with modal verbs or words such as rather, perhaps). 2 Mostly, the student does not respond emotionally to the peer’s work. Only very minor and weak emotions or personal emotional statements are integrated. The student writes mostly objectively (e.g., “Okay”, “This should be added”, “The task was done correctly”, etc.). In comparison to level 1, she might be using modal verbs (might, could, etc.) or words to show insecurity in her feedback (rather, maybe, possibly). 1 The student does not respond emotionally to the peer’s work at all. She does not show her feelings towards the peer and writes objectively (e.g., no “I feel”, “Personally” “I find this...” and no emotions such as “good”, “great”, “fantastic”, “concerned”, etc.). Typical examples would be “Add a picture.” or “The value gap XY is missing.”. Table 2: Description of the emotional empathy scores. and Gurevych (2017), our guidelines consisted of 14 pages, including definitions and rules for how the review components should be composed, which annotation scheme was to be used, and how the cognitive and emotional empathy level were to be judged. Several individual training sessions and eight team workshops were performed to resolve disagreements among the annotators and to reach a common understanding of the annotation guidelines on the cognitive and emotional empathy structures. We used the tagtog annotation tool,5 which offers an environment for cloud-based annotation in a team. First, a text was classified into peer review components (strengths, weaknesses, suggestions for improvement, or none) by the trained annotators. Second, the same annotator then scored the cognitive and emotional empathy levels of each component based on our annotation guideline on a one to five scale. After the first 92 reviews were 5https://tagtog.net/ annotated by all three annotators, we calculated the inter-annotator agreement (IAA) scores (see Section 4.1).6 As we obtained satisfying results, we proceeded with two annotators annotating 130 remaining documents each and the senior annotator annotating 148 peer reviews, resulting in 408 additional annotated documents. Together with the 92 annotations of the annotation study of the senior annotator (the annotator with the most reviewing experience), we counted 500 annotated documents in our final corpus. 4 Corpus Analysis 4.1 Inter-Annotator Agreement To evaluate the reliability of the review components and empathy level annotations, we followed the approach of Stab and Gurevych (2014). 6Our intention was to capture the annotation of 100 randomly selected essays. However, we discarded 8 of the 100 essays as they contained less than 2 review components. 4069 Review Components Concerning the review components, two strategies were used. Since there were no predefined markables, annotators not only had to identify the type of review component but also its boundaries. In order to assess the latter, we use Krippendorff’s αU (Krippendorff, 2004), which allows for an assessment of the reliability of an annotated corpus, considering the differences in the markable boundaries. To evaluate the annotators’ agreement in terms of the selected category of a review component for a given sentence, we calculated the percentage agreement and two chance-corrected measures, multi-π (Fleiss, 1971) and Krippendorff’s α (Krippendorff, 1980). Since each annotation always covered a full sentence (or a sequence of sentences), we operated at the sentence level for calculating the reliability of the annotations in terms of the IAA. % Multi-π Kripp. α Kripp. αU Strength 0.9641 0.8871 0.8871 0.5181 Weakness 0.8893 0.7434 0.7434 0.3109 Suggestions 0.8948 0.6875 0.6875 0.3512 None 0.9330 0.8312 0.8312 0.9032 Table 3: IAA of review component annotations. Table 3 displays the resulting IAA scores. The obtained scores for Krippendorff’s α indicated an almost perfect agreement for the strengths components and a substantial agreement for both the weaknesses and the suggestions for improvement components. The unitized α of strengths, weaknesses and suggestions for improvement annotations was slightly smaller compared to the sentence-level agreement. Thus, the boundaries of review components were less precisely identified in comparison to the classification into review components. Yet the scores still suggest that there was a moderate level of agreement between the annotators for the strengths and a fair agreement for the weaknesses and the suggestions for improvement. With a score of αU=90.32%, the boundaries of the nonannotated text units were more reliably detected, indicating an almost perfect agreement between the annotators. Percentage agreement, multi-π, and Krippendorff’s α were considerably higher for the non-annotated spans as compared to the strengths, weaknesses, and suggestions for improvement, indicating an almost perfect agreement between the annotators. Hence, we conclude that the annotation of the review components in student-written peer reviews is reliably possible . Empathy Level To assess the reliability of the cognitive and emotional empathy level annotations, we calculated the multi-π for both scales. For the cognitive empathy level, we received a multi-π of 0.41 for both the emotional and cognitive empathy level, suggesting a moderate agreement between the annotators in both cases. Thus, we conclude that the empathy level can also be reliably annotated in student-generated peer reviews. To analyze the disagreement between the three annotators, we created a confusion probability matrix (CPM) (Cinkov´a et al., 2012) for the review components and the empathy level scores. The results can be found in Section C of the appendix. 4.2 Corpus Statistics The corpus we compiled consists of 500 studentwritten peer reviews in German that were composed of 9,614 sentences with 126,887 tokens in total. Hence, on average, each document had 19 sentences and 254 tokens. A total of 2,107 strengths, 3,505 weaknesses and 2,140 suggestions for improvement were annotated. Tables 4, 5, and 6 present some detailed statistics on the final corpus. total mean std dev min max median Sentences 9,614 19.23 10.39 1 85 17 Tokens 126,887 253.77 134.18 10 1026 228 Table 4: Distribution of sentences and tokens in the created corpus. Mean, std dev, min, max and median refer to the number of sentences and tokens per document. total mean std dev min max median % Str. 2,107 4.21 2.71 1 20 4 0.27 Weak. 3,505 7.01 6.10 0 41 5 0.45 Sug. 2,140 4.28 5.49 0 59 3 0.28 Table 5: Distribution of the review components. mean std dev min max median Cognitive EL 2.94 0.99 1 5 3 Emotional EL 3.22 1.03 1 5 3 Table 6: Distribution of the empathy level (EL) scores. Moreover, Figure 3 displays the distribution of the empathy scores in the annotated dataset. Both the cognitive and the emotional empathy levels approximately follow a normal distribution with a mean score of 2.94 and 3.22, respectively (see Table 6). We measured only a low correlation of 0.38 between the scores of cognitive and emotional empathy. 4070 Figure 3: Distribution of the cognitive (left) and emotional (right) empathy scores (1-5 scale). 5 Providing Students Adaptive Feedback Modelling Cognitive and Emotional Empathy The empathy detection task is considered a paragraph-based, multi-class classification task, where each paragraph is either considered to be a strength, weakness, or a suggestion for improvement and has a “non-empathic”, “neutral”, or “empathic” cognitive and emotional empathy level. Therefore, we assigned the levels of our cognitive and emotional empathy scores to three different labels: level 1 and 2 were assigned to a “nonempathic” text label, level 3 to a “neutral” label, and levels 4 and 5 to an“empathic” label . We split the data into 70% training, 20% validation, and 10% test data. To apply the model, the corpus texts were split into word tokens. The model performances were measured in terms of accuracy, precision, recall, and f1-score. We trained a predictive model following the architecture of Bidirectional Encoder Representations from Transformers (BERT) proposed by Devlin et al. (2018). We used the BERT model from deepset,7 since it is available in German and provides a deep pretrained model that was unsupervised while training on domain-agnostic German corpora (e.g., the German Wikipedia). The best performing paramenter combination for our BERT model incorporated a dropout probability of 10% and a learning rate of 3e-5, and the number of epochs were 3. After several iterations, we reached a micro f1-score of 74.96% for the detection of the emotional empathy level and 69.98% for the detection of the cognitive empathy level of a text paragraph. Moreover, we reached an f1-score of 94.83% to predict a text paragraph as a strength, a 64.28% to predict a text paragraph as a weakness, 7https://github.com/deepset-ai/FARM and 59.79% to predict suggestions for improvement. To ensure the validity of our BERT model, we benchmarked against bidirectional Long-ShortTerm-Memory-Conditional-Random-Fields classifiers (BiLSTM-CRF). In combination with the corresponding embeddings vocabulary (GloVe) (Pennington et al., 2014), our LSTM reached an unsatisfying f1-score of 61% for detecting the emotional empathy level and 51% for detecting the cognitive empathy level. Evaluation in a Peer Learning Setting We designed and built an adaptive writing support system that provides students with individual feedback on their cognitive and emotional empathy skills. The application is illustrated in Figure 4. We embedded our system into a peer writing exercise where students were asked to write a peer review on a business model. During this writing task, they received adaptive feedback on the cognitive and emotional empathy level based on our model. The evaluation was conducted as a web experiment facilitated by the behavioral lab of our university, and thus, designed and reviewed according to the ethical guidelines of the lab and the university. We received 58 valid results (mean age = 23.89, SD= 3.07, 30 were male, 28 female). The participants were told to read an essay about a business model of a peer student. Afterwards, they were asked to write a business model review for the peer by providing feedback on the strengths, weaknesses, and suggestions for improvement of the particular business model. After the treatment, we measured the intention to use (ITU) (Venkatesh and Bala, 2008) by asking three items. We also asked the participants to judge their perceived empathy skill learning (PESL) by asking two items that covered cognitive and emotional empathy skills (Spreng 4071 Figure 4: Screenshot of a trained model on our corpus as an adaptive writing support system. et al., 2009; Davis, 1983). Finally, we surveyed the perceived feedback accuracy (PFA) (Podsakoff and Farh, 1989) to control the accuracy of our model. All constructs were measured with a 1-to-7 point Likert scale (1: totally disagree to 7: totally agree, with 4 being a neutral statement).8 Furthermore, we asked three qualitative questions: “What did you particularly like about the use of the tool?”, “What else could be improved?”, and “Do you have any other ideas?” and captured the demographics. In total, we asked 13 questions. All participants were compensated with an equivalent of about 12 USD for a 25 to 30 minute experiment. Results Participants judged their empathy skill learning with a mean of 5.03 (SD= 1.05). Concerning the PFA, the subjects rated the construct with a mean of 4.93 (SD= 0.94). The mean value of intention to use of the participants using our application as a writing support tool in peer learning scenarios was 5.14 (SD= 1.14). The mean values of all three constructs were very promising when comparing the results to the midpoints. All results were better than the neutral value of 4, indicating a positive evaluation of our application for peer learning tasks. We also asked open questions in our survey to receive the participants’ opinions about the tool they used. The general attitude was very positive. Participants positively mentioned the simple and easy interaction, the distinction between cognitive and emotional empathy feedback, and the overall empathy score together with the adaptive feedback message several times. However, participants also said that the tool should provide even more detailed feedback based on more categories and should pro8The exact items are listed in the appendix. vide concrete text examples on how to improve their empathy score. We translated the responses from German and clustered the most representative responses in Table 16 in the appendix. 6 Conclusion We introduce a novel empathy annotation scheme and an annotated corpus of student-written peer reviews extracted from a real-world learning scenario. Our corpus consisted of 500 student-written peer reviews that were annotated for review components and their emotional and cognitive empathy levels. Our contribution is threefold: 1) we derived a novel annotation scheme for empathy modeling based on psychological theory and previous work for empathy modeling (Buechel et al., 2018); 2) we present an annotation study based on 92 student peer reviews and three annotators to show that the annotation of empathy in student peer reviews is reliably possible ; and 3) to the best of our knowledge, we present the second freely available corpus for empathy detection and the first corpus for empathy detection in the educational domain based on 500 student peer reviews in German. For future research, this corpus could be leveraged to support students’ learning processes, e.g., through a conversational interaction (Zierau et al., 2020). However, we would also encourage research on the ethical considerations of empathy detection models in userbased research (i.e., Wambsganss et al. (2021a)). We, therefore, hope to encourage future research on student-generated empathetic texts and on writing support systems to train empathy skills of students based on NLP towards quality education independent of a student’s location or instructors. 4072 References Ahmed Abbasi, Hsinchun Chen, and Arab Salem. 2008. Sentiment analysis in multiple languages: Feature selection for opinion classification in Web forums. ACM Transactions on Information Systems, 26(3):1– 34. Muhammad Abdul-Mageed and Lyle Ungar. 2017. EmoNet: Fine-grained emotion detection with gated recurrent neural networks. ACL 2017 - 55th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference (Long Papers), 1:718–728. Firoj Alam, Morena Danieli, and Giuseppe Riccardi. 2018. Annotating and modeling empathy in spoken conversations. Computer Speech and Language, 50:40–61. Simon Baron-Cohen and Sally Wheelwright. 2004. The Empathy Quotient: An Investigation of Adults with Asperger Syndrome or High Functioning Autism, and Normal Sex Differences. Technical Report 2. C. Daniel Batson, Jim Fultz, and Patricia A. Schoenrade. 1987. Distress and Empathy: Two Qualitatively Distinct Vicarious Emotions with Different Motivational Consequences. Journal of Personality, 55(1):19–39. Emily Teding van Berkhout and John M. Malouff. 2016. The efficacy of empathy training: A metaanalysis of randomized controlled trials. Journal of Counseling Psychology, 63(1):32–41. Paul Black and Dylan Wiliam. 2009. Developing the theory of formative assessment. Educational Assessment, Evaluation and Accountability, 21(1):5–31. Laura Ana Maria Bostan and Roman Klinger. 2018. An Analysis of Annotated Corpora for Emotion Classification in Text Title and Abstract in German. Proceedings of the 27th International Conference on Computational Linguistics, pages 2104–2119. Sven Buechel, Anneke Buffone, Barry Slaff, Lyle Ungar, and Jo˜ao Sedoc. 2018. Modeling empathy and distress in reaction to news stories. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, EMNLP 2018, pages 4758–4765. Sven Buechel and Udo Hahn. 2018. Emotion Representation Mapping for Automatic Lexicon Construction (Mostly) Performs on Human Level. pages 2892–2904. Sheng-Yeh Chen, Chao-Chun Hsu, Chuan-Chun Kuo, Ting-Hao, Huang, and Lun-Wei Ku. 2018. EmotionLines: An Emotion Corpus of Multi-Party Conversations. LREC 2018 - 11th International Conference on Language Resources and Evaluation, pages 1597–1601. Silvie Cinkov´a, Martin Holub, and Vincent Kr´ıˇz. 2012. Managing uncertainty in semantic tagging. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 840–850, Avignon, France. Association for Computational Linguistics. Mark H. Davis. 1983. Measuring individual differences in empathy: Evidence for a multidimensional approach. Journal of Personality and Social Psychology, 44(1):113–126. Jean Decety and Philip L. Jackson. 2004. The functional architecture of human empathy. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Pierre Dillenbourg, Sanna J¨arvel¨a, and Frank Fischer. 2009. The Evolution of Research on ComputerSupported Collaborative Learning. In Nicolas Balacheff, Sten Ludvigsen, Ton de Jong, Ard Lazonder, and Sally Barnes, editors, Technology-Enhanced Learning: Principles and Products, pages 3–19. Springer Netherlands, Dordrecht. Paul Ekman. 1992. An Argument for Basic Emotions. COGNITION AND EMOTION, 6(3/4):169–200. J.L. Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological Bulletin, 76(5):378–382. John Hattie and Helen Timperley. 2007. The Power of Feedback. Review of Educational Research, 77(1):81–112. Hamed Khanpour, Cornelia Caragea, and Prakhar Biyani. 2017. Identifying Empathetic Messages in Online Health Communities. Technical report. Sara H. Konrath, Edward H. O’Brien, and Courtney Hsing. 2011. Changes in dispositional empathy in American college students over time: A metaanalysis. Personality and Social Psychology Review, 15(2):180–198. Klaus Krippendorff. 1980. Content Analysis: An Introduction to Methodology. Sage Publications, Inc., Beverly Hills, CA. Klaus Krippendorff. 2004. Measuring the reliability of qualitative text analysis data. Quality and Quantity, 38(6):787–800. E. J. Lawrence, P. Shaw, D. Baker, S. Baron-Cohen, and Anthony S. David. 2004. Measuring empathy: Reliability and validity of the Empathy Quotient. Psychological Medicine, 34(5):911–919. Bing Liu. 2015. Sentiment analysis: Mining opinions, sentiments, and emotions. Cambridge University Press. 4073 Benjamin Lok and Adriana E. Foster. 2019. Can Virtual Humans Teach Empathy? In Teaching Empathy in Healthcare, pages 143–163. Springer International Publishing. Joseph Luca and Pina Tarricone. 2001. Does Emotional Intelligence Affect Successful Teamwork? Proceedings of the 18th Annual Conference of the Australasian Society for Computers in Learning in Tertiary Education, (December 2001):367–376. Scott W. McQuiggan and James C. Lester. 2007. Modeling and evaluating empathy in embodied companion agents. International Journal of Human Computer Studies, 65(4):348–360. Saif M. Mohammad and Felipe Bravo-Marquez. 2017. Emotion intensities in tweets. *SEM 2017 - 6th Joint Conference on Lexical and Computational Semantics, Proceedings, pages 65–77. Tb Moyers and T Martin. 2010. Revised Global Scales: Motivational Interviewing Treatment Integrity 3.1.1 (MITI 3.1.1). University of New ... , 1(January):1– 29. OECD. 2018. The Future of Education and Skills - Education 2030. Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global vectors for word representation. In EMNLP 2014 - 2014 Conference on Empirical Methods in Natural Language Processing, Proceedings of the Conference, pages 1532–1543. Association for Computational Linguistics (ACL). Ver´onica P´erez-Rosas, Rada Mihalcea, Kenneth Resnicow, Satinder Singh, and Lawrence An. 2017. Understanding and predicting empathic behavior in counseling therapy. ACL 2017 - 55th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference (Long Papers), 1:1426–1435. Philip M. Podsakoff and Jiing Lih Farh. 1989. Effects of feedback sign and credibility on goal setting and task performance. Organizational Behavior and Human Decision Processes, 44(1):45–67. Mathis Poser and Eva A. C. Bittner. 2020. Hybrid Teamwork: Consideration of Teamwork Concepts to Reach Naturalistic Interaction between Humans and Conversational Agents. In WI2020. GITO Verlag. Roman Rietsche and Matthias S¨ollner. 2019. Insights into Using IT-Based Peer Feedback to Practice the Students Providing Feedback Skill. Proceedings of the 52nd Hawaii International Conference on System Sciences. Carolyn Ros´e, Yi Chia Wang, Yue Cui, Jaime Arguello, Karsten Stegmann, Armin Weinberger, and Frank Fischer. 2008. Analyzing collaborative learning processes automatically: Exploiting the advances of computational linguistics in computer-supported collaborative learning. International Journal of Computer-Supported Collaborative Learning, 3(3):237–271. Sara Rosenthal, Noura Farra, and Preslav Nakov. 2018. SemEval-2017 Task 4: Sentiment Analysis in Twitter. pages 502–518. Breno Santana Santos, Methanias Colaqo Junior, and Janisson Gois De Souza. 2018. An Experimental Evaluation of the NeuroMessenger: A Collaborative Tool to Improve the Empathy of Text Interactions. Proceedings - IEEE Symposium on Computers and Communications, 2018-June:573–579. Klaus R. Scherer and Harald G. Wallbott. 1994. Evidence for Universality and Cultural Variation of Differential Emotion Response Patterning. Journal of Personality and Social Psychology, 66(2):310–328. Julia E. Seaman, I. E. Allen, and Jeff Seaman. 2018. Higher Education Reports - Babson Survey Research Group. Technical report. Ashish Sharma, Adam S. Miner, David C. Atkins, and Tim Althoff. 2020. A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support. pages 5263–5276. R. Nathan Spreng, Margaret C. McKinnon, Raymond A. Mar, and Brian Levine. 2009. The Toronto empathy questionnaire: Scale development and initial validation of a factor-analytic solution to multiple empathy measures. Journal of Personality Assessment, 91(1):62–71. Christian Stab and Iryna Gurevych. 2014. Annotating Argument Components and Relations in Persuasive Essays. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers ,, pages 1501–1510. Christian Stab and Iryna Gurevych. 2017. Parsing Argumentation Structures in Persuasive Essays. Computational Linguistics, 43(3):619–659. Carlo Strapparava and Rada Mihalcea. 2007. SemEval2007 task 14: Affective text. ACL 2007 - SemEval 2007 - Proceedings of the 4th International Workshop on Semantic Evaluations, (June):70–74. Stephen E. Toulmin. 1984. Introduction to Reasoning. Viswanath Venkatesh and Hillol Bala. 2008. Technology acceptance model 3 and a research agenda on interventions. Decision Sciences, 39(2):273–315. Thiemo Wambsganss, Anne H¨och, Naim Zierau, and Matthias S¨ollner. 2021a. Ethical Design of Conversational Agents: Towards Principles for a ValueSensitive Design. In Proceedings of the 16th International Conference on Wirtschaftsinformatik (WI). Thiemo Wambsganss, Tobias K¨ung, Matthias S¨ollner, and Jan Marco Leimeister. 2021b. ArgueTutor: An 4074 Adaptive Dialog-Based Learning System for Argumentation Skills. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. Thiemo Wambsganss, Christina Niklaus, Matthias Cetto, Matthias S¨ollner, Jan Marco Leimeister, and Siegfried Handschuh. 2020a. AL : An Adaptive Learning Support System for Argumentation Skills. In ACM CHI Conference on Human Factors in Computing Systems, pages 1–14. Thiemo Wambsganss, Christina Niklaus, Matthias S¨ollner, Siegfried Handschuh, and Jan Marco Leimeister. 2020b. A corpus for argumentative writing support in German. In Proceedings of the 28th International Conference on Computational Linguistics, pages 856–869, Barcelona, Spain (Online). International Committee on Computational Linguistics. Thiemo Wambsganss, Florian Weber, and Matthias S¨ollner. 2021c. Design and Evaluation of an Adaptive Empathy Learning Tool. In Hawaii International Conference on System Sciences (HICSS). Jin Wang, Liang Chih Yu, K. Robert Lai, and Xuejie Zhang. 2016. Dimensional sentiment analysis using a regional CNN-LSTM model. 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016 - Short Papers, pages 225–230. Bo Xiao, Dogan Can, Panayiotis G Georgiou, David Atkins, and Shrikanth S Narayanan. 2012. Analyzing the Language of Therapist Empathy in Motivational Interview based Psychotherapy. Signal and Information Processing Association Annual Summit and Conference (APSIPA), ... Asia-Pacific. AsiaPacific Signal and Information Processing Association Annual Summit and Conference, 2012. N. Zierau, T Wambsganss, Andreas Janson, Sofia Sch¨obel, and Jan Marco Leimeister. 2020. The Anatomy of User Experience with Conversational Agents : A Taxonomy and Propositions of Service Clues. In International Conference on Information Systems (ICIS)., pages 1–17. A Details on the Description of the Annotation Scheme9 A more detailed description of the cognitive and emotional empathy scores can be found in Table 7 and Table 8. B Details on the Annotation Process The annotation process was split into three steps: 1. Reading of the entire peer review: The annotators are confronted with the studentwritten peer review and are asked to read the 9Further examples and descriptions can be found in our annotation guideline. whole document. This helps to get a first impression of the review and get an overview of the single components and the structure of it. 2. Labeling the components and elaborations: After reading the entire student-written peer review, the annotator is asked to label the three different components (strengths, weaknesses and suggestions for improvement). Every supporting sentence (such as explanation, example, etc.) is annotated together with the referred component. 3. Classification of the cognitive and emotional empathy levels: Each component is assessed on its level of cognitive and emotional empathy by giving a number between 1-5. Each category is carefully defined and delimited according to Table 7 and Table 8. C Disagreement Analysis To analyze the disagreement between the three annotators, we created a confusion probability matrix (CPM) (Cinkov´a et al., 2012) for the review components and the empathy level scores. A CPM contains the conditional probabilities that an annotator assigns to a certain category (column) given that another annotator has chosen the category in the row for a specific item. In contrast to traditional confusion matrices, a CPM also allows for the evaluation of confusions if more than two annotators are involved in an annotation study (Stab and Gurevych, 2014). Table 9 shows that there is a broad agreement between the annotators in distinguishing between the different types of review components. The major disagreement is between suggestions and weaknesses, though with a score of 60%, the agreement is still fairly high. Consequently, the annotation of review components in terms of strengths, weaknesses, and suggestions for improvements yields highly reliable results. The CPMs for the empathy levels (see Tables 10 and 11 reveal that there is a higher confusion between the scores assigned by the three reviewers, as compared to the annotation of the review components. However, when analyzed more closely, one can see that the scores mostly vary only within a small window of two or three neighboring scores. Therefore, we conclude that the annotation of cognitive and emotional empathy scores is reliably possible, too. 4075 Score Description 5 = strong The student fully understands the peer’s thoughts. She completely steps outside her own perspective and thinks from the peer’s perspective. She does that by carefully evaluating the peer’s idea with rich explanations. Questions, personal pronouns, or direct addressing of the author can be used in order to better understand and elaborate on the peer’s perspective. Strengths: The student fully grasps the idea of the peer. She elaborates on strengths that are important for the peer for her continuation of the task and adds explanations, thoughts, or examples to her statements and reasons why the strength is/strengths are important for the business idea. Weaknesses: The student thinks completely from the peer’s perspective and what would help him/her to further succeed with the task. The student explains the weakness in a very detailed manner and describes why the weakness is important to consider. He can also give counterarguments or ask questions to illustrate the weakness. Suggestions for improvement: The student suggests improvements as if he were in the peer’s position in creating the best possible solution. The student completes his suggestions with rich explanations on why he/she would do so and elaborates on the improvements in a very concrete and detailed way. Almost every suggestion is supported by further explanations. 4 = fairly strong The student thinks from the perspective of the peer. She elaborates in a way that serves the peer best to further establish the idea or activity. Each component is affirmed with further explanations. Strengths: The student is able to recognize one or more strengths that are helpful for the peer to affirm their business idea and activity. He/She highlights contextual strengths rather than formal strengths. The student supports most statements with examples or further personal thoughts on the topic but might still be missing some reasonings. Weaknesses: The student thinks from the peer’s perspective and what would help him/her to further succeed with the task. This could be demonstrated by stating various questions and establishing further thoughts. The student explains the weakness and adds examples, but he/she is still missing some reasonings. Suggestions for improvement: The student suggests one or more improvements that are relevant for the further establishment of the activity and idea from the perspective of the peer. Most suggestions are written concretely and, if applicable, supported by examples. In most cases, the student explains why he/she suggests a change. 3 = slightly weak / equal The student tries to understand the perspective of the peer and adds further elaborations to her statements. However, her elaborations are not completely thought through and her feedback is missing some essential explanations, examples, or questions to make sure she understood everything correctly. Strengths: The student mentions one or more strengths and explains some of them with minor explanations or examples on why it is seen as a strength. However, most strengths focus on formal aspects rather than contextual aspects. Weaknesses: The student states one or more weaknesses and explains some of them with minor explanations or examples. The student could also just state questions to illustrate the weakness in the peer’s business idea. Most weaknesses are not explained why they are such. Suggestions from improvements: The student suggests one or more improvements that are mostly relevant for the further establishment of the activity. The suggestions are written only on a high-level and most of them do not include further explanations or examples. The student explains only occasionally why he/she suggests a change or how it could be implemented. 2 = very weak The student does not try to understand the peer’s perspective. The student rather just tries to accomplish the task of giving feedback. Strengths: The student mentions one or more strengths. They could be relevant for the peer. However, he does not add any further explanation or details. Weaknesses: The student states one or more weaknesses without explaining why they are seen as such. They could be relevant for the peer. However, the statements do not include any further elaboration on the mentioned weakness. Suggestions for improvement: The student suggests one or more improvements that could be relevant for the peer. However, the student does not explain why he/she suggests the change or how the suggestions for improvement could be implemented. 1 = absolutely weak The student’s feedback is very short and does not include the peer’s perspective. She does not add any further elaboration in her thoughts. Strengths: The student only mentions one strength. This might not be relevant at all and lacks any further explanation, detail, or example. Weakness: The student only mentions one weakness. This might not be relevant at all and lacks any further explanation, detail, or example. Suggestions for improvement: The student only mentions one suggestion. The suggestion is not followed by any explanation or example and might not be relevant for the further revision of the peer. Table 7: Detailed description of the cognitive empathy scores. 4076 Score Description 5 = strong The student is able to respond very emotionally to the peer’s work and fully represents the affectional state in her entire review. She illustrates this by writing in a very emotional and personal manner and expresses her feelings (positive or negative) throughout the review. Strong expressions include exclamation marks (!). Typical feedback in this category includes phrases such as “brilliant!”, “fantastic”, “excellent”, “I am totally on the same page as you”, “I am very convinced”, “Personally, I find this very important, too”, “I am very unsure”, “I find this critical”, “I am very sure you feel”, “This is compelling for me”, etc. 4 = fairly strong The student is able to respond emotionally to the peer’s submitted activity with suitable emotions (positive or negative). She returns emotions in her feedback on various locations and expresses her feelings by using the personal pronoun (“I”, “You”). Some sentences might include exclamations marks (!). Typical feedback in this category includes phrases such as “I am excited”, “This is very good!”, “I am impressed by your idea”, “I feel concerned about”, “I find this very...”, “In my opinion”, “Unfortunately, I do not understand”, “I am very challenged by your submission”, “I am missing”, “You did a very good job”, etc. 3 = slightly weak / equal The student occasionally includes emotions or personal emotional statements in the peer review. They could be quite strong. However, the student’s review is missing personal pronouns (“I”, “You”) and is mostly written in third person. Emotions can both be positive or negative. Negative emotions can be demonstrated with concern, missing understanding or insecurity (e. g., with modal verbs or words such as rather, perhaps). Typically, scale 3 includes phrases such as “it’s important”, “the idea is very good”, ”the idea is comprehensible”, “it would make sense”, “the task was done very nicely”, “It could probably be that”, etc. 2 = very weak Mostly, the student does not respond emotionally to the peer’s work. Only very minor and weak emotions or personal emotional statements are integrated. The student writes mostly objectively (e.g., “Okay”, “This should be added”, “The task was done correctly”, etc.). In comparison to level 1, she might use modal verbs (might, could, etc.) or words to show insecurity in her feedback (rather, maybe, possibly). 1 = absolutely weak The student does not respond emotionally to the peer’s work at all. She does not show her feelings towards the peer and writes objectively (e.g., no “I feel”, “personally” “I find this..” and no emotions, such as “good”, “great”, “fantastic”, “concerned”, etc.). Typical examples would be “Add a picture.” or “The value gap XY is missing.” Table 8: Detailed description of the emotional empathy scores. Suggestions Weakness Strength None Suggestions 0.6056 0.2970 0.0214 0.0759 Weakness 0.2139 0.7009 0.0203 0.0648 Strength 0.0264 0.0347 0.8340 0.1049 None 0.0662 0.0784 0.0742 0.7812 Table 9: CPM for review component annotations. 1 2 3 4 5 1 .113 .387 .175 .165 .160 2 .125 .266 .362 .211 .035 3 .025 .159 .223 .482 .112 4 .014 .054 .283 .300 .349 5 .021 .014 .105 .556 .303 Table 10: CPM for cognitive empathy level annotations. 1 2 3 4 5 1 .106 .459 .286 .086 .063 2 .154 .234 .455 .128 .029 3 .059 .282 .350 .240 .068 4 .026 .115 .347 .295 .218 5 .043 .061 .227 .501 .168 Table 11: CPM for emotional empathy level annotations. precision recall f1-score support non-empathic 0.5746 0.5662 0.5704 136 empathic 0.6364 0.5625 0.5972 112 neutral 0.5240 0.5707 0.5464 191 None 0.9863 0.9729 0.9795 295 micro avg 0.7322 0.7302 0.7482 734 macro avg 0.6803 0.6681 0.6734 734 weighted avg 0.7363 0.7302 0.7327 734 samples avg 0.7248 0.7302 0.7266 734 Table 12: BERT model results for emotional empathy. precision recall f1-score support non-empathic 0.5739 0.3587 0.4415 184 empathic 0.6434 0.5490 0.5925 286 neutral 0.3062 0.4747 0.3723 198 None 0.9841 0.9802 0.9822 506 micro avg 0.6949 0.6925 0.6937 1174 macro avg 0.6269 0.5907 0.5971 1174 weighted avg 0.7225 0.6925 0.6996 1174 samples avg 0.6861 0.6925 0.6882 1174 Table 13: BERT model results for cognitive empathy. D Details on Application and Evaluation of Writing Support Tool To ensure the validity of our BERT model, we benchmarked against bidirectional Long-Short4077 precision recall f1-score support non-empathic 0.5739 0.3587 0.4415 184 neutral 0.3062 0.4747 0.3723 198 empathic 0.6434 0.5490 0.5925 286 None 0.9841 0.9802 0.9822 506 f1 avg 0.64 0.64 0.64 368 weighted avg 0.73 0.73 0.73 368 Table 14: Results for the LSTM for emotional empathy. precision recall f1-score support non-empathic 0.74 0.28 0.40 83 neutral 0.43 0.55 0.49 60 empathic 0.35 0.63 0.45 57 None 0.99 0.94 0.97 168 f1 avg 0.63 0.60 0.58 368 weighted avg 0.75 0.68 0.68 368 Table 15: Results for the LSTM for cognitive empathy. Term-Memory-Conditional-Random-Fields classifiers (BiLSTM-CRF). In combination with the corresponding embeddings vocabulary (GloVe) (Pennington et al., 2014), our LSTM reached an unsatisfying f1-score of 61% for detecting the emotional empathy level and 51% for detecting the cognitive empathy level. More information on the results of our BERT model and the LSTM for emotional and cognitive empathy detection can be found in the Tables 12, 13, 15, and 15. In the post-survey, we measured perceived usefulness following the technology acceptance model (Venkatesh and Bala, 2008). The items for the constructs were: ”Imagine the tool was available in your next course, would you use it?”, ”Assuming the learning tool would be available at a next course, I would plan to use it.”, or ”Using the learning tool helps me to write more emotional and cognitive empathic reviews. ” Moreover, we asked the participants to judge their perceived empathy skill learning (PESL) by asking two items that cover cognitive and emotional empathy skills (Spreng et al., 2009; Davis, 1983): “I assume that the tool would help me improve my ability to give appropriate emotional feedback.” and “I assume that the tool would help me improve my ability to empathize with others when writing reviews.” Finally, we surveyed the perceived feedback accuracy (PFA) (Podsakoff and Farh, 1989) of both learning tools by asking three items: “The feedback I received reflected my true performance.”, “The tool accurately evaluated my performance.”, and “The feedback I received from the tool was an accurate evaluation of my performance”. All constructs were measured with a 1- to 7-point Likert scale (1: totally disagree to 7: totally agree, with 4 being a neutral statement). Cluster Feature On empathy feedback reaction ”I think that this tool could help me not only to put myself in the position of a person in terms of content and make suggestions but also to communicate to them better” On the feedback for skill learning ”The empathy feedback was clear and could be easily implemented. I had the feeling I learned something.Would use it again!” On cognitive and emotional empathy ”It was helpful that a distinction was made between the two categories of empathy. This again clearly showed me that I do not show emotional empathy enough. It was also useful that the tool said how to show emotional empathy (feelings when reading the business idea etc.).” Improvements on feedback granularity ”It would be better if the feedback was more s elective or with detailed categories about empathy.” Improvements on feedback recommendations ”Even more detailed information on how I can improve my empathy writing would be helpful, e.g., with review examples.” Table 16: Representative examples of qualitative user responses after the usage of our empathy support tool.
|
2021
|
314
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4078–4088 August 1–6, 2021. ©2021 Association for Computational Linguistics 4078 Dual Reader-Parser on Hybrid Textual and Tabular Evidence for Open Domain Question Answering Alexander Hanbo Li, Patrick Ng, Peng Xu, Henghui Zhu, Zhiguo Wang, Bing Xiang AWS AI Labs, Amazon {hanboli, patricng, pengx, henghui, zhiguow, bxiang}@amazon.com Abstract The current state-of-the-art generative models for open-domain question answering (ODQA) have focused on generating direct answers from unstructured textual information. However, a large amount of world’s knowledge is stored in structured databases, and need to be accessed using query languages such as SQL. Furthermore, query languages can answer questions that require complex reasoning, as well as offering full explainability. In this paper, we propose a hybrid framework that takes both textual and tabular evidence as input and generates either direct answers or SQL queries depending on which form could better answer the question. The generated SQL queries can then be executed on the associated databases to obtain the final answers. To the best of our knowledge, this is the first paper that applies Text2SQL to ODQA tasks. Empirically, we demonstrate that on several ODQA datasets, the hybrid methods consistently outperforms the baseline models that only take homogeneous input by a large margin. Specifically we achieve state-of-theart performance on OpenSQuAD dataset using a T5-base model. In a detailed analysis, we demonstrate that the being able to generate structural SQL queries can always bring gains, especially for those questions that requires complex reasoning. 1 Introduction Open-domain question answering (ODQA) is a task to answer factoid questions without a prespecified domain. Recently, generative models (Roberts et al., 2020; Lewis et al., 2020; Min et al., 2020; Izacard and Grave, 2020) have achieved the state-of-the-art performance on many ODQA tasks. These approaches all share the common pipeline where the first stage is retrieving evidence from the free-form text in Wikipedia. However, a large amount of world’s knowledge is not stored as plain text but in structured databases, and need to be accessed using query languages such as SQL. Furthermore, query languages can answer questions that require complex reasoning, as well as offering full explainability. In practice, an ideal ODQA model should be able to retrieve evidence from both unstructured textual and structured tabular information sources, as some questions are better answered by tabular evidence from databases. For example, the current state-of-the-art ODQA models struggle on questions that involve aggregation operations such as counting or averaging. One line of research on accessing databases, although not open domain, is translating natural language questions into SQL queries (Zhong et al., 2017; Xu et al., 2017; Yu et al., 2018c; Guo et al., 2019; Wang et al., 2018a, 2020; Yu et al., 2018a; Guo and Gao, 2019; Choi et al., 2020). These methods all rely on knowing the associated table for each question in advance, and hence are not trivially applicable to the open-domain setting, where the relevant evidence might come from millions of tables. In this paper, we provide a solution to the aforementioned problem by empowering the current generative ODQA models with the Text2SQL ability. More specifically, we propose a dual readerparser (DUREPA) framework that can take both textual and tabular data as input, and generate either direct answers or SQL queries based on the context1. If the model chooses to generate a SQL query, we can then execute the query on the corresponding database to get the final answer. Overall, our framework consists of three stages: retrieval, joint ranking and dual reading-parsing. First we retrieve supporting candidates of both textual and tabular types, followed by a joint reranker that predicts how relevant each supporting candidate is to 1Our code is available at https://github.com/ AlexanderYogurt/Hybrid-Open-QA 4079 the question, and finally we use a fusion-in-decoder model (Izacard and Grave, 2020) for our readerparser, which takes all the reranked candidates in addition to the question to generate direct answers or SQL queries. To evaluate the effectiveness of our DUREPA, we construct a hybrid dataset that combines SQuAD (Rajpurkar et al., 2016) and WikiSQL (Zhong et al., 2017) questions. We also conduct experiments on NaturalQuestions (NQ) (Kwiatkowski et al., 2019) and OTT-QA (Chen et al., 2020a) to evaluate DuRePa performance. As textual and tabular open-domain knowledge, we used textual and tabular data from Wikipedia via Wikidumps (from Dec. 21, 2016) and Wikitables (Bhagavatula et al., 2015). We study the model performance on different kinds of questions, where some of them only need one supporting evidence type while others need both textual and tabular evidence. On all question types, DUREPA performs significantly better than baseline models that were trained on a single evidence type. We also demonstrate that DUREPA can generate humaninterpretable SQLs that answer questions requiring complex reasoning, such as calculations and superlatives. Our highlighted contributions are as follows: • We propose a multi-modal framework that incorporates hybrid knowledge sources with the Text2SQL ability for ODQA tasks. To the best of our knowledge, this is the first work that investigates Text2SQL in the ODQA setting. • We propose a simple but effective generative approach that takes both textual and tabular evidence and generates either direct answers or SQL queries, automatically determined by the context. With that, we achieve the state-of-the-art performance on OpenSQuAD using a T5-base model. • We conduct comprehensive experiments to demonstrate the benefits of Text2SQL for ODQA tasks. We show that interpretable SQL generation can effectively answer questions that require complex reasoning in the ODQA setting. 2 Related Work Open Domain Question Answering ODQA has been extensively studied recently including extractive models (Chen et al., 2017; Clark and Gardner, 2018; Wang et al., 2019; Min et al., 2019; Yang et al., 2019) that predict spans from evidence passages, and generative models (Raffel et al., 2020; Roberts et al., 2020; Min et al., 2020; Lewis et al., 2020; Izacard and Grave, 2020) that directly generate the answers. Wang et al. (2018b,c); Nogueira and Cho (2019) proposed to rerank the retrieved passages to get higher top-n recall. Table Parsing Text2SQL is a task to translate natural questions to executable SQL queries. Brad et al. (2017) proposed SENLIDB dataset which only contains 29 tables and lacks annotation in their training set. Recently, with datasets like WikiSQL (Zhong et al., 2017), Spider (Yu et al., 2018c) and CoSQL (Yu et al., 2019) being introduced, many works have shown promising progress on these dataset (Yu et al., 2018b; He et al., 2019; Hwang et al., 2019; Min et al., 2019; Wang et al., 2020; Choi et al., 2020; Guo et al., 2019; Lyu et al., 2020; Zhang et al., 2019; Zhong et al., 2020; Shi et al., 2020). Another line of work proposes to reason over tables without generating logical forms (Neelakantan et al., 2015; Lu et al., 2016; Herzig et al., 2020; Yin et al., 2020). However, they are all closed-domain and each question is given the associated table. Hybrid QA Chen et al. (2020a) also proposed an open-domain QA problem with textual and tabular evidence. Unlike our problem, they generate an answer directly from the tabular evidence instead of generating an SQL query. In addition, they assume some contextual information about table is available during retrieval stage (e.g. their fusion-retriever is pretrained using hyperlinks between tables and paragraphs), whereas we don’t use any link information between tables and passages. Moreover, Chen et al. (2020b) proposed a closed-domain hybrid QA dataset where each table is linked to on average 44 passages. Different from ours, their purpose is to study multi-hop reasoning over both forms of information, and each question is still given the associated table. 3 Method In this section, we describe our method for hybrid open-domain question answering. It mainly consists of three components: (1) a retrieval system; (2) a joint reranker and (3) a dual Seq2Seq model that uses fusion-in-decoder (Izacard and Grave, 2020) to generate direct answer or SQL query. 4080 Figure 1: The pipeline of our proposed hybrid model. The candidates are retrieved from knowledge source such as Wikipedia including both paragraphs and tables. Then a generative Seq2Seq model reads the question and all the candidates, and produces k outputs using beam search. Each output can be either a final answer or an intermediate SQL query. The types and order of the outputs are automatically determined by the model itself. 3.1 Retrieval For the hybrid open-domain setting, we build two separate search indices – one for textual input and another for tabular input. For paragraphs, we split them into passages of at most 100 words. For tables, we flattened each table into passages by concatenating cell values along each row. If the flattened table exceeds 100 words, we split it into a separate passage, respecting row boundaries. The column headers are concatenated to each tabular passage. Some examples of flattened tables are given in the Appendix A.1. Given a natural language question, the retrieval system retrieves 100 textual and 100 tabular passages as the support candidates from the textual and tabular indices, respectively, using BM25 (Robertson et al., 1995) ranking function. 3.2 Joint Reranking The purpose of our reranking model is to produce a score si of how relevant a candidate (either an unstructured passage or table) is to a question. Specifically, the reranker input is the concatenation of question, a retrieved candidate-content, and its corresponding title if available2, separated by special tokens shown in Figure 1. The candidate content can be either the unstructured 2Wikipedia passages have page titles, and tables have table titles. text or flattened table. We use BERTbase model in this paper. Following Nogueira and Cho (2019), we finetune the BERT (Devlin et al., 2019) model using the following loss: L = − X i2Ipos log(si) − X i2Ineg log(1 −si). (1) The Ipos is sampled from all relevant BM25 candidates, and the set Ineg is sampled from all non-relevant BM25 candidates. Different from Nogueira and Cho (2019), during training, for each question, we sample 64 candidates including one positive candidate and 63 negative candidates, that is, |Ipos| = 1 and |Ineg| = 63. If none of the 200 candidates is relevant, we skip the question. During inference, we use the hybrid reranker to assign a score to each of the 200 candidates, and choose the top 50 candidates as the input to the next module – the reader-parser model. For the top 50 candidates, we choose them from the joint pool of all candidates, according to the scores assigned by the reranker. 3.3 Dual Reading-Parsing Our dual reader-parser model is based on the fusionin-decoder (FID) proposed in Izacard and Grave (2020), and is initialized using the pretrained T5 (Raffel et al., 2020) model. The overall pipeline of the reader-parser is shown in Figure 1. Each 4081 retrieved candidate is represented by its title and content, in the following formats: Textual Candidate We represent each textual candidate as the concatenation of the passage title and content, appended by special tokens [text title] and [text content] respectively. Tabular Candidate In order to represent a structured table as a passage, we first flatten each table into the following format: each flattened table starts with the complete header names and then followed by rows. Figure 1 presents an example for this conversion. Finally, a tabular candidate is the concatenation of the table title and content flattened as a passage, appended by special tokens [table title] and [table content] respectively. We use the table ID as the title so that it can be copied to the generated SQL queries by the model. Prefix of the Targets During training, we also add special tokens answer: or sql: to a targeted sentence depending on whether it is a plain text or a SQL query. For those questions that have both textual answer and SQL query annotations (for example, WikiSQL questions), we create two training examples for each question. During inference, the generated outputs will also contain these two special prefixes, indicating which output type the model has generated. Dual Reader-Parser Our generative Seq2Seq model has reader-parser duality. During inference, the model reads the question and all the candidates, and produces k outputs using beam search. Each output can be either a final answer or an intermediate SQL query. Depending on the context, the types and order of the outputs are automatically determined by the model itself. All the generated SQL queries will then be executed to produce the final answers. In this paper, we fix k = 3 and always generate three outputs for each question. 4 Experiments In this section, we report the performance of the proposed method on several hybrid open-domain QA datasets. 4.1 Datasets In this section, we describe all the datasets we use in our experiments. First we summarize the statistics of the open-domain QA datasets we use in Table 1. Dataset #Train&Dev #Test OpenSQuAD 82,599 5,000 OpenNQ 87,925 3,610 OTT-QA 41,469 2,214 OpenWikiSQL 52,026 7,764 Mix-SQuWiki 134,625 12,764 WikiSQL-both – 3,029 Table 1: Statistics of Datasets OpenSQuAD is an open-domain QA dataset constructed from the original SQuAD-v1.1 (Rajpurkar et al., 2016), which was designed for the reading comprehension task, consisting of 100,000+ questions posed by annotators on a set of Wikipedia articles, where the answer to each question is a span from the corresponding paragraph. OpenNQ is an open-domain QA datasets constructed from the NaturalQuestions (Kwiatkowski et al., 2019), which was desgined for the end-toend question answering task. The questions were from real google search queries and the answers were from Wikipedia articles annotated by humans. OTT-QA (Chen et al., 2020a) is a large-scale open table-and-text question answering dataset for evaluating open QA over both tabular and textual data. The questions were constructed through “decontextualization” from HybridQA (Chen et al., 2020b) with additional 2200 new questions mainly used in dev/test set. OTT-QA also provides its own corpus which contains over 5 million passages and around 400k tables. OpenWikiSQL is an open-domain Text2SQL QA dataset constructed from the original WikiSQL (Zhong et al., 2017). WikiSQL is a dataset of 80,654 annotated questions and SQL queries distributed across 24,241 tables from Wikipedia. Mix-SQuWiki is the union of OpenSQuAD and OpenWikiSQL datasets. WikiSQL-both is a subset of OpenWikiSQL evaluation data that contains the questions that can be answered by both textual and tabular evidences. The purpose of this dataset is to study when both types of evidence are possible to answer a question, whether the hybrid model can still choose the better one. We select these questions in a weaklysupervised way by only keeping a question if the 4082 Model Evidence Corpus Type OpenSQuAD OpenNQ OTT-QA OpenWikiSQL FiD(T5-base) Text-only 53.4 48.2 FiD(T5-large) Text-only 56.7 51.4 IR+CR Text+Table w/o SQL 14.4 FR+CR Text+Table w/o SQL 28.13 Unified Model Text+NQ Table w/o SQL 54.64 Ours FID+ Text-only 56.4 45.2 14.5 13.9 FID+ Table-only w/o SQL 2.5 14.3 4.1 30.3 DUREPA Table-only with SQL 2.7 14.8 4.7 40.2 FID+ Text+Table w/o SQL 56.4 46.7 15.0 30.9 DUREPA Text+Table with SQL 57.0 48.0 15.8 42.6 Table 2: Comparison to the state-of-the-art on open-domain QA datasets. The numbers reported are in EM metric. FiD(T5-base & T5-large) is reported from (Izacard and Grave, 2020), IR+CR (Iterative Retrieval+Cross-block Reader) and FR+CR (Fusion Retrieval+Cross-block Reader) are from (Chen et al., 2020a), Unified Model is from (Oguz et al., 2020). Comparing DUREPA with FID+ , we observe that having the ability to generate structural queries is always beneficial even for questions with mostly extractive answers like SQuAD and NQ. groundtruth answer is contained in both textual and tabular BM25 candidates. For example in Figure 1, the answer “Richard Marquand” can be found in both types of passages. We filter out some trivial cases where the answer shows up in more than half of the candidates. 5 Wikipedia Passages and Tables For the textual evidences, we process the Wikipedia 2016 dump and split the articles into overlapping passages of 100 words following (Wang et al., 2019). To create the tabular evidences, we combine 1.6M Wikipedia tables (Bhagavatula et al., 2015) and all the 24,241 WikiSQL tables, and flatten and split each table into passages not exceeding 100 words, in the same format mentioned in the previous section. We use these two collections as the evidence sources for all the QA datasets except for OTT-QA, where we use its own textual and tabular collections. 4.2 Implementation Details Retriever and Reranker. We conduct BM25 retrieval using Elasticsearch 7.7 6 with the default settings. And we use a BERT reranker initialized with pretrained BERT-base-uncased model. Dual Reader and Parser with fusion-in-decoder. Similar to (Izacard and Grave, 2020), we initialize the fusion-in-decoders with the pretrained T5 model (Raffel et al., 2020). We only explore T5base model in this paper, which has 220M parameters. 5For example, some numerical number like ”1” is a very common substring and shows up in most of the candidates. 6https://www.elastic.co/ For both reranker and FiD models, we use Adam optimizer (Kingma and Ba, 2014) with a maximum learning rate of 10−4 and a dropout rate of 10%. The learning rate linearly warms up to 10−4 and then linearly anneals to zero. We train models for 10k gradient steps with a batch size of 32, and save a checkpoint every 1k steps. For the FiD model, when there are multiple answers for one question, we randomly sample one answer from the list. For the FiD model, during inference, we generate 3 answers for each question using beam search with beam size 3. 4.3 Main Results We present the end-to-end results on the opendomain QA task comparing with the baseline methods as show in Table 2. We build models with 5 different settings based on the source evidence modality as well as the format of model prediction. Specifically, we consider single modality settings with only textual evidence or tabular evidence and the hybrid setting with both textual and tabular evidence available. For tabular evidence, the models either predict direct answer text or generate structure SQL queries. Note we also consider a baseline model, FID+ , a FiD model that only generates direct answer text, but can make use of both textual and tabular evidence. 3Chen et al. (2020a) uses a fusion-retriever to retrieved table-passages blocks as evidences. To construct the fusion blocks, they train a GPT-2 model using extra hyperlink information to link table cell to passages. In contrast, we do not use any hyperlink information. 4Oguz et al. (2020) uses tables provided by NQ training data (less than 500k in total), whereas we use all the tables extracted from Wikipedia dumps (around 1.6M in total). 4083 BM25 Reranker BM25 Reranker Reranker Index textual textual tabular tabular hybrid R@1 34.40 69.76 1.60 10.16 69.92 R@10 59.38 80.30 6.34 18.88 80.90 R@25 65.92 81.64 8.84 21.20 82.42 R@50 72.16 82.50 12.36 22.62 83.26 R@100 76.50 83.44 15.04 23.72 84.10 Table 3: Recalls on top-k textual, tabular or the hybrid candidates for SQuAD questions. The recalls on hybrid inputs are almost the same as or even better than the best recalls on individual textual or tabular inputs, meaning that the reranker is able to jointly rank both types of candidates and provide better evidences to the next component – the reader-parser. First, in the single modality setting, we observe that for OpenSQuAD, OpenNQ and OTT-QA datasets, textual QA model is performing significantly better than tabular QA models, while for OpenWikiSQL, it is the opposite. This is expected due to the nature of the construction process of those datasets. In the hybrid setting, the hybrid models outperform single modality models consistently across all these datasets. This indicates hybrid models are more robust and flexible when dealing with questions of various types in practice. Comparing DUREPA with FID+ , we observe that having the ability to generate structural queries is always beneficial even for extractive questions like SQuAD and NQ. And for WikiSQL-type questions, the gain of SQL generation is significant. On OpenSQuAD dataset, our DUREPA model using hybrid evidences achieves a new state-ofthe-art EM score of 57.0. It is worth noting that the previous best score was attained by FiD using T5-large model, while our model is using T5base, which has much fewer parameters. On NQ dataset, FID+ with text-only evidences has lower EM score compared with FiD-base, despite having the same underlying model and inputs. We suspect that this is because (1) we truncate all passages into at most 150 word pieces while in FiD paper they keep 250 word pieces, so the actual input (top-100 passages) to our FiD model is much less than that in the FiD paper; and (2) we use BM25 to retrieve the initial pool of candidates instead of trained embedding-based neural retrieval model(Karpukhin et al., 2020; Izacard and Grave, 2020). Nevertheless, the DUREPA model with hybrid evidences still improve the EM by 2.8 points compared to FID+ using only text inputs. On OTTQA questions, our full model also outperforms the IR+CR baseline by 1.4 points. The FR+CR model is using a different setting where they use hyperlinks between tables and passages to train the fusion-retriever (FR), so the result is not directly comparable to ours. We provide more analysis on OTT-QA in the Appendix. On OpenWikiSQL dataset, enabling SQL generation brings more than 10 points improvement on the EM scores. This is because many questions therein require complex reasoning like COUNT, AVERAGE or SUM on the table evidences. We provide more in-depth analysis in Section 5.2 including some complex reasoning examples in Table 7. 5 Analysis 5.1 Retrieval and Reranking Performance In this section, we investigate the performance of the BM25 retriever and the BERT reranker using top-k recalls as our evaluation metric. During both training and inference, for each question, the textual and tabular passages are reranked jointly using a single reranker. On the Mix-SQuWiki dataset, we report the reranking results on SQuAD questions in Table 3. The result on WikiSQL questions is in Table 9 in Appendix. To provide better insights on the reranker’s performance, we show the top-k recalls on textual, tabular and hybrid evidences separately. From Table 3, on both textual and tabular candidates, recall@25 of the ranker is even higher than recall@100 of the BM25 retriever. This suggest that during inference, instead of providing 100 BM25 candidates to the fusion-in-decoder (FiD), only 25 reranked candidates would suffice. In Table 9 and 10 in Appendix, we observe similar trend with top-25 recalls comparable to top100 recalls on both WikiSQL and NQ questions. Finally, across all datasets, the recalls on hybrid inputs are almost the same as or even better than the best recalls on individual textual or tabular inputs, meaning that the reranker is able to jointly rank both types of candidates and provide better 4084 evidences to the next component – the dual readerparser. 5.2 Performance of the Reader-Parser In this section, we discuss the performance of the dual reader-parser on different kinds of questions. SQL prediction helps with complex reasoning. In Table 4, we compare the top-1 EM execution accuracy of DUREPA and FID+ on OpenWikiSQL. If DUREPA generated a SQL, we execute the SQL to obtain its answer prediction. If the ground-truth answer is a list (e.g., What are the names of Simpsons episodes aired in 2008?), we use set-equivalence to evaluate accuracy. DUREPA outperforms FID+ on the test set in most of the settings. We also compare their performance under a breakdown of different categories based on the ground-truth SQL query. DUREPA achieved close to 3x and 5x improvements on WikiSQL questions that have superlative (MAX/MIN) and calculation (SUM/AVG) operations, respectively. For COUNT queries, FID+ often predicted either 0 or 1. Thus, these results support our hypothesis that the SQL generation helps in complex reasoning and explainability for tabular question answering. DUREPA FID+ #Test All 47.1 29.3 7764 COUNT 2 {0,1} 78.0 82.9 770 COUNT ≥2 44.4 0.0 9 MIN/MAX 26.6 9.3 654 SUM/AVG 22.6 4.7 314 Comparison (< or >) 45.8 32.0 939 AND-condition 53.0 31.8 2045 Answer is a list 34.3 0.0 160 Direct answers 78.7 75.6 933 Table 4: Comparison of DUREPA and FID+ on OpenWikiSQL dataset. We compare their accuracy under a breakdown of different categories based on the ground-truth SQL query. “Direct answers” stands for the questions that DUREPA predicts direct answers. DUREPA significantly outperforms on questions that require complex reasoning such as superlatives and calculations. Using hybrid evidence types leads to better performance. Shown in Table 5 is the model performance on the Mix-SQuWiki questions. As the baseline models, if we only use a single evidence type, the best top-1 EM is 34.0, achieved by the model FID+ using only textual candidates. However, if we use both evidence types, the hybrid model DUREPA attains a significantly better top1 EM of 47.9, which implies that including both textual and tabular evidences leads a better model performance on Mix-SQuWiki. Furthermore, we observe that the model DUREPA has a better top-1 EM compared to FID+, suggesting that the answers for some of these questions need to be obtained by executing SQL queries instead of generated directly. In Table 7, we samples some questions on which the model DUREPA predicts the correct answers but the model FID+ fails. What if the questions can be answered by both textual and tabular evidences? Table 6 shows the model performance on WikiSQL-both dataset. Recall that all these questions in the dataset can be answered by both type of evidence. First of all, the DUREPA model using tabular evidences behaves better than the FID+ model using textual evidences. This implies on WikiSQL questions, using tabular information leads to better answers. Next, when using only one type of evidence, both DUREPA and FID+ models behave significantly worse than their hybrid counterparts. This indicates that the hybrid model can again figure out which evidence type should be used to provide the correct final answer. 6 Discussion and Future Work Our experiments consistently show that the proposed framework DUREPA brings significant improvement on answering questions using hybrid types of evidence. Especially on the questions that can be answered by both supporting evidence types, our multi-modal method still shows clear advantage over models using single-type knowledge, implying that our approach could figure out the most relevant evidence to answer a question. We also demonstrate that the dual reader-parser is essential to the good performance of DUREPA; the ability of generating both direct answers and structural SQL queries help DUREPA perform much better than FID+ and other baselines on questions that require complex reasoning like counting or averaging. We believe that our methods can be improved in two aspects. First, our general framework Fig. 1 can be improved by a better retrieval system. For example, instead of using BM25, we can use more powerful neural retrieval models (Karpukhin et al., 2020). On the hybrid evidence, one can also use an entity linking module to link the entities between the tables and passages (Chen et al., 2020a) and utilize the structure information for better multi4085 Model Evidence Corpus Type % of SQL Answers Acc of SQL Answers (%) % of Direct Answers Acc of Direct Answers (%) EM (Overall) FID+ Text-only 0.0 100.0 34.0 34.0 FID+ Table-only w/o SQL 0.0 100.0 19.3 19.3 DUREPA Table-only with SQL 53.9 42.5 46.1 8.4 26.8 FID+ Text+Table w/o SQL 0.0 100.0 40.0 40.0 DUREPA Text+Table with SQL 33.5 44.1 66.5 49.8 47.9 Table 5: Detailed results on Mix-SQuWiki dataset under various settings. Model Evidence Corpus Type % of SQL Answers Acc of SQL Answers (%) % of Direct Answers Acc of Direct Answers (%) EM (Overall) FID+ Text-only 0.0 100.0 38.7 38.7 FID+ Table-only w/o SQL 0.0 100.0 38.4 38.4 DUREPA Table-only with SQL 38.6 30.4 61.4 57.2 46.8 FID+ Text+Table w/o SQL 0.0 100.0 43.2 43.2 DUREPA Text+Table with SQL 39.8 35.5 60.2 64.0 53.6 Table 6: Model Performance on WikiSQL-both dataset. The models are trained on Mix-SQuWiki training data. Question: Which party won in the election in voting district Kentucky 5? Groundtruth: [’democratic’] Top-1 generation by DUREPA: sql: SELECT Party FROM table 1-1342218-17 WHERE District = "Kentucky 5" Execution result: [’democratic’] Top-1 generation by DUREPA– answer: republican Question: Which Condition has an unaffected Partial thromboplastin time, Platelet count, and a Prothrombin time? Groundtruth: [’aspirin’, ’uremia’, ”glanzmann’s thrombasthenia”] Top-1 generation by DUREPA: sql: SELECT Condition FROM table 1-14006-1 WHERE Partial thromboplastin time = "Unaffected" AND Platelet count = "Unaffected" AND Prothrombin time = "Unaffected" Execution result: [’uremia’, ”glanzmann’s thrombasthenia”, ’aspirin’] Top-1 generation by DUREPA– answer: vitamin k deficiency or warfarin Analysis: Answer is a list of medical conditions Question: How many Wins have Goals against smaller than 30, and Goals for larger than 25, and Draws larger than 5? Groundtruth: [’3’] Top-1 generation by DUREPA: sql: SELECT COUNT(Wins) FROM table 2-18017970-2 WHERE Goals against < 30 AND Goals for > 25 AND Draws > 5 Execution result: [3] Top-1 generation by DUREPA– answer: 0 Analysis: COUNT operation Question: What is the highest Rd that Tom Sneva had the pole position in? Groundtruth: [’7’] Top-1 generation by DUREPA: sql: SELECT MAX(Rd) FROM table 1-10706961-2 WHERE Pole Position = "Tom Sneva" Execution result: [7] Top-1 generation by DUREPA– answer: 2.0 Analysis: MAX operation Question: Name the average ERP W and call sign of w237br Groundtruth: [110] Top-1 generation by DUREPA: sql: SELECT AVG(ERP W) FROM table 2-14208614-1 WHERE Call sign = "w237br" Execution result: [110] Top-1 generation by DUREPA– answer: 1.0 Analysis: AVG calculation Table 7: Examples of the SQuWiki and OpenWikiSQL questions that are answered correctly by model DUREPA but incorrectly by model FID+. hop reasoning. Second, as we have demonstrated, having the ability of generating structural SQL queries is a very powerful and necessary feature for answering questions that require complex rea4086 soning. Given the limited Text2SQL data and the difficulty of obtaining such SQL supervision, two interesting future work include (1) getting SQL annotations more efficiently and (2) adapting weaklysupervised approaches like discrete EM (Min et al., 2019) for model training. References Chandra Sekhar Bhagavatula, Thanapon Noraset, and Doug Downey. 2015. Tabel: entity linking in web tables. In International Semantic Web Conference, pages 425–441. Springer. Florin Brad, Radu Iacob, Ionel Hosu, and Traian Rebedea. 2017. Dataset for a neural natural language interface for databases (nnlidb). arXiv preprint arXiv:1707.03172. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading wikipedia to answer opendomain questions. In 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, pages 1870–1879. Association for Computational Linguistics (ACL). Wenhu Chen, Ming-Wei Chang, Eva Schlinger, William Wang, and William W Cohen. 2020a. Open question answering over tables and text. arXiv preprint arXiv:2010.10439. Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. 2020b. Hybridqa: A dataset of multi-hop question answering over tabular and textual data. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 1026– 1036. DongHyun Choi, Myeong Cheol Shin, EungGyun Kim, and Dong Ryeol Shin. 2020. Ryansql: Recursively applying sketch-based slot fillings for complex textto-sql in cross-domain databases. arXiv preprint arXiv:2004.03125. Christopher Clark and Matt Gardner. 2018. Simple and effective multi-paragraph reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 845–855. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Jiaqi Guo, Zecheng Zhan, Yan Gao, Yan Xiao, Jian-Guang Lou, Ting Liu, and Dongmei Zhang. 2019. Towards complex text-to-sql in cross-domain database with intermediate representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4524– 4535. Tong Guo and Huilin Gao. 2019. Content enhanced bert-based text-to-sql generation. arXiv preprint arXiv:1910.07179. Pengcheng He, Yi Mao, Kaushik Chakrabarti, and Weizhu Chen. 2019. X-sql: reinforce schema representation with context. arXiv preprint arXiv:1908.08113. Jonathan Herzig, Pawel Krzysztof Nowak, Thomas Mueller, Francesco Piccinno, and Julian Eisenschlos. 2020. Tapas: Weakly supervised table parsing via pre-training. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4320–4333. Wonseok Hwang, Jinyeong Yim, Seunghyun Park, and Minjoon Seo. 2019. A comprehensive exploration on wikisql with table-aware word contextualization. arXiv preprint arXiv:1902.01069. Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282. Vladimir Karpukhin, Barlas O˘guz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453–466. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. arXiv preprint arXiv:2005.11401. Zhengdong Lu, Hang Li, and Ben Kao. 2016. Neural enquirer: learning to query tables in natural language. IEEE Data Eng. Bull., 39(3):63–73. Qin Lyu, Kaushik Chakrabarti, Shobhit Hathi, Souvik Kundu, Jianwen Zhang, and Zheng Chen. 2020. Hybrid ranking network for text-to-sql. arXiv preprint arXiv:2008.04759. 4087 Sewon Min, Danqi Chen, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019. A discrete hard em approach for weakly supervised question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2844– 2857. Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. Ambigqa: Answering ambiguous open-domain questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5783– 5797. Arvind Neelakantan, Quoc V Le, and Ilya Sutskever. 2015. Neural programmer: Inducing latent programs with gradient descent. arXiv preprint arXiv:1511.04834. Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage re-ranking with bert. arXiv preprint arXiv:1901.04085. Barlas Oguz, Xilun Chen, Vladimir Karpukhin, Stan Peshterliev, Dmytro Okhonko, Michael Schlichtkrull, Sonal Gupta, Yashar Mehdad, and Scott Yih. 2020. Unified open-domain question answering with structured and unstructured knowledge. arXiv preprint arXiv:2012.14610. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392. Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418–5426. Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu, Mike Gatford, et al. 1995. Okapi at trec-3. Nist Special Publication Sp, 109:109. Peng Shi, Patrick Ng, Zhiguo Wang, Henghui Zhu, Alexander Hanbo Li, Jun Wang, Cicero Nogueira dos Santos, and Bing Xiang. 2020. Learning contextual representations for semantic parsing with generation-augmented pre-training. arXiv preprint arXiv:2012.10309. Bailin Wang, Richard Shin, Xiaodong Liu, Oleksandr Polozov, and Matthew Richardson. 2020. Rat-sql: Relation-aware schema encoding and linking for text-to-sql parsers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7567–7578. Chenglong Wang, Kedar Tatwawadi, Marc Brockschmidt, Po-Sen Huang, Yi Mao, Oleksandr Polozov, and Rishabh Singh. 2018a. Robust text-to-sql generation with execution-guided decoding. arXiv preprint arXiv:1807.03100. Shuohang Wang, Mo Yu, Xiaoxiao Guo, Zhiguo Wang, Tim Klinger, Wei Zhang, Shiyu Chang, Gerry Tesauro, Bowen Zhou, and Jing Jiang. 2018b. R 3: Reinforced ranker-reader for open-domain question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32. Shuohang Wang, Mo Yu, Jing Jiang, Wei Zhang, Xiaoxiao Guo, Shiyu Chang, Zhiguo Wang, Tim Klinger, Gerald Tesauro, and Murray Campbell. 2018c. Evidence aggregation for answer re-ranking in opendomain question answering. In International Conference on Learning Representations. Zhiguo Wang, Patrick Ng, Xiaofei Ma, Ramesh Nallapati, and Bing Xiang. 2019. Multi-passage bert: A globally normalized bert model for opendomain question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5881–5885. Xiaojun Xu, Chang Liu, and Dawn Song. 2017. Sqlnet: Generating structured queries from natural language without reinforcement learning. arXiv preprint arXiv:1711.04436. Wei Yang, Yuqing Xie, Aileen Lin, Xingyu Li, Luchen Tan, Kun Xiong, Ming Li, and Jimmy Lin. 2019. End-to-end open-domain question answering with bertserini. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 72–77. Pengcheng Yin, Graham Neubig, Wen-tau Yih, and Sebastian Riedel. 2020. Tabert: Pretraining for joint understanding of textual and tabular data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8413– 8426. Tao Yu, Zifan Li, Zilin Zhang, Rui Zhang, and Dragomir Radev. 2018a. Typesql: Knowledgebased type-aware neural text-to-sql generation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 588–594. Tao Yu, Michihiro Yasunaga, Kai Yang, Rui Zhang, Dongxu Wang, Zifan Li, and Dragomir Radev. 2018b. Syntaxsqlnet: Syntax tree networks for complex and cross-domain text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1653–1663. 4088 Tao Yu, Rui Zhang, Heyang Er, Suyi Li, Eric Xue, Bo Pang, Xi Victoria Lin, Yi Chern Tan, Tianze Shi, Zihan Li, et al. 2019. Cosql: A conversational text-to-sql challenge towards cross-domain natural language interfaces to databases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1962–1979. Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, et al. 2018c. Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3911–3921. Rui Zhang, Tao Yu, Heyang Er, Sungrok Shim, Eric Xue, Xi Victoria Lin, Tianze Shi, Caiming Xiong, Richard Socher, and Dragomir Radev. 2019. Editing-based sql query generation for cross-domain context-dependent questions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5341–5352. Victor Zhong, Mike Lewis, Sida I Wang, and Luke Zettlemoyer. 2020. Grounded adaptation for zeroshot executable semantic parsing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6869– 6882. Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103.
|
2021
|
315
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4089–4100 August 1–6, 2021. ©2021 Association for Computational Linguistics 4089 Generation-Augmented Retrieval for Open-Domain Question Answering Yuning Mao1∗, Pengcheng He2, Xiaodong Liu3, Yelong Shen2, Jianfeng Gao3, Jiawei Han1, Weizhu Chen2 1University of Illinois, Urbana-Champaign 2Microsoft Azure AI 3Microsoft Research 1{yuningm2, hanj}@illinois.edu 2,3{penhe, xiaodl, yeshe, jfgao,wzchen}@microsoft.com Abstract We propose Generation-Augmented Retrieval (GAR) for answering open-domain questions, which augments a query through text generation of heuristically discovered relevant contexts without external resources as supervision. We demonstrate that the generated contexts substantially enrich the semantics of the queries and GAR with sparse representations (BM25) achieves comparable or better performance than state-of-the-art dense retrieval methods such as DPR (Karpukhin et al., 2020). We show that generating diverse contexts for a query is beneficial as fusing their results consistently yields better retrieval accuracy. Moreover, as sparse and dense representations are often complementary, GAR can be easily combined with DPR to achieve even better performance. GAR achieves state-of-the-art performance on Natural Questions and TriviaQA datasets under the extractive QA setup when equipped with an extractive reader, and consistently outperforms other retrieval methods when the same generative reader is used.1 1 Introduction Open-domain question answering (OpenQA) aims to answer factoid questions without a pre-specified domain and has numerous real-world applications. In OpenQA, a large collection of documents (e.g., Wikipedia) are often used to seek information pertaining to the questions. One of the most common approaches uses a retriever-reader architecture (Chen et al., 2017), which first retrieves a small subset of documents using the question as the query and then reads the retrieved documents to extract (or generate) an answer. The retriever is crucial as it is infeasible to examine every piece of information in the entire document collection (e.g., millions of Wikipedia passages) and the retrieval accuracy bounds the performance of the (extractive) reader. ∗Work was done during internship at Microsoft Azure AI. 1Our code is available at https://github.com/ morningmoni/GAR. Early OpenQA systems (Chen et al., 2017) use classic retrieval methods such as TF-IDF and BM25 with sparse representations. Sparse methods are lightweight and efficient, but unable to perform semantic matching and fail to retrieve relevant passages without lexical overlap. More recently, methods based on dense representations (Guu et al., 2020; Karpukhin et al., 2020) learn to embed queries and passages into a latent vector space, in which text similarity beyond lexical overlap can be measured. Dense retrieval methods can retrieve semantically relevant but lexically different passages and often achieve better performance than sparse methods. However, the dense models are more computationally expensive and suffer from information loss as they condense the entire text sequence into a fixed-size vector that does not guarantee exact matching (Luan et al., 2020). There have been some recent studies on query reformulation with text generation for other retrieval tasks, which, for example, rewrite the queries to context-independent (Yu et al., 2020; Lin et al., 2020; Vakulenko et al., 2020) or well-formed (Liu et al., 2019) ones. However, these methods require either task-specific data (e.g., conversational contexts, ill-formed queries) or external resources such as paraphrase data (Zaiem and Sadat, 2019; Wang et al., 2020) that cannot or do not transfer well to OpenQA. Also, some rely on timeconsuming training process like reinforcement learning (RL) (Nogueira and Cho, 2017; Liu et al., 2019; Wang et al., 2020) that is not efficient enough for OpenQA (more discussions in Sec. 2). In this paper, we propose GenerationAugmented Retrieval (GAR), which augments a query through text generation of a pre-trained language model (PLM). Different from prior studies that reformulate queries, GAR does not require external resources or downstream feedback via RL as supervision, because it does not rewrite the query but expands it with heuristically discov4090 ered relevant contexts, which are fetched from PLMs and provide richer background information (Table 2). For example, by prompting a PLM to generate the title of a relevant passage given a query and appending the generated title to the query, it becomes easier to retrieve that relevant passage. Intuitively, the generated contexts explicitly express the search intent not presented in the original query. As a result, GAR with sparse representations achieves comparable or even better performance than state-of-the-art approaches (Karpukhin et al., 2020; Guu et al., 2020) with dense representations of the original queries, while being more lightweight and efficient in terms of both training and inference (including the cost of the generation model) (Sec. 6.4). Specifically, we expand the query (question) by adding relevant contexts as follows. We conduct seq2seq learning with the question as the input and various freely accessible in-domain contexts as the output such as the answer, the sentence where the answer belongs to, and the title of a passage that contains the answer. We then append the generated contexts to the question as the generationaugmented query for retrieval. We demonstrate that using multiple contexts from diverse generation targets is beneficial as fusing the retrieval results of different generation-augmented queries consistently yields better retrieval accuracy. We conduct extensive experiments on the Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (Trivia) (Joshi et al., 2017) datasets. The results reveal four major advantages of GAR: (1) GAR, combined with BM25, achieves significant gains over the same BM25 model that uses the original queries or existing unsupervised query expansion (QE) methods. (2) GAR with sparse representations (BM25) achieves comparable or even better performance than the current state-of-the-art retrieval methods, such as DPR (Karpukhin et al., 2020), that use dense representations. (3) Since GAR uses sparse representations to measure lexical overlap2, it is complementary to dense representations: by fusing the retrieval results of GAR and DPR, we obtain consistently better performance than either method used individually. (4) GAR outperforms DPR in the end-to-end QA performance (EM) when the same extractive reader is used: EM=41.8 (43.8 when combining with DPR) 2Strictly speaking, GAR with sparse representations handles semantics before retrieval by enriching the queries, while maintaining the advantage of exact matching. on NQ and 62.7 on Trivia, creating new state-ofthe-art results for extractive OpenQA. GAR also outperforms other retrieval methods under the generative setup when the same generative reader is used: EM=38.1 (45.3 when combining with DPR) on NQ and 62.2 on Trivia. Contributions. (1) We propose GenerationAugmented Retrieval (GAR), which augments queries with heuristically discovered relevant contexts through text generation without external supervision or time-consuming downstream feedback. (2) We show that using generation-augmented queries achieves significantly better retrieval and QA results than using the original queries or existing unsupervised QE methods. (3) We show that GAR, combined with a simple BM25 model, achieves new state-of-the-art performance on two benchmark datasets in extractive OpenQA and competitive results in the generative setting. 2 Related Work Conventional Query Expansion. GAR shares some merits with query expansion (QE) methods based on pseudo relevance feedback (Rocchio, 1971; Abdul-Jaleel et al., 2004; Lv and Zhai, 2010) in that they both expand the queries with relevant contexts (terms) without the use of external supervision. GAR is superior as it expands the queries with knowledge stored in the PLMs rather than the retrieved passages and its expanded terms are learned through text generation. Recent Query Reformulation. There are recent or concurrent studies (Nogueira and Cho, 2017; Zaiem and Sadat, 2019; Yu et al., 2020; Vakulenko et al., 2020; Lin et al., 2020) that reformulate queries with generation models for other retrieval tasks. However, these studies are not easily applicable or efficient enough for OpenQA because: (1) They require external resources such as paraphrase data (Zaiem and Sadat, 2019), search sessions (Yu et al., 2020), or conversational contexts (Lin et al., 2020; Vakulenko et al., 2020) to form the reformulated queries, which are not available or showed inferior domain-transfer performance in OpenQA (Zaiem and Sadat, 2019); (2) They involve time-consuming training process such as RL. For example, Nogueira and Cho (2017) reported a training time of 8 to 10 days as it uses retrieval performance in the reward function and conducts retrieval at each iteration. In contrast, GAR uses freely accessible in-domain contexts like 4091 passage titles as the generation targets and standard seq2seq learning, which, despite its simplicity, is not only more efficient but effective for OpenQA. Retrieval for OpenQA. Existing sparse retrieval methods for OpenQA (Chen et al., 2017) solely rely on the information of the questions. GAR extends to contexts relevant to the questions by extracting information inside PLMs and helps sparse methods achieve comparable or better performance than dense methods (Guu et al., 2020; Karpukhin et al., 2020), while enjoying the simplicity and efficiency of sparse representations. GAR can also be used with dense representations to seek for even better performance, which we leave as future work. Generative QA. Generative QA generates answers through seq2seq learning instead of extracting answer spans. Recent studies on generative OpenQA (Lewis et al., 2020a; Min et al., 2020; Izacard and Grave, 2020) are orthogonal to GAR in that they focus on improving the reading stage and directly reuse DPR (Karpukhin et al., 2020) as the retriever. Unlike generative QA, the goal of GAR is not to generate perfect answers to the questions but pertinent contexts that are helpful for retrieval. Another line in generative QA learns to generate answers without relevant passages as the evidence but solely the question itself using PLMs (Roberts et al., 2020; Brown et al., 2020). GAR further confirms that one can extract factual knowledge from PLMs, which is not limited to the answers as in prior studies but also other relevant contexts. 3 Generation-Augmented Retrieval 3.1 Task Formulation OpenQA aims to answer factoid questions without pre-specified domains. We assume that a large collection of documents C (i.e., Wikipedia) are given as the resource to answer the questions and a retriever-reader architecture is used to tackle the task, where the retriever retrieves a small subset of the documents D ⊂C and the reader reads the documents D to extract (or generate) an answer. Our goal is to improve the effectiveness and efficiency of the retriever and consequently improve the performance of the reader. 3.2 Generation of Query Contexts In GAR, queries are augmented with various heuristically discovered relevant contexts in order to retrieve more relevant passages in terms of both quantity and quality. For the task of OpenQA where the query is a question, we take the following three freely accessible contexts as the generation targets. We show in Sec. 6.2 that having multiple generation targets is helpful in that fusing their results consistently brings better retrieval accuracy. Context 1: The default target (answer). The default target is the label in the task of interest, which is the answer in OpenQA. The answer to the question is apparently useful for the retrieval of relevant passages that contain the answer itself. As shown in previous work (Roberts et al., 2020; Brown et al., 2020), PLMs are able to answer certain questions solely by taking the questions as input (i.e., closedbook QA). Instead of using the generated answers directly as in closed-book QA, GAR treats them as contexts of the question for retrieval. The advantage is that even if the generated answers are partially correct (or even incorrect), they may still benefit retrieval as long as they are relevant to the passages that contain the correct answers (e.g., cooccur with the correct answers). Context 2: Sentence containing the default target. The sentence in a passage that contains the answer is used as another generation target. Similar to using answers as the generation target, the generated sentences are still beneficial for retrieving relevant passages even if they do not contain the answers, as their semantics is highly related to the questions/answers (examples in Sec. 6.1). One can take the relevant sentences in the ground-truth passages (if any) or those in the positive passages of a retriever as the reference, depending on the trade-off between reference quality and diversity. Context 3: Title of passage containing the default target. One can also use the titles of relevant passages as the generation target if available. Specifically, we retrieve Wikipedia passages using BM25 with the question as the query, and take the page titles of positive passages that contain the answers as the generation target. We observe that the page titles of positive passages are often entity names of interest, and sometimes (but not always) the answers to the questions. Intuitively, if GAR learns which Wikipedia pages the question is related to, the queries augmented by the generated titles would naturally have a better chance of retrieving those relevant passages. While it is likely that some of the generated query contexts involve unfaithful or nonfactual information due to hallucination in text generation (Mao et al., 2020) and introduce noise during re4092 trieval, they are beneficial rather than harmful overall, as our experiments show that GAR improve both retrieval and QA performance over BM25 significantly. Also, since we generate 3 different (complementary) query contexts and fuse their retrieval results, the distraction of hallucinated content is further alleviated. 3.3 Retrieval with Generation-Augmented Queries After generating the contexts of a query, we append them to the query to form a generation-augmented query.3 We observe that conducting retrieval with the generated contexts (e.g., answers) alone as queries instead of concatenation is ineffective because (1) some of the generated answers are rather irrelevant, and (2) a query consisting of the correct answer alone (without the question) may retrieve false positive passages with unrelated contexts that happen to contain the answer. Such low-quality passages may lead to potential issues in the following passage reading stage. If there are multiple query contexts, we conduct retrieval using queries with different generated contexts separately and then fuse their results. The performance of one-time retrieval with all the contexts appended is slightly but not significantly worse. For simplicity, we fuse the retrieval results in a straightforward way: an equal number of passages are taken from the top-retrieved passages of each source. One may also use weighted or more sophisticated fusion strategies such as reciprocal rank fusion (Cormack et al., 2009), the results of which are slightly better according to our experiments.4 Next, one can use any off-the-shelf retriever for passage retrieval. Here, we use a simple BM25 model to demonstrate that GAR with sparse representations can already achieve comparable or better performance than state-of-the-art dense methods while being more lightweight and efficient (including the cost of the generation model), closing the gap between sparse and dense retrieval methods. 4 OpenQA with GAR To further verify the effectiveness of GAR, we equip it with both extractive and generative readers for end-to-end QA evaluation. We follow the 3One may create a title field during document indexing and conduct multi-field retrieval but here we append the titles to the questions as other query contexts for generalizability. 4We use the fusion tools at https://github.com/ joaopalotti/trectools. reader design of the major baselines for a fair comparison, while virtually any existing QA reader can be used with GAR. 4.1 Extractive Reader For the extractive setup, we largely follow the design of the extractive reader in DPR (Karpukhin et al., 2020). Let D = [d1, d2, ..., dk] denote the list of retrieved passages with passage relevance scores D. Let Si = [s1, s2, ..., sN] denote the top N text spans in passage di ranked by span relevance scores Si. Briefly, the DPR reader uses BERT-base (Devlin et al., 2019) for representation learning, where it estimates the passage relevance score Dk for each retrieved passage dk based on the [CLS] tokens of all retrieved passages D, and assigns span relevance scores Si for each candidate span based on the representations of its start and end tokens. Finally, the span with the highest span relevance score from the passage with the highest passage relevance score is chosen as the answer. We refer the readers to Karpukhin et al. (2020) for more details. Passage-level Span Voting. Many extractive QA methods (Chen et al., 2017; Min et al., 2019b; Guu et al., 2020; Karpukhin et al., 2020) measure the probability of span extraction in different retrieved passages independently, despite that their collective signals may provide more evidence in determining the correct answer. We propose a simple yet effective passage-level span voting mechanism, which aggregates the predictions of the spans in the same surface form from different retrieved passages. Intuitively, if a text span is considered as the answer multiple times in different passages, it is more likely to be the correct answer. Specifically, GAR calculates a normalized score p(Si[j]) for the j-th span in passage di during inference as follows: p(Si[j]) = softmax(D)[i] × softmax(Si)[j]. GAR then aggregates the scores of the spans with the same surface string among all the retrieved passages as the collective passage-level score.5 4.2 Generative Reader For the generative setup, we use a seq2seq framework where the input is the concatenation of the question and top-retrieved passages and the target output is the desired answer. Such generative readers are adopted in recent methods such as SpanSe5We find that the number of spans used for normalization in each passage does not have significant impact on the final performance (we take N = 5) and using the raw or normalized strings for aggregation also perform similarly. 4093 qGen (Min et al., 2020) and Longformer (Beltagy et al., 2020). Specifically, we use BART-large (Lewis et al., 2019) as the generative reader, which concatenates the question and top-retrieved passages up to its length limit (1,024 tokens, 7.8 passages on average). Generative GAR is directly comparable with SpanSeqGen (Min et al., 2020) that uses the retrieval results of DPR but not comparable with Fusion-in-Decoder (FID) (Izacard and Grave, 2020) since it encodes 100 passages rather than 1,024 tokens and involves more model parameters. 5 Experiment Setup 5.1 Datasets We conduct experiments on the open-domain version of two popular QA benchmarks: Natural Questions (NQ) (Kwiatkowski et al., 2019) and TriviaQA (Trivia) (Joshi et al., 2017). The statistics of the datasets are listed in Table 1. Dataset Train / Val / Test Q-len A-len #-A NQ 79,168 / 8,757 / 3,610 12.5 5.2 1.2 Trivia 78,785 / 8,837 / 11,313 20.2 5.5 13.7 Table 1: Dataset statistics that show the number of samples per data split, the average question (answer) length, and the number of answers for each question. 5.2 Evaluation Metrics Following prior studies (Karpukhin et al., 2020), we use top-k retrieval accuracy to evaluate the performance of the retriever and the Exact Match (EM) score to measure the performance of the reader. Top-k retrieval accuracy is defined as the proportion of questions for which the top-k retrieved passages contain at least one answer span, which is an upper bound of how many questions are “answerable” by an extractive reader. Exact Match (EM) is the proportion of the predicted answer spans being exactly the same as (one of) the ground-truth answer(s), after string normalization such as article and punctuation removal. 5.3 Compared Methods For passage retrieval, we mainly compare with BM25 and DPR, which represent the most used state-of-the-art methods of sparse and dense retrieval for OpenQA, respectively. For query expansion, we re-emphasize that GAR is the first QE approach designed for OpenQA and most of the recent approaches are not applicable or efficient enough for OpenQA since they have task-specific objectives, require external supervision that was shown to transfer poorly to OpenQA, or take many days to train (Sec. 2). We thus compare with a classic unsupervised QE method RM3 (Abdul-Jaleel et al., 2004) that does not need external resources for a fair comparison. For passage reading, we compare with both extractive (Min et al., 2019a; Asai et al., 2019; Lee et al., 2019; Min et al., 2019b; Guu et al., 2020; Karpukhin et al., 2020) and generative (Brown et al., 2020; Roberts et al., 2020; Min et al., 2020; Lewis et al., 2020a; Izacard and Grave, 2020) methods when equipping GAR with the corresponding reader. 5.4 Implementation Details Retriever. We use Anserini (Yang et al., 2017) for text retrieval of BM25 and GAR with its default parameters. We conduct grid search for the QE baseline RM3 (Abdul-Jaleel et al., 2004). Generator. We use BART-large (Lewis et al., 2019) to generate query contexts in GAR. When there are multiple desired targets (such as multiple answers or titles), we concatenate them with [SEP] tokens as the reference and remove the [SEP] tokens in the generation-augmented queries. For Trivia, in particular, we use the value field as the generation target of answer and observe better performance. We take the checkpoint with the best ROUGE-1 F1 score on the validation set, while observing that the retrieval accuracy of GAR is relatively stable to the checkpoint selection since we do not directly use the generated contexts but treat them as augmentation of queries for retrieval. Reader. Extractive GAR uses the reader of DPR with largely the same hyperparameters, which is initialized with BERT-base (Devlin et al., 2019) and takes 100 (500) retrieved passages during training (inference). Generative GAR concatenates the question and top-10 retrieved passages, and takes at most 1,024 tokens as input. Greedy decoding is adopted for all generation models, which appears to perform similarly to (more expensive) beam search. 6 Experiment Results We evaluate the effectiveness of GAR in three stages: generation of query contexts (Sec. 6.1), retrieval of relevant passages (Sec. 6.2), and passage reading for OpenQA (Sec. 6.3). Ablation studies are mostly shown on the NQ dataset to understand the drawbacks of GAR since it achieves 4094 Question: when did bat out of hell get released? Answer: September 1977 {September 1977} Sentence: Bat Out of Hell is the second studio album and the major - label debut by American rock singer Meat Loaf ... released in September 1977 on Cleveland International / Epic Records. {The album was released in September 1977 on Cleveland International / Epic Records.} Title: Bat Out of Hell {Bat Out of Hell} Question: who sings does he love me with reba? Answer: Brooks & Dunn {Linda Davis} Sentence: Linda Kaye Davis ( born November 26, 1962 ) is an American country music singer. {“ Does He Love You ” is a song written by Sandy Knox and Billy Stritch, and recorded as a duet by American country music artists Reba McEntire and Linda Davis.} Title: Does He Love Me [SEP] Does He Love Me (Reba McEntire song) [SEP] I Do (Reba McEntire album) {Linda Davis [SEP] Greatest Hits Volume Two (Reba McEntire album) [SEP] Does He Love You} Question: what is the name of wonder womans mother? Answer: Mother Magda {Queen Hippolyta} Sentence: In the Amazonian myths, she is the daughter of the Amazon queen Sifrat and the male dwarf Shuri, and is the mother of Wonder Woman. {Wonder Woman’s origin story relates that she was sculpted from clay by her mother Queen Hippolyta and given life by Aphrodite.} Title: Wonder Woman [SEP] Diana Prince [SEP] Wonder Woman (2011 TV pilot) {Wonder Woman [SEP] Orana (comics) [SEP] Wonder Woman (TV series)} Table 2: Examples of generated query contexts. The issue of generating wrong answers is alleviated by generating other contexts highly related to the question/answer. Ground-truth references are shown in the {braces}. better performance on Trivia. 6.1 Query Context Generation Automatic Evaluation. To evaluate the quality of the generated query contexts, we first measure their lexical overlap with the ground-truth query contexts. As suggested by the nontrivial ROUGE scores in Table 3, GAR does learn to generate meaningful query contexts that could help the retrieval stage. We next measure the lexical overlap between the query and the ground-truth passage. The ROUGE-1/2/L F1 scores between the original query and ground-truth passage are 6.00/2.36/5.01, and those for the generation-augmented query are 7.05/2.84/5.62 (answer), 13.21/6.99/10.27 (sentence), 7.13/2.85/5.76 (title) on NQ, respectively. Such results further demonstrate that the generated query contexts significantly increase the word overlap between the queries and the positive passages, and thus are likely to improve retrieval results.6 Context ROUGE-1 ROUGE-2 ROUGE-L Answer 33.51 20.54 33.30 Sentence 37.14 24.71 33.91 Title 43.20 32.11 39.67 Table 3: ROUGE F1 scores of the generated query contexts on the validation set of the NQ dataset. 6We use F1 instead of recall to avoid the unfair favor of (longer) generation-augmented query. Case Studies. In Table 2, we show several examples of the generated query contexts and their ground-truth references. In the first example, the correct album release date appears in both the generated answer and the generated sentence, and the generated title is the same as the Wikipedia page title of the album. In the last two examples, the generated answers are wrong but fortunately, the generated sentences contain the correct answer and (or) other relevant information and the generated titles are highly related to the question as well, which shows that different query contexts are complementary to each other and the noise during query context generation is thus reduced. 6.2 Generation-Augmented Retrieval Comparison w. the state-of-the-art. We next evaluate the effectiveness of GAR for retrieval. In Table 4, we show the top-k retrieval accuracy of BM25, BM25 with query expansion (+RM3) (Abdul-Jaleel et al., 2004), DPR (Karpukhin et al., 2020), GAR, and GAR +DPR. On the NQ dataset, while BM25 clearly underperforms DPR regardless of the number of retrieved passages, the gap between GAR and DPR is significantly smaller and negligible when k ≥100. When k ≥500, GAR is slightly better than DPR despite that it simply uses BM25 for retrieval. In contrast, the classic QE method RM3, while showing 4095 Method NQ Trivia Top-5 Top-20 Top-100 Top-500 Top-1000 Top-5 Top-20 Top-100 Top-500 Top-1000 BM25 (ours) 43.6 62.9 78.1 85.5 87.8 67.7 77.3 83.9 87.9 88.9 BM25 +RM3 44.6 64.2 79.6 86.8 88.9 67.0 77.1 83.8 87.7 88.9 DPR 68.3 80.1 86.1 90.3 91.2 72.7 80.2 84.8 GAR 60.9 74.4 85.3 90.3 91.7 73.1 80.4 85.7 88.9 89.7 GAR +DPR 70.7 81.6 88.9 92.0 93.2 76.0 82.1 86.6 Table 4: Top-k retrieval accuracy on the test sets. All baselines are evaluated by ourselves and better than reported in Karpukhin et al. (2020). GAR helps BM25 to achieve comparable or better performance than DPR. marginal improvement over the vanilla BM25, does not achieve comparable performance with GAR or DPR. By fusing the results of GAR and DPR in the same way as described in Sec. 3.3, we further obtain consistently higher performance than both methods, with top-100 accuracy 88.9% and top1000 accuracy 93.2%. On the Trivia dataset, the results are even more encouraging – GAR achieves consistently better retrieval accuracy than DPR when k ≥5. On the other hand, the difference between BM25 and BM25 +RM3 is negligible, which suggests that naively considering top-ranked passages as relevant (i.e., pseudo relevance feedback) for QE does not always work for OpenQA. Results on more cutoffs of k can be found in App. A. Effectiveness of diverse query contexts. In Fig. 1, we show the performance of GAR when different query contexts are used to augment the queries. Although the individual performance when using each query context is somewhat similar, fusing their retrieved passages consistently leads to better performance, confirming that different generation-augmented queries are complementary to each other (recall examples in Table 2). Performance breakdown by question type. In Table 5, we show the top-100 accuracy of the compared retrieval methods per question type on the NQ test set. Again, GAR outperforms BM25 on all types of questions significantly and GAR +DPR achieves the best performance across the board, which further verifies the effectiveness of GAR. 6.3 Passage Reading with GAR Comparison w. the state-of-the-art. We show the comparison of end-to-end QA performance of extractive and generative methods in Table 6. Extractive GAR achieves state-of-the-art performance among extractive methods on both NQ and Trivia datasets, despite that it is more lightweight and computationally efficient. Generative GAR outper1 5 10 20 50 100 200 300 500 1000 k: # of retrieved passages 30 40 50 60 70 80 90 Top-k Accuracy (%) Answer+Sentence+Title Answer+Sentence Answer+Title Answer Title Sentence Figure 1: Top-k retrieval accuracy on the test set of NQ when fusing retrieval results of different generation-augmented queries. Type Percentage BM25 DPR GAR GAR +DPR Who 37.5% 82.1 88.0 87.5 90.8 When 19.0% 73.1 86.9 83.8 88.6 What 15.0% 76.5 82.6 81.5 86.0 Where 10.9% 77.4 89.1 87.0 90.8 Other 9.1% 79.3 78.1 81.8 84.2 How 5.0% 78.2 83.8 83.2 85.5 Which 3.3% 89.0 90.7 94.1 94.9 Why 0.3% 90.0 90.0 90.0 90.0 Table 5: Top-100 retrieval accuracy breakdown of question type on NQ. Best and second best methods in each category are bold and underlined, respectively. forms most of the generative methods on Trivia but does not perform as well on NQ, which is somewhat expected and consistent with the performance at the retrieval stage, as the generative reader only takes a few passages as input and GAR does not outperform dense retrieval methods on NQ when k is very small. However, combining GAR with DPR achieves significantly better performance than both methods or baselines that use DPR as input such as SpanSeqGen (Min et al., 2020) and RAG (Lewis et al., 2020a). Also, GAR outperforms BM25 significantly under both extractive and generative se4096 Method NQ Trivia Extractive Hard EM (Min et al., 2019a) 28.1 50.9 Path Retriever (Asai et al., 2019) 32.6 ORQA (Lee et al., 2019) 33.3 45.0 Graph Retriever (Min et al., 2019b) 34.5 56.0 REALM (Guu et al., 2020) 40.4 DPR (Karpukhin et al., 2020) 41.5 57.9 BM25 (ours) 37.7 60.1 GAR 41.8 62.7 74.8 GAR +DPR 43.8 Generative GPT-3 (Brown et al., 2020) 29.9 71.2 T5 (Roberts et al., 2020) 36.6 60.5 SpanSeqGen (Min et al., 2020) 42.2 RAG (Lewis et al., 2020a) 44.5 56.1 68.0 FID (Izacard and Grave, 2020) 51.4 67.6 80.1 BM25 (ours) 35.3 58.6 GAR 38.1 62.2 GAR +DPR 45.3 Table 6: End-to-end comparison with the state-ofthe-art methods in EM. For Trivia, the left column denotes the open-domain test set and the right is the hidden Wikipedia test set on the public leaderboard. tups, which again shows the effectiveness of the generated query contexts, even if they are heuristically discovered without any external supervision. The best performing generative method FID (Izacard and Grave, 2020) is not directly comparable as it takes more (100) passages as input. As an indirect comparison, GAR performs better than FID when FID encodes 10 passages (cf. Fig. 2 in Izacard and Grave (2020)). Moreover, since FID relies on the retrieval results of DPR as well, we believe that it is a low-hanging fruit to replace its input with GAR or GAR +DPR and further boost the performance.7 We also observe that, perhaps surprisingly, extractive BM25 performs reasonably well, especially on the Trivia dataset, outperforming many recent state-of-the-art methods.8 Generative BM25 also performs competitively in our experiments. Model Generalizability. Recent studies (Lewis et al., 2020b) show that there are significant question and answer overlaps between the training and test sets of popular OpenQA datasets. Specifically, 60% to 70% test-time answers also appear in the training set and roughly 30% test-set questions have a near-duplicate paraphrase in the training set. Such observations suggest that many questions might have been answered by simple question or 7This claim is later verified by the best systems in the NeurIPS 2020 EfficientQA competition (Min et al., 2021). 8We find that taking 500 passages during reader inference instead of 100 as in Karpukhin et al. (2020) improves the performance of BM25 but not DPR. answer memorization. To further examine model generalizability, we study the per-category performance of different methods using the annotations in Lewis et al. (2020b). Method Total Question Overlap Answer Overlap Only No Overlap DPR 41.3 69.4 34.6 19.3 GAR +DPR (E) 43.8 66.7 38.1 23.9 BART 26.5 67.6 10.2 0.8 RAG 44.5 70.7 34.9 24.8 GAR +DPR (G) 45.3 67.9 38.1 27.0 Table 7: EM scores with question-answer overlap category breakdown on NQ. (E) and (G) denote extractive and generative readers, respectively. Results of baseline methods are taken from Lewis et al. (2020b). The observations on Trivia are similar and omitted. As listed in Table 7, for the No Overlap category, GAR +DPR (E) outperforms DPR on the extractive setup and GAR +DPR (G) outperforms RAG on the generative setup, which indicates that better endto-end model generalizability can be achieved by adding GAR for retrieval. GAR +DPR also achieves the best EM under the Answer Overlap Only category. In addition, we observe that a closed-book BART model that only takes the question as input performs much worse than additionally taking topretrieved passages, i.e., GAR +DPR (G), especially on the questions that require generalizability. Notably, all methods perform significantly better on the Question Overlap category, which suggests that the high Total EM is mostly contributed by question memorization. That said, GAR +DPR appears to be less dependent on question memorization given its lower EM for this category.9 6.4 Efficiency of GAR GAR is efficient and scalable since it uses sparse representations for retrieval and does not involve time-consuming training process such as RL (Nogueira and Cho, 2017; Liu et al., 2019). The only overhead of GAR is on the generation of query contexts and the retrieval with generationaugmented (thus longer) queries, whose computational complexity is significantly lower than other methods with comparable retrieval accuracy. We use Nvidia V100 GPUs and Intel Xeon Platinum 8168 CPUs in our experiments. As listed in 9The same ablation study is also conducted on the retrieval stage and similar results are observed. More detailed discussions can be found in App. A. 4097 Training Indexing Retrieval DPR 24h w. 8 GPUs 17.3h w. 8 GPUs 30 min w. 1 GPU GAR 3 ∼6h w. 1 GPU 0.5h w. 35 CPUs 5 min w. 35 CPUs Table 8: Comparison of computational cost between DPR and GAR at different stages. The training time of GAR is for one generation target but different generators can be trained in parallel. Table 8, the training time of GAR is 3 to 6 hours on 1 GPU depending on the generation target. As a comparison, REALM (Guu et al., 2020) uses 64 TPUs to train for 200k steps during pre-training alone and DPR (Karpukhin et al., 2020) takes about 24 hours to train with 8 GPUs. To build the indices of Wikipedia passages, GAR only takes around 30 min with 35 CPUs, while DPR takes 8.8 hours on 8 GPUs to generate dense representations and another 8.5 hours to build the FAISS index (Johnson et al., 2017). For retrieval, GAR takes about 1 min to generate one query context with 1 GPU, 1 min to retrieve 1,000 passages for the NQ test set with answer/title-augmented queries and 2 min with sentence-augmented queries using 35 CPUs. In contrast, DPR takes about 30 min on 1 GPU. 7 Conclusion In this work, we propose Generation-Augmented Retrieval and demonstrate that the relevant contexts generated by PLMs without external supervision can significantly enrich query semantics and improve retrieval accuracy. Remarkably, GAR with sparse representations performs similarly or better than state-of-the-art methods based on the dense representations of the original queries. GAR can also be easily combined with dense representations to produce even better results. Furthermore, GAR achieves state-of-the-art end-to-end performance on extractive OpenQA and competitive performance under the generative setup. 8 Future Extensions Potential improvements. There is still much space to explore and improve for GAR in future work. For query context generation, one can explore multi-task learning to further reduce computational cost and examine whether different contexts can mutually enhance each other when generated by the same generator. One may also sample multiple contexts instead of greedy decoding to enrich a query. For retrieval, one can adopt more advanced fusion techniques based on both the ranking and score of the passages. As the generator and retriever are largely independent now, it is also interesting to study how to jointly or iteratively optimize generation and retrieval such that the generator is aware of the retriever and generates query contexts more beneficial for the retrieval stage. Last but not least, it is very likely that better results can be obtained by more extensive hyper-parameter tuning. Applicability to other tasks. Beyond OpenQA, GAR also has great potentials for other tasks that involve text matching such as conversation utterance selection (Lowe et al., 2015; Dinan et al., 2020) or information retrieval (Nguyen et al., 2016; Craswell et al., 2020). The default generation target is always available for supervised tasks. For example, for conversation utterance selection one can use the reference utterance as the default target and then match the concatenation of the conversation history and the generated utterance with the provided utterance candidates. For article search, the default target could be (part of) the ground-truth article itself. Other generation targets are more taskspecific and can be designed as long as they can be fetched from the latent knowledge inside PLMs and are helpful for further text retrieval (matching). Note that by augmenting (expanding) the queries with heuristically discovered relevant contexts extracted from PLMs instead of reformulating them, GAR bypasses the need for external supervision to form the original-reformulated query pairs. Acknowledgments We thank Vladimir Karpukhin, Sewon Min, Gautier Izacard, Wenda Qiu, Revanth Reddy, and Hao Cheng for helpful discussions. We thank the anonymous reviewers for valuable comments. References Nasreen Abdul-Jaleel, James Allan, W Bruce Croft, Fernando Diaz, Leah Larkey, Xiaoyan Li, Mark D Smucker, and Courtney Wade. 2004. Umass at trec 2004: Novelty and hard. Computer Science Department Faculty Publication Series, page 189. Akari Asai, Kazuma Hashimoto, Hannaneh Hajishirzi, Richard Socher, and Caiming Xiong. 2019. Learning to retrieve reasoning paths over wikipedia graph for question answering. arXiv preprint arXiv:1911.10470. Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. 4098 Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer opendomain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870– 1879, Vancouver, Canada. Association for Computational Linguistics. Gordon V Cormack, Charles LA Clarke, and Stefan Buettcher. 2009. Reciprocal rank fusion outperforms condorcet and individual rank learning methods. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, pages 758–759. Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the trec 2019 deep learning track. arXiv preprint arXiv:2003.07820. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Emily Dinan, Varvara Logacheva, Valentin Malykh, Alexander Miller, Kurt Shuster, Jack Urbanek, Douwe Kiela, Arthur Szlam, Iulian Serban, Ryan Lowe, et al. 2020. The second conversational intelligence challenge (convai2). In The NeurIPS’18 Competition, pages 187–208. Springer. Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909. Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282. Jeff Johnson, Matthijs Douze, and Herv´e J´egou. 2017. Billion-scale similarity search with gpus. arXiv preprint arXiv:1702.08734. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–1611, Vancouver, Canada. Association for Computational Linguistics. Vladimir Karpukhin, Barlas O˘guz, Sewon Min, Ledell Wu, Sergey Edunov, Danqi Chen, and Wentau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452–466. Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6086–6096, Florence, Italy. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461. Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, et al. 2020a. Retrieval-augmented generation for knowledge-intensive nlp tasks. arXiv preprint arXiv:2005.11401. Patrick Lewis, Pontus Stenetorp, and Sebastian Riedel. 2020b. Question and answer test-train overlap in open-domain question answering datasets. arXiv preprint arXiv:2008.02637. Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Query reformulation using query history for passage retrieval in conversational search. arXiv preprint arXiv:2005.02230. Ye Liu, Chenwei Zhang, Xiaohui Yan, Yi Chang, and Philip S Yu. 2019. Generative question refinement with deep reinforcement learning in retrieval-based qa system. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pages 1643–1652. Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle Pineau. 2015. The ubuntu dialogue corpus: A large dataset for research in unstructured multi-turn dialogue systems. arXiv preprint arXiv:1506.08909. Yi Luan, Jacob Eisenstein, Kristina Toutanova, and Michael Collins. 2020. Sparse, dense, and attentional representations for text retrieval. arXiv preprint arXiv:2005.00181. 4099 Yuanhua Lv and ChengXiang Zhai. 2010. Positional relevance model for pseudo-relevance feedback. In Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, pages 579–586. Yuning Mao, Xiang Ren, Heng Ji, and Jiawei Han. 2020. Constrained abstractive summarization: Preserving factual consistency with constrained generation. arXiv preprint arXiv:2010.12723. Sewon Min, Jordan Boyd-Graber, Chris Alberti, Danqi Chen, Eunsol Choi, Michael Collins, Kelvin Guu, Hannaneh Hajishirzi, Kenton Lee, Jennimaria Palomaki, et al. 2021. Neurips 2020 efficientqa competition: Systems, analyses and lessons learned. arXiv preprint arXiv:2101.00133. Sewon Min, Danqi Chen, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019a. A discrete hard EM approach for weakly supervised question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2851– 2864, Hong Kong, China. Association for Computational Linguistics. Sewon Min, Danqi Chen, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2019b. Knowledge guided text retrieval and reading for open domain question answering. arXiv preprint arXiv:1911.03868. Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. Ambigqa: Answering ambiguous open-domain questions. arXiv preprint arXiv:2004.10645. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human-generated machine reading comprehension dataset. Rodrigo Nogueira and Kyunghyun Cho. 2017. Taskoriented query reformulation with reinforcement learning. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 574–583, Copenhagen, Denmark. Association for Computational Linguistics. Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? arXiv preprint arXiv:2002.08910. Joseph Rocchio. 1971. Relevance feedback in information retrieval. The Smart retrieval systemexperiments in automatic document processing, pages 313–323. Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2020. Question rewriting for conversational question answering. arXiv preprint arXiv:2004.14652. Xiao Wang, Craig Macdonald, and Iadh Ounis. 2020. Deep reinforced query reformulation for information retrieval. arXiv preprint arXiv:2007.07987. Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1253–1256. Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. arXiv preprint arXiv:2006.05009. Salah Zaiem and Fatiha Sadat. 2019. Sequence to sequence learning for query expansion. In Proceedings of the AAAI Conference on Artificial Intelligence, Student Abstract Track, volume 33, pages 10075–10076. 4100 A More Analysis of Retrieval Performance We show the detailed results of top-k retrieval accuracy of the compared methods in Figs. 2 and 3. GAR performs comparably or better than DPR when k ≥100 on NQ and k ≥5 on Trivia. 1 5 10 20 50 100 200 300 500 1000 k: # of retrieved passages 20 30 40 50 60 70 80 90 Top-k Accuracy (%) GAR +DPR DPR GAR BM25 +RM3 BM25 Figure 2: Top-k retrieval accuracy of sparse and dense methods on the test set of NQ. GAR improves BM25 and achieves comparable or better performance than DPR when k ≥100. 1 5 10 20 50 100 k: # of retrieved passages 50 55 60 65 70 75 80 85 Top-k Accuracy (%) GAR +DPR DPR GAR BM25 +RM3 BM25 Figure 3: Top-k retrieval accuracy on the Trivia test set. GAR achieves better results than DPR when k ≥5. We show in Table 9 the retrieval accuracy breakdown using the question-answer overlap categories. The most significant gap between BM25 and other methods is on the Question Overlap category, which coincides with the fact that BM25 is unable to conduct question paraphrasing (semantic matching). GAR helps BM25 to bridge the gap by providing the query contexts and even outperform DPR in this category. Moreover, GAR consistently improves over BM25 on other categories and GAR +DPR outperforms DPR as well. Method Total Question Overlap Answer Overlap Only No Overlap BM25 78.8 81.2 85.1 70.6 DPR 86.1 93.2 89.5 76.8 GAR 85.3 94.1 87.9 73.7 GAR +DPR 88.9 96.3 91.7 79.8 Table 9: Top-100 retrieval accuracy by questionanswer overlap categories on the NQ test set.
|
2021
|
316
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4101–4110 August 1–6, 2021. ©2021 Association for Computational Linguistics 4101 Check It Again: Progressive Visual Question Answering via Visual Entailment Qingyi Si1,2, Zheng Lin1∗, Mingyu Zheng1,2, Peng Fu1, Weiping Wang1 1Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China 2School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China {siqingyi,linzheng,zhengmingyu,fupeng,wangweiping}@iie.ac.cn Abstract While sophisticated Visual Question Answering models have achieved remarkable success, they tend to answer questions only according to superficial correlations between question and answer. Several recent approaches have been developed to address this language priors problem. However, most of them predict the correct answer according to one best output without checking the authenticity of answers. Besides, they only explore the interaction between image and question, ignoring the semantics of candidate answers. In this paper, we propose a select-and-rerank (SAR) progressive framework based on Visual Entailment. Specifically, we first select the candidate answers relevant to the question or the image, then we rerank the candidate answers by a visual entailment task, which verifies whether the image semantically entails the synthetic statement of the question and each candidate answer. Experimental results show the effectiveness of our proposed framework, which establishes a new state-of- the-art accuracy on VQA-CP v2 with a 7.55% improvement.1 1 Introduction Visual Question Answering (VQA) task is a multimodal problem which requires the comprehensive understanding of both visual and textual information. Presented with an input image and a question, the VQA system tries to determine the correct answer in the large prediction space. Recently, some studies (Jabri et al., 2016; Agrawal et al., 2016; Zhang et al., 2016; Goyal et al., 2017) demonstrate that VQA systems suffer from the superficial correlation bias (i.e. language priors) caused by accidental correlations between answers and questions. As a result, traditional VQA models always output the ∗Corresponding author: Zheng Lin. 1The code is available at https://github.com/ PhoebusSi/SAR Figure 1: (a) We evaluate the performance of UpDn, LMH, SSL on the VQA-CP v2 test. topN represents the topN accuracy. (b) Visual verification utilizing answer semantics. most common answer(Selvaraju et al., 2019) of the input sample’s question category, no matter what image is given. To address this language priors problem, various approaches have been developed. However, through exploring the characteristics of the existing methods, we find that whether the general VQA models such as UpDn(Anderson et al., 2018) and LXMERT(Tan and Bansal, 2019) or models carefully designed for language priors, as LMH(Clark et al., 2019) and SSL(Zhu et al., 2020), yield a non-negligible problem. Both models predict the correct answer according to one best output without checking the authenticity of answers. Besides, these models have not made good use of the semantics information of answers that could be helpful for alleviating the language-priors. As presented in Figure 1(a), quite a few correct answers usually occur at top N candidates rather than top one. Meanwhile, if the top N candidate answers are given, the image can further verify the visual presence/absence of concepts based on the combination of the question and the candidate 4102 answer. As shown in Figure 1(b), the question is about the color of the bat and two candidate answers are “yellow” and “black”. After checking the correctness of candidate answers, the wrong answer “yellow” which is contradicted with the image can be excluded and the correct answer “black” which is consistent with the image is confirmed. Nevertheless, this visual verification, which utilizes answer semantics to alleviate language priors, has not been fully investigated. In this paper, we propose a select-and-rerank (SAR) progressive framework based on Visual Entailment. The intuition behind the proposed framework comes from two observations. First, after excluding the answers unrelated to the question and image, the prediction space is shrunken and we can obtain a small number of candidate answers. Second, on the condition that a question and one of its candidate answer is bridged into a complete statement, the authenticity of this statement can be inferred by the content of the image. Therefore, after selecting several possible answers as candidates, we can utilize the visual entailment, consisting of image-text pairs, to verify whether the image semantically entails the synthetic statement. Based on the entailment degree, we can further rerank candidate answers and give the model another chance to find the right answer. To summarize, our contributions are as follows: 1. We propose a select-and-rerank progressive framework to tackle the language priors problem, and empirically investigate a range of design choices for each module of this framework. In addition, it is a generic framework, which can be easily combined with the existing VQA models and further boost their abilities. 2. We highlight the verification process between text and image, and formulate the VQA task as a visual entailment problem. This process makes full use of the interactive information of image, question and candidate answers. 3. Experimental results demonstrate that our framework establishes a new state-of-the-art accuracy of 66.73%, outperforming the existing methods by a large margin. 2 Related Work Language-Priors Methods To address the language prior problem of VQA models, a lot of approaches have been proposed, which can be roughly categorized into two lines: (1) Designing Specific Debiasing Models to Reduce Biases. Most works of this line are ensemble-based methods (Ramakrishnan et al., 2018; Grand and Belinkov, 2019; Belinkov et al., 2019; Cadene et al., 2019; Clark et al., 2019; Mahabadi and Henderson, 2019), among these, LMH(Clark et al., 2019) reduces all biases between question-answer pairs by penalizing the samples that can be answered without utilizing image content. (2) Data Augmentation to Reduce Biases. The main idea of such works (Zhang et al., 2016; Goyal et al., 2017; Agrawal et al., 2018) is to carefully construct more balanced datasets to overcome priors. For example, the recent method SSL(Zhu et al., 2020) first automatically generates a set of balanced question-image pairs, then introduces an auxiliary self-supervised task to use the balanced data. CSS(Chen et al., 2020a) balances the data by adding more complementary samples which are generated by masking objects in the image or some keywords in the question. Based on CSS, CL(Liang et al., 2020) forces the model to utilize the relationship between complementary samples and original samples. Unlike SSL and CSS which do not use any extra manual annotations, MUTANT(Gokhale et al., 2020) locates critical objects in the image and critical words in the question by utilizing the extra object-name labels, which directly helps the model to ground the textual concepts in the image. However, above methods only explore the interaction between the image and the question, ignoring the semantics of candidate answers. In this paper, we propose a progressive VQA framework SAR which achieves better interaction among the question, the image and the answer. Answer Re-ranking Although Answer Reranking is still in the infancy in VQA task, it has been widely studied for QA tasks like open-domain question answering, in which models need to answer questions based on a broad range of opendomains knowledge sources. Recent works (Wang et al., 2018b,a; Kratzwald et al., 2019) address this task in a two-stage manner: extract candidates from all passages, then focus on these candidate answers and rerank them to get a final answer. RankVQA(Qiao et al., 2020) introduces Answer Re-ranking method to VQA task. They design an auxiliary task which reranks candidate answers according to their matching degrees with the input image and off-line generated image captions. However, RankVQA still predicts the final answer from 4103 Figure 2: Overview of the progressive framework SAR. the huge prediction space rather than selected candidate answers. 3 Method Figure 2 shows an overview of the proposed selectand-rerank (SAR) framework, which consists of a Candidate Answer Selecting module and an Answer Re-ranking module. In the Candidate Answer Selecting module, given an image and a question, we first use a current VQA model to get a candidate answer set consisting of top N answers. In this module, the answers irrelevant to the question can be filtered out. Next, we formulate the VQA as a VE task in the Answer Re-ranking module, where the image is premise and the synthetic dense caption(Johnson et al., 2016) (combination of the answer and the question ) is hypothesis. We use the cross-domain pre-trained model LXMERT(Tan and Bansal, 2019) as VE scorer to compute the entailment score of each image-caption pair, and thus the answer corresponding to the dense caption with the highest score is our final prediction. 3.1 Candidate Answer Selecting The Candidate Answer Selector (CAS) selects several answers from all possible answers as candidates and thus shrinks the huge prediction space. Given a VQA dataset D = {Ii, Qi}M i=1 with M samples, where Ii ∈I, Qi ∈Q are the image and question of the ith sample and A is the whole prediction space consisting of thousands of answer categories. Essentially, the VQA model applied as CAS is a |A|-class classifier, and is a free choice in our framework. Given an image Ii and a question Qi, CAS first gives the regression scores over all optional answers: P(A|Qi, Ii). Then CAS chooses N answers A∗ i with top N scores as candidates, which is concluded as follows: A∗ i = topN(argsort(P(A|Qi, Ii))) (1) N (hyper-parameter) candidate answers A∗ i = [A1 i , A2 i , ..., AN i ] are selected for each (Ii, Qi) pair by CAS, forming a dataset D ′ = {Ii, Qi, An i }M ,N i=1,n=1 with M ∗N instances, where An i ∈A∗ i , for the next Answer Re-ranking module. In this paper, we mainly use SSL as our CAS. We also conduct experiments to analyze the impact of different CAS and different N. 3.2 Answer Re-ranking 3.2.1 Visual Entailment Visual Entailment (VE) task is proposed by Xie et al. (2019), where the premise is a real-world image, denoted by Pimage, and the hypothesis is a text, denoted by Htext. Given a sample of (Pimage, Htext), the goal of VE task is to determine whether the Htext can be concluded based on the information of Pimage. According to following protocols, the label of the sample is assigned to (1) Entailment, if there is enough evidence in Pimage to conclude Htext is true. (2) Contradiction, if there is enough evidence in Pimage to conclude Htext is false. (3) Neutral, if there is no sufficient evidence in Pimage to give a conclusion about Htext. 3.2.2 VQA As Visual Entailment A question Qi and each of its candidate answers A∗ i can be bridged into a complete statement, and then the image could verify the authenticity of each statement. More specifically, the visual presence of concepts (e.g. “black bat”/“yellow bat”) based on the combination of the question and the correct/wrong candidate answer can be entailed/contradicted by the content of the image. In this way, we achieve better interaction among question, image and answer. Therefore, we formulate VQA as a VE problem, in which the image Ii is premise, and the synthetic statement of an answer An i in A∗ i and question Qi, represented as (Qi,An i ), is hypothesis. For an image, synthetic statements of different 4104 questions describe different regions of the same image. Following Johnson et al. (2016), we also refer to the synthetic statement as “dense caption”. We use A+ i to represent the An i if An i is the correct answer of Qi, use A− i otherwise. There is enough evidence in Ii to prove (Qi,A+ i ) is true, i.e. the visual linguistic semantically entails (Qi,A+ i ). And there is enough evidence in Ii to prove (Qi, A− i ) is false, i.e. the visual linguistic semantically contradicts (Qi, A− i ). Note that, there is no Neutral label in our VE task and we only have two labels: Entailment and Contradiction. 3.2.3 Re-Ranking based on VE We re-rank dense captions by contrastive learning, that is, (Qi,A+ i ) should be more semantically similar to Ii than (Qi,A− i ). The right part of Figure 2 illustrates this idea. The more semantically similar Ii to (Qi,An i ), the deeper the visual entailment degree is. We score the visual entailment degree of Ii to each (Qi,An i ) ∈(Qi,A∗ i ) and rerank the candidate answers A∗ i by this score. The ranking-first answer is our final output. Question-Answer Combination Strategy The answer information makes sense only when combine it with the question. We encode the combination of question and answer text to obtain the joint concept. We design three question-answer combination strategies: R, C and R→C to combine question and answer into synthetic dense caption Ci: R: Replace question category prefix with answer. The prefix of each question is the question category such as “are there”, “what color”, etc. For instance, given a question “How many flowers in the vase?”, its answer “8” and its question category “how many”, the resulting dense caption is “8 flowers in the vase”. Similarly, “No a crosswalk” is the result of question “ Is this a crosswalk?” and answer “No”. We build a dictionary of all question categories of the train set, then we adopt a Forward Maximum Matching algorithm to determine the question category for every test sample. C: Concatenate question and answer directly. For two cases above, the resulting dense captions are “8 How many flowers in the vase?” and “No Is this a crosswalk?”. The resulting dense captions after concatenation are actually rhetorical questions. We deliberately add answer text to the front of question text in order to avoid the answer being deleted when trimming dense captions to the same length. R→C: We first use strategy R at training, which is aimed at preventing the model from excessively focusing on the co-occurrence relation between question category and answer, and then use strategy C at testing to introduce more information for inference. Adopting any strategy above, we combine Qi and each answer in A∗ i to derive the dense captions C∗ i . And thus we have a dataset D ′′ = {Ii, Cn i }M ,N i=1,n=1with M ∗N instances for VE task. VE Scorer We use the pre-trained model LXMERT to score the visual entailment degree of (Ii, Cn i ). LXMERT separately encodes image and caption text in two streams. Next, the separate streams interact through co-attentional transformer layers. In the textual stream, the dense caption is encoded into a high-level concept. Then the visual representations from visual stream can verify the visual presence/absence of the high-level concept. We represent the VE score for the ith image and its nth candidate caption as: sigmoid(Trm(Ii, Cn i )), where Trm() is the 1-demensional output from the dense layers following LXMERT, δ() denotes the sigmoid function. The larger score represents higher entailment degree. We optimize parameters by minimizing the multi-label soft loss: LV E = −1 M ∗N M X i=1 N X n=1 [tn i log(δ(Trm(Ii, Cn i ))) + (1 −tn i )log(1 −δ(Trm(Ii, Cn i )))] (2) where tn i is the soft target score of the nth answer. Combination with Language-Priors Method After Candidate Answer Selecting, the amount of candidate answers decreases from all possible answers to top N. Although some unrelated answers are filtered out, the dataset D ′′ for VE system is still biased. Therefore, we can optionally apply existing language-priors methods to our framework for further reducing language priors. Take the SSL as an example, we apply the loss function of its self-supervised task to our framework by adjusting the loss function to: Lssl = α M ∗N M X i=1 N X n=1 P(I′ i, Cn i ) (3) where (I′ i, Cn i ) denotes the irrelevant imagecaption pairs, α is a down-weighting coefficients. 4105 The probability P(I′ i, Cn i ) could be considered as the confidence of (I′ i, Cn i ) being a relevant pair. We can reformulate the overall loss function: L = LV E + Lssl (4) 3.3 Inference Process Question Type Discriminator Intuitively, most “Yes/No” questions can be answered by the answer “Yes” or “No”. There is no need to provide too many candidate answers for “Yes/No” questions at the test stage. Therefore, we propose a Question Type Discriminator(QTD) to determine the question type and then correspondingly set different numbers of candidate answers, denoted as N′. Specifically, we roughly divided question types (including “Yes/No”, “Num” and “Other”) into yes/no and non-yes/no. A GRU binary classifier is trained with cross-entropy loss and evaluated with 5-fold cross-validation on the train split of each dataset. Then, the trained QTD model with an accuracy about 97% is implemented as an off-line module during the test stage. We will further investigate the effect of N′ on each question type in the next section. Final Prediction In the inference phase, we search for the best dense caption ˆCi among all candidates C∗ i for the ith image. ˆCi = argmax n∈N′ δ(Trm(Ii, Cn i )) (5) The answer ˆAi corresponding to ˆCi is the final prediction. 4 Experiments 4.1 Setting Datasets Our models are trained and evaluated on the VQA-CP v2(Agrawal et al., 2018) dataset, which is well-crafted by re-organizing VQA v2(Goyal et al., 2017) training and validation sets such that answers for each question category (65 categories according to the question prefix) have different distributions in the train and test sets. Therefore, VQA-CP v2 is a natural choice for evaluating VQA model’s generalizability. The questions of VQA-CP v2 include 3 types: “Yes/No”, “Num” and “Other”. Note that the question type and question category (e.g.“what color”) are different. Besides, we also evaluate our models on the VQA v2 validation set for completeness, and compare the accuracy difference between two datasets with the standard VQA evaluation metric(Antol et al., 2015). Baselines We compare our method with the following baseline methods: UpDn(Anderson et al., 2018), AReg(Ramakrishnan et al., 2018), RUBi(Cadene et al., 2019), LMH(Clark et al., 2019), RankVQA(Qiao et al., 2020), SSL(Zhu et al., 2020), CSS(Chen et al., 2020a), CL(Liang et al., 2020) and LXMERT(Tan and Bansal, 2019). Most of them are designed for the language priors problem, while LXMERT represents the recent trend towards utilizing BERT-like pre-trained models(Li et al., 2019; Chen et al., 2020b; Li et al., 2020) which have top performances on various downstream vision and language tasks (including VQA-v2). Note that MUTANT(Gokhale et al., 2020) uses the extra object-name label to ground the textual concepts in the image. For fair comparison, we do not compare with MUTANT. 4.2 Implementation Details In this paper, we mainly choose SSL as our CAS and set N=12 and N=20 for training. To extract image features, we follow previous work and use the pre-trained Faster R-CNN to encode each image as a set of fixed 36 objects with 2048-dimensional feature vectors. We use the tokenizer of LXMERT to segment each dense caption into words. All the questions are trimmed to the same length of 15 or 18, respectively for R or C question-answer combination strategy. In the Answer Re-ranking Module, we respectively incorporate two languagepriors methods, SSL and LMH, into our proposed framework SAR, which is dubbed as SAR+SSL and SAR+LMH. Our models are trained on two TITAN RTX 24GB GPUs. We train SAR+SSL for 20 epochs with batch size of 32, SAR and SAR+LMH for 10 epochs with batch size of 64. For SAR+SSL, we follow the same setting as the original paper(Zhu et al., 2020), except that we don’t need to pre-train the model with the VQA loss before fine-tuning it with the self-supervised loss. The Adam optimizer is adopted with the learning rate 1e–5. For Question Type Discriminator, we use 300dimensional Glove(Pennington et al., 2014) vectors to initialize word embeddings and feed them into a unidirectional GRU with 128 hidden units. When testing on the VAQ-CP v2, N′ ranges from 1-2 for yes/no questions and 5-15 for non-yes/no questions. As for VQA v2, N′ ranges from 1-2 for yes/no 4106 Model VQA-CP v2 test(%)↑ VQA-v2 val(%)↑ GAP ALL Yes/No Num Other All Yes/No Num Other (%)↓ UpDN(Anderson et al., 2018) 39.74 42.27 11.93 46.05 63.48 81.18 42.14 55.66 23.74 Areg(Ramakrishnan et al., 2018) 41.17 65.49 15.48 35.48 62.75 79.84 42.35 55.16 21.58 RUBI(Cadene et al., 2019) 47.11 68.65 20.28 43.18 61.16 14.05 LMH(Clark et al., 2019) 52.45 69.81 44.46 45.54 61.64 77.85 40.03 55.04 9.19 RankVQA(Qiao et al., 2020) 43.05 42.53 13.91 51.32 65.42 82.51 57.75 45.35 22.37 LXMERT(Tan and Bansal, 2019) 46.23 42.84 18.91 55.51 74.16 89.31 56.85 65.14 27.93 SSL(Zhu et al., 2020) 57.59 86.53 29.87 50.03 63.73 6.14 CSS(Chen et al., 2020a) 58.95 84.37 49.42 48.21 59.91 73.25 39.77 55.11 0.96 CL(Liang et al., 2020) 59.18 86.99 49.89 47.16 Top12-SAR(R→C) (Ours) 64.55 83.03 50.05 58.8 70.41 87.87 54.34 61.38 5.86 Top20-SAR(R→C) (Ours) 65.44 83.13 54.52 59.16 70.63 87.91 54.93 61.64 5.19 Top12-SAR+SSL(R→C) (Ours) 64.29 82.86 51.98 57.94 69.84 87.22 54.41 60.70 5.55 Top20-SAR+SSL(R→C) (Ours) 65.32 83.41 54.32 58.85 70.03 87.47 54.59 60.85 4.71 Top12-SAR+LMH(R) (Ours) 65.93 85.38 62.30 56.73 69.13 87.61 50.43 60.03 3.20 Top20-SAR+LMH(R) (Ours) 66.73 86.00 62.34 57.84 69.22 87.46 51.20 60.12 2.49 Table 1: Results on VQA-CP v2 test and VQA-v2 validation set. Overall best scores are bold, our best are underlined. The gap represents the accuracy difference between VQA v2 and VQA-CP v2. questions and 2-5 for non-yes/no questions. 4.3 Results and Analysis 4.3.1 Main Results Performance on two benchmarks VQA-CP-v2 and VQA-v2 is shown in Table 1. We report the best results of SAR, SAR+SSL and SAR+LMH among 3 question-answer combination strategies respectively. “TopN-” represents that N candidate answers (selected by CAS) feed into the Answer Reranking Module for training. Our approach is evaluated with two settings of N (12 and 20). From the results on VQA-CP v2 shown in Table 1, we can observe that: (1) Top20-SAR+LMH establishes a new state-of-the-art accuracy of 66.73% on VQA-CP v2, beating the previous bestperforming method CL by 7.55%. Even without combining language-priors methods in Answer Re-ranking module, our model Top20-SAR outperforms CL by 6.26%. These show the outstanding effectiveness of our proposed SAR framework. (2) SAR+SSL and SAR+LMH achieve much better performance than SSL and LMH, which demonstrates that SAR is compatible with current language-priors methods and could realize their full potential. (3) Compared with another reranking-based model RankVQA, our method elevates the performance by a large margin of 23.68%. This shows the superiority of our proposed progressive select-and-rerank framework over RankVQA which only uses the answer reranking as an auxiliary task. (4) Previous models did not generalize well on all question types. CL is the previous best on the “Yes/No”, “Num” questions and LXMERT on the “Other” questions. In comparison, our model not only rivals the previous best model on the “Yes/No” questions but also improves the best performance on the “Num” and “Other” questions by 12.45% and 3.65%. The remarkable performance on all question types demonstrates that our model makes a significant progress toward a truly comprehensive VQA model. We also evaluate our method on the VQA v2 which is deemed to have strong language biases. As shown in Table 1, our method achieves the best accuracy of 70.63% amongst baselines specially designed for overcoming language priors, and is the closest to the SOTA established by LXMERT which is trained explicitly for the biased data setting. For completeness, the performance gap between two datasets is also compared in Table 1 with the protocol from Chen et al. (2020a). Compared with most previous models which suffer severe performance drops between VQA v2 and VQA-CP v2 (e.g., 27.93% in LXMERT), the Top20-SAR+LMH significantly decreases the performance drop to 2.49%, which demonstrates the effectiveness of our framework to further overcome the language biases. Though CSS achieves a better performance gap, it sacrifices the performance on the VQA v2. Meanwhile, as N rises from 12 to 20, our models achieve better accuracy on both datasets along with a smaller performance gap. This demonstrates that, unlike previous methods, our method can alleviate language priors while maintaining an excellent capability of answering questions. Nonetheless, we 4107 Figure 3: Results from model SAR+SSL(R→C) in VQA-CP v2 with different N during training. Model/CAS UpDn LMH SSL w/o SAR∗ 41.04 53.03 57.66 SAR 61.71 61.65 64.55 SAR+SSL 63.52 61.78 64.29 SAR+LMH 64.98 62.72 65.14 Table 2: Results based on different CAS in VQA-CP v2. We set N=12. ∗indicates the results come from our reimplementation using official released codes. believe that, how to improve the model’s generality and further transform the trade-off between eliminating language priors and answering questions into win–win outcomes, is a promising research direction in the future. 4.3.2 The Effect of N From Figure 3, we can observe that the overall performance is getting better as N increases. The performance improvement on the “Num” and “Other” questions is especially obvious, and there is a very slight drop on the “Yes/No” questions. We believe that SAR can further get better performance by properly increasing N. Due to the resource limitation, the largest N we use is 20 in this paper. 4.3.3 The Effect of Different CAS To find out the potential performance limitation of CAS models, we show the accuracy of 3 CAS models on the VQA-CP v2 test set. As shown in Figure 1 (a), the Top3 accuracy (acc) of 3 models is about 70% and Top6 acc is 80%, which guarantees that sufficient correct answers are recalled by CAS. And thus, the performance limitation of CAS is negligible. We also conduct experiments to investigate the effect of different CAS on SAR. From the results shown in Table 2, we can observe that: (1) Choosing a better VQA model as CAS does not guarantee a better performance, e.g. performance based on Top N Model R C R→C Top12 SAR 59.51 60.24 64.55 SAR+SSL 62.12 62.87 64.29 SAR+LMH 65.93 65.23 65.14 Top20 SAR 60.43 61.86 65.44 SAR+SSL 62.29 63.94 65.32 SAR+LMH 66.73 65.19 66.71 Table 3: Results on the VQA-CP v2 test set based on different question-answer combination strategies: R, C and R→C. The major difference between R and C is whether keeping question prefix which includes 65 categories. UpDn outperforms that based on LMH, but LMH is a better VQA model in overcoming language priors compared with UpDn. This is because a good Candidate Answer Selector has two requirements: (a) It should be able to recall more correct answers. (b) Under the scenario of language biases, wrong answers recalled by CAS at training time should have superficial correlations with the question as strong as possible. However, the ensemble methods, such as LMH, are trained to pay more attention to the samples which are not correctly answered by the question-only model. This seriously reduces the recall rate of those language-priors wrong answers, which leads to the training data for VE is too simple and thus hurts the model’s capability of reducing language priors. (2) If CAS is the general VQA model UpDn rather than LMH and SSL, the improvement brought from the combination with language-priors method in Answer Re-ranking module is more obvious. (3) Even we choose the UpDn, a backbone model of most current works, as our CAS and do not involve any language-priors methods, SAR still achieves a much better accuracy than the previous SOTA model CL by 2.53%, which shows that our basic framework already possesses outstanding capability of reducing language priors. 4.3.4 The Effect of Question-Answer Combination Strategies From the results shown in Table 3, we can observe that: (1) From overall results, R→C achieves or rivals the best performance on three models. On average, R→C outperforms C by 2.02% which demonstrates avoiding the co-occurrence of question category and answer during training time could effectively alleviate language priors; R→C outperforms R by 2.41% which indicates that the informa4108 Model All Yes/No Num Other LXM 46.23 42.84 18.91 55.51 LXM+SSL 53.09 55.07 29.60 58.50 CAS+LXM(R) 55.58 70.91 29.14 54.81 CAS+LXM+SSL(R) 59.41 76.60 40.81 55.51 CAS+LXM+QTD(R) 59.51 83.20 29.17 55.42 CAS+LXM+SSL+QTD(R) 62.12 85.14 41.63 55.68 Table 4: Ablation study to investigate the effect of each component of Top12-SAR+SSL: Candidate Answer Selector (CAS), LXMERT (LXM), Question Type Discriminator (QTD) and SSL. tion of question category is useful in inference. (2) On the SAR and SAR+SSL, C consistently outperforms R, but on the SAR+LMH, we see opposite results. This is probably because our method and the balancing-data method SSL could learn the positive bias resulted from the superficial correlations between question category and answer, which is useful for generalization, but the ensemble-based method LMH will attenuate positive bias during de-biasing process. (3) Even without language priors method, SAR with R→C rivals or outperforms the SAR+SSL and SAR+LMH with R or C, which shows that R→C strategy could help the model to alleviate language priors. As a result, compared with R or C, our framework with R→C only gains a slight performance improvement after using the same language-priors methods. 4.3.5 Ablation Study “CAS+” represents we use the select-and-rerank framework. From Table 4, we can find that: (1) LXM+SSL represents directly applying SSL to LXMERT. Its poor performance shows that the major contribution of our framework does not come from the combination of the language-priors method SSL and pre-trained model LXMERT. (2) Compared with LXM and LXM+SSL, CAS+LXM and CAS+LXM+SSL respectively gain prominent performance boost of 9.35% and 6.32%, which demonstrates the importance and effectiveness of our proposed selectand-rerank procedure. (3) CAS+LXM+QTD(R) and CAS+LXM+SSL+QTD(R) respectively outperform CAS+LXM(R) and CAS+LXM+SSL(R) by 3.93% and 2.71%, which shows the contribution of QTD module. This further demonstrates that choosing appropriate N′ for different question types is a useful step for model performance. (4) CAS+LXM+SSL+QTD improves the performance of CAS+LXM+QTD by 2.61%, which shows that Figure 4: Results from SAR(R), SAR+SSL(R), SAR(R→C) and SAR+LMH(R) with different N ′ during test. To better investigate the impact of N ′ on each question type, we report the results without Question Type Discriminator. Figure 5: Qualitative comparison between our Top20SAR(R→C) and the baseline SSL. The green/red bounding boxes indicate the most important regions resulting from ours/SSL. G-T is ground-truth. current language-priors methods fit our framework well and could further improve performance. 4.3.6 The Effect of N′ From Figure 4, we can find that: (1) The best N′ for yes/no questions is smaller than that for nonyes/no questions due to the nature of yes/no question. (2) As N′ increases, the accuracy of “Num” and “Other” questions rises first and then decreases. There is a trade-off behind this phenomenon: when N′ is too small, the correct answer may not be recalled by CAS; when N′ is too large, the distraction from wrong answers makes it more difficult for model to choose the correct answer. 4.3.7 Qualitative Examples We qualitatively evaluate the effectiveness of our framework. As shown in Figure 5, compared with SSL, SAR performs better not only in question answering but also in visual grounding. With the 4109 help of answer semantics, SAR can focus on the region relevant to the candidate answer and further use the region to verify its correctness. 5 Conclusion In this paper, we propose a select-and-rerank (SAR) progressive framework based on Visual Entailment. Specifically, we first select candidate answers to shrink the prediction space, then we rerank candidate answers by a visual entailment task which verifies whether the image semantically entails the synthetic statement of the question and each candidate answer. Our framework can make full use of the interactive information of image, question and candidate answers. In addition, it is a generic framework, which can be easily combined with the existing VQA models and further boost their abilities. We demonstrate advantages of our framework on the VQA-CP v2 dataset with extensive experiments and analyses. Our method establishes a new state-of-the-art accuracy of 66.73% with an improvement of 7.55% on the previous best. Acknowledgments This work was supported by National Natural Science Foundation of China (No. 61976207, No. 61906187) References Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. 2016. Analyzing the behavior of visual question answering models. In EMNLP. Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. 2018. Don’t just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4971–4980. Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. 2018. Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6077–6086. Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. 2015. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425–2433. Yonatan Belinkov, Adam Poliak, Stuart M Shieber, Benjamin Van Durme, and Alexander M Rush. 2019. Don’t take the premise for granted: Mitigating artifacts in natural language inference. In ACL (1). Remi Cadene, Corentin Dancette, Matthieu Cord, Devi Parikh, et al. 2019. Rubi: Reducing unimodal biases for visual question answering. Advances in Neural Information Processing Systems, 32:841–852. Long Chen, Xin Yan, Jun Xiao, Hanwang Zhang, Shiliang Pu, and Yueting Zhuang. 2020a. Counterfactual samples synthesizing for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10800–10809. Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. 2020b. Uniter: Universal image-text representation learning. In European Conference on Computer Vision, pages 104–120. Springer. Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. 2019. Don’t take the easy way out: Ensemble based methods for avoiding known dataset biases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4060–4073. Tejas Gokhale, Pratyay Banerjee, Chitta Baral, and Yezhou Yang. 2020. Mutant: A training paradigm for out-of-distribution generalization in visual question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 878–892. Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2017. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6904–6913. Gabriel Grand and Yonatan Belinkov. 2019. Adversarial regularization for visual question answering: Strengths, shortcomings, and side effects. NAACL HLT 2019, page 1. Allan Jabri, Armand Joulin, and Laurens Van Der Maaten. 2016. Revisiting visual question answering baselines. In European conference on computer vision, pages 727–739. Springer. Justin Johnson, Andrej Karpathy, and Li Fei-Fei. 2016. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Bernhard Kratzwald, Anna Eigenmann, and Stefan Feuerriegel. 2019. Rankqa: Neural question answering with answer re-ranking. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6076–6085. 4110 Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. 2019. Visualbert: A simple and performant baseline for vision and language. arXiv preprint arXiv:1908.03557. Xiujun Li, Xi Yin, Chunyuan Li, Pengchuan Zhang, Xiaowei Hu, Lei Zhang, Lijuan Wang, Houdong Hu, Li Dong, Furu Wei, et al. 2020. Oscar: Objectsemantics aligned pre-training for vision-language tasks. In European Conference on Computer Vision, pages 121–137. Springer. Zujie Liang, Weitao Jiang, Haifeng Hu, and Jiaying Zhu. 2020. Learning to contrast the counterfactual samples for robust visual question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3285–3292. Rabeeh Karimi Mahabadi and James Henderson. 2019. Simple but effective techniques to reduce biases. arXiv preprint arXiv:1909.06321, 2(3):5. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Yanyuan Qiao, Zheng Yu, and Jing Liu. 2020. Rankvqa: Answer re-ranking for visual question answering. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE. Sainandan Ramakrishnan, Aishwarya Agrawal, and Stefan Lee. 2018. Overcoming language priors in visual question answering with adversarial regularization. In NeurIPS. Ramprasaath R Selvaraju, Stefan Lee, Yilin Shen, Hongxia Jin, Shalini Ghosh, Larry Heck, Dhruv Batra, and Devi Parikh. 2019. Taking a hint: Leveraging explanations to make vision and language models more grounded. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2591–2600. Hao Tan and Mohit Bansal. 2019. Lxmert: Learning cross-modality encoder representations from transformers. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5103–5114. Shuohang Wang, Mo Yu, Jing Jiang, Wei Zhang, Xiaoxiao Guo, Shiyu Chang, Zhiguo Wang, Tim Klinger, Gerald Tesauro, and Murray Campbell. 2018a. Evidence aggregation for answer re-ranking in opendomain question answering. In International Conference on Learning Representations. Zhen Wang, Jiachen Liu, Xinyan Xiao, Yajuan Lyu, and Tian Wu. 2018b. Joint training of candidate extraction and answer selection for reading comprehension. In ACL (1). Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. 2019. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706. Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. 2016. Yin and yang: Balancing and answering binary visual questions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5014–5022. Xi Zhu, Zhendong Mao, Chunxiao Liu, Peng Zhang, Bin Wang, and Yongdong Zhang. 2020. Overcoming language priors with self-supervised learning for visual question answering.
|
2021
|
317
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4111–4124 August 1–6, 2021. ©2021 Association for Computational Linguistics 4111 A Mutual Information Maximization Approach for the Spurious Solution Problem in Weakly Supervised Question Answering Zhihong Shao1, Lifeng Shang2, Qun Liu2, Minlie Huang1∗ 1The CoAI group, DCST, Tsinghua University, Institute for Artificial Intelligence; 1State Key Lab of Intelligent Technology and Systems; 1Beijing National Research Center for Information Science and Technology; 1Tsinghua University, Beijing 100084, China 2Huawei Noah’s Ark Lab [email protected], [email protected] {shang.lifeng, qun.liu}@huawei.com Abstract Weakly supervised question answering usually has only the final answers as supervision signals while the correct solutions to derive the answers are not provided. This setting gives rise to the spurious solution problem: there may exist many spurious solutions that coincidentally derive the correct answer, but training on such solutions can hurt model performance (e.g., producing wrong solutions or answers). For example, for discrete reasoning tasks as on DROP, there may exist many equations to derive a numeric answer, and typically only one of them is correct. Previous learning methods mostly filter out spurious solutions with heuristics or using model confidence, but do not explicitly exploit the semantic correlations between a question and its solution. In this paper, to alleviate the spurious solution problem, we propose to explicitly exploit such semantic correlations by maximizing the mutual information between question-answer pairs and predicted solutions. Extensive experiments on four question answering datasets show that our method significantly outperforms previous learning methods in terms of task performance and is more effective in training models to produce correct solutions. 1 Introduction Weakly supervised question answering is a common setting of question answering (QA) where only final answers are provided as supervision signals while the correct solutions to derive them are not. This setting simplifies data collection, but exposes model learning to the spurious solution problem: there may exist many spurious ways to derive the correct answer, and training a model with spurious solutions can hurt model performance (e.g., misleading the model to produce unreasonable solutions or wrong answers). As shown in Fig 1, ∗*Corresponding author: Minlie Huang. Multi-mention Reading Comprehension Question: In the television series ‘Thunderbirds’, what is Lady Penelope’s surname? Answer: Creighton Ward Document(s): Born on 24 December 2039, Lady Penelope is the 26-year old daughter of aristocrat Lord Hugh Creighton Ward and his wife, Amelia. The early years of her life were spent at Creighton Ward Mansion. … Lady Penelope Creighton Ward is a fictional character introduced in the British mid-1960s Supermarionation television series Thunderbirds, … Perce is the gardener for the 2000 acre Creighton Ward estate and a friend of Parker. … Possible Solution(s): ``Creighton Ward’’ across the document(s), only the third one is correct Discrete Reasoning over Paragraphs Question: How many years after the Battle of Powder River did Powerville Montana become the first establishment in the county? Answer: 2 Paragraph: … From September 1-15, 1865, the Powder River Expedition (1865) battled Native Americans in the Powder River Battles (1865) near the future site of Broadus. On March 17, ①1876, the Battle of Powder River occurred in the south-central part of the county, about southwest of Broadus. In June ②1876 six companies of the 7th Cavalry Regiment (United States) led by Major Marcus Reno marched along the Powder River … On November 1, ③1878, Powderville, Montana became the first establishment in the county, … On April 5, 1879, the Mizpah Creek Incidents … Possible Solution(s): ③1878 - ①1876 ✓ ③1878 - ②1876 ✗ Semantic Parsing Question: Give me the kickoff time of the game that was aired on CBS against the St. Louis Cardinals. Answer: 1:00 Table Header: | week | date | opponent | result | kickoff[a] | game site | tv | attendance | … Possible Solution(s): SELECT (kickoff[a]) WHERE tv=CBS AND opponent=St. Louis Cardinals ✓ SELECT (kickoff[a]) WHERE opponent=St. Louis Cardinals ✗ Figure 1: Examples from three weakly supervised QA tasks, i.e., multi-mention reading comprehension, discrete reasoning, and semantic parsing. Spans in dark gray and green denote semantic correlations between a question and its solution, while spans in orange are spurious information and should not be used in a solution. for multi-mention reading comprehension, many mentions of an answer in the document(s) are irrelevant to the question; for discrete reasoning tasks or text2SQL tasks, an answer can be produced by the equations or SQL queries that do not correctly match the question in logic. Some previous works heuristically selected one possible solution per question for training, e.g., the first answer span in the document (Joshi et al., 2017; Tay et al., 2018; Talmor and Berant, 2019); some treated all possible solutions equally and maximized the sum of their likelihood (maximum marginal likelihood, or MML) (Swayamdipta et al., 2018; Clark and Gardner, 2018; Lee et al., 2019); many others selected solutions according to model confidence (Liang et al., 2018; Min et al., 2019), 4112 i.e., the likelihood of the solutions being derived by the model. A drawback of these methods is that they do not explicitly consider the mutual semantic correlations between a question and its solution when selecting solutions for training. Intuitively speaking, a question often contains vital clues about how to derive the answer, and a wrong solution together with its context often fails to align well with the question. Take the discrete reasoning case in Fig 1 as an example. To answer the question, we need to know the start year of the Battle of Powder River, which is answered by the first 1876; the second 1876 is irrelevant as it is the year of an event that happened during the battle. To exploit the semantic correlations between a question and its solution, we propose to maximize the mutual information between question-answer pairs and model-predicted solutions. As demonstrated by Min et al. (2019), for many QA tasks, it is feasible to precompute a modestly-sized, taskspecific set of possible solutions containing the correct one. Therefore, we focus on handling the spurious solution problem under this circumstance. Specifically, we pair a task-specific model with a question reconstructor and repeat the following training cycle (Fig 2): (1) sample a solution from the solution set according to model confidence, train the question reconstructor to reconstruct the question from that solution, and then (2) train the task-specific model on the most likely solution according to the question reconstructor. During training, the question reconstructor guides the taskspecific model to predict those solutions consistent with the questions. For the question reconstructor, we devise an effective and unified way to encode solutions in different tasks, so that solutions with subtle differences (e.g., different spans with the same surface form) can be easily discriminated. Our contributions are as follows: (1) We propose a mutual information maximization approach for the spurious solution problem in weakly supervised QA, which exploits the semantic correlations between a question and its solution; (2) We conducted extensive experiments on four QA datasets. Our approach significantly outperforms strong baselines in terms of task performance and is more effective in training models to produce correct solutions. 2 Related Work Question answering has raised prevalent attention and has achieved great progress these years. A lot of challenging datasets have been constructed to advance models’ reasoning abilities, such as (1) reading comprehension datasets with extractive answer spans (Joshi et al., 2017; Dhingra et al., 2017), with free-form answers (Kocisk´y et al., 2018), for multi-hop reasoning (Yang et al., 2018), or for discrete reasoning over paragraphs (Dua et al., 2019), and (2) datasets for semantic parsing (Pasupat and Liang, 2015; Zhong et al., 2017; Yu et al., 2018). Under the weakly supervised setting, the specific solutions to derive the final answers (e.g., the correct location of an answer text, or the correct logic executing an answer) are not provided. This setting is worth exploration as it simplifies annotation and makes it easier to collect large-scale corpora. However, this setting introduces the spurious solution problem, and thus complicates model learning. Most existing approaches for this learning challenge include heuristically selecting one possible solution per question for training (Joshi et al., 2017; Tay et al., 2018; Talmor and Berant, 2019), training on all possible solutions with MML (Swayamdipta et al., 2018; Clark and Gardner, 2018; Lee et al., 2019; Wang et al., 2019), reinforcement learning (Liang et al., 2017, 2018), and hard EM (Min et al., 2019; Chen et al., 2020). All these approaches either use heuristics to select possibly reasonable solutions, rely on model architectures to bias towards correct solutions, or use model confidence to filter out spurious solutions in a soft or hard way. They do not explicitly exploit the semantic correlations between a question and its solution. Most relevantly, Cheng and Lapata (2018) focused on text2SQL tasks; they modeled SQL queries as the latent variables for question generation, and maximized the evidence lower bound of log likelihood of questions. A few works treated solution prediction and question generation as dual tasks and introduced dual learning losses to regularize learning under the fully-supervised or the semi-supervised setting (Tang et al., 2017; Cao et al., 2019; Ye et al., 2019). In dual learning, a model generates intermediate outputs (e.g., the task-specific model predicts solutions from a question) while the dual model gives feedback signals (e.g., the question reconstructor computes the likelihood of the question conditioned on predicted solutions). This method is featured in three aspects. First, both models need training on fully-annotated data so that they can produce reasonable intermediate outputs. Second, the intermediate outputs can 4113 introduce noise during learning as they are sampled from models but not restricted to solutions with correct answer or valid questions. Third, this method typically updates both models with reinforcement learning while the rewards provided by a dual model can be unstable or of high variance. By contrast, we focus on the spurious solution problem under the weakly supervised setting and propose a mutual information maximization approach. Solutions used for training are restricted to those with correct answer. What’s more, though a taskspecific model and a question reconstructor interact with each other, they do not use the likelihood from each other as rewards, which can stabilize learning. 3 Method 3.1 Task Definition For a QA task, each instance is a tuple ⟨d, q, a⟩, where q denotes a question, a is the answer, and d is reference information such as documents for reading comprehension, or table headers for semantic parsing. A solution z is a task-specific derivation of the answer, e.g., a particular span in a document, an equation, or a SQL query (as shown in Fig 1). Let f(·) be the task-specific function that maps a solution to its execution result, e.g., by returning a particular span, solving an equation, or executing a SQL query. Our goal is to train a task-specific model Pθ(z|d, q) that takes ⟨d, q⟩as input and predicts a solution z satisfying f(z) = a. Under the weakly supervised setting, only the answer a is provided for training while the groundtruth solution ¯z is not. We denote the set of possible solutions as Z = {z|f(z) = a}. In cases where the search space of solution is large, we can usually approximate Z so that it contains the ground-truth solution ¯z with a high probability (Min et al., 2019; Wang et al., 2019). Note that Z is task-specific, which will be instantiated in section 4. During training, we pair the task-specific model Pθ(z|d, q) with a question reconstructor Pφ(q|d, z) and maximize the mutual information between ⟨q, a⟩and z. During test, given ⟨d, q⟩, we use the taskspecific model to predict a solution and return the execution result. 3.2 Learning Method Given an instance ⟨d, q, a⟩, the solution set Z usually contains only one solution that best fits the instance while the rest are spurious. We propose to exploit the semantic correlations between a quesA Case of Discrete Reasoning over Paragraphs Question q = “How many years after the Battle of Powder River did Powerville Montana become the first establishment in the county?” Answer a = “2” Paragraph d = “… On March 17, ①1876, the Battle of Powder River occurred in the southcentral part of the county ... In June ②1876 six companies of … On November 1, ③1878, Powderville, Montana became the first establishment in the county…” Solution Set Z = { z1 = ③1878 - ①1876 , z2 = ③1878 - ②1876 } 0.45 0.55 0 0.2 0.4 0.6 Category 1 Category 2 Question Reconstructor % Sample &! Maximize log *"(,|., &!) -0.6 -0.3 0 0.2 0.4 0.6 0.8 Category 1 Category 2 Task-specific Model 1 Score Maximize log *#(&!!|., ,) Score Training Cycle 1. 2. log *"(,|., &) *#(&|., ,, 2) &$ &% &$ &% &!! = 245 627 &∈( *"(,|., &) Figure 2: Illustration of the learning method. tion and its solution to alleviate the spurious solution problem via mutual information maximization. Our objective is to obtain the optimal taskspecific model θ∗that maximizes the following conditional mutual information: θ∗= arg max θ Iθ(⟨q, a⟩; z|d) = arg max θ H(⟨q, a⟩|d) −Hθ(⟨q, a⟩|d, z) = arg max θ −Hθ(⟨q, a⟩|d, z) = arg max θ EP (d,q,a)EPθ(z|d,q,a) log Pθ(q, a|d, z) (1) where Iθ(⟨q, a⟩; z|d) denotes conditional mutual information between ⟨q, a⟩and z over P(d, q, a)Pθ(z|d, q, a). H(·|·) is conditional entropy of random variable(s). P(d, q, a) is the probability of an instance from the training distribution. Pθ(z|d, q, a) is the posterior prediction probability of z (∈Z) which is the prediction probability Pθ(z|d, q) normalized over Z: Pθ(z|d, q, a) = ( Pθ(z|d,q) P z′ ∈Z Pθ(z′ |d,q) z ∈Z 0 z /∈Z (2) Note that computing Pθ(q, a|d, z) is intractable. We therefore introduce a question reconstructor Pφ(q|d, z) and approximate Pθ(q, a|d, z) with I(f(z) = a)Pφ(q|d, z) where I(·) denotes indicator function. Eq. 1 now becomes: θ∗= arg max θ L1 + L2 L1 = EP (d,q,a)EPθ(z|d,q,a) log Pφ(q|d, z) L2 = EP (d,q,a)EPθ(z|d,q,a) log Pθ(q, a|d, z) Pφ(q|d, z) (3) To optimize Eq. 3 is to repeat the following training cycle, which is analogous to the EM algorithm: 1. Minimize L2 w.r.t. the question reconstructor φ to draw Pφ(q|d, z) close to Pθ(q, a|d, z), by sampling a solution z ′ ∈Z according to its posterior prediction probability Pθ(z|d, q, a) (see Eq. 2) and maximizing log Pφ(q|d, z ′). 4114 Bart Encoder <s> a b <sol> op1 <span> op2 </s> Refers to Reference Infomation Solution (a) BART Encoder Inputs (b) BART Encoder Attention Mask <s> a b c <sol> op1 <span> op2 </s> <s> a <sol> op1 <span> op2 </s> Queries Targets c a b d b c a b d a b d Figure 3: Solution encoding. (a) For BART encoder inputs, ⟨s⟩and ⟨/s⟩denote start and end of input sequence, respectively. ⟨sol⟩denotes start of solution. ⟨span⟩is the placeholder of the referred span in reference information (e.g., the second ab in this figure. (b) For attention mask, gray circles block attention. ⟨span⟩retrieves the contextual representation(s) of the referred span by only attending to the referred span. reference information and the solution (except for the token ⟨span⟩) are kept from attending to each other. 2. Maximize L1 w.r.t. the task-specific model θ. L1 can be seen as a reinforcement learning objective with log Pφ(q|d, z) being the reward function. During training, the reward function is dynamically changing and may be of high variance. As we can compute the reward for all z ∈Z, we therefore adopt a greedy but more stable update method, i.e., to maximize log Pθ(z ′′|d, q) where z ′′ = arg maxz∈Z log Pφ(q|d, z) is the best solution according to the question reconstructor. We illustrate the above training cycle in Fig 2. 3.3 Question Reconstructor The question reconstructor Pφ(q|d, z) takes reference information d and a solution z as input, and reconstructs the question q. We use BARTbase, a pre-trained Seq2Seq model, as the question reconstructor so that semantic correlations between questions and solutions can be better captured. A solution typically consists of task-specific operation token(s) (e.g., COUNT for discrete reasoning or semantic parsing), literal(s) (e.g., numeric constants for discrete reasoning or semantic parsing), or span(s) from a question or reference information (e.g., for most QA tasks). It is problematic to just feed the concatenation of d and the surface form of z to the BART encoder; otherwise, different spans with the same surface form can no longer be discriminated as their contextual semantics are lost. To effectively encode d and z, we devise a unified solution encoding as in Fig 3 which is applicable to solutions of various types. Specifically, we leave most of the surface form of z unchanged, except that we replace any span from reference information with a placeholder ⟨span⟩. The representation of ⟨span⟩is computed by forcing it to only attend to the contextual representation(s) of the referred span. To obtain disentangled and robust representations of reference information and a solution, we keep reference information and the solution (except for the token ⟨span⟩) from attending to each other. Intuitively speaking, semantics of reference information should not be affected by a solution, and the representations of a solution should largely determined by its internal logic. 3.4 Solution Set While our learning method and question reconstructor are task-agnostic, solutions are usually taskspecific. Precomputing solution sets needs formal definitions of solutions which define the search space of solutions. A possible search method is to exhaustively enumerate all solutions that produce the correct answer. We will introduce the definitions of solutions for different tasks in section 4. 4 Experiments Datasets # Examples |Z| Train Dev Test Avg Median Multi-mention Reading Comprehension Quasar-T 37,012 3,000 3,000 8.1 4 WebQuestions 3,778 2,032 52.1 36 Discrete Reasoning over Paragraphs DROP 69,669 7,740 9,535 5.1 2 Semantic Parsing WikiSQL 56,355 8,421 15,878 315.4 4 Table 1: Statistics of the datasets we used. Statistics of the size of solution set |Z| are computed on Train sets. Following Min et al. (2019), we conducted experiments on three QA tasks, namely multi-mention reading comprehension, discrete reasoning over paragraphs, and semantic parsing. This section introduces baselines, the definitions of solutions in different tasks, how the solution set can be precomputed, and our experimental results. Statistics of the datasets we used are presented in Table 1. 4115 For convenience, we denote reference information as d = [d1, d2, ..., d|d|] and denote a question as q = [q1, q2, ..., q|q|] where di and qj are a token of d and q respectively. A span from reference information and a question span is represented as (s, e)d and (s, e)q respectively, where s and e are start and end index of the span respectively. 4.1 Baselines First Only (Joshi et al., 2017) which trains a reading comprehension model by maximizing log Pθ(z|d, q) where z is the first answer span in d. MML (Min et al., 2019) which maximizes log P z∈Z Pθ(z|d, q). HardEM (Min et al., 2019) which maximizes log maxz∈ZPθ(z|d, q). HardEM-thres (Chen et al., 2020): a variant of HardEM that optimizes only on confident solutions, i.e., to maximize maxz∈ZI(Pθ(z|d, q) > γ) log Pθ(z|d, q) where γ is an exponentially decaying threshold. γ is initialized such that a model is trained on no less than half of training data at the first epoch. We halve γ after each epoch. VAE (Cheng and Lapata, 2018): a method that views a solution as the latent variable for question generation and adopts the training objective of Variational Auto-Encoder (VAE) (Kingma and Welling, 2014) to regularize the task-specific model. The overall training objective is given by: θ∗, φ∗= arg max θ,φ L(θ, φ) L(θ, φ) = Lmle(θ) + λLvae(θ, φ) = X z∈B log Pθ(z|d, q) + λEPθ(z|d,q) log Pφ(q|d, z) Pθ(z|d, q) where θ denotes a task-specific model and φ is our question reconstructor. Lmle(θ) is the total log likelihood of the set of model-predicted solutions (denoted by B) which derive the correct answer. Lvae(θ, φ) is the evidence lower bound of the log likelihood of questions. λ is the coefficient of Lvae(θ, φ). This method needs pre-training both θ and φ before optimizing the overall objective L(θ, φ). Notably, model θ optimizes on Lvae(θ, φ) via reinforcement learning. We tried stabilizing training by reducing the variance of rewards and setting a small λ. 4.2 Multi-Mention Reading Comprehension Multi-mention reading comprehension is a natural feature of many QA tasks. Given a document d and a question q, a task-specific model is required to locate the answer text a which is usually mentioned many times in the document(s). A solution is defined as a document span. The solution set Z is computed by finding exact match of a: Z = {z = (s, e)d|[ds, ..., de] = a} We experimented on two open domain QA datasets, i.e., Quasar-T (Dhingra et al., 2017) and WebQuestions (Berant et al., 2013). For Quasar-T, we retrieved 50 reference sentences from ClueWeb09 for each question; for WebQuestions, we used the 2016-12-21 dump of Wikipedia as the knowledge source and retrieved 50 reference paragraphs for each question using a Lucene index system. We used the same BERTbase (Devlin et al., 2019) reading comprehension model and data preprocessing from (Min et al., 2019). Quasar-T WebQuestions Dev Test Test EM F1 EM F1 EM F1 First Only 36.0 43.9 35.6 42.8 16.7 22.6 MML 40.1 47.4 39.1 46.5 18.4 25.0 HardEM 41.5 49.1 40.7 47.7 18.0 24.2 HardEM-thres 42.8 50.2 41.9 49.4 19.0 25.3 Ours 44.7‡ 52.6‡ 44.0‡ 51.5‡ 20.4‡ 27.2‡ Table 2: Evaluation on multi-mention reading comprehension datasets. Numbers marked with ‡ are significantly better than the others (t-test, p-value < 0.05). Results: Our method outperforms all baselines on both datasets (Table 2). The improvements can be attributed to the effectiveness of solution encoding, as solutions for this task are typically different spans with the same surface form, e.g., in Qusart-T, all z ∈Z share the same surface form. 4.3 Discrete Reasoning over Paragraphs Some reading comprehension tasks pose the challenge of comprehensive analysis of texts by requiring discrete reasoning (e.g., arithmetic calculation, sorting, and counting) (Dua et al., 2019). In this task, given a paragraph d and a question q, an answer a can be one of the four types: numeric value, a paragraph span or a question span, a sequence of paragraph spans, and a date from the paragraph. The definitions of z depend on answer types (Table 4). These solutions can be searched by following Chen et al. (2020). Note that some solutions involve numbers in d. We treated those numbers as spans while reconstructing q from z. We experimented on DROP (Dua et al., 2019). As the original test set is hidden, for convenience of 4116 Overall Test Number (61.97%) Span (31.47%) Spans (4.99%) Date (1.57%) EM F1 EM F1 EM F1 EM F1 EM F1 MML 58.99‡ 62.30‡ 55.38 55.58 69.96 75.51 39.29 66.01 42.57 49.05 HardEM 68.52‡ 71.88‡ 68.40 68.70 73.50 79.25 44.79 69.63 49.32 56.87 HardEM-thres 69.06 72.35‡ 69.05 69.39 74.61 79.79 39.50 66.38 52.67 58.75 VAE 32.34‡ 36.28‡ 51.65 52.35 0.37 10.01 0.00 8.89 0.00 4.11 Ours 69.35 72.92 69.96 70.27 73.38 79.32 42.86 70.42 48.67 57.47 Table 3: Evaluation on DROP. We used the public development set of DROP as our test set. We also provide performance breakdown of different question types on our test set. Results on the overall test set marked with ‡ are significantly worse than the best one (t-test, p-value < 0.05). Numeric Answers Arithmetic z =n1[, o1, n2[, o2, n3]], s.t. o1, o2 ∈{+, −}, n1, n2, n3 ∈Nd ∪S Sorting z =o{nk}k≥1, s.t. o ∈{max, min}, nk ∈Nd Counting z =|{(sk, ek)d}k≥1| Non-numeric Answers Span(s) z = {(sk, ek)t}k≥1, s.t. t ∈{d, q} Sorting z =o{kv⟨(sk, ek)d, nk⟩}k≥1, s.t. o ∈{argmax, argmin}, nk ∈Nd Table 4: Definitions of solutions for numeric answers and non-numeric answers. Nd is the set of numbers in d, and S is a set of pre-defined numbers. For arithmetic solutions for numeric answers, z = n1[, o1, n2[, o2, n3]] denotes equations with no more than three operands. For solutions of sorting type for non-numeric answers, kv⟨·, ·⟩is a key-value pair where the key is a span in d and the value is its associated number from d. argmax (argmin) returns the key with the largest (smallest) value. analysis, we used the public development set as our test set, and split the public train set into 90%/10% for training and development. We used Neural Symbolic Reader (NeRd) (Chen et al., 2020) as the taskspecific model. NeRd is a Seq2Seq model which encodes a question and a paragraph, and decodes a solution (e.g., count (paragraph span(s1, e1), paragraph span(s2, e2)) where paragraph span(si, ei) means a paragraph span starting at si and ending at ei). We used the precomputed solution sets provided by Chen et al. (2020)1. Data preprocessing 1Our implementation of NeRd has four major differences from that of (Chen et al., 2020). (1) Instead of choosing BERTlarge as encoder, we chose the discriminator of Electrabase (Clark et al., 2020) which is of a smaller size. (2) We did not use moving averages of trained parameters. (3) We did not use the full public train set for training but used 10% of it for development. (4) For some questions, it is hard to guarantee that a precomputed solution set covers the ground-truth solution. For example, the question How many touchdowns did was also kept the same. Results: As shown in Table 3, our method significantly outperforms all baselines in terms of F1 score on our test set. We also compared our method with the baseline VAE which uses a question reconstructor φ to adjust the task-specific model θ via maximizing a variational lower bound of log P(q|d) as the regularization term Lvae(θ, φ). To pre-train the task-specific model for this method, we simply obtained the best task-specific model trained with HardEM-thres. VAE optimizes the task-specific model on Lvae(θ, φ) with reinforcement learning where Pφ(q|d, z) is used as learning signals for the task-specific model. Despite our efforts to stabilize training, the F1 score still dropped to 36.28 after optimizing the overall objective L(θ, φ) for 1,000 steps. By contrast, our method does not use Pφ(q|d, z) to compute learning signals for the taskspecific model but rather uses it to select solutions to train the task-specific model, which makes a better use of the question reconstructor. 4.4 Semantic Parsing Text2SQL is a popular semantic parsing task. Given a question q and a table header d = [h1, ..., hL] where hl is a multi-token column, a parser is required to parse q into a SQL query z and return the execution results. Under the weakly supervised setting, only the final answer is provided while the SQL query is not. Following Min et al. (2019), Z is approximated as a set of non-nested SQL queries with no more than three conditions: Z = {z = (zsel, zagg, {zcond k }3 k=1)|f(z) = a, zsel ∈{h1, ..., hL}, zcond k ∈{none} ∪C, zagg ∈{none, sum, mean, max, min, count}} Brady throw? needs counting, but the related mentions are not known. (Chen et al., 2020) partly solved this problem by adding model-predicted solutions (with correct answer) into the initial solution sets as learning proceeds. In this paper, we kept the initial solution sets unchanged during training, so that different QA tasks share the same experimental setting. 4117 where zagg is an aggregating operator and zsel is the operated column (a span of d). C = {(h, o, v)} is the set of all possible conditions, where h is a column, o ∈{=, <, >}, and v is a question span. We experimented on WikiSQL (Zhong et al., 2017) under the weakly supervised setting2. We chose SQLova (Hwang et al., 2019) as the taskspecific model which is a competitive text2SQL parser on WikiSQL. Hyperparameters were kept the same as in (Hwang et al., 2019). We used the solution sets provided by Min et al. (2019). Results: All models in Table 5 do not apply execution-guided decoding during inference. Our method achieves new state-of-the-art results under the weakly supervised setting. Though without supervision of ground-truth solutions, our execution accuracy (i.e., accuracy of execution results) on the test set is close to that of the fully supervised SQLova. Notably, GRAPPA focused on representation learning and used a stronger task-specific model while we focus on the learning method and outperform GRAPPA with a weaker model. 5 Ablation Study 5.1 Performance on Test Data with Different Size of Solution Set Fig 4 shows the performance on test data with different size of solution set3. Our method consistently outperforms HardEM-thres and by a large margin when test examples have a large solution set. 5.2 Effect of |Z| at Training The more complex a question is, the larger the set of possible solutions tends to be, the more likely a model will suffer from the spurious solution problem. We therefore investigated whether our learning method can deal with extremely noisy solution sets. Specifically, we extracted a hard train set from the original train set of WikiSQL. The hard train set consists of 10K training data with the largest Z. The average size of Z on the hard train set is 1,554.6, much larger than that of the original train set (315.4). We then compared models trained on the original train set and the hard train set using different learning methods. 2WikiSQL has annotated ground-truth SQL queries. We only used them for evaluation but not for training. 3In this experiment, |Z| is only seen as a property of an example. Evaluated solutions are predicted by the task-specific model but not from Z. Model Execution Accuracy Dev Test Fully-supervised Setting SQLova (Hwang et al., 2019) 87.2 86.2 HydraNet (Lyu et al., 2020) 89.1 89.2 Weakly-supervised Setting MeRL (Agarwal et al., 2019) 74.9 74.8 GRAPPA (Yu et al., 2021) 85.9 84.7 MML(Min et al., 2019) 70.6 70.5 HardEM 84.5‡ 84.1‡ HardEM-thres 85.2† 84.1‡ Ours 85.9 85.6 Table 5: Evaluation on WikiSQL. Accuracy that is significantly lower than the highest one is marked with † for p-value < 0.1, and ‡ for p-value < 0.05 (t-test). 0 20 40 60 80 100 62 65 68 71 74 77 [0,3) [3,5) [5,7) [7,9) [9,+∞) % of Data F1 Score |Z| % of Data HardEM-thres Ours 0 25 50 75 100 65 68 71 74 77 [0,3) [3,5) [5,7) [7,9) [9,+∞) % of Data F1 Score |Z| % of Data HardEM-thres Ours Figure 4: Performance on test examples with different size of Z on DROP. Figure 5: Logical form accuracy (left) and execution accuracy (right) on dev set and test set of WikiSQL. A method marked with Ori. Train or Hard Train means the evaluated model is trained on the original train set or a hard subset of training data, respectively. The hard train set consists of 10K training data with the largest solution set; the average size of solution set is 1,554.6. As shown in Fig 5, models trained with our method consistently outperform baselines in terms of logical form accuracy (i.e., accuracy of predicted solutions) and execution accuracy. When using the hard train set, the logical form accuracy of models trained with HardEM or HardEM-thres drop to below 14%. Compared with HardEM, HardEM-thres is better when trained on the original train set but is worse when trained on the hard train set. These indicate that model confidence can be unreliable and thus insufficient to filter out spurious solutions. By contrast, our method explicitly exploits the semantic correlations between a question and a solution, thus much more resistant to spurious solutions. 4118 Training Epochs 2 4 6 8 10 BARTbase w/ HardEM 65.1 60.8 59.7 58.6 61.0 SQLova w/ HardEM 61.3 62.2 61.8 61.8 61.7 SQLova w/ Ours 79.7 82.8 79.8 81.2 87.4 Table 6: Accuracy on the SQL selection task. The hard train set was used for training. BARTbase w/ HardEM and SQLova w/ HardEM are a BARTbase parser and SQLova, respectively; both were trained with HardEM. SQLova w/ Ours is SQLova trained with the proposed mutual information maximization approach (using BARTbase question reconstructor). 5.3 Effect of the Question Reconstructor As we used BARTbase as the question resconstructor, we investigated how our question reconstructor contributes to performance improvements. We first investigated whether BARTbase itself is less affected by the spurious solution problem than the task-specific models. Specifically, we viewed text2SQL as a sequence generation task and finetuned a BARTbase on the hard train set of WikiSQL with HardEM. The input of BART shares the same format as that of SQLova, which is the concatenation of a question and a table header. The output of BART is a SQL query. Without constraints on decoding, BART might not produce valid SQL queries. We therefore evaluated models on a SQL selection task instead: for each question in the development set of WikiSQL, a model picks out the correct SQL from at most 10 candidates by selecting the one with the highest prediction probability. As shown in Table 6, when trained with HardEM, both BARTbase parser and SQLova perform similarly, and underperform our method by a large margin. This indicates that using BARTbase as a task-specific model can not avoid the spurious solution problem. It is our mutual information maximization objective that makes a difference. DROP WikiSQL (Hard Train Set) Dev Test Dev Test EM F1 EM F1 LF. Acc Exe. Acc LF. Acc Exe. Acc T-scratch 61.5 66.3 69.0 72.4 24.7 67.9 24.9 67.5 T-DAE 61.5 66.3 69.4 72.7 49.4 68.9 48.5 68.4 BARTbase 61.5 66.4 69.3 72.9 45.8 69.1 45.6 68.4 Table 7: Results with different question reconstructors. LF. Acc and Exe. Acc are logical form accuracy and execution accuracy, respectively. T-scratch is a Transformer without pre-training. T-DAE is a Transformer pre-trained as a denoising auto-encoder of questions. We further investigated the effect of the choice of question reconstructor. We compared BARTbase with two alternatives: (1) T-scratch: a three-layer Transformer (Vaswani et al., 2017) without pretraining and (2) T-DAE: a three-layer Transformer pre-trained as a denoising auto-encoder of questions on the train set; the text infilling pre-training task for BART was used. As shown in Table 7, our method with either of the three question reconstructors outperforms or is at least competitive with baselines, which verifies the effectiveness of our mutual information maximization objective. What’s more, using T-DAE is competitive with BARTbase, indicating that our training objective is compatible with other choices of question reconstructor besides BART, and that using a denoising auto-encoder to initialize the question reconstructor may be beneficial to exploit the semantic correlations between a question and its solution. 6 Evaluation of Solution Prediction As solutions with correct answer can be spurious, we further analyzed the quality of predicted solutions. We randomly sampled 50 test examples from DROP for which our method produced the correct answer, and found that our method also produced the correct solution for 92% of them. To investigate the effect of different learning methods on models’ ability to produce correct solutions, we manually analyzed another 50 test samples for which HardEM, HardEM-thres, and our method produced the correct answer with different solutions. The percentage of samples for which our method produced the correct solution is 58%, much higher than that of HardEM (10%) and HardEMthres (30%). For experimental details, please refer to the appendix. 7 Case Study Fig 6 compares NeRd predictions on four types of questions from DROP when using different learning methods. An observation is that NeRd using our method shows more comprehensive understanding of questions, e.g., in the Arithmetic case, NeRd using our method is aware of the two key elements in the question including the year when missionaries arrived in Ayutthaya and the year when the Seminary of Saint Joseph was built, while NeRd using HardEM-thres misses the first element. What’s more, NeRd using our method is more precise in locating relevant information, e.g., in the first Sorting case, NeRd with our method locates the second appearance of 2 whose contextual semantics matches the question, while NeRd using HardEM-thres locates the first appearance of 2 which is irrelevant. 4119 Span(s) Question: Which team attempted a 2-point conversion? Answer: Rams Paragraph: Hoping to rebound from their road loss to the Patriots, the ①Rams went home for a Week 9 NFC West duel with the Arizona Cardinals … In the second quarter, the Cardinals responded with a vengeance as safety Antrel Rolle returned an interception 40 yards for a touchdown, kicker Neil Rackers got a 36-yard field goal, RB Tim Hightower got a 30-yard TD run, and former ②Rams QB Kurt Warner completed a 56-yard TD pass to WR Jerheme Urban. In the third quarter, Arizona increased its lead as Warner completed a 7-yard TD pass to WR Anquan Boldin. In the fourth quarter, the ③Rams tried to come back as Bulger completed a 3-yard TD pass to WR Torry Holt (with a failed 2-point conversion). However, the Cardinals flew away as Rackers nailed a 30-yard field goal. During the game, the ④Rams inducted former Head Coach Dick Vermeil (who helped the franchise win Super Bowl XXXIV) onto the ⑤ Rams Ring of Honor. Model Prediction: Ours: ③Rams ✓ HardEM-thres: ⑤Rams ✗ Arithmetic Question: How many years after the missionaries arrived in Ayutthaya did the build the Seminary of Saint Joseph? Answer: 2 or 1 Paragraph: In 1664, a group of missionaries led by Franois Pallu, Bishop of Heliopolis, also of the Paris Foreign Missions Society, joined Lambert in the capital city of Ayutthaya after 24 months overland travel and started missionary work. In 1665-66 they built a seminary in Ayutthaya with the approval of King Narai, the Seminary of Saint Joseph. In 1669, Louis Laneau, Bishop of Motella, also a member of the Paris Foreign Missions Society, … Model Prediction: Ours: 1666 - 1664 ✓ HardEM-thres: 1666 - 1665 ✗ Sorting Question: How many yards was the shortest touchdown pass? Answer: 2 Paragraph: The Giants played their Week ①2 home opener against the Green Bay Packers … The Giants responded with a 26-yard scoring strike by Eli Manning to Plaxico Burress. The Giants got a Lawrence Tynes field goal and a 10-7 half time lead. In the second half, the Packers drove 51 yards to start the second half. Favre capped off the scoring drive with a ②2-yard pass to Bubba Franks for a 14-10 lead the Packers would not relinquish… Model Prediction: Ours: min{②2} ✓ HardEM-thres: ①2 ✗ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------Question: How many yards was Sebastian Janikowski's longest field goal? Answer: 49 Paragraph: … The Seahawks immediately trailed on a scoring rally by the Raiders with kicker Sebastian Janikowski nailing a 31-yard field goal. This was followed in the second quarter by QB Jason Campbell's 30-yard TD pass to FB Marcel Reece. Then in the third quarter Janikowski made a 36-yard field goal. Then he made a 22-yard field goal in the fourth quarter to put the Raiders up 16-0 ... with kicker Olindo Mare hitting a 47-yard field goal. However, they continued to trail as Janikowski made a 49-yard field goal … Model Prediction: Ours: max{49, 36} Incomplete HardEM-thres: max{49, 31} Incomplete Counting Question: How many passed did Houshmandzadeh catch? Answer: 2 Paragraph: … In the third quarter, Cincinnati tried to rally as QB Carson Palmer completed an 18-yard TD pass to WR T. J. Houshmandzadeh... Cincinnati tried to come back as Palmer completed a 10-yard TD pass to Houshmandzadeh (with a failed 2-point conversion), but Dallas pulled away with Romo completing a 15-yard TD pass to WR Patrick Crayton. Model Prediction: Ours: |{18-yard TD pass, 10-yard}| ✓ HardEM-thres: 2 ✗ Figure 6: NeRd predictions on four types of questions from DROP when using different learning methods. Spans in dark gray and green denote semantic correlations between a question and its solution, while spans in orange are spurious information and should not be used in a solution. These two observations can be attributed to our mutual information maximization objective which biases a task-specific model towards those solutions that align well with the questions. However, we also observed that when there are multiple mentions of relevant information of the same type, NeRd trained with HardEM-thres or our method has difficulty in recalling them all, e.g., in the second Sorting case, the correct solution should locate all four mentions of Sebastian Janikowski’s field goals while NeRd using either method locates only two of them. We conjecture that this is because the solution sets provided by Chen et al. (2020) are noisy. For example, all precomputed solutions of sorting type for numeric answers involve up to two numbers from reference information, which makes it hard for a model to learn to sort more than two numbers. 8 Conclusion To alleviate the spurious solution problem in weakly supervised QA, we propose to explicitly exploit the semantic correlations between a question and its solution via mutual information maximization. During training, we pair a task-specific model with a question reconstructor which guides the task-specific model to predict solutions that are consistent with the questions. Experiments on four QA datasets demonstrate the effectiveness of our learning method. As shown by automatic and manual analyses, models trained with our method are more resistant to spurious solutions during training, and are more precise in locating information that is relevant to the questions during inference, leading to higher accuracy of both answers and solutions. 9 Acknowledgements This work was partly supported by the NSFC projects (Key project with No. 61936010 and regular project with No. 61876096). This work was also supported by the Guoqiang Institute of Tsinghua University, with Grant No. 2019GQG1 and 2020GQG0005. 4120 References Rishabh Agarwal, Chen Liang, Dale Schuurmans, and Mohammad Norouzi. 2019. Learning to generalize from sparse and underspecified rewards. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 130– 140. PMLR. Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on freebase from question-answer pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, EMNLP 2013, 18-21 October 2013, Grand Hyatt Seattle, Seattle, Washington, USA, A meeting of SIGDAT, a Special Interest Group of the ACL, pages 1533–1544. ACL. Ruisheng Cao, Su Zhu, Chen Liu, Jieyu Li, and Kai Yu. 2019. Semantic parsing with dual learning. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 51–64. Association for Computational Linguistics. Xinyun Chen, Chen Liang, Adams Wei Yu, Denny Zhou, Dawn Song, and Quoc V. Le. 2020. Neural symbolic reader: Scalable integration of distributed and symbolic representations for reading comprehension. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Jianpeng Cheng and Mirella Lapata. 2018. Weaklysupervised neural semantic parsing with a generative ranker. In Proceedings of the 22nd Conference on Computational Natural Language Learning, CoNLL 2018, Brussels, Belgium, October 31 - November 1, 2018, pages 356–367. Association for Computational Linguistics. Christopher Clark and Matt Gardner. 2018. Simple and effective multi-paragraph reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 845–855. Association for Computational Linguistics. Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: pretraining text encoders as discriminators rather than generators. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171–4186. Association for Computational Linguistics. Bhuwan Dhingra, Kathryn Mazaitis, and William W. Cohen. 2017. Quasar: Datasets for question answering by search and reading. CoRR, abs/1707.03904. Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACLHLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 2368– 2378. Association for Computational Linguistics. Wonseok Hwang, Jinyeung Yim, Seunghyun Park, and Minjoon Seo. 2019. A comprehensive exploration on wikisql with table-aware word contextualization. CoRR, abs/1902.01069. Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. 2017. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 1601–1611. Association for Computational Linguistics. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Diederik P. Kingma and Max Welling. 2014. Autoencoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings. Tom´as Kocisk´y, Jonathan Schwarz, Phil Blunsom, Chris Dyer, Karl Moritz Hermann, G´abor Melis, and Edward Grefenstette. 2018. The narrativeqa reading comprehension challenge. Trans. Assoc. Comput. Linguistics, 6:317–328. Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6086–6096. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pretraining for natural language generation, translation, 4121 and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 7871–7880. Association for Computational Linguistics. Chen Liang, Jonathan Berant, Quoc V. Le, Kenneth D. Forbus, and Ni Lao. 2017. Neural symbolic machines: Learning semantic parsers on freebase with weak supervision. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 August 4, Volume 1: Long Papers, pages 23–33. Association for Computational Linguistics. Chen Liang, Mohammad Norouzi, Jonathan Berant, Quoc V. Le, and Ni Lao. 2018. Memory augmented policy optimization for program synthesis and semantic parsing. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr´eal, Canada, pages 10015–10027. Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Qin Lyu, Kaushik Chakrabarti, Shobhit Hathi, Souvik Kundu, Jianwen Zhang, and Zheng Chen. 2020. Hybrid ranking network for text-to-sql. CoRR, abs/2008.04759. Sewon Min, Danqi Chen, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2019. A discrete hard EM approach for weakly supervised question answering. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 2851– 2864. Association for Computational Linguistics. Panupong Pasupat and Percy Liang. 2015. Compositional semantic parsing on semi-structured tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing of the Asian Federation of Natural Language Processing, ACL 2015, July 26-31, 2015, Beijing, China, Volume 1: Long Papers, pages 1470– 1480. The Association for Computer Linguistics. Swabha Swayamdipta, Ankur P. Parikh, and Tom Kwiatkowski. 2018. Multi-mention learning for reading comprehension with neural cascades. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. Alon Talmor and Jonathan Berant. 2019. Multiqa: An empirical investigation of generalization and transfer in reading comprehension. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 4911–4921. Association for Computational Linguistics. Duyu Tang, Nan Duan, Tao Qin, and Ming Zhou. 2017. Question answering and question generation as dual tasks. CoRR, abs/1706.02027. Yi Tay, Anh Tuan Luu, Siu Cheung Hui, and Jian Su. 2018. Densely connected attention propagation for reading comprehension. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montr´eal, Canada, pages 4911–4922. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 49, 2017, Long Beach, CA, USA, pages 5998–6008. Bailin Wang, Ivan Titov, and Mirella Lapata. 2019. Learning semantic parsers from denotations with latent structured alignments and abstract programs. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3772– 3783. Association for Computational Linguistics. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 2369–2380. Association for Computational Linguistics. Hai Ye, Wenjie Li, and Lu Wang. 2019. Jointly learning semantic parser and natural language generator via dual information maximization. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 2090–2101. Association for Computational Linguistics. Tao Yu, Chien-Sheng Wu, Xi Victoria Lin, bailin wang, Yi Chern Tan, Xinyi Yang, Dragomir Radev, richard socher, and Caiming Xiong. 2021. Gra{pp}a: Grammar-augmented pre-training for table semantic parsing. In International Conference on Learning Representations. Tao Yu, Rui Zhang, Kai Yang, Michihiro Yasunaga, Dongxu Wang, Zifan Li, James Ma, Irene Li, Qingning Yao, Shanelle Roman, Zilin Zhang, and 4122 Dragomir R. Radev. 2018. Spider: A largescale human-labeled dataset for complex and crossdomain semantic parsing and text-to-sql task. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 - November 4, 2018, pages 3911–3921. Association for Computational Linguistics. Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103. A Implementation Details A.1 Learning Methods HardEM: We followed Min et al. (2019) to apply annealing to HardEM on reading comprehension tasks: at the training step t, a model optimizes MML objective with a probability of min(t/τ, 0.8) and optimizes HardEM objective otherwise. τ was chosen from {10K, 20K, 30K, 40K, 50K} based on model performance on the development set. HardEM-thres: We set the confidence threshold as γ = 0.5n where n was initialized as follows: we first computed the prediction probability of each solution with a task-specific model, and then set n to a value such that the model was trained on no less than half of training data at the first epoch. We halved γ after each epoch. VAE(Cheng and Lapata, 2018): A method that views a solution as the latent variable for question generation and adopts the training objective of Variational Auto-Encoder (VAE) to regularize the task-specific model. The overall training objective is given by: θ∗, φ∗= arg max θ,φ L(θ, φ) L(θ, φ) = Lmle(θ) + λLvae(θ, φ) = X z∈B log Pθ(z|d, q) + λEPθ(z|d,q) log Pφ(q|d, z) Pθ(z|d, q) where Lmle(θ) is the total log likelihood of the set of model-predicted solutions (denoted by B) with correct answer. Lvae(θ, φ) is the evidence lower bound of the log likelihood of questions. λ is the coefficient of Lvae(θ, φ). The optimization process is divided into three stages: (1) the 1st stage pre-trains a task-specific model θ with HardEMthres on solution sets4; (2) the 2nd stage pairs the task-specific model with our question reconstructor φ to optimize L(θ, φ) for one epoch, except that Lvae(θ, φ) is used to pre-train φ and is kept from back-propagating to θ; (3) the 3rd stage optimizes L(θ, φ) while allowing Lvae(θ, φ) to backpropagate to θ. The gradient of Lvae(θ, φ) w.r.t. θ is given by: ▽θLvae(θ, φ) = EPθ(z|d,q)R ▽θ log Pθ(z|d, q) R = log Pφ(q|d, z) Pθ(z|d, q) where R is the reward function. To stabilize training, we use the average reward of 5 sampled so4Cheng and Lapata (2018) pre-trained the task-specific model θ by maximizing Lmle(θ). We enhanced their method by pre-training θ with HardEM-thres. 4123 lutions as a baseline b and re-define the reward function as R ′ = R −b. λ is set to 0.1. In section 4.3, we report performance of the best model in the 3rd stage. At the 2nd stage, as the task-specific model optimized on both correct solutions and spurious solutions equally, the F1 score dropped from 72.35 to 67.93 at the end of this stage, indicating that correct training solutions is vital for generalization. At the 3rd stage, model learning was further regularized with Lvae(θ, φ) which was optimized via reinforcement learning. Despite our efforts to stabilize training, the F1 score still dropped to 36.28 after training for 1,000 steps at the 3rd stage. A.2 Experimental Settings For all experiments, we used previously proposed task-specific models and optimized them with their original optimizer. We chose the best task-specific model according to its performance on the development set. As for our learning method, we used BARTbase as the question reconstructor. AdamW optimizer (Loshchilov and Hutter, 2019) was used to update the question reconstructor with learning rate set to 5e-5. A.2.1 Multi-mention Reading Comprehension We adopted the reading comprehension model, data preprocessing, and training configurations from Min et al. (2019). Task-specific model: The model is based on uncased version of BERTbase, which takes as input the concatenation of a question and a paragraph, and outputs the probability distribution of the start and end position of the answer span. To deal with multiparagraph reading comprehension, it also trains a paragraph selector; during inference, it outputs a span from the paragraph ranked 1st. Data Preprocessing: Documents are split to segments up to 300 tokens. For Quasar-T, as retrieved sentences are short, we concatenated all sentences into one document in decreasing order of retrieval score (i.e., relevance with the question); for WebQuestions, we concatenated 5 retrieved paragraphs into one document, resulting in 10 reference documents per question. Training: Batch size is 20. BertAdam optimizer was used to update the reading comprehension model with learning rate set to 5e-5. The number of training epochs is 10. A.2.2 Discrete Reasoning over Paragraphs We used NeRd (Chen et al., 2020) for discrete reasoning. The major differences with its original implementation have been discussed in section 4.3. Task-specific Model: Chen et al. (2020) have designed a domain-specific language for discrete reasoning on DROP. The definitions of solutions for discrete reasoning introduced in section 4.3 are also expressed in this language except that we use different symbols (e.g., the minus sign “-” in our definitions has the same meaning as the symbol “DIFF” in their paper). NeRd is a Seq2Seq model which tasks as input the concatenation of a question and a paragraph, and generates the solution as a sequence. The answer is obtained by executing the solution. Data Preprocessing: The input of the task-specific model is truncated to up to 512 words. We used the solution sets provided by Chen et al. (2020), which cover 93.2% of examples in the train set. Training: Batch size is 32. Adam optimizer (Kingma and Ba, 2015) was used to update NeRd with learning rate set to 5e-5. The number of training epochs is 20. A.2.3 Semantic Parsing Following Min et al. (2019), we used SQLova (Hwang et al., 2019) on WikiSQL. Task-specific Model: SQLova encodes the concatenation of a question and a table header with uncased BERTbase, and outputs a SQL query via slot filling with an NL2SQL (natural language to SQL) layer. Data Preprocessing: Data preprocessing was kept the same as in (Min et al., 2019). We also used the solution sets provided by Min et al. (2019) which cover 98.8% of examples in the train set. Training: Following Min et al. (2019), we set the batch size to 10. Following Hwang et al. (2019), Adam optimizer was used to update SQLova with learning rate of BERTbase and NL2SQL layer set to 1e-5 and 1e-3, respectively. The number of training epochs is 15 and 20 when using the original train set and the hard train set of WikiSQL, respectively. A.3 Computing Infrastructure We conducted experiments on 24GB Quadro RTX 6000 GPUs. Most experiments used 1 GPU except that experiments on DROP used 4 GPUs in parallel. 4124 B Details of Ablation Study B.1 SQL Selection Task We defined a SQL selection task on the development set of WikiSQL. Specifically, for each question, we randomly sampled min(10, |Z|) solution candidates from the solution set Z without replacement while ensuring the ground-truth solution was one of the candidates. A model was required to pick out the ground-truth solution by selecting the candidate with the highest prediction probability. In section 5.3, we only show model accuracy in the first 10 training epochs because for BARTbase w/ HardEM, SQLova w/ HardEM, and SQLova w/ Ours, model confidence (computed as the average log likelihood of selected SQLs) showed a downward trend after the 2nd, 4th, and ≥10th epoch, respectively. B.2 Choice of Question Reconstructor We investigated how the choice of the question reconstructor affects results. One alternative choice is a Transformer pre-trained as a denoising autoencoder of questions on the train set. This question reconstructor is the same as BARTbase except that the number of encoder layers and the number of decoder layers are 3 respectively. We pre-trained the question reconstructor for one epoch to reconstruct original questions from corrupted ones. For 50% of the time, the input question is the original question; otherwise, we followed Lewis et al. (2020) to corrupt the original question by randomly masking a number of text spans with span lengths drawn from a Poisson distribution (λ = 3). Batch size is 4. AdamW optimizer was used with learning rate set to 5e-5.
|
2021
|
318
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4125–4140 August 1–6, 2021. ©2021 Association for Computational Linguistics 4125 Breaking Down Walls of Text: How Can NLP Benefit Consumer Privacy? Abhilasha Ravichander♦Alan W Black♦Thomas Norton♠ Shomir Wilson♥Norman Sadeh♦ ♦Carnegie Mellon University, Pittsburgh, PA ♠Fordham Law School, New York, NY ♥Penn State University, University Park, PA {aravicha, awb, sadeh}@cs.cmu.edu {shomir}@psu.edu, {tnorton1}@law.fordham.edu Abstract Privacy plays a crucial role in preserving democratic ideals and personal autonomy. The dominant legal approach to privacy in many jurisdictions is the “Notice and Choice” paradigm, where privacy policies are the primary instrument used to convey information to users. However, privacy policies are long and complex documents that are difficult for users to read and comprehend. We discuss how language technologies can play an important role in addressing this information gap, reporting on initial progress towards helping three specific categories of stakeholders take advantage of digital privacy policies: consumers, enterprises, and regulators. Our goal is to provide a roadmap for the development and use of language technologies to empower users to reclaim control over their privacy, limit privacy harms, and rally research efforts from the community towards addressing an issue with large social impact. We highlight many remaining opportunities to develop language technologies that are more precise or nuanced in the way in which they use the text of privacy policies. 1 Introduction Privacy is a fundamental right central to a democratic society, in which individuals can operate as autonomous beings free from undue interference from other individuals or entities (Assembly, 1948). However, certain functions of privacy, such as the power to grant or deny access to one’s personal information, are eroded by modern commercial and business practices that involve vast collection, linking, sharing, and processing of digital personal information through an opaque network, often without data subjects’ knowledge or consent. In many jurisdictions, online privacy is largely governed by “Notice and Choice” (Federal Trade Commission, 1998). Under this framework, data-collecting and data-processing entities publish privacy policies that disclose their data practices. Theoretically, users are free to make choices about which services and products they use based on the disclosures made in these policies. Thus, the legitimacy of this framework hinges on users reading a large number of privacy policies to understand what data can be collected and how that data can be processed before making informed privacy decisions. In practice, people seldom read privacy policies, as this would require prohibitive amounts of their time (McDonald and Cranor, 2008; Cate, 2010; Cranor, 2012; Reidenberg et al., 2015; Schaub et al., 2015; Jain et al., 2016). Thus, an opportunity exists for language technologies to bridge this gap by processing privacy policies to meet the needs of Internet and mobile users. NLP has made inroads in digesting large amounts of text in domains such as scientific publications and news (Jain et al., 2020; Cachola et al., 2020; Kang et al., 2018; Rush et al., 2015; See et al., 2017), with several practical tools based on these technologies helping users every day (Cachola et al., 2020; TLDR, 2021; News, 2021). These domains have also received considerable research attention: several benchmark datasets and technologies are based in texts from these domains (Nallapati et al., 2016; See et al., 2017; Narayan et al., 2018; Beltagy et al., 2019). We highlight that the privacy domain can also benefit from increased research attention from the community. Moreover, technologies developed in the privacy domain have potential for significant and large-scale positive social impact—the affected population includes virtually every Internet or mobile user (Sadeh et al., 2013). Automated processing of privacy policies opens the door to a number of scenarios where language technologies can be developed to support users in the context of different tasks. This includes saving data subjects the trouble of having to read the 4126 entire text of policies when they are typically only concerned about one or a small number of issues (e.g., determining whether they can opt out of some practices or whether some of their data might be shared with third parties). It includes helping companies ensure that they are compliant and that their privacy policies are consistent with what their code actually does. It also includes supporting regulators, as they face the daunting task of enforcing compliance across an ever-growing collection of software products and processes, including sophisticated data collection and use practices. In this work, we conduct an extensive survey of initial progress in applying NLP to address limitations of the Notice and Choice model. We expect our work to serve as a useful starting point for practitioners to familiarize themselves with technological progress in this domain, by providing both an introduction to the basic privacy concerns and frameworks surrounding privacy policies, as well as an account of applications for which language technologies have been developed. Finally, we highlight many remaining opportunities for NLP technologies to extract more precise, more nuanced, and ultimately more useful information from privacy policy text— describing key challenges in this area and laying out a vision for the future. 2 Privacy as a Social Good In 1890, Warren and Brandeis defined the right to privacy as “the right to be let alone”(Warren and Brandeis, 1890). More recently, Westin defined the right as “the claim of individuals, groups, or institutions to determine for themselves when, how, and to what extent information about them is communicated to others” (Westin, 1968). A primary aspiration of privacy is to allow for the separation of individual and society as a means of fostering personal autonomy. To that end, privacy “protects the situated practices of boundary management through which the capacity for self-determination develops,” and further “shelters dynamic, emergent subjectivity from the efforts of commercial and government actors to render individuals and communities fixed, transparent, and predictable” (Cohen, 2012). Privacy, therefore, is “foundational to the practice of informed and reflective citizenship,” and serves as “an indispensable structural feature of liberal democratic political systems” (Cohen, 2012). When privacy is threatened, we risk losing the chance for critical self-reflection of political processes and social norms. Indeed, privacy undergirds the concepts of human dignity and other key values, such as the freedoms of association and speech. For these reasons and others, privacy is regarded as a fundamental human right (Assembly, 1948). In the digital age, privacy is threatened by aggressive, rapid, and largely automated collection, linking, sharing, and processing of digital personal information. Digital privacy is intrinsically linked to the fundamental ethical principles of transparency, fairness and agency. • Transparency: Users have a right to know how information about them is collected and used. Entities collecting user data stay clear of manipulative schemes designed to influence the data subject’s willingness to disclose their data (e.g. overemphasizing benefits while remaining silent about potential risks associated with the disclosure of data in a given context). • Fairness: Users should receive perceived value commensurate to the perceived loss of privacy associated with disclosure and use of their data. • Agency: Users should have a choice about what data is collected about them and how it is used. The dominant paradigm to address these principles in the United States and most legal jurisdictions around the world, is the ’Notice and Choice’ regulatory framework (Westin, 1968; Federal Trade Commission, 1998). ’Notice and Choice’ regimes are based on the presupposition that consumers will adequately manage their privacy, if provided sufficient information about how their data will be collected, used and managed, as well as offered meaningful choices. Today, ’Notice’ is often practically realized through publishing privacy policies, which are long and verbose documents that users are expected to read and understand. ‘Choice’ is often limited to the user clicking ‘I agree’ to the privacy policy, or even interpreting their continued use of the service as some sort of meaningful consent to the terms of the policy. The ’Notice and Choice’ framework is fundamentally broken. In practice, users seldom read privacy policies (McDonald and Cranor, 2008; Cate, 2010; US Federal Trade Commission et al., 2012) and it is prohibitively expensive for them to even do so. McDonald and Cranor (2008) estimate that if internet users were to actually read the privacy policies of the websites they visited, they would have to spend roughly 250 hours each year just reading 4127 Challenge Example Ambiguity We may also use aggregate personal information for regulatory compliance, industry and market analysis, research, demographic profiling, marketing and advertising, and other business purposes. Vagueness [X] collects, or may have a third-party service providers collect, non-personally-identifying information of the sort that mobile applications typically make available, such as the type of device using the Application, the operating system, location information, and aggregated user statistics. Modality If you use our services to make and receive calls or send and receive messages, we may collect call and message log information like your phone number, calling-party number, receiving-party number... Negation No apps have access to contact information, nor do they read or store any contact information Lists and Document Structure We may collect data or ask you to provide certain data when you visit and use our websites, products and services. The sources from which we collect Personal Data include: • Data collected directly from you or your device .... ; • If we link other data relating to you with your Personal Data, we will treat that linked data as Personal Data; and • We may also collect Personal Data from trusted third-party sources.... Tabular Understanding Reasons we Can Share Your Personal Information Does X share? Can you limit this sharing? For our everyday business purposes ... Yes No For our everyday marketing purposes ... Yes No For joint marketing with other companies No We don’t share Table 1: Examples of some challenging aspects for language understanding in privacy policies, including reasoning over ambiguity and vagueness, modality, negation (including scope),lists and document structure, and tables. privacy policies. A 2014 report from the Presidents Council of Advisors on Science and Technology stated that “only in some fantasy world” were users reading and understanding privacy policies before giving their consent (of the President’s Council of Advisors on Science and Technology, 2014). Indeed, 91% of people in the U.S have reported feeling like they have lost control over their information (Madden et al., 2014). Moreover, recent privacy laws such as the EU’s General Data Protection Regulation (GDPR) (Regulation, 2016) still fail to address the critical limitation of notice and choice: the continued reliance on users to read and understand a large number of privacy policies. Studies have shown that GDPR requirements have actually resulted in longer privacy policies (Linden et al., 2020), and users still encounter unreadable privacy policies (Becher and Benoliel, 2019). The lack of respect for individuals’ rights to privacy also has implications for society. With social platforms in particular having access to an unprecedented scale of information about human behaviour, Vicario et al. (2019) discuss that users’ polarization and confirmation bias can play a role in spreading misinformation on social platforms. Madden et al. (2017) report that particular groups of lessprivileged users on the internet are uniquely vulnerable to various forms of surveillance and privacy harms, which could widen existing economic gaps. Introna (1997) describe privacy as central to human autonomy in social relationships. In this work, we examine the potential of language technologies in enabling people to derive the benefits of their rights to transparency, fairness and agency. 3 Can NLP Help Privacy? Privacy policies present interesting challenges for NLP practitioners, as they often feature characteristic aspects of language that remain under-examined or difficult to process (Table. 1). For example, while many policies discuss similar issues surrounding how user data is collected, managed and stored, policy silence about certain data practices may carry great weight from a legal, policy, and regulatory perspective.1 In the privacy policy domain, understanding what has not been said in a privacy policy (policy silence) is just as important as understanding what is said (Zimmeck et al., 2019a; Marotta-Wurgler, 2019). Further, though policies tend to feature literal language (compared to more subjective domains like literature or blog posts), processing them ef1For example, in United States v. Path, the defendant’s (Path) privacy policy described that its app collected ”certain information such as your Internet Protocol (IP) address, your operating system, the browser type.” The Federal Trade Commission found this disclosure to be incomplete and insufficient to provide notice about the collection of users’ contact data (FTC, 2013). 4128 Task Goal Consumer Regulator Enterprise Data Practice Identification (Wilson et al., 2016b) Annotate segments of privacy policies with described data practices. Opt-Out Identification (Sathyendra et al., 2017; Bannihatti Kumar et al., 2020) Extract opt-out choices buried in privacy policy text. Compliance Analysis (Zimmeck et al., 2017, 2019a) Analyze mobile app code and privacy policy to identify potential compliance issues. Privacy Question-Answering (Ravichander et al., 2019; Ahmad et al., 2020) Allow consumers to selectively query privacy policies for issues that are important to them. Policy Summarization (Zaeem et al., 2018; Keymanesh et al., 2020) Construct summaries to aid consumers to quickly digest the content of privacy policies. Readability Analysis (Massey et al., 2013; Meiselwitz, 2013) Characterize the ease of understanding or comprehension of privacy policies. Table 2: Overview of some applications of NLP to privacy policies, and primary stakeholders they are intended to benefit. fectively also requires several additional capabilities such as reasoning over vagueness and ambiguity, understanding elements such as lists (including when they are intended to be exhaustive and when they are not (Bhatia et al., 2016)), effectively incorporating ‘co-text’- aspects of web document structure such as document headers that are meaningful semantically to the content of privacy policies(Mysore Gopinath et al., 2018) and incorporating domain knowledge (for example, understanding whether information is sensitive requires background knowledge in the form of applicable regulation). Privacy policies also differ from several closely related domains, such as legal texts which are largely meant to be processed by domain experts. In contrast, privacy policies are legal documents with legal effects—generally drafted by experts—that are ostensibly meant to be understood by everyday users. NLP applications in the privacy domain also need to be designed with end user requirements in mind. For example, from a legal standpoint, when generating answers to a user’s question about the content of a privacy policy, it is generally advisable to include disclaimers, but users may prefer to be presented with shorter answers, where disclaimers are kept as short as possible. Challenges are described in more detail in (§4). We survey current efforts to apply NLP in the privacy domain, discussing both existing task formulations as well as future areas in this domain where language technologies can have impact. 2 2Our survey includes relevant papers from major NLP venues, including ACL, EMNLP, NAACL, EACL, COLING, CoNLL, SemEval, TACL, and CL. We supplemented these publications with a review of the literature at venues such as SOUPS, PETS, WWW, ACM, and NDSS. We also included relevant legal venues, such as law reviews and journals. 3.1 Data Practice Identification Initial efforts in applying NLP in the privacy domain have largely focused on discovering or identifying data practice categories in privacy policies (Costante et al., 2012a; Ammar et al., 2012; Costante et al., 2012b; Liu et al., 2014b; Ramanath et al., 2014a; Wilson et al., 2016b). Automating the identification of such data practices could potentially support users in navigating privacy policies more effectively3, as well as automate analysis for regulators who currently do not have techniques to assess a large number of privacy policies. Wilson et al. (2016b) create a corpus of 115 website privacy policies annotated with detailed information of the privacy policies described. The corpus and associated taxonomy have been of utility in the development of several subsequent privacy-enhancing language technologies (Mysore Sathyendra et al., 2017a; Zimmeck et al., 2017; Ravichander et al., 2019; Ahmad et al., 2020). 3.2 Choice Identification Studies have shown that consumers desire control over the use of their information for marketing communication, and object to the use of their information for web tracking or marketing purposes including targeted advertising (Cranor et al., 2000; Turow et al., 2009; Ur et al., 2012; Bleier and Eisenbeiss, 2015). However, McDonald and Cranor (2010) find that many people are unaware of the opt-out choices available to them. These choices are often buried in policy text, and thus there has been interest in applying NLP to extract choice language. Mysore Sathyendra et al. (2017b) automatically identify choice instances within a privacy 3For example, through the data exploration tool developed by the Usable Privacy Policy Project: https://explore. usableprivacy.org/?view=machine 4129 Figure 1: The results from Opt-Out Easy, a browser extension to extract opt-out choices from privacy policies, for Overleaf.com (Bannihatti Kumar et al., 2020). policy, labeling different types of opt-out choices, with a particular emphasis on extracting actionable choices in the policy, i.e. those associated with hyperlinks. Bannihatti Kumar et al. (2020) develop a web-browser extension to present extracted choice instances to users (Figure. 1), finding that the tool can considerably increase awareness of choices available to users and reduce the time taken to identify actions the users can take. 3.3 Compliance Analysis In 2012, six major mobile app stores entered into an agreement with the California Attorney General, where they agreed to adopt privacy principles that require mobile apps to have privacy policies(Justice, 2012). Regulations such as the the EU General Data Protection Directive (GDPR) and the California Consumer Protection Act (CCPA) impose further requirements on what entities collecting and using personal data need to disclose in their privacy policies and what rights they need to offer to their users (e.g. privacy controls, option to request deletion of one’s data). However, regulators lack the necessary resources to systematically check that these requirements are satisfied. In fact, even app stores lack the resources to systematically check that disclosures made in privacy policies are consistent with the code of apps and comply with relevant regulatory requirements. Thus, there has been interest in developing technologies to automatically identify potential compliance issues (Enck et al., 2014; Zimmeck et al., 2017; Wang et al., 2018; Libert, 2018a; Zimmeck et al., 2019b). A first application of language technologies to aid compliance analysis is detailed by Zimmeck et al. (2017), including results of a systematic analysis of 17,991 apps using both natural language processing and code analysis techniques. Classifiers are trained to identify data practices based on the OPP-115 ontology (Wilson et al., 2016b), and static code analysis techniques are employed to extract app’s privacy behaviors. The results from the two procedures are compared to identify potential compliance issues. The system was piloted with personnel at the California Office of the Attorney General. Users reported that the system could significantly increase productivity, and decrease the effort and time required to analyze practices in apps and audit compliance. Zimmeck et al. (2019b) review 1,035,853 apps from the Google Play Store for compliance issues. Their system identifies disclosed privacy practices in policies using classifiers trained on the APP-350 corpus (Story et al., 2019), and static code analysis techniques to identify apps’ privacy behaviors. Results of the analysis of this large corpus of privacy policies revealed a particularly large number of potential compliance problems, with a subset of results shared with the Federal Trade Commission. The system was also reported to have been used by a large electronics manufacturer to verify compliance of legacy mobile apps prior to the introduction of GDPR. 3.4 Policy Summarization Due to the lengthy and verbose nature of privacy policies, it is appealing to attempt to develop automated text summarization techniques to generate short and concise summaries of a privacy policy’s contents (Liu et al., 2015). Tomuro et al. (2016) develop an extractive summarization system that identifies important sentences in a privacy policy along five categories: purpose, third parties, limited collection, limited use and data retention. Zaeem et al. (2018, 2020) identify ten questions about privacy policies, and automatically categorize ‘risk levels’ associated with each of the questions, as shown in Table. 3. Keymanesh et al. (2020) focus on extractive summarization approaches to identify ‘risky sections’ of the privacy policy, which are sentences that are likely to describe a privacy risk posed to the end-user. However, while automated summarization seems like a promising application of language technologies, identifying which parts of a policy should be shown to users is exceedingly difficult, and studies by privacy experts have shown 4130 # Question Green Risk Level Yellow Risk Level Red Risk Level (1) How well does this website protect your email address? Not asked for Used for intended service Shared w/ third parties (2) How well does this website protect your credit card information and address? Not asked for Used for intended service Shared w/ third parties (3) How well does this website handle your social security number? Not asked for Used for intended service Shared w/ third parties (4) Does this website use or share your PII for marketing purposes? PII not used for marketing PII used for marketing PII shared for marketing (5) Does this website track or share your location? Not tracked Used for intended service Shared w/ third parties (6) Does this website collect PII from children under 13? Not collected Not mentioned Collected (7) Does this website share your information with law enforcement? PII not recorded Legal docs required Legal docs not required (8) Does this website notify or allow you to opt-out after changing their privacy policy? Posted w/ opt-out option Posted w/o opt-out option Not posted (9) Does this website allow you to edit or delete your information from its records? Edit/delete Edit only No edit/delete (10) Does this website collect or share aggregated data related to your identity or behavior? Not aggregated Aggregated w/o PII Aggregated w/ PII Table 3: Ten privacy questions used for summarization, and associated ‘risk levels’ from (Zaeem et al., 2018). that such ‘one-size-fits-all’ approaches are unlikely to be effective (Gluck et al., 2016; Rao et al., 2016). 3.5 Privacy Question-Answering A desire to move away from ‘one-size-fits-all’ approaches has led to increased interest in supporting automated privacy question-answering (QA) capabilities. If realized, such functionality will help users selectively and iteratively explore issues that matter most to them. Table 4 lists current efforts to develop resources for privacy question-answering. Amongst the initial explorations in this area, Harkous et al. (2018) examine privacy questions asked by Twitter users to companies, with answers annotated by the paper’s authors. Ravichander et al. (2019) collect questions asked by crowdworkers about a mobile app without seeing the app’s privacy policy, and hire legal experts to identify sentences in the privacy policy relevant for each question. (Ahmad et al., 2020) provide ‘skilled annotators’ with privacy policy segments drawn from the OPP-115 corpus (Wilson et al., 2016b), and ask them to construct questions based on the provided span of text. Ravichander et al. (2019) and Ahmad et al. (2020) both find that current QA baselines based on pretrained language models(Devlin et al., 2019) are inadequate for answering privacy questions. Ahmad et al. (2020) indicate that identifying longer evidence spans are challenging and describe transfer learning as a potential direction to improve performance. Ravichander et al. (2019) examine unanswerability as a challenge to privacy QA systems, highlighting the many facets of unanswerable questions that can be asked. It is worth noting that all three resources formulate ground truth based in the text of the privacy policy, but policy language is difficult for non-experts to understand (Reidenberg et al., 2015). Future QA dataset architects could consider abstractive answers as ground truths, which are validated by legal experts for correctness and evaluated by users for helpfulness. It may also be desirable for benchmarks to aim for ecological validity (de Vries et al., 2020), with users asking questions, and legal experts constructing answers. 3.6 Other Applications In this section, we survey further tasks where NLP has been applied to consumer privacy, including analyzing privacy policy readability, with the goal of aiding writers of privacy policies (Fabian et al., 2017; Massey et al., 2013; Meiselwitz, 2013; Ermakova et al., 2015), and understanding data practice categories are described in a policy, known as measuring policy coverage (Linden et al., 2020; Shvartzshnaider et al., 2020). A significant amount of recent work has also focused on information extraction from privacy policies (Costante et al., 2012a). Shvartzshanider et al. (2018); Shvartzshnaider et al. (2019, 2020) identify contextual integrity parameters (Nissenbaum, 2004) in policy text. Studies have also tried to extract other, more specific kinds of information from policies, such as third party entities (Libert, 2018b; Bokaie Hosseini et al., 2020) and information about regulated information types (Bhatia et al., 2016; Evans et al., 2017) as well as their similarity (Hosseini et al., 2016). There have also been efforts to analyze vague statements in privacy policies (Liu et al., 2016b; Lebanoff and Liu, 2018), and explore how benchmarks in this domain can be constructed through crowdsourcing (Ramanath et al., 2014b; Wilson et al., 2016c; Audich et al., 2018). Lastly, there has been research focused on identifying header information in privacy policies (Mysore Gopinath et al., 2018) and generating them (Gopinath et al., 2020). Techniques to 4131 Dataset #Questions Question Scenario Legal Expert Annotator Asker Cannot See Evidence Unanswerable Questions Non-Contiguous Answer Polisis (Harkous et al., 2018) 120 Twitter users ask questions to a company. PrivacyQA (Ravichander et al., 2019) 1750 Crowdworkers ask questions about a mobile app. PolicyQA (Ahmad et al., 2020) 714 Skilled annotators are shown a text span and data practice, and asked to construct a question. Table 4: Comparison of Polisis (Harkous et al., 2018), PrivacyQA (Ravichander et al., 2019) and PolicyQA (Ahmad et al., 2020) QA datasets. Question Scenario describes conditions under which the questions were generated. ‘Asker Cannot See Evidence’ indicates the asker of the question was not shown evidence from the document when formulating questions. Unanswerable questions indicates if the corpus includes unanswerable questions. ‘Non Contriguous Answer’ indicates the answers are allowed to be from non-adjacent segments of the privacy policy. process privacy policies have largely followed successful approaches elsewhere in NLP, starting from feature-based approaches (Sathyendra et al., 2017; Zimmeck et al., 2019a), training domain-specific word embeddings (Kumar et al., 2019) and finetuning pretrained language models on privacy policies (Nejad et al., 2020; Mustapha et al., 2020). 3.7 Towards New Tasks and Formulations We discuss a vision of future applications of NLP in aiding consumer privacy. We believe these applications present interesting opportunities for the community to develop technologies, both because of the technical challenges they offer and the impact they are likely to have. Detecting surprising statements: Since users do not read privacy policies, their expectations for the data practices of services might not align with services’ actual practices. These mismatches may result in unexpected privacy risks which lead to loss of user trust (Rao et al., 2016). Identifying such ‘surprising’ statements will require understanding social context and domain knowledge of privacy information types. For example, it is natural for a banking website to collect payment information, but not health information. Moreover, understanding what statements will be surprising for each individual user requires understanding their personal, social and cultural backrounds (Rao et al., 2016). We speculate that NLP can potentially be leveraged to increase transparency by identifying discordant statements within privacy policies. Detecting missing information: In contrast to detecting surprising statements, privacy policies may be underspecified. Story et al. (2018) find that many policies contain language appearing in unrelated privacy policies, indicating that policy writers may use privacy policy generators not suited to their application, potentially resulting in missing information. Techniques from compliance analysis could help in flagging some of these issues (Zimmeck et al., 2017, 2019a). Generating privacy nutrition labels: One proposal to overcome the gap in communicating privacy information to users has been the privacy ‘nutrition label’ approach (Kelley et al., 2009, 2013), as shown in Fig. 2. The proposal draws from industries such as nutrition, warning and energy labeling where information has to be communicated to consumers in a standardized way. Recently, Apple announced that developers will be required to provide information for these labels (Campbell, 2020), which disclose to the user the information a company and third parties collect.4 This approach could potentially be helpful to users to understand privacy information at a glance, but presents challenges to both developers and app platforms. Developers need to ensure their nutrition label is accurate and platforms need to enforce compliance to these requirements. Potentially, early successes of language technologies in compliance systems can be extended to analyzing a specified nutrition label, policy and application code. NLP may also be used to generate nutrition labels which developers inspect, as opposed to the more costly process of developers specifying nutrition labels from scratch which may hinder adoption (Fowler, 2021). Personalized privacy summaries: One approach to mitigating inadequacies of policy summarization—where generic summaries may not be sufficiently complete —is personalized summarization (D´ıaz and Gerv´as, 2007; Hu et al., 4An example of such a nutrition label can be found in Appendix. A 4132 2012). In this formulation, policies are summarized for each user based on issues that matter most to them. This formulation may alleviate some downsides of QA approaches, which require the user know how to manage their privacy by asking the right questions. Personalized summarization systems would benefit from modeling users’ level of knowledge, as well as their beliefs, desires and goals. In NLP, there has been effort towards addressing similar challenges for personalized learning in intelligent tutoring (McLaren et al., 2006; Malpani et al., 2011). Assistive Policy Writing: We speculate advances in natural language generation and compliance analysis techniques may jointly be leveraged to help app developers create more accurate privacy policies, rather than relying on policy generators (Story et al., 2018). Privacy policies generally cover a known set of data practices (Wilson et al., 2016a), providing potential statistical commonalities to aid natural language generation. Code analysis can be leveraged to constrain generation to accurately describe data practices of a service. 4 Progress and Challenges Although privacy policies have legal effects for most Internet users, these types of texts constitute an underserved domain in NLP. NLP has the potential to play a role in easing user burden in understanding salient aspects of privacy policies, help regulators enforce compliance and help developers enhance the quality of privacy policies by reducing the effort required to construct them. Yet, the privacy domain presents several challenges that require specialized resources to deal with effectively. We describe some of these distinctive challenges, as well as the capabilities that will need to be developed to process policies satisfactorily. • Disagreeable privacy policies: Privacy policies are complex, but are the most important source of information about how user data is collected, managed and used. Reidenberg et al. (2015) find that sometimes discrepancies can arise in the interpretation of policy language, even between experts. This additional complexity should be taken into consideration by those developing language technologies in this domain. • Difficulty or validity of collecting annotations: Privacy policies are legal documents that have legal effects on how user data is collected and used. While crowdworkers have been found to provide non-trivial annotations for some tasks in this domain (Wilson et al., 2016c), individual practitioners constructing applications must carefully consider the consequences of sourcing non-expert annotations in the context of their task and the impacted stakeholders, and not rely on crowdsourced annotation simply because it is cheaper or easier to scale. • Difficult for users to articulate their needs and questions: Developing effective privacy QA functionality will require understanding the kinds of questions users ask and quantifying to what extent privacy literacy affects users’ ability to ask the right questions. Ravichander et al. (2019) find many questions collected from crowdworkers were either incomprehensible, irrelevant or atypical. Understanding these factors could lead to the development of more proactive QA functionality- for example, rather than wait for users to form questions, the QA system could prompt users to reflect on certain privacy issues. • Challenges to QA: Additionally, privacy question-answering systems themselves will require several capabilities in order to have larger impact. These systems must be capable of doing question-answering iteratively, working with the user towards resolving information-seeking needs. They will also need to consider unanswerability(Rajpurkar et al., 2018; Ravichander et al., 2019; Asai and Choi, 2020) as a graded problem, recognizing to what extent the privacy policy contains an answer and communicating both what is known and what is not known to the user. QA systems must also consider what kinds of answers are useful, identifying appropriate response format and tailoring answers to the user’s level of knowledge and individual preferences. • Domain Knowledge: It remains an open question how to best incorporate expert knowledge into the processing of privacy policies. Although privacy policies are intended to be read by everyday users, experts and users often disagree on their interpretations (Reidenberg et al., 2015). • Combining Disparate Sources of Information: While privacy policies are the single most important source of information about collection and sharing practices surrounding user data, technologies to address users’ personalized concerns could leverage additional sources of informationsuch as analyzing the code of a given technology 4133 such as a mobile app, news articles, or background knowledge of a legal, technical or statistical nature. For example, when the policy is silent on an issue- a QA system could report the practices of other similiar services to the user, or if a user asks about the likelihood of a data breach, the QA system could refer to news sources for information about the service. • User Modeling: Personalized privacy approaches will also need to model individual user’s personal, social and cultural contexts to deliver impact. This could include information about the issues likely to matter most to users, their background knowledge, privacy preferences and expectations (Liu et al., 2014a; Lin et al., 2014; Liu et al., 2016a). • Accessibility: Efforts to help users understand privacy policies by breaking through walls of text to identify salient aspects, are expected to help users with a range of visual impairments navigate their privacy. Future work would conduct user studies to determine the extent to which developed technologies ease visually-impaired users’ accessibility to learn about the content of policies, related to their interests or concerns. 5 Ethical Considerations While NLP has the potential to benefit consumer privacy, we emphasize there are also ethical considerations to be taken in account. These include: Bias of agent providing technology: A factor that must be considered in the practical deployment of NLP systems in this domain is the incentives of the entity creating or providing the technology. For example, the incentives of a company that develops a QA system to answer questions about its own privacy policy may not align with those of a trusted third-party privacy assistant that reviews the privacy policies of many different companies. This information also needs to be communicated in an accurate and unbiased fashion to users. User Trust: While NLP systems have the potential to digest policy text and present information to users, NLP systems are seldom completely accurate, and therefore it is important that users be appropriately informed of these limitations. For example, if a QA system communicates a data practice incorrectly in response to a users’ question and the user encounters privacy harms contrary to their expectations as a result, they may lose trust in the system. It is important to also identify appropriate disclaimers to accompany NLP systems to manage user expectations. Discriminatory Outcomes: It is possible that different populations will benefit to different extents from the developed technologies, and we are yet unable to anticipate precisely where the benefits will accrue. For example, users with higher degrees of privacy literacy may be able to take better advantage of a developed QA system. Technological Solutionism: It is important to consider that while language technologies have the potential to considerably alleviate user burden in reading privacy policies, they are unlikely to completely resolve the issue that users are unable to read and review a multitude of privacy policies everyday. Advances toward addressing the limitations of notice and choice will also require progress in regulation and enforcement by regulatory bodies to ensure that enterprises are more accurate in their disclosures and use clearer language, in tandem with creative technological solutions. 6 Conclusion Privacy is about the right of people to control the collection and use of their data. Today privacy relies on the ’Notice and Choice’ framework, which assumes that people actually read the text of privacy policies. This is a fantasy as users do not have the time to do so. In this article, we summarize how language technologies can help overcome this challenge and support the development of solutions that assist customers, technology providers and regulators. We reviewed early successes and presented a vision of how NLP could further help in the future. We hope this article will motivate NLP researchers to contribute to this vision and empower people to regain control over their privacy. Acknowledgements This research was supported in part by grants from the National Science Foundation Secure and Trustworthy Computing program (CNS-1330596, CNS1330214, CNS-15-13957, CNS-1801316, CNS1914486, CNS-1914444) and DARPA(FA8750-152-0277). Part of the work summarized in this paper was conducted by the Usable Privacy Policy Project(https://usableprivacy.org). The authors would like to thank Siddhant Arora, Rex Chen and Aakanksha Naik for valuable discussion. 4134 References Wasi Ahmad, Jianfeng Chi, Yuan Tian, and Kai-Wei Chang. 2020. PolicyQA: A reading comprehension dataset for privacy policies. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 743–749, Online. Association for Computational Linguistics. Waleed Ammar, Shomir Wilson, Norman Sadeh, and Noah A Smith. 2012. Automatic categorization of privacy policies: A pilot study. Technical Report CMU-LTI-12-019, Carnegie Mellon University. Akari Asai and Eunsol Choi. 2020. Challenges in information seeking qa: Unanswerable questions and paragraph retrieval. arXiv preprint arXiv:2010.11915. UN General Assembly. 1948. Universal declaration of human rights. UN General Assembly, 302(2). Dhiren A Audich, Rozita Dara, and Blair Nonnecke. 2018. Privacy policy annotation for semi-automated analysis: A cost-effective approach. In IFIP International Conference on Trust Management, pages 29– 44. Springer. Vinayshekhar Bannihatti Kumar, Roger Iyengar, Namita Nisal, Yuanyuan Feng, Hana Habib, Peter Story, Sushain Cherivirala, Margaret Hagan, Lorrie Cranor, Shomir Wilson, et al. 2020. Finding a choice in a haystack: Automatic extraction of optout statements from privacy policy text. In Proceedings of The Web Conference 2020, pages 1943– 1954. Shmuel I Becher and Uri Benoliel. 2019. Law in books and law in action: the readability of privacy policies and the gdpr. In Consumer Law and Economics, pages 179–204. Springer. Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3615– 3620, Hong Kong, China. Association for Computational Linguistics. Jaspreet Bhatia, Morgan C Evans, Sudarshan Wadkar, and Travis D Breaux. 2016. Automated extraction of regulated information types using hyponymy relations. In 2016 IEEE 24th International Requirements Engineering Conference Workshops (REW), pages 19–25. IEEE. Alexander Bleier and Maik Eisenbeiss. 2015. The importance of trust for personalized online advertising. Journal of Retailing, 91(3):390–409. Mitra Bokaie Hosseini, Pragyan K C, Irwin Reyes, and Serge Egelman. 2020. Identifying and classifying third-party entities in natural language privacy policies. In Proceedings of the Second Workshop on Privacy in NLP, pages 18–27, Online. Association for Computational Linguistics. Isabel Cachola, Kyle Lo, Arman Cohan, and Daniel Weld. 2020. TLDR: Extreme summarization of scientific documents. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4766–4777, Online. Association for Computational Linguistics. Ian Carlos Campbell. 2020. Apple will require apps to add privacy ‘nutrition labels’ starting december 8th. Fred H Cate. 2010. The limits of notice and choice. IEEE Security & Privacy, 8(2). Julie E Cohen. 2012. What privacy is for. Harv. L. Rev., 126:1904. Elisa Costante, Jerry den Hartog, and Milan Petkovi´c. 2012a. What websites know about you. In Data Privacy Management and Autonomous Spontaneous Security, pages 146–159. Springer. Elisa Costante, Yuanhao Sun, Milan Petkovi´c, and Jerry den Hartog. 2012b. A machine learning solution to assess privacy policy completeness: (short paper). In Proceedings of the 2012 ACM Workshop on Privacy in the Electronic Society, WPES ’12, page 91–96, New York, NY, USA. Association for Computing Machinery. Lorrie Faith Cranor. 2012. Necessary but not sufficient: Standardized mechanisms for privacy notice and choice. J. on Telecomm. & High Tech. L., 10:273. Lorrie Faith Cranor, Joseph Reagle, and Mark S Ackerman. 2000. Beyond concern: Understanding net users’ attitudes about online privacy. The Internet upheaval: raising questions, seeking answers in communications policy, pages 47–70. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Alberto D´ıaz and Pablo Gerv´as. 2007. User-model based personalized summarization. Information Processing and Management: an International Journal, 43(6):1715–1734. William Enck, Peter Gilbert, Seungyeop Han, Vasant Tendulkar, Byung-Gon Chun, Landon P Cox, Jaeyeon Jung, Patrick McDaniel, and Anmol N Sheth. 2014. Taintdroid: an information-flow tracking system for realtime privacy monitoring on smartphones. ACM Transactions on Computer Systems (TOCS), 32(2):1–29. 4135 Tatiana Ermakova, Benjamin Fabian, and E. Babina. 2015. Readability of privacy policies of healthcare websites. In 12. Internationale Tagung Wirtschaftsinformatik. Morgan C Evans, Jaspreet Bhatia, Sudarshan Wadkar, and Travis D Breaux. 2017. An evaluation of constituency-based hyponymy extraction from privacy policies. In 2017 IEEE 25th International Requirements Engineering Conference (RE), pages 312–321. IEEE. Benjamin Fabian, Tatiana Ermakova, and Tino Lentz. 2017. Large-scale readability analysis of privacy policies. In Proceedings of the International Conference on Web Intelligence, WI ’17, page 18–25, New York, NY, USA. Association for Computing Machinery. Federal Trade Commission. 1998. Privacy online: A report to congress. Washington, DC, June, pages 10– 11. Geoffrey Fowler. 2021. I checked apple’s new privacy ‘nutrition labels.’ many were false. FTC. 2013. Path social networking app settles ftc charges it deceived consumers and improperly collected personal information from users’ mobile address books. https://www.ftc.gov/newsevents/press-releases/2013/02/path-socialnetworking-app-settles-ftc-charges-it-deceived. Joshua Gluck, Florian Schaub, Amy Friedman, Hana Habib, Norman Sadeh, Lorrie Faith Cranor, and Yuvraj Agarwal. 2016. How short is too short? implications of length and framing on the effectiveness of privacy notices. In 12th Symposium on Usable Privacy and Security (SOUPS), pages 321–340. Abhijith Athreya Mysore Gopinath, Vinayshekhar Bannihatti Kumar, Shomir Wilson, and Norman Sadeh. 2020. Automatic section title generation to improve the readability of privacy policies. Hamza Harkous, Kassem Fawaz, R´emi Lebret, Florian Schaub, Kang G Shin, and Karl Aberer. 2018. Polisis: Automated analysis and presentation of privacy policies using deep learning. arXiv preprint arXiv:1802.02561. Mitra Bokaei Hosseini, Sudarshan Wadkar, Travis D Breaux, and Jianwei Niu. 2016. Lexical similarity of information type hypernyms, meronyms and synonyms in privacy policies. Po Hu, Donghong Ji, Chong Teng, and Yujing Guo. 2012. Context-enhanced personalized social summarization. In Proceedings of COLING 2012, pages 1223–1238, Mumbai, India. The COLING 2012 Organizing Committee. Lucas D Introna. 1997. Privacy and the computer: why we need privacy in the information society. Metaphilosophy, 28(3):259–275. Priyank Jain, Manasi Gyanchandani, and Nilay Khare. 2016. Big data privacy: a technological perspective and review. Journal of Big Data, 3(1):25. Sarthak Jain, Madeleine van Zuylen, Hannaneh Hajishirzi, and Iz Beltagy. 2020. Scirex: A challenge dataset for document-level information extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7506–7516. California Department of Justice. 2012. Attorney general kamala d. harris secures global agreement to strengthen privacy protections for users of mobile applications. Dongyeop Kang, Waleed Ammar, Bhavana Dalvi, Madeleine van Zuylen, Sebastian Kohlmeier, Eduard Hovy, and Roy Schwartz. 2018. A dataset of peer reviews (PeerRead): Collection, insights and NLP applications. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1647–1661, New Orleans, Louisiana. Association for Computational Linguistics. Patrick Gage Kelley, Joanna Bresee, Lorrie Faith Cranor, and Robert W Reeder. 2009. A nutrition label for privacy. In Proceedings of the 5th Symposium on Usable Privacy and Security, page 4. ACM. Patrick Gage Kelley, Lorrie Faith Cranor, and Norman Sadeh. 2013. Privacy as Part of the App DecisionMaking Process, page 3393–3402. Association for Computing Machinery, New York, NY, USA. Moniba Keymanesh, Micha Elsner, and Srinivasan Parthasarathy. 2020. Toward domain-guided controllable summarization of privacy policies. In Natural Legal Language Processing Workshop. KDD. Vinayshekhar Bannihatti Kumar, Abhilasha Ravichander, Peter Story, and Norman Sadeh. 2019. Quantifying the effect of in-domain distributed word representations: A study of privacy policies. In AAAI Spring Symposium on Privacy Enhancing AI and Language Technologies: PAL 2019. Logan Lebanoff and Fei Liu. 2018. Automatic detection of vague words and sentences in privacy policies. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3508–3517, Brussels, Belgium. Association for Computational Linguistics. Timothy Libert. 2018a. An automated approach to auditing disclosure of third-party data collection in website privacy policies. In Proceedings of the 2018 World Wide Web Conference, WWW ’18, page 207–216, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee. 4136 Timothy Libert. 2018b. An automated approach to auditing disclosure of third-party data collection in website privacy policies. In Proceedings of the 2018 World Wide Web Conference, pages 207–216. Jialiu Lin, Bin Liu, Norman Sadeh, and Jason I Hong. 2014. Modeling users’ mobile app privacy preferences: Restoring usability in a sea of permission settings. In 10th Symposium On Usable Privacy and Security ({SOUPS} 2014), pages 199–212. Thomas Linden, Rishabh Khandelwal, Hamza Harkous, and Kassem Fawaz. 2020. The privacy policy landscape after the gdpr. Proceedings on Privacy Enhancing Technologies, 2020(1):47–64. Bin Liu, Mads Schaarup Andersen, Florian Schaub, Hazim Almuhimedi, SA Zhang, Norman Sadeh, Alessandro Acquisti, and Yuvraj Agarwal. 2016a. Follow my recommendations: A personalized privacy assistant for mobile app permissions. In Symposium on Usable Privacy and Security. Bin Liu, Jialiu Lin, and Norman Sadeh. 2014a. Reconciling mobile app privacy and usability on smartphones: Could user privacy profiles help? In Proceedings of the 23rd international conference on World wide web, pages 201–212. Fei Liu, Nicole Lee Fella, and Kexin Liao. 2016b. Modeling language vagueness in privacy policies using deep neural networks. In 2016 AAAI Fall Symposium Series. Fei Liu, Jeffrey Flanigan, Sam Thomson, Norman Sadeh, and Noah A Smith. 2015. Toward abstractive summarization using semantic representations. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1077–1086. Fei Liu, Rohan Ramanath, Norman Sadeh, and Noah A. Smith. 2014b. A step towards usable privacy policy: Automatic alignment of privacy statements. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 884–894, Dublin, Ireland. Dublin City University and Association for Computational Linguistics. Mary Madden, Michele Gilman, Karen Levy, and Alice Marwick. 2017. Privacy, poverty, and big data: A matrix of vulnerabilities for poor americans. Wash. UL Rev., 95:53. Mary Madden, Lee Rainie, Kathryn Zickuhr, Maeve Duggan, and Aaron Smith. 2014. Public perceptions of privacy and security in the post-snowden era. Pew Research Center, 12. Ankit Malpani, Balaraman Ravindran, and Hema A Murthy. 2011. Personalized intelligent tutoring system using reinforcement learning. Florencia Marotta-Wurgler. 2019. Does “notice and choice” disclosure regulation work? an empirical study of privacy policies,”. Aaron K Massey, Jacob Eisenstein, Annie I Ant´on, and Peter P Swire. 2013. Automated text mining for requirements analysis of policy documents. In 2013 21st IEEE International Requirements Engineering Conference (RE), pages 4–13. IEEE. Aleecia M McDonald and Lorrie Faith Cranor. 2008. The cost of reading privacy policies. ISJLP, 4:543. Aleecia M McDonald and Lorrie Faith Cranor. 2010. Americans’ attitudes about internet behavioral advertising practices. In Proceedings of the 9th annual ACM workshop on Privacy in the electronic society, pages 63–72. Bruce M McLaren, Sung-Joo Lim, France Gagnon, David Yaron, and Kenneth R Koedinger. 2006. Studying the effects of personalized language and worked examples in the context of a web-based intelligent tutor. In International Conference on Intelligent Tutoring Systems, pages 318–328. Springer. Gabriele Meiselwitz. 2013. Readability assessment of policies and procedures of social networking sites. In International Conference on Online Communities and Social Computing, pages 67–75. Springer. Majd Mustapha, Katsiaryna Krasnashchok, Anas Al Bassit, and Sabri Skhiri. 2020. Privacy policy classification with xlnet (short paper). In Data Privacy Management, Cryptocurrencies and Blockchain Technology, pages 250–257. Springer. Abhijith Athreya Mysore Gopinath, Shomir Wilson, and Norman Sadeh. 2018. Supervised and unsupervised methods for robust separation of section titles and prose text in web documents. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 850–855, Brussels, Belgium. Association for Computational Linguistics. Kanthashree Mysore Sathyendra, Shomir Wilson, Florian Schaub, Sebastian Zimmeck, and Norman Sadeh. 2017a. Identifying the provision of choices in privacy policy text. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2774–2779, Copenhagen, Denmark. Association for Computational Linguistics. Kanthashree Mysore Sathyendra, Shomir Wilson, Florian Schaub, Sebastian Zimmeck, and Norman Sadeh. 2017b. Identifying the provision of choices in privacy policy text. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2774–2779, Copenhagen, Denmark. Association for Computational Linguistics. 4137 Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, C¸ a˘glar Gulc¸ehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence RNNs and beyond. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, pages 280–290, Berlin, Germany. Association for Computational Linguistics. Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don’t give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797–1807, Brussels, Belgium. Association for Computational Linguistics. Najmeh Mousavi Nejad, Pablo Jabat, Rostislav Nedelchev, Simon Scerri, and Damien Graux. 2020. Establishing a strong baseline for privacy policy classification. IFIP International Conference on ICT Systems Security and Privacy Protection. Sansa News. 2021. Sansa news. https://sansa. news/. Helen Nissenbaum. 2004. Privacy as contextual integrity. Wash. L. Rev., 79:119. Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for SQuAD. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 784– 789, Melbourne, Australia. Association for Computational Linguistics. Rohan Ramanath, Fei Liu, Norman Sadeh, and Noah A. Smith. 2014a. Unsupervised alignment of privacy policies using hidden markov models. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 605–610, Baltimore, Maryland. Association for Computational Linguistics. Rohan Ramanath, Florian Schaub, Shomir Wilson, Fei Liu, Norman Sadeh, and Noah Smith. 2014b. Identifying relevant text fragments to help crowdsource privacy policy annotations. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 2. Ashwini Rao, Florian Schaub, Norman Sadeh, Alessandro Acquisti, and Ruogu Kang. 2016. Expecting the unexpected: Understanding mismatched privacy expectations online. In Twelfth Symposium on Usable Privacy and Security ({SOUPS} 2016), pages 77– 96. Abhilasha Ravichander, Alan W Black, Shomir Wilson, Thomas Norton, and Norman Sadeh. 2019. Question answering for privacy policies: Combining computational and legal perspectives. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4947–4958, Hong Kong, China. Association for Computational Linguistics. General Data Protection Regulation. 2016. Regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive 95/46. Official Journal of the European Union (OJ), 59(1-88):294. Joel R Reidenberg, Travis Breaux, Lorrie Faith Cranor, Brian French, Amanda Grannis, James T Graves, Fei Liu, Aleecia McDonald, Thomas B Norton, and Rohan Ramanath. 2015. Disagreeable privacy policies: Mismatches between meaning and users’ understanding. Berkeley Tech. LJ, 30:39. Alexander M. Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 379–389, Lisbon, Portugal. Association for Computational Linguistics. Norman Sadeh, Ro Acquisti, Travis D Breaux, Lorrie Faith Cranor, Aleecia M Mcdonalda, Joel R Reidenbergb, Noah A Smith, Fei Liu, N Cameron Russellb, Florian Schaub, et al. 2013. The usable privacy policy project: Combining crowdsourcing, machine learning and natural language processing to semi-automatically answer those privacy questions users care about. Technical Report CMU-ISR-13119, Carnegie Mellon University. Kanthashree Mysore Sathyendra, Abhilasha Ravichander, Peter Garth Story, Alan W Black, and Norman Sadeh. 2017. Helping Users Understand Privacy Notices with Automated Query Answering Functionality: An Exploratory Study. Technical report. Florian Schaub, Rebecca Balebako, Adam L Durity, and Lorrie Faith Cranor. 2015. A design space for effective privacy notices. In Eleventh Symposium On Usable Privacy and Security (SOUPS 2015), pages 1–17. Executive Office of the President’s Council of Advisors on Science and Technology. 2014. Big data and privacy: A technological perspective. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073– 1083, Vancouver, Canada. Association for Computational Linguistics. Yan Shvartzshanider, Ananth Balashankar, Thomas Wies, and Lakshminarayanan Subramanian. 2018. RECIPE: Applying open domain question answering to privacy policies. In Proceedings of the Workshop on Machine Reading for Question Answering, pages 71–77, Melbourne, Australia. Association for Computational Linguistics. 4138 Yan Shvartzshnaider, Noah Apthorpe, Nick Feamster, and Helen Nissenbaum. 2019. Going against the (appropriate) flow: a contextual integrity approach to privacy policy analysis. In Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, volume 7, pages 162–170. Yan Shvartzshnaider, Ananth Balashankar, Vikas Patidar, Thomas Wies, and Lakshminarayanan Subramanian. 2020. Beyond the text: Analysis of privacy statements through syntactic and semantic role labeling. arXiv preprint arXiv:2010.00678. Peter Story, Sebastian Zimmeck, Abhilasha Ravichander, Daniel Smullen, Ziqi Wang, Joel Reidenberg, N Cameron Russell, and Norman Sadeh. 2019. Natural language processing for mobile app privacy compliance. In AAAI Spring Symposium on Privacy Enhancing AI and Language Technologies: PAL 2019. Peter Story, Sebastian Zimmeck, and Norman Sadeh. 2018. Which apps have privacy policies? In Annual Privacy Forum, pages 3–23. Springer. Auto TLDR. 2021. Auto tl;dr. http://autotldr. io/. Noriko Tomuro, Steven Lytinen, and Kurt Hornsburg. 2016. Automatic summarization of privacy policies using ensemble learning. In Proceedings of the Sixth ACM Conference on Data and Application Security and Privacy, CODASPY ’16, page 133–135, New York, NY, USA. Association for Computing Machinery. Joseph Turow, Jennifer King, Chris Jay Hoofnagle, Amy Bleakley, and Michael Hennessy. 2009. Americans reject tailored advertising and three activities that enable it. Available at SSRN 1478214. Blase Ur, Pedro Giovanni Leon, Lorrie Faith Cranor, Richard Shay, and Yang Wang. 2012. Smart, useful, scary, creepy: perceptions of online behavioral advertising. In proceedings of the eighth symposium on usable privacy and security, pages 1–15. FTC US Federal Trade Commission et al. 2012. Protecting consumer privacy in an era of rapid change: Recommendations for businesses and policymakers. FTC Report. Michela Del Vicario, Walter Quattrociocchi, Antonio Scala, and Fabiana Zollo. 2019. Polarization and fake news: Early warning of potential misinformation targets. ACM Trans. Web, 13(2). Harm de Vries, Dzmitry Bahdanau, and Christopher Manning. 2020. Towards ecologically valid research on language user interfaces. arXiv preprint arXiv:2007.14435. Xiaoyin Wang, Xue Qin, Mitra Bokaei Hosseini, Rocky Slavin, Travis D. Breaux, and Jianwei Niu. 2018. Guileak: Tracing privacy policy claims on user input data for android applications. In Proceedings of the 40th International Conference on Software Engineering, ICSE ’18, page 37–47, New York, NY, USA. Association for Computing Machinery. Samuel D Warren and Louis D Brandeis. 1890. The right to privacy. Harvard law review, pages 193– 220. Alan F Westin. 1968. Privacy and freedom. Washington and Lee Law Review, 25(1):166. S Wilson, F Schaub, A Dara, F Liu, S Cherivirala, P G Leon, M S Andersen, S Zimmeck, K Sathyendra, N C Russell, T B Norton, E Hovy, J R Reidenberg, and N Sadeh. 2016a. The creation and analysis of a website privacy policy corpus. In Annual Meeting of the Association for Computational Linguistics, Aug 2016. ACL. Shomir Wilson, Florian Schaub, Aswarth Abhilash Dara, Frederick Liu, Sushain Cherivirala, Pedro Giovanni Leon, Mads Schaarup Andersen, Sebastian Zimmeck, Kanthashree Mysore Sathyendra, N Cameron Russell, et al. 2016b. The creation and analysis of a website privacy policy corpus. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), volume 1, pages 1330–1340. Shomir Wilson, Florian Schaub, Rohan Ramanath, Norman Sadeh, Fei Liu, Noah A Smith, and Frederick Liu. 2016c. Crowdsourcing annotations for websites’ privacy policies: Can it really work? In Proceedings of the 25th International Conference on World Wide Web, pages 133–143. Razieh Nokhbeh Zaeem, Safa Anya, Alex Issa, Jake Nimergood, Isabelle Rogers, Vinay Shah, Ayush Srivastava, and K Suzanne Barber. 2020. Privacycheck v2: A tool that recaps privacy policies for you. In 29th ACM International Conference on Information and Knowledge Management (CIKM). ACM. To appear. Razieh Nokhbeh Zaeem, Rachel L German, and K Suzanne Barber. 2018. Privacycheck: Automatic summarization of privacy policies using data mining. ACM Transactions on Internet Technology (TOIT), 18(4):1–18. Sebastian Zimmeck, Peter Story, Daniel Smullen, Abhilasha Ravichander, Ziqi Wang, Joel Reidenberg, N Cameron Russell, and Norman Sadeh. 2019a. Maps: Scaling privacy compliance analysis to a million apps. Proceedings on Privacy Enhancing Technologies, 2019(3):66–86. Sebastian Zimmeck, Peter Story, Daniel Smullen, Abhilasha Ravichander, Ziqi Wang, Joel R. Reidenberg, N. Russell, and N. Sadeh. 2019b. Maps: Scaling privacy compliance analysis to a million apps. Proceedings on Privacy Enhancing Technologies, 2019:66 – 86. 4139 Sebastian Zimmeck, Ziqi Wang, Lieyong Zou, Roger Iyengar, Bin Liu, Florian Schaub, Shomir Wilson, Norman M Sadeh, Steven M Bellovin, and Joel R Reidenberg. 2017. Automated analysis of privacy requirements for mobile apps. In NDSS. 4140 Figure 2: Example of privacy nutrition labels, disclosing information collected by companies and third parties through an application. Source: Apple. A Privacy Nutrition Labels Figure.2 includes an example of a privacy nutrition label, intended to disclose to a user the information a company and any third parties collect through an app. Apple requires developers to self-report the information for these nutrition labels.
|
2021
|
319
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 377–387 August 1–6, 2021. ©2021 Association for Computational Linguistics 377 Multi-TimeLine Summarization (MTLS): Improving Timeline Summarization by Generating Multiple Summaries Yi Yu1, Adam Jatowt2, Antoine Doucet3 Kazunari Sugiyama1, Masatoshi Yoshikawa1 1Kyoto University, Japan 2University of Innsbruck, Austria, 3University of La Rochelle, France [email protected] [email protected], [email protected] {kaz.sugiyama, yoshikawa}@i.kyoto-u.ac.jp Abstract In this paper, we address a novel task, Multiple TimeLine Summarization (MTLS), which extends the flexibility and versatility of TimeLine Summarization (TLS). Given any collection of time-stamped news articles, MTLS automatically discovers important yet different stories and generates a corresponding timeline for each story. To achieve this, we propose a novel unsupervised summarization framework based on the two-stage affinity propagation process. We also introduce a quantitative evaluation measure for MTLS based on the previous TLS evaluation methods. Experimental results show that our MTLS framework demonstrates high effectiveness and MTLS task can provide better results than TLS. 1 Introduction Nowadays, online news articles are one of the most popular Web documents. However, due to a huge amount of news articles available online, it is getting difficult for users to effectively search, understand, and track the entire news stories. To solve this problem, a research area of TimeLine Summarization (TLS) has been established, which can alleviate the redundancy and complexity inherent in news article collections thereby helping users better understand the news landscape. After the influential work on temporal summaries by Swan and Allan (2000), TLS has attracted researchers’ attention. Most of works on TLS (Martschat and Markert, 2018; Steen and Markert, 2019; Gholipour Ghalandari and Ifrim, 2020) have focused on improving the performance of summarization. However, their drawbacks are as follows: (a) the methods work essentially on a homogeneous type of datasets such as ones compiled from the search results of an unambiguous query (e.g., “BP Oil Spill”). The requirements imposed on the input dataset make it hard for TLS systems to generalize; (b) the output is usually a single timeline regardless of the size and the complexity of the input dataset. We propose here the Multiple TimeLine Summarization (MTLS) task that enhances and further generalizes TLS. MTLS automatically generates a set of timelines that summarize disparate yet important stories, rather than always generating a single timeline as is in the case of TLS. An effective MTLS framework should: (a) detect key events including both short- and long-term events, (b) link events related to the same story and separate events belonging to other stories, and (c) provide informative summaries of constituent events to be incorporated into the generated timelines. MTLS can also help to deal with the ambiguity, which is common in information retrieval. For example, suppose that a user wants to get an overview of news about a basketball player, Michael Jordan, from a large collection of news articles. However, when a search engine over such a collection takes “Michael Jordan” as a query, it would likely return documents constituting a mixture of news about different persons having the same name. Then, how can a typical TLS system return meaningful results if only a single timeline can be generated? Similarly, ambiguous queries such as “Apple”, “Amazon”, “Java” require MTLS solutions to produce high quality results. To address this task, we further propose a TwoStage Affinity Propagation Summarization framework (2SAPS). It uses temporal information embedded in sentences to discover important events, and their linking information latent in news articles to construct timelines. 2SAPS has several advantages: firstly, it is entirely unsupervised which is especially suited to TLS-related tasks as there are very few gold summaries available for training supervised systems; secondly, both the number of events and the number of generated timelines are 378 self-determined. This allows our framework to be dependent only on the input document collection, instead of on human efforts. Furthermore, the current TLS evaluation measures allow only 1-to-1 comparison (system- to human-generated timeline), which is not suitable for MTLS task where multiple timelines must be compared to (typically) multiple ground-truth timelines. Therefore, we also propose a quantitative evaluation measure for MTLS based on the adaptation of the previous TLS evaluation framework. Given these points, our contributions in this work are summarized as follows: 1. We propose a novel task (MTLS), which automatically generates multiple, informative, and diverse timelines from an input time-stamped document collection. 2. We introduce a superior MTLS model that outperforms all TLS-adapted MTLS baselines. 3. We design an evaluation measure for MTLS systems by extending the original TLS evaluation framework. 2 Related Work 2.1 Timeline Summarization Since the first work on timeline summarization (Swan and Allan, 2000; Allan et al., 2001), this topic has received much attention over the years (Alonso et al., 2009; Yan et al., 2011a; Zhao et al., 2013; Tran et al., 2013; Li and Li, 2013; Suzuki and Kobayashi, 2014; Wang et al., 2016; Takamura et al., 2011; Pasquali et al., 2019, 2021). In the following, we review the major approaches. Chieu and Lee (2004) constructed timeline by directly selecting the top ranked sentences based on the summed similarities within n-day long window. Yan et al. (2011b) proposed evolutionary timeline summarization (ETS) to return the evolution trajectory along the timeline, consisting of individual but correlated summaries of each date. Shahaf et al. (2012) created information maps (Maps) to help users understand domain-specific knowledge. However, the output consists of a set of storylines that have intersections or overlaps, which is not appropriate for a dataset that may contain quite different topics. Nguyen et al. (2014) proposed a pipeline to generate timelines consisting of date selection, sentence clustering and sentence ranking. Recently, Martschat and Markert (2018) adapted a submodular function model for TLS task, which is originally used for multi-document summarization (MDS). Duan et al. (2020) introduced the task of Comparative Timeline Summarization (CTS), which captures important comparative aspects of evolutionary trajectories in two input sets of documents. The output of the CTS system is, however, always two timelines generated in a contrastive way. Then, Gholipour Ghalandari and Ifrim (2020) examined different TLS strategies and categorized TLS frameworks into the following three types: direct summarization approaches, date-wise approaches, and event detection approaches. To the best of our knowledge, the idea of multiple timeline summarization has not been formally proposed yet. Table 1 compares the related tasks. 2.2 Timeline Evaluation Some works (Yan et al., 2011b; Chen et al., 2019; Duan et al., 2020) evaluate timeline by only computing ROUGE scores (Lin, 2004). This way ignores the temporal aspect of a timeline, which is important in timeline summarization. Martschat and Markert (2017) then proposed a framework, called tilse, to assess timelines from both textual and temporal aspects. Subsequently, TLS works (Steen and Markert, 2019; Gholipour Ghalandari and Ifrim, 2020; Born et al., 2020) have followed this framework to evaluate their models. Some researches (Tran et al., 2015; Shahaf et al., 2012; Alonso and Shiells, 2013) also involved user studies, in which users are required to score systemgenerated timelines based on varying criteria such as relevance and understandability. In Section 5, we will adapt the tilse framework to MTLS task. 3 Problem Definition We formulate MTLS task as follows: Input: A time-stamped news article collection D = {d1, d2, ..., d|D|}. The collection can be standalone or compiled from search results returned by a news search engine. Output: A set of timelines, T = {T1, T2, . . . , Tk} is generated based on D, so that each timeline Ti includes a sequence of time/date1 and summary pairs (tTi 1 , sTi 1 ), . . . , (tTi l , sTi l ) where sTi j (i = 1, . . . , k) are the summary sentences for the time tTi j (j = 1, . . . , l) and l is the length of Ti. Each timeline in T should be consistent and coherent, yet different from other timelines. 1In this paper, time and date are used as synonyms. 379 Tasks Output 1 timeline Output ≥2 timelines Automatically Determine k Input Heterogeneous Collection Quantitatively Evaluate TLS (Most of which in Section 2.1) ✓ ✓ CTS (Duan et al., 2020) ✓(always 2) ✓ ETS (Yan et al., 2011b) ✓ ✓ Maps (Shahaf et al., 2012) ✓ ✓ MTLS (Proposed task) ✓ ✓ ✓ ✓ ✓ Table 1: Comparison between different TLS related tasks (k is the number of generated timelines). We note that while the traditional TLS task is limited as a document collection for it is typically coherent and homogeneous, MTLS is more flexible as the input news collection can be diverse. For example, the input collection can be generated using a search query q composed of multiple entities or concepts like q = {“egypt”, “h1n1”, “iraq”} or by using an ambiguous query like q = {“michael”, “jordan”}, or it can also consist of news articles crawled over a certain time span from multiple news sources. Generally, the more heterogeneous D is, the more timelines could be produced. The intuition behind this idea is that users will need more structured information to help them understand a relatively complex document collection. 4 Framework Next, we present two key components of our framework: event generation module (Sec. 4.1) and timeline generation module (Sec. 4.2). We first make the following two assumptions: Assumption 1: News articles sometimes retrospectively mention past events for providing necessary context to the target event, for underlying continuation, causality, etc. Assumption 2: Sentences mentioning similar dates have higher probability to refer to the same event than sentences with different dates. 4.1 Event Generation Module In this module, we extract important historical events from a document collection. Gholipour Ghalandari and Ifrim (2020) constructed events by simply grouping articles with close publication dates into clusters, resulting in lower accuracy. Note that Assumption 1 implies that a single news article may contain multiple events. Accordingly, in our work, the concept of event is more fine-grained. We define event as a set of sentences that describe the same real-world occurrence, typically using the same identifying information (e.g., actions, entities, locations). This information is captured by sentence-BERT (Reimers and Gurevych, 2019): a pre-trained model on a transformer network where similar meanings are positioned nearby in semantic vector space. We then employ Affinity Propagation (AP) (Frey and Dueck, 2007) following Steen and Markert (2019) for clustering similar sentences. AP algorithm groups data points by selecting a set of exemplars along with their followers due to message passing. It operates over an affinity matrix S, where S(i, j) denotes similarity between data points xi and xj. We observe that high semantic similarity does not always guarantee that sentences refer to the same event. Especially, for some periodic events, similar happenings might have occurred several times. For example, a news article could include sentences reporting that Brazil won the gold medal in the World Cup (in 2002) while some other sentences in this document could recall that Brazil has won the first place in the World Cup in 1994. It is clear that those sentences describe two distinct events, which would be grouped into one event if only semantic similarity is considered. Therefore, based on Assumption 2, we introduce another key factor, temporal similarity, which enhances the confidence of how likely two sentences will refer to the same event. We define each element S1(vi, vj) of affinity matrix S1 as follows: S1(vi, vj) = α1·Sdate(ti, tj)+(1 −α1)·Scos(vi, vj), (1) where vi and vj denote different sentences, and ti and tj denote dates mentioned by vi and vj, respectively.2 In addition, Sdate and Scos denote the temporal and semantic similarities, respectively. While we employ cosine similarity for the semantic similarity, we define temporal similarity Sdate(i, j) to quantify how similar two dates are using Equation (2): Sdate(ti, tj) = 1 expγ·|ti−tj| , (2) where γ3 is the decay rate of the exponential func2We use Heideltime (Strötgen and Gertz, 2013) for resolving temporal expressions. If a sentence does not explicitly mention any date, we assume it refers to the publication date of the article. 3We set γ = 0.05 in the experiments. 380 tion. The larger the time gap between two dates, the smaller the value of Sdate. By passing messages of both semantic and temporal information between sentences, clusters consisting of exemplar and non-exemplar sentences are constructed to form the candidate event set E. Each cluster represents an event. Event Selection. In a timeline, it is not necessary to show all events of a story as users usually care about the most important events only. We design an event selection step that is helpful for handling excessive number of events. The selection relies on two measures: Salience and Consistency defined by Equations (3) and (4), respectively: Salience(e) = log(| e |) log(| D |), (3) Consistency(e) = P vi∈e,vi̸=veScos(vi,ve) | e | −1 , (4) where ve is the exemplar sentence in event e; | e | and | D | denote the number of sentences in e and document collection D, respectively. Intuitively, important historical events would often be mentioned by future news reports. Salience of event is used to evaluate such importance and is computed as the relative frequency of sentences about that event compared with all sentences in the collection. On the other hand, Consistency ensures high quality of events. We then rank all candidate events based on the weighted summed score of these two measures. Hereafter, we denote the weight of Event Salience as ζ1 and that of Event Consistency as 1 −ζ1. We select the top-scored events obtaining a new event set E∗by setting a threshold. To avoid tuning its value, we set the value to one standard deviation from the mean (lower end). 4.2 Timeline Generation Module While TLS systems directly link all the identified events, MTLS requires their deeper understanding. As described in Section 1, an effective MTLS framework should link events related to the same story and separate other unrelated events to different timelines. To achieve this, we explain the following steps in this module: Event Linking, Timeline Selection, and Timeline Summarizing. Event Linking. According to Assumption 1, current events can refer to related past events. We thus define a reference matrix R, in which each element R(ei, ej) denotes the degree of reference between two events ei and ej. As events in our work are represented by sentences and a sentence belongs to a single event, R(ei, ej) can be reflected by counting patterns of sentence co-occurrences in documents. Formally, R(vi, vj) represents the case where two sentences vj and vi refer to each other as defined by Equation (5): R(vi, vj)= 1 vi,vj ∈d ∧vi ∈ek,vj ∈el, ek ̸= el 0 otherwise, (5) where d is an article, ek and el are elements in E∗. The degree of reference between ei and ej is then defined as follows: R(ei, ej) = P v1∈ei P v2∈ej R(v1, v2) | ei | · | ej | , (6) where |ei| and |ej| are sizes of ei, ej, respectively. We then construct a graph of events where each node is an e ∈E∗, and the value of an edge reflects the connection degree between a pair of two events. We reuse AP algorithm to detect the community of events over the affinity matrix S2 defined by Equation (7): S2(ei, ej) = α2 · R(ei, ej) + (1 −α2) · Scos(ei, ej), (7) where Scos(ei, ej) denotes cosine similarity between ei and ej to capture semantic similarity. Based on the affinity matrix S2, AP finally generates clusters, i.e., the initial timeline set, T . Timeline Selection. In order to ensure the quality of constructed timelines, we define criteria to select high-quality timelines from T . Similar to event selection described in Section 4.1, we also use two indicators to evaluate the quality of a timeline. We define Timeline Salience as the average score of Event Salience of all events within the timeline, and Timeline Coherence as the average of semantic similarity scores between any chronologically4 adjacent events defined by Equation (8): Coherence(T) = P ei,ei+1∈T Scos(ei, ei+1) | T | −1 , (8) where | T | is the size of a timeline, i.e., the number of events in this timeline. Intuitively, important timelines, which reflect important stories in the document collection, are more likely to be preferred by users. Timeline Salience captures this importance by passing the importance of its components (i.e., events), while Timeline Coherence ensures that the story expressed by the timeline is consistent. 4The time of an event e is given by its exemplar sentence. 381 We rank timelines based on a weighted sum of Timeline Salience and Timeline Coherence. The weight of Timeline Salience is denoted as ζ2; thus the weight of Timeline Coherence is 1−ζ2. We then select the top-scored elements from the timeline set T based on a threshold. Same as before, we set the value to one standard deviation from the mean. Timeline Summarizing. By previous steps, we have now obtained multiple timelines {T1, T2, ...}, where T is a list of events {e1, e2, ...}. However, it is not feasible to show all contents of each e as it usually contains many sentences. We use only the exemplar sentence in event since exemplar is the most typical and representative member in the group. In addition, it is possible that two events ei and ej occur on the same day. In this case, we concatenate their exemplar sentences. Timeline Tagging. This step is an add-on to MTLS systems. To better understand the stories of constructed timelines, we believe that it should be helpful for users to also obtain a label for each timeline. As described in Section 1, the input document collection may be composed of different topics or of one topic discussed through different aspects. For example, among the timelines generated based on the topic syria, one timeline might summarize the story about Syrian civil war while another might be about Syrian political elections. A label should then help people understand the story of the timeline. We simply select the 3 most frequent words among events (excluding stopwords) for each timeline as its label. 5 Evaluation Framework 5.1 TLS Evaluation TLS evaluation relies on ROUGE score and its variants as follows: Concatenation-based ROUGE (concat). It considers only the textual overlap between concatenated system summaries and ground-truth, while ignoring all date information of timeline (Yan et al., 2011b; Nguyen et al., 2014; Wang et al., 2016). Date-agreement ROUGE (agreement). It measures both textual and temporal information overlap by computing ROUGE score only when the date in the system-generated timelines matches the one of the ground-truth timeline (Tran et al., 2013). Otherwise, its value is 0. Alignment-based ROUGE. It linearly penalizes the ROUGE score by the distances of dates or/and summary contents. Martschat and Markert (2017) proposed three types of this metric: align, align+, align+m:1 (align by date, align by date and contents, align by date and contents where the map function is non-injective, respectively). Date selection (d-select). It evaluates how well the model works in selecting correct dates in the ground-truth (Martschat and Markert, 2018). 5.2 MTLS evaluation The evaluation methods for TLS cannot directly assess the performance of MTLS systems as there are multiple output timelines and multiple ground-truth timelines. Concretely, given an input collection D, corresponding ground-truth timeline set G = {G1, G2, ...Gk1} (k1 ≥1), and system-generated timeline set T = {T1, T2, ..., Tk2} (k2 ≥1), evaluation metrics need information to automatically “match” the ground-truth timeline when evaluating Ti. Therefore, we make the system find the closest ground-truth G∗to timeline T as follows: G∗= arg max G∈G fm(T, G), (9) where fm is the TLS evaluation function to compute the score between T and G based on metric m, which can be either concat, agreement, align, align+, align+m:1, or d-select. Then, the overall performance of the MTLS models is computed by taking the average of all the members in T . 6 Experimental Setup The goal of our experiments is to answer the following research questions (RQs): RQ1: Do MTLS models produce more meaningful output than TLS models? RQ2: How does 2SAPS framework perform on MTLS task compared with other MTLS baselines? RQ3: How effective are the components of the modules in 2SAPS? How do parameter changes in the model affect the results? 6.1 Datasets We note that there is no available dataset for MTLS task, thus we construct MTLS datasets5 extending existing TLS datasets. Tran et al. released Timeline17 (Binh Tran et al., 2013) and Crisis (Tran et al., 2015) datasets for TLS over news articles. 5The datasets are now available at https://yiyualt.github.io/mtlsdata/. 382 Name #Topics #Groundtruth Avg.Timespan #Docs. #Sents. Timeline17 9 17 250 days 4,650 183,782 Crisis 4 22 343 days 9,242 331,044 Table 2: Statistics on TLS datasets. L=1 D1:egypt D2:finan D3:haiti D4:h1n1 D5:libya L=2 D6:egypt+libya D7:haiti+iraq D8:h1n1+haiti D9:finan+mj D10:egypt+mj L=3 D11:egypt+h1n1+iraq D12:finan+iraq+syria D13:egypt+ iraq+mj D14:finan+h1n1+mj D15:finan+libya+mj L=4 D16:egypt+finan+haiti+iraq D17:finan+h1n1+ iraq+mj D18:h1n1+haiti+iraq+mj D19:finan+ h1n1+haiti+mj D20:egypt+haiti+iraq+mj L=5 D21:finan+h1n1+haiti+iraq+mj D22:h1n1+haiti+iraq+mj+syria D23:egypt+finan+haiti+mj+syria D24:egypt+finan+ iraq+mj+syria D25:egypt+finan+h1n1+haiti+mj Table 3: MTLS datasets used for our experiments. Table 2 shows their statistics. To assure high complexity of data, we generate multiple datasets from TLS datasets by varying degree of story mixtures. We construct MTLS datasets based on combining TLS datasets, according to the following procedure: (1) set the number of topics L used to generate a new dataset; (2) from TLS datasets, randomly choose L topics, then merge their document collections into a new dataset D along with grouping their associated ground-truth timelines into G.6 (3) repeat steps (1) and (2). Here, the value of L reflects the complexity of the dataset. The more topics the dataset contains, the more complex it is. We repeated the steps (1)~(3) on Timeline177 and finally created 25 datasets as shown in Table 3. Timeline17 contains 9 document collections, covering the following topics: “BP Oil Spill” (bpoil), “Influenza H1N1” (h1n1), “Michael Jackson death” (mj), “Libyan War” (libya), “Egyptian Protest” (egypt), “Financial Crisis” (finan), “Haiti Earthquake” (haiti), “Iraq War” (iraq), “Syrian Crisis” (syria). 6.2 Baselines As there are no ready models for MTLS task, we design the baselines as “divide-and-summarize” approaches. The underlying idea is: first segment the input dataset into sub-datasets (subsequently called 6If a topic has multiple ground-truth timelines, we pick one that has length closest to the average length of the timelines for that topic. 7We note that Crisis contains only 4 topics, resulting in few possible combinations, so we finally decided to skip it. segments) by partition/division algorithms; then adopt TLS techniques to generate a timeline for each sub-dataset (segment). We now describe the choices for each step. Dataset Division Approaches: • Random. We randomly decide the number of segments from 1 to 10. Then, we assign a news article to a random segment. • LDA (Latent Dirichlet Allocation) (Blei et al., 2003). Given a dataset, we first use LDA to detect the main topics in the dataset. Then, we assign each news article to its dominant topic. • K-means (MacQueen et al., 1967). We use k-means algorithm in scikit-learn.8 TLS Approaches: • CHIEU2004 (Chieu and Lee, 2004): It is a frequently used unsupervised TLS baseline which selects the top-ranked sentences based on summed similaries within n-day window. • MARTSCHAT2018 (Martschat and Markert, 2018): It is one of the state-of-the-art TLS models and is also the first work to establish formal experimental settings for TLS task. We use the implementation given by the authors.9 • GHALANDARI2020 (Gholipour Ghalandari and Ifrim, 2020): It constructs timeline by first predicting the important dates via a simple regression model and then selecting important sentences for each date.10 We combine the above 3 dataset division approaches and 3 TLS approaches and thus yield 9 baselines. 6.3 Experimental Settings Concerning the characteristics of MTLS task and our datasets, the experimental settings differ from the TLS settings applied in Martschat and Markert (2018). In particular, the settings are: • When generating timelines, none of the compared models knows the actual value of L (i.e., L is not an input data). The stratification given in Table 3 is shown only for the reader to explain the datasets’ construction method. 8https://scikit-learn.org/ 9https://github.com/smartschat/tilse. 10https://github.com/complementizer/ news-tls. 383 • For the dataset-division algorithms, LDA and k-means, we use different techniques to find optimal number of segments. For LDA, we evaluate topic coherence measure (Cv score) (Röder et al., 2015) for topic numbers ranging from 1 to 10, and then choose the optimal number. For k-means, we use silhouette value (Rousseeuw, 1987) to determine the optimal number of segments. • All the compared methods do not take the information of the ground-truth as input. That is, the number of dates, the average number of summary sentences per date, the total number of summary sentences, the ground-truth start dates, and end dates are all unknown. • We set the length of timelines to 20 and summary length to 2 sentences per date. 7 Results and Discussion 7.1 MTLS vs. TLS We first address RQ1 to show the necessity of MTLS and to demonstrate that TLS performs poorly when an input dataset contains mixture of documents on different stories. To achieve this, we compare results of MTLS baselines with a standard TLS approach. Table 4 shows the performance comparison between TLS and MTLS baselines based on MARTCHAT2018. For fair comparison in this first experiment, we select only one timeline from MTLS outputs that is most similar to the timeline generated by TLS. We observe that when L = 1, 2, MTLS underperforms TLS by 15.1%, 4.8% in terms of align+m:1 ROUGE-1, respectively. However, it outperforms TLS by 150%, 117.1%, and 94.7% when L equals 3,4,5, respectively. This indicates that as the complexity of input document collection increases (higher L values), TLS systems do not produce good results when compared to MTLS ones. In real world scenarios, it is rather rare that the input dataset is clean enough to contain only a single topic. Thus, these results suggest that MTLS approach should in practice be more useful than TLS. The results for the other two TLS algorithms introduced in Section 6.2 show a similar trend, too. Furthermore, the example outputs of TLS and MTLS systems are also available as supplementary materials. 7.2 Performance of 2SAPS We now investigate the performance of our framework to answer RQ2. Table 5 shows the overall performance of MTLS systems. We observe that 2SAPS achieves the best performance in terms of all ROUGE metrics. In particular, when compared with CHIEU2004, MARTSCHAT2018 and GHALANDARI2020 in terms of concat ROUGE1 score, it outperforms them by 52.9%, 12.2%, and 16.4%, respectively. We also observe that GHALANDARI2020 method still achieves the best performance among baselines except for concat ROUGE-1. Furthermore, it is worth noticing that kmeans works best in dividing datasets. On average, k-means outperforms Random and LDA by 15% and 7.2%, respectively, in terms of concat ROUGE1. Finally, compared with the best-performing baseline, k-means-GHALANDARI2020, our 2SAPS outperforms it by 9.9%, 15.1%, 0%, 10%, 4.7%, 3.6%, 19.1%, in terms of concat (ROUGE-1,ROUGE2), align+m:1 (ROUGE-1,ROUGE-2), agreement (ROUGE-1,ROUGE-2) and d-select, respectively. 7.3 Ablation Study We turn to the first part of RQ3. We conduct ablation tests on Event Selection (ES) and Timeline Selection (TS) components. Table 6 shows the changes of different models. We observe that without ES, d-select and align+m:1 ROUGE-2 scores decrease 14.6% and 42.2% compared with 2SAPS. The plausible reason is that without ES, many unimportant dates and events are included in a timeline, resulting in low recall of correct dates. On the other hand, without TS component, the generated timeline set tends to contain noisy timelines, causing low ROUGE-1 as the performance drops by 18.8%. 7.4 Parameter Impact We now analyze the impact of key parameters, α1, α2, ζ1, ζ2. α1 and α2 directly influence the quality of generated events and timelines, while ζ1 and ζ2 indirectly affect the model’s performance by controlling the selection steps. Figure 1 shows the performance of 2SAPS under concat ROUGE-1, align+m:1 ROUGE-1, and agreement ROUGE-1. In particular, we observe that: a smaller value of α1 (from 0.1 to 0.4) gives better results than a larger value (Figure 1a). When α1 turns to 1, AP algorithm does not converge, and the values of all measures become 0. The plausible reason for this could be that when sentence dates are very 384 Model Metric L=1 L=2 L=3 L=4 L=5 TLS (MARTSCHAT2018) concat (ROUGE-1) 0.287 0.310 0.214 0.261 0.202 concat (ROUGE-2) 0.061 0.069 0.038 0.044 0.035 align+m:1 (ROUGE-1) 0.053 0.063 0.032 0.041 0.038 align+m:1 (ROUGE-2) 0.011 0.017 0.011 0.007 0.007 MTLS (k-means-MARTSCHAT2018) concat (ROUGE-1) 0.272 0.364 0.362 0.400 0.390 concat (ROUGE-2) 0.056 0.084 0.085 0.100 0.084 align+m:1 (ROUGE-1) 0.046 0.063 0.082 0.097 0.082 align+m:1 (ROUGE-2) 0.009 0.014 0.026 0.034 0.024 MTLS (LDA-MARTSCHAT2018) concat (ROUGE-1) 0.274 0.332 0.363 0.335 0.273 concat (ROUGE-2) 0.054 0.074 0.089 0.079 0.059 align+m:1 (ROUGE-1) 0.043 0.057 0.078 0.080 0.065 align+m:1 (ROUGE-2) 0.007 0.009 0.027 0.024 0.018 Table 4: Performance comparison between TLS and MTLS systems. For fair comparisons, we compare the single timeline generated by TLS model with the most related timeline generated by MTLS models. MTLS Methods concat align+m:1 agreement d-select ROUGE-1 ROUGE-2 ROUGE-1 ROUGE-2 ROUGE-1 ROUGE-2 F1 Baselines CHIEU2004 Random 0.191 0.027 0.019 0.004 0.010 0.002 0.075 LDA 0.192 0.035 0.023 0.005 0.013 0.004 0.089 k-means 0.229 0.046 0.027 0.006 0.014 0.004 0.096 MARTSCHAT2018 Random 0.254 0.049 0.044 0.009 0.037 0.007 0.352 LDA 0.289 0.068 0.062 0.017 0.052 0.015 0.387 k-means 0.291 0.071 0.061 0.017 0.051 0.015 0.376 GHALANDARI2020 Random 0.253 0.048 0.068 0.015 0.058 0.013 0.414 LDA 0.268 0.062 0.085 0.025 0.076 0.024 0.440 k-means 0.284 0.073 0.096 0.030 0.085 0.028 0.467 Our method 2SAPS 0.312 0.084 0.096 0.033 0.089 0.029 0.556 Table 5: Overall performance obtained by the baselines and the proposed methods over D1 ~D25 datasets. d-select ROUGE-1 ROUGE-2 2SAPS w/o ES 0.475 0.085 0.019 2SAPS w/o TS 0.502 0.078 0.023 2SAPS 0.556 0.096 0.033 Table 6: Ablation results of 2SAPS model, showing changes of align+m:1 ROUGE and d-select F1 scores. close, the elements of transition matrix differ only slightly, resulting in non-convergence. Figure 1b shows the impact of the reference relation in linking events. The values of all metrics increase as α2 increases. It makes sense that reference relation exerts an important role in linking events into timelines, thus a higher value is necessary. However, when α2 is over 0.9, the performance drops because when news articles provide few contextual events (e.g., background events, related events, etc.), then the reference relation between events becomes unreliable. ζ1 controls the impact of Event Salience described in Section 4.1. Another corresponding factor is Event Consistency, which is weighted by 1-ζ1. Figure 1c shows that the model with larger values of ζ1 underperforms the ones with relatively small values of ζ1 (from 0.2 to 0.4), indicating that con(a) α1: Temporal similarity (b) α2: Reference relation (c) ζ1: Event salience (d) ζ2: Timeline salience Figure 1: Impact of parameters on F1 score. sistency of event matters more than its salience in selecting high-quality events. Finally, in Figure 1d, we observe that along with the increase of ζ2, the performance of all metrics decrease, suggesting that the coherence of timeline is more effective than salience in selecting good timelines. 7.5 Limitations Our 2SAPS model works essentially on the unit of sentences and constructs a graph where each sentence is a node and edge is the relation between 385 sentences. It has then a complexity of O(n2). Future work could address this by simplifying graph structure and providing approximate solutions to cover also the cases of processing large datasets. Another solution is to select only important sentences from news articles using the combination of classification, summarization or filtering. 8 Conclusions We introduced MTLS task to generalize the timeline summarization problem. MTLS improves the performance of timeline summarization by generating multiple summaries. We conducted experiments to first show that given a heterogeneous time-stamped news article collection, TLS usually does not produce satisfactory result. We further proposed 2SAPS, a two-stage clustering-based framework, to effectively solve MTLS task. Furthermore, we extended TLS datasets to MTLS datasets, as well as introduced a novel evaluation measure for MTLS. Experimental results show that 2SAPS outperforms MTLS baselines which follow the “divide-and-summarize” strategy. Our work significantly improves the generalization ability of timeline summarization and can provide users with easier access to news collections. As an unsupervised approach that does not require costly training data, it can be applied to any potential datasets and languages. In future work, we plan to test our approach on additional MTLS datasets. We will also investigate scenarios in which MTLS can enhance information retrieval systems operating over news article collections. For users searching over large temporal collections, structuring the returned results into a series of timelines could prove beneficial, instead of returning a usual list of interwoven documents that relate to different stories or periods. Acknowledgments We greatly appreciate the authors in CoNLL’18 paper (Martschat and Markert, 2018) for making their data public. In particular, we wish to thank Sebastian Martschat for his great support in discussions about the experiment setup and reproduction. We also want to thank anonymous reviewers for their invaluable feedback. References James Allan, Rahul Gupta, and Vikas Khandelwal. 2001. Temporal Summaries of New Topics. In Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’01), pages 10–18. Omar Alonso, Michael Gertz, and Ricardo BaezaYates. 2009. Clustering and Exploring Search Results Using Timeline Constructions. In Proceedings of the 18th ACM Conference on Information and Knowledge Management (CIKM ’09), pages 97– 106. Omar Alonso and Kyle Shiells. 2013. Timelines as Summaries of Popular Scheduled Events. In Proceedings of the 22nd International Conference on World Wide Web (WWW ’13), pages 1037–1044. Giang Binh Tran, Mohammad Alrifai, and Dat Quoc Nguyen. 2013. Predicting Relevant News Events for Timeline Summaries. In Proceedings of the 22nd International Conference on World Wide Web (WWW ’13), pages 91–92. David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent Dirichlet Allocation. Journal of Machine Learning Research, 3(Jan):993–1022. Leo Born, Maximilian Bacher, and Katja Markert. 2020. Dataset Reproducibility and IR Methods in Timeline Summarization. In Proceedings of the 12th Language Resources and Evaluation Conference (LREC’20), pages 1763–1771. Xiuying Chen, Zhangming Chan, Shen Gao, MengHsuan Yu, Dongyan Zhao, and Rui Yan. 2019. Learning Towards Abstractive Timeline Summarization. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI19), pages 4939–4945. Hai Leong Chieu and Yoong Keok Lee. 2004. Query Based Event Extraction Along a Timeline. In Proceedings of the 27th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’04), pages 425–432. Yijun Duan, Adam Jatowt, and Masatoshi Yoshikawa. 2020. Comparative Timeline Summarization via Dynamic Affinity-Preserving Random Walk. In Proceedings of the 24th European Conference on Artificial Intelligence (ECAI’20), pages 1778–1785. Brendan J Frey and Delbert Dueck. 2007. Clustering by Passing Messages Between Data Points. Science, 315(5814):972–976. Demian Gholipour Ghalandari and Georgiana Ifrim. 2020. Examining the State-of-the-Art in News Timeline Summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL’20), pages 1322–1334. 386 Jiwei Li and Sujian Li. 2013. Evolutionary Hierarchical Dirichlet Process for Timeline Summarization. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (ACL’13), pages 556–560. Chin-Yew Lin. 2004. ROUGE: A Package for Automatic Evaluation of Summaries. In Proceedings of the 42th Annual Meeting of the Association for Computational Linguistics (ACL’04), pages 74–81. James MacQueen et al. 1967. Some methods for classification and analysis of multivariate observations. In Proceedings of the 5th Berkeley Symposium on Mathematical Statistics and Probability, pages 281– 297. Sebastian Martschat and Katja Markert. 2017. Improving Rouge for Timeline Summarization. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics (EACL’17), pages 285–290. Sebastian Martschat and Katja Markert. 2018. A Temporally Sensitive Submodularity Framework for Timeline Summarization. In Proceedings of the 22nd Conference on Computational Natural Language Learning (CONLL’18), pages 230–240. Kiem-Hieu Nguyen, Xavier Tannier, and Véronique Moriceau. 2014. Ranking Multidocument Event Descriptions for Building Thematic Timelines. In Proceedings of the 25th International Conference on Computational Linguistics (COLING 2014), pages 1208–1217. Arian Pasquali, Ricardo Campos, Alexandre Ribeiro, Brenda Santana, Alípio Jorge, and Adam Jatowt. 2021. TLS-Covid19: A New Annotated Corpus for Timeline Summarization. In Proceedings of the 43rd European Conference on Information Retrieval (ECIR 2021), pages 497 – 512. Arian Pasquali, Vítor Mangaravite, Ricardo Campos, Alípio Mário Jorge, and Adam Jatowt. 2019. Interactive System for Automatically Generating Temporal Narratives. In Proceedings of the 41st European Conference on Information Retrieval (ECIR 2019), pages 251–255. Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence Embeddings using Siamese BERTNetworks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP-IJCNLP 2019), pages 3982–3992. Michael Röder, Andreas Both, and Alexander Hinneburg. 2015. Exploring the Space of Topic Coherence Measures. In Proceedings of the 8th ACM International Conference on Web Search and Data Mining (WSDM ’15), pages 399–408. Peter J Rousseeuw. 1987. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20:53–65. Dafna Shahaf, Carlos Guestrin, and Eric Horvitz. 2012. Trains of Thought: Generating Information Maps. In Proceedings of the 21st International Conference on World Wide Web (WWW ’12), pages 899–908. Julius Steen and Katja Markert. 2019. Abstractive Timeline Summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization (NewSum’19), pages 21–31. Jannik Strötgen and Michael Gertz. 2013. Multilingual and Cross-Domain Temporal Tagging. Language Resources and Evaluation, 47(2):269–298. Satoko Suzuki and Ichiro Kobayashi. 2014. On-line Summarization of Time-Series Documents Using a Graph-Based Algorithm. In Proceedings of the 28th Pacific Asia Conference on Language, Information and Computing (PACLIC’14), pages 470–478. Russell Swan and James Allan. 2000. Automatic Generation of Overview Timelines. In Proceedings of the 23rd Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’00), pages 49–56. Hiroya Takamura, Hikaru Yokono, and Manabu Okumura. 2011. Summarizing a Document Stream. In Proceedings of the 33rd European Conference on Information Retrieval (ECIR 2011), pages 177–188. Giang Tran, Mohammad Alrifai, and Eelco Herder. 2015. Timeline Summarization From Relevant Headlines. In Proceedings of the 37th European Conference on Information Retrieval (ECIR 2015), pages 245–256. Giang Binh Tran, Tuan A Tran, Nam-Khanh Tran, Mohammad Alrifai, and Nattiya Kanhabua. 2013. Leveraging Learning to Rank in an Optimization Framework for Timeline Summarization. In Proceedings of SIGIR 2013 Workshop on Time-aware Information Access (#TAIA’13). William Yang Wang, Yashar Mehdad, Dragomir Radev, and Amanda Stent. 2016. A Low-Rank Approximation Approach to Learning Joint Embeddings of News Stories and Images for Timeline Summarization. In Proceedings of the 15th Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT 2016), pages 58–68. Rui Yan, Liang Kong, Congrui Huang, Xiaojun Wan, Xiaoming Li, and Yan Zhang. 2011a. Timeline Generation Through Evolutionary Trans-temporal Summarization. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing (EMNLP ’11), pages 433–443. Rui Yan, Xiaojun Wan, Jahna Otterbacher, Liang Kong, Xiaoming Li, and Yan Zhang. 2011b. Evolutionary Timeline Summarization: a Balanced Optimization Framework via Iterative Substitution. In Proceedings of the 34th international ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’11), pages 745–754. 387 Xin Wayne Zhao, Yanwei Guo, Rui Yan, Yulan He, and Xiaoming Li. 2013. Timeline Generation with Social Attention. In Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’13), pages 1061–1064.
|
2021
|
32
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4141–4152 August 1–6, 2021. ©2021 Association for Computational Linguistics 4141 Supporting Land Reuse of Former Open Pit Mining Sites using Text Classification and Active Learning Christopher Schr¨oder1,5, Kim B¨urgl1,4,5, Yves Annanias2,5, Andreas Niekler1,5, Lydia M¨uller1,4,5, Daniel Wiegreffe2,5, Christian Bender3,5, Christoph Mengs3,5, Gerik Scheuermann2,4,5, and Gerhard Heyer1,4,5 1Natural Language Processing Group 2Image and Signal Processing Group 3Institute of Public Finance and Public Management 4Institute for Applied Informatics (InfAI), Leipzig, Germany 5Leipzig University, Germany {schroeder,buergl,annanias,aniekler}@informatik.uni-leipzig.de {lydia,daniel,scheuermann,heyer}@informatik.uni-leipzig.de {bender,mengs}@wifa.uni-leipzig.de Abstract Open pit mines left many regions worldwide inhospitable or uninhabitable. Many sites are left behind in a hazardous or contaminated state, show remnants of waste, or have other restrictions imposed upon them, e.g., for the protection of human or nature. Such information has to be permanently managed in order to reuse those areas in the future. In this work we present and evaluate an automated workflow for supporting the post-mining management of former lignite open pit mines in the eastern part of Germany, where prior to any planned land reuse, aforementioned information has to be acquired to ensure the safety and validity of such an endeavor. Usually, this information is found in expert reports, either in the form of paper documents, or in the best case as digitized unstructured text—all of them in German language. However, due to the size and complexity of these documents, any inquiry is tedious and time-consuming, thereby slowing down or even obstructing the reuse of related areas. Since no training data is available, we employ active learning in order to perform multi-label sentence classification for two categories of restrictions and seven categories of topics. The final system integrates optical character recognition (OCR), active-learningbased text classification, and geographic information system visualization in order to effectively extract, query, and visualize this information for any area of interest. Active learning and text classification results are twofold: Whereas the restriction categories were reasonably accurate (>0.85 F1), the seven topicoriented categories seemed to be complex even for human annotators and achieved mediocre evaluation scores (<0.70 F1). 1 Introduction In many parts of the world, raw materials were mined in open pit mines during the last century, leaving many of these regions inhospitable or uninhabitable. To put these regions back into use, entire stretches of land must be renaturalized, which means that land must be ecologically restored with the aim to ultimately increase biodiversity, or recultivated, which means its productivity must be restored, e.g., reused for agriculture, recreational areas, industrial parks, solar and wind farms, or as building land (Luc et al., 2015). In the following, we subsume both renaturalization and recultivation under land reuse. For land reuse, it is essential that all relevant information about the sites is retained, which used to be recorded in the form of textual reports. Such reports include information such as, among others, hazards, soil composition, or environmental factors. Therefore, having access to all these reports, it can be determined if a site can be reused immediately, only under certain conditions, or not at all in the foreseeable future. For reaching a sustainable future, the United Nations (2015) has defined objectives, called sustainable development goals (SDGs). Land reuse is a shared common denominator among several of those goals such as “Zero hunger”, “Clean water and sanitation”, “Sustainable cities and communities”, “Climate action”, “Life below water”, and “Life on land”. Moreover, it provides co-benefit to all SDGs as shown by Herrick et al. (2019) and directly supports “Life on Land”. By implication, anything that obstructs land reuse also impedes the fulfillment of several SDGs. 4142 This work deals with the real-world use case of post-mining management (Kretschmann, 2020) of former lignite open pit mines in the eastern part of Germany. Here, a large number of such documents exist, and moreover, there is metadata maintained, which maps each document to its related area. Apart from that, before any land can be reused in these areas, it is legally required that local authorities must be consulted before proceeding any further. This process includes seeing through numerous legacy documents, which is laborious, timeconsuming and delays a subsequent reuse of such areas. We address this issue by demonstrating and evaluating a workflow consisting of optical character recognition (OCR), text classification and active learning, whose results are then visualized by a Geographic Information System (GIS). By automating information extraction and making extracted results available through a GIS, we increase efficiency by which information about a specific location of interest can be queried. This can accelerate the reuse of land by supporting the efficiency of employees managing these areas, and thereby contributes towards the fulfillment of several SDGs. This necessary review of a multitude of documents, which is obligatory prior to any land reuse, is aggravated even more by Germany’s federal structure (German Federal Government, 2016) due to which land management is a task of the municipalities. The federal government, as well as the states are responsible for the SDGs’ implementation, which is then passed on to the municipalities, which therefore are effectively responsible for supporting SDGs. Municipalities, however, do not have a standardized software infrastructure (Zern-Breuer et al., 2020), which results in a heterogeneous data management landscape and thereby makes the implementation of SDGs challenging, especially for small municipalities. Information about former lignite open pit mines is stored in independent GISes, related unstructured documents are stored in dedicated storage systems (either in form of piles of paper, scanned documents, or even as digitized text), and the connections between documents and geographic coordinates are stored in yet other databases. In order to obtain information about an area of interest, all information must be contextualized, compiled, and manually evaluated. Although the presented approach is tailored towards the post-mining management in Eastern Germany, this is relevant to many other countries in the world, which are also concerned with stopping lignite and coal mining to reduce CO2 emissions. To give a few examples, Belgium performed coal phase-out in 2016, Sweden and Austria in 2020; Canada will follow in 2030, and Germany in 2038. All of these countries will need to post-manage former mining sites in order to reuse the affected areas. Apart from lignite and coal mining, this is also true for other mining sites. Once resources are exhausted or are no longer needed, land has to be renaturalized or recultivated or will stay deserted for unknown time. 2 Foundations and Related Work Sustainability issues have long been politically ignored, but became much more relevant in recent years. As a result, this societal challenge has recently started to get traction in computer science (Gomes et al., 2019) and natural language processing (Conforti et al., 2020), where only few previous works study methods to support SDGs: Conforti et al. (2020) classify user-perceived values on unstructured interview text in order to gather structured data about the people’s subjective values. This is performed in developing countries to increase the success of sustainability projects, each targeted at one or more SDGs, by aligning them to the encountered values, so that the projects will be more likely to be continued by the community after their initial implementation. Similar to us, they also performs sentence classification to support SDGs, however, besides using data from a completely different domain, we perform multi-label classification, use more recent transformer-based models, and integrate additional geospatial information. Pincet et al. (2019) support the automatic classification of SDGs in reporting documents in the form of an official API for the OECD (Organisation for Economic Co-operation and Development), which is responsible for implementing SDGs and monitoring the progress thereof. This clearly shows the problem of an increasing number of documents relevant for implementing SDGs, and also the need for tools to support such processes. There are a variety of OCR engines available, with Tesseract (Smith, 1987) being a good starting point. Tesseract offers a number of pre-processing mechanisms for document images, however, it does not implement the full range of state-of-the-art OCR. Image pre-processing as proposed and implemented by the OCR-D project (Binmakhashen 4143 and Mahmoud, 2019; Neudecker et al., 2019), is beneficial to additionally extend the tool with the latest developments in OCR. In recent years, text classification, like many other fields in natural language processing, has experienced a paradigm shift towards transformerbased models (Vaswani et al., 2017; Devlin et al., 2019), which raised the state-of-the-art results on many tasks. Besides the impressive performance gains, the main advantage of using a pre-trained model is that its performance can be translated to low-data scenarios, which were previously challenging due to deep models overfitting on small data. Transformers, however, have been shown to work well on small data (Ein-Dor et al., 2020; Yuan et al., 2020), and consequently open up new possibilities on previously challenging tasks. Geographic information systems are a common technological choice to visualize spatial data on cartographic maps and have been shown to be invaluable for supporting SDGs (Avtar et al., 2020). Using a GIS, one can combine a database storing textual information with the spatial data to support experts in the decision-making process or to enable the exploration of data. To support renaturalization of a river valley, Matysik and Absalon (2012) used a GIS to analyze hydrological aspects in the area and develop a plan for renaturalization. Similarly to our work, they also combined several layers of features in the GIS. A recent toolbox of the commercial ARCGIS software called LocateXT (ESRI) can connect a larger number of unstructured datasets into a running GIS, however, although it can automatically link information and coordinates from the data, it does not support the extraction and processing of unstructured information and other attributes providing further information. 3 Data We use data from the Lausitzer und Mitteldeutsche Bergbauverwaltungsgesellschaft mbH (LMBV)1, who are responsible for the management and reuse of abandoned mining sites in the eastern part of Germany. For this purpose, they archive and manage all documents related to sites in this area, and issue new documents if required. Moreover, they are obligated to provide reliable information about the managed lands for the public on request. Such requests require, among others, to inform about any restrictions for the specific area, which can be 1https://de.wikipedia.org/wiki/LMBV found in the associated documents. An illustrated example is shown in Figure 1. For research, the LMBV provided us with 31,605 of such documents (16,883 for the region Lausitz2, 14,722 for the region Mitteldeutschland3). The oldest documents date back to the 1960s, but scans were only produced within the last 20 years. The documents encompass several different types, for example, reports, drilling logs, expert opinions, statements, plans, maps, and correspondences. The quality of the scans varies from excellent to fair quality. Moreover, some documents are stored in other digital formats (.doc, .docx, .odf) and others are stored as scanned images. The documents have different origins: They are authored by the companies mining the open pit mine, by the LMBV managing the closed open pit mines, by companies responsible for certain subtasks such as building infrastructure, or by other experts. They include documents from the time when open pit mines were actively mined but also documents created after the mines were closed. Besides the documents, our dataset contains over 30,000 geographic features. These features are described as points, lines, polygons and multipolygons, and can be visualized in a GIS. In addition, these data are provided with additional non-spatial information, such as the geographical affiliation of the documents mentioned. 3.1 Labels In this work, the goal is to find restrictions and topics, which are described in Table 1 and which will be used as labels during text classification. Restrictions are formulated in many different ways, e.g., a specific action may be forbidden, an action may require specific preceding steps to be allowed, or the action may be explicitly allowed under certain circumstances. Moreover, a restriction may refer to certain topics, e.g., restricting a construction method depending on the weather. Regarding topic labels, due to the different types of documents, they vary largely. For example, geotechnical issues can be frequently found in experts’ opinions from geotechnical experts but may also appear in reports, statements or correspondence. Thus, topics are not limited to certain type of document and within one document or even one sentence more than one topic may appear. Likewise, restrictions can be found in 2https://en.wikipedia.org/wiki/Lusatia 3https://en.wikipedia.org/wiki/ Central_Germany_(cultural_area) 4144 Figure 1: Example of a typical documentation. Part of the textual reports are passages about restrictions or prohibitions in the described area. (The background image consists of two photos, one by sludgeulper (left background, CC BY-SA 2.0), and the other by Johannes Kazah (right background, CC BY-SA 2.0). The resulting image changes the originals only by adding overlays (to the front) and is also licensed under the CC BY-SA 2.0 license.) most types of documents and describe known issues with the associated area. Those labels are deductively defined and reflect the requirements of the most frequent requests to the LMBV. Since this label system is specifically defined for this novel approach, no training or pre-labeled data could be provided by the LMBV, but for each label example sentences and common keywords were defined. 3.2 OCR For documents which are not digitized yet, the text is extracted using Tesseract4 and best practices regarding German language from the OCR-D community (Smith, 1987; Neudecker et al., 2019; Binmakhashen and Mahmoud, 2019). The major challenges here were: (1) The vast number of documents make it infeasible to optimize OCR parameters for each document, therefore OCR has to be optimized with regard to the whole collection. (2) There is no manually transcribed evaluation data. (3) The documents are written by humans without any review process making erroneous words or grammar very likely. For these reasons, and because of the many varying document types, investigating OCR quality is impractical and therefore outside the scope of this work. However, we use the built-in Tesseract evaluation procedure to judge the overall quality of the process and apply further filtering to cope with difficult documents and insufficient OCR quality. OCR pre-processing steps for the images included orientation analysis and rotation, resizing of the image (400dpi), denoising, lighting intensity correction, binarization, and deskewing. Light4https://github.com/tesseract-ocr/ tesseract ing intensity correction only improved the result in some cases but worsened the result in others. It is therefore only used if it improves the result based on the confidence score from Tesseract as explained below. Denoising converts the images into grayscales, applies a dilution filter, an erosion filter, and finally, a median blur filter. The quality of the results is measured by evaluating the confidence score as produced by Tesseract which provides a word level confidence score reflecting the OCR quality. We aggregated the word level confidence score to a page level confidence score by averaging over all recognized words, resulting in a score between 0 and 100%, and assigning pages without recognized text a score of 0%. Again, the document quality and layout is very heterogeneous and annotating a test set for the OCR process would lack completeness. We identified 45,141 (Mitteldeutschland) and 35,256 (Lausitz) pages in the document dataset. We accept all pages for our experimental dataset which are evaluated with a confidence score of more than 75%. Without pre-processing only 45% of the pages are detected with a confidence score of at least 75%. The correct pre-processing improved the OCR result to 97% of pages exceeding the defined threshold for the region Lausitz (from 44% to 93% for region Mitteldeutschland, respectively). In 93% (Lausitz) and 83% (Mitteldeutschland) of the pages with a confidence below 75% the original documents do not contain any recognizable text. Hence, the majority of unrecognized or insufficiently recognized documents does not contain proper amounts of text. 3.3 Datasets and Splits In order to obtain both a point of reference for evaluation and an initial set of labeled data for the initial 4145 LABEL DESCRIPTION EXAMPLE Restrictions Prohibition Statements which actively prohibit or restrict actions in general or conditionally. Machines heavier than 30t are forbidden, landslide hazard. Requirement Requirements limit usages and/or are directives what is to be done. The area must be secured with ’no trespassing’signs. Topics Weather Weather-related phenomena, consequences, and protection measures. Shore areas must be avoided during heavy rain. Construction Statements about construction plans, construction sites, or construction procedures. Only one-storey buildings should be placed around the marina. Geotechnics Information related to the ground, e.g., about soil, stability, or slopes. Slopes must be protected against the effects of the weather. Restricted area Indicates a limited accessibility, mostly due to hazards, soil stability, or safety precautions. Always keep a distance of at least 50m to the shore. Planting Plans, reports, or specific details about the type of plant and location of plantings. Native species of bushes must be planted on the slope, to stabilize it against rupture. Environment For renaturalization, it is often strictly regulated where to plant, what, types etc. Forest operations are forbidden during breeding season. Disposal Instructions concerning storage and disposal of (building) materials. Contaminated soil must be cleaned and provably be disposed of. Table 1: Description of restrictions and topics, illustrated by examples (translated from German into English). active learning model, we manually labeled a subset of 2000 sentences. For each label, we defined a set of keywords (see Appendix Table 3), which are used to find sentences in the unlabeled data, that likely belong to that label. This is necessary because of the high ratio of unlabeled to labeled sentences in most documents, i.e., a majority of the sentences in the complete dataset will not have any label assigned. Keywords were used to locate restrictions and prohibition candidates. From this candidate pool, we select candidates for the topic categories utilizing further keyword matching. In doing so, we take a maximum of 150 examples per topic category. If no more than 300 candidates can be found for a topic category, we only include half of them in the candidate list to leave examples of such rare categories for active learning in our unlabeled dataset. Since we want to demonstrate the capabilities of active learning this is a necessary decision. Additionally, we added more than 700 randomly drawn sentences, resulting in a dataset of 2000 sentences in total. This dataset was annotated by three different (non-expert) annotators with the help of a guideline describing each label. We measured the inter-annotator agreement with Krippendorff’s α (Krippendorff, 2011) which resulted in values between 0.91 and 0.7, with Restricted Area as the most agreed label and Construction the least agreed label between annotators. This confirms our observation that labeling in this domain is challenging and needs domain expertise. We combine the annotations of all annotators by majority voting in order to obtain more stable judgments of our non-expert annotators (Nowak and R¨uger, 2010). “Requirement” is the most frequent label, while “Weather” is the least frequent. The true label distribution is unknown, but at least some of the labels seem to occur very rarely. We split the annotated dataset into training (500 samples), validation (500 samples), and test (1000 samples) set using iterative stratification (Sechidis et al., 2011) to preserve the label distribution in all three sets (see Appendix, Table 5). 3.4 Geospatial Connection The linkage of the non-spatial data (i.e., the documents and predicted restriction and topic labels) and the spatial data (e.g., coordinates for certain areas) can be represented as a graph. For this, the documents and the associated areas are expressed as nodes. Then, edges are used to link these document nodes to their corresponding area node. The graph serves as an efficient data structure, which is necessary to make the data available in a GIS and to enable the answering of requests by linking and collecting the required information. The predicted labels can be integrated into the data model with the following procedure: For each topic (see Table 1), an additional node is created carrying the label description as a node property. 4146 Restrictions that have not been classified more precisely by a topic are grouped together under a generic topic node. Then, edges are created for each restriction, pointing from the associated topic node to the document node from which the restriction originated. Additional information about the restrictions is available as attributes of the respective edges. This includes the sentence from which the restriction is derived and the confidence value from the text classification algorithm. These attributes are attached to the edge instead of the document node itself, since a document can lead to several restrictions either related to the same topic or a different topic (e.g., “large installations may not be built” [construction-related] and “may not enter shore areas during heavy rain” [weather, restricted area]). Thus, a document node may be connected to (one or more) topic nodes via several edges that contain more detailed information about restriction and the corresponding sentence. Many queries can be realized with this data structure. For example, it is possible to efficiently query which restrictions exist in the same topic, in the same document, or in the same geographic area, since only the corresponding nodes need to be followed in the data model. This enables, in particular, an exploratory search that incorporates information from existing projects that may be relevant for a given request (see the use case in Section 6). 4 Approach The goal of our approach is to detect restriction and topic labels at the sentence level. Subsequently, we can map predicted labels to geospatial data, which is already available in a structured format. This means, with a process chain of OCR, text classification, and GIS, we can effectively detect the presence or absence of labels at geographic coordinates of interest. In the end, this can be directly used to manage land reuse efforts, thereby supporting aforementioned SDGs. As existing OCR solutions are tried and tested, and the geospatial link is already given, the main challenge of this method is text classification, namely: (1) There is no predefined industrial standard for the labels which are not formally defined but given by some exemplary formulations and keywords provided by the LMBV. Consequently, the definitions are incomplete and new formulation not using the keywords are expected. (2) The documents exhibit a domain-specific, often convoluted, vocabulary. 4.1 Text Pre-processing Since the following text classification depends on the quality of the raw text obtained through the OCR step, which we observed to be rather noisy owing to the structure of some documents, we applied a series of pre-processing steps: We detect word wraps and remove the hyphen, convert line breaks into white space, and finally trim repeated sequences of white space. Subsequently, sentence segmentation was performed using syntok5. In order to filter out sentences which are obviously erroneous, e.g., sentences containing only gibberish words, we filtered all sentences which violate the properties of a valid sentence (Goldhahn et al., 2012). This was achieved by a set of regular expressions and filter rules, which detect improper sentences, e.g., sentences which contain too many special characters, start with a lowercase letter, or are missing a terminal punctuation character. 4.2 Text Classification and Active Learning Using the extracted sentences described in Section 4.1 as input, our goal is to classify restriction and topic labels. In contrast to standard text classification datasets, the LMBV data, like most real-world data, provides no labels. Manually labeling documents, however, is time-consuming and therefore costly, especially when some labels are very rare. For this reason, we use active learning (Lewis and Gale, 1994), which works as follows: In an iterative process the active learner presents unlabeled data to a user, which the user has to label. The purpose of this is to reduce the total labeling effort, by identifying samples that add the most value to the current model. The key for this is the query strategy, which selects examples to be labeled by the user. After labeling the presented samples, a new model is trained, and the loop is repeated, either for a specific number of rounds, or until a stopping criterion is met. We assume the pool-based scenario (Settles, 2010), in which the active learner has access to all unlabeled data. Since no labels are provided, and the percentage of sentences having at least one label is quite small, randomly sampling data is not an option, and AL is the obvious choice. Because it is easier for the human annotator to focus on only a single set of labels during the AL process, the text classification is realized using one independent classifier each for restric5https://github.com/fnl/syntok 4147 tions and topics (see Table 1). As the single labels under both restrictions and topics are not mutually exclusive, we train a multi-hot-encoded multi-label classification for both label sets. 5 Experiments We evaluate multi-label active learning performed by three human annotators, who each train a sentence classification model for classifying restrictions and topics, resulting in two runs per person. 5.1 Pre-processing and Experimental Setup Starting from the initial model, which is trained on the train set (described in Section 3), active learning is performed iteratively: (1) 10 unlabeled sentences are presented to the annotator; (2) The annotator may assign zero, one, or multiple labels per sentence; (3) The newly-assigned labels are added to the train set, and a new model is trained. This process is repeated for 50 iterations. Data We use train, validation and test splits, as defined in Section 4.1, and an unlabeled pool consisting of 312,299 sentences. Query Strategy For the query strategy, which selects the sentences to be labeled, we use predictionentropy-based (Roy and McCallum, 2001) uncertainty sampling (Lewis and Gale, 1994), which selects the most uncertain samples, e.g., in this case those whose predicted class posterior exhibits the highest entropy. Since inference on transformers is computationally expensive, and we aim to keep the waiting times at a minimum, at the beginning of each iteration, we subsample the whole unlabeled pool randomly by selecting 4096 examples (Mukherjee and Awadallah, 2020). Moreover, because the ratio of unlabeled sentences to sentences having at least one label is quite large, we adapt the query strategy to balance classes, by considering the class predictions and sampling evenly over the labels. In case this is not possible, e.g., when there is no prediction for a certain label, we fill the remainder with the remaining most uncertain samples, regardless of the predicted class. 5.2 Model and Training Regarding the classification, we fine-tune the pretrained gbert-base model (Chan et al., 2020), which has 110M parameters and is the best performing German transformer model for text classification at this number of parameters. While there is a larger gbert model available, we opted for the base variant due to its efficiency, which results in lower turnaround times of an AL step for the practitioner. We encode the labels as multi-hot encoded vectors. The model is trained using a softmax binary cross-entropy loss. During each active learning iteration, the previous model is fine-tuned for 40 epochs using a learning rate of 5e−5 on the data that has been labeled to this point. To avoid overfitting, we stop early when the validation loss has not changed for more than 5 epochs. 5.3 Results Table 2 shows the classification scores of aforementioned setting evaluated by three annotators and compared to an automated text classification baseline. The baseline is a gbert-base model F1 B. F1 AL LABEL A1 A2 A3 AVG. RESTRICTIONS Prohibition 0.93 0.96 0.96 0.94 0.95 Requirement 0.84 0.86 0.86 0.84 0.85 MICRO 0.87 0.88 0.88 0.86 0.87 MACRO 0.90 0.91 0.91 0.89 0.90 TOPICS Weather 0.53 0.71 0.73 0.77 0.74 Construction 0.58 0.63 0.64 0.63 0.63 Geotechnics 0.58 0.50 0.54 0.53 0.52 Restr. Area 0.89 0.92 0.91 0.90 0.91 Planting 0.78 0.73 0.69 0.61 0.68 Environment 0.73 0.79 0.77 0.73 0.76 Disposal 0.73 0.72 0.74 0.72 0.73 MICRO 0.70 0.72 0.72 0.70 0.71 MACRO 0.69 0.71 0.72 0.70 0.71 Table 2: Active learning experiments, performed by three human annotators. “AVG.” is the annotator average over all three runs. “F1 AL” shows the final scores, broken down by annotator. “F1 B.” is a text classification baseline that is trained on the initial training set. For each label and annotator, we used McNemar’s test (McNemar, 1947) with α = 0.05 to test for significant change in the predictions compared to the baseline: We report obtained p-values, indicated by an underlined result for p < 0.05, and bold text for p < 0.01. trained on the initial data, i.e., without using AL at all. AL improves overall both micro-F1 and macroF1 for topics by up to 3 percentage points, whereas improvements for restrictions seem marginal. While the overall result improves just slightly, looking at the single labels, we can see consider4148 able changes between plain text classification and active learning. Previously underperforming labels like “Weather” and “Construction” improve on average by 5 to 21 percentage points in F1. Smaller improvements can also be seen for “Restricted Area” and “Environment”, and “Disposal” stays about the same. Unfortunately, “Geotechnics” and “Planting” and also drop in performance by 6 and 10 percentage points respectively. Interestingly, when we compare the difference in the relative quantities of co-occurring labels before and after the AL process, we find that the labeled pool changed notably during AL. We observed that (1) the average number of labels per sentence increases; (2) label co-occurrences shift considerably and some combinations even appear for the first time; (3) every combination of topic labels occurs together in the data, which is not the case for our keyword-bootstrapped train set. (The exact numbers be seen in the Appendix, Figure 4-6). All in all, this indicates that AL is beneficial and improves classification metrics by a small amount, and moreover, many samples with previously rarely or even unseen label combinations are found. Apparently, as these notable changes only lead to a small difference in F1, this new value of having more diverse label combinations is difficult to measure here against our keyword-bootstrapped test set. The only solution to a more representative test set, however, would require massive annotation efforts, since labels may be very sparse. 6 Visualization and Interaction Use Case As an example, we present a workflow regarding areas which may not be entered during heavy rains for safety reasons. To answer a request (see Section 3), which e.g., is asking if a specific area may be entered, the expert uses the GIS, centers the map on the corresponding area, and displays the associated features (e.g., active dismantling areas, see Figure 2 A). To enable the expert to analyze the different feature categories, the displayed features are colored by category as suggested by Ware (2012). Since areas can overlap in the map display, all features are colored only semi-transparently. Information immediately prohibiting certain activities is identified by clicking on a feature, which displays the non-spatial data in an information panel (Figure 2 A-B). To keep the expert’s overview of the selected features, they are represented with a striped texture. All restrictions that result from the documents linked to the selected feature are Figure 2: (A) Two geographic features of type ”active dismantling” are displayed on a map. One feature was selected by mouse click (orange striped texture). The weather map is shown as isobands, with precipitation values represented by shades of blue (dark blue tones indicate areas with high precipitation values). The information panel is displayed on the right hand side. It contains the non-spatial data of the selected geographic feature (B), as well as the usage restrictions together with a list of other features with similar usage restrictions (C). 4149 listed (Figure 2 C). The entries are grouped by the restriction type and sorted by a confidence value (Figure 2 C1 and C2). The document title and the sentence from which the restriction is derived are indicated. A click on the title opens a new window for reading the document. This list provides the expert with direct feedback on which documents might be relevant and, without reading them completely, an overview of which usage restrictions are present. This information is crucial for the experts, as it can have a significant impact on planned projects and their planning time. Additionally, the area described by a document can be superimposed with weather data. In this way, decisions regarding conditional restrictions (e.g., “may not enter shore areas during heavy rain”) can also be made more quickly and directly on the basis of the system. The selected features overlap with a heavy rain area represented by isobands, therefore the request is directly answered and access to the area is currently prohibited (Figure 2 A). However, for other restrictions, more information may be necessary, because experts often compare the region of interest with similar regions. Therefore, we provide a filter to highlight all features within the same restriction topic (Figure 2 C). By analyzing similar regions, the expert can derive recommendations for action, which might be necessary for the land reuse of an area. Recommendations for possible usage restrictions can also be derived in this way. Furthermore, this comparison can prevent actions from not being taken or from being taken too late, because the current information does not make them appear necessary, but it is clear from similar projects that they may nevertheless become necessary. This leads to a safe and quick reuse of regions maintained in that manner since precautions can be taken in advance. 7 Conclusions and Future Work In this work, we have presented and evaluated a system which automates information requests related to the post-management of former open pit mines by leveraging unstructured and geospatial data. We used active learning for multi-label text classification to extract restrictions and topics from unstructured text in legacy documents and visualized the results using a GIS. As a result, targeted queries about restrictions and topics at specific geographic locations can be obtained much more efficiently, thereby speeding up the process of land reuse, which directly contributes to several SDGs. Further research is needed to shift recall towards 100% to minimize false negatives, then correcting false positives in the system. Acknowledgments We thank the anonymous reviewers for their valuable and constructive feedback. We also thank the LMBV for many interesting and fruitful discussions. This research was partially funded by the Development Bank of Saxony (SAB) under project numbers 100335729 and 100400221. Ethical Considerations This work presents a workflow for the automatic information extraction in reports related to mining, construction and nature conservation. The collected information represents issues such as access restrictions or hazards. We are aware that misclassification in the application can lead to people being endangered or prevented from entering these regions for no reason. Misuse cannot be ruled out, but currently no specific example is known. To ensure that misclassifications do not impact the stakeholders of the application, a quality assurance process will be used in the operating company so that employees in the piloting phase manually check where errors or information losses can be detected. In addition, there will be quality assurance for the application so that the probability of missing restrictions is minimized. Furthermore, our results could in theory lead to a decline in employees needed to read and check old documents, possibly resulting in job losses. Our scenario, however, requires specialists, who are not easily replaceable. References Ram Avtar, Ridhika Aggarwal, Ali Kharrazi, Pankaj Kumar, and Tonni Agustiono Kurniawan. 2020. Utilizing geospatial information to implement SDGs and monitor their progress. Environmental Monitoring and Assessment, 192:1–22. Galal M. Binmakhashen and Sabri A. Mahmoud. 2019. Document layout analysis: A comprehensive survey. ACM Comput. Surv., 52(6). Branden Chan, Stefan Schweter, and Timo M¨oller. 2020. German’s next language model. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6788–6796, Barcelona, Spain (Online). International Committee on Computational Linguistics. 4150 Costanza Conforti, Stephanie Hirmer, Dai Morgan, Marco Basaldella, and Yau Ben Or. 2020. Natural language processing for achieving sustainable development: the case of neural labelling to enhance community profiling. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8427–8444, Online. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Liat Ein-Dor, Alon Halfon, Ariel Gera, Eyal Shnarch, Lena Dankin, Leshem Choshen, Marina Danilevsky, Ranit Aharonov, Yoav Katz, and Noam Slonim. 2020. Active Learning for BERT: An Empirical Study. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7949–7962, Online. Association for Computational Linguistics. ESRI. Entity Extraction Software Unstructured Data Analysis ArcGIS LocateXT. https://www.esri.com/en-us/arcgis/ products/locatext/overview, (last accessed on 01/21/2021). German Federal Government. 2016. German sustainable development strategy. https://www.bundesregierung.de/ resource/blob/998220/455740/ 7d1716e5d5576bec62c9d16ca908e80e/ 2017-06-20-langfassung-n-en-data.pdf, (last accessed on 05/27/2021). Dirk Goldhahn, Thomas Eckart, and Uwe Quasthoff. 2012. Building large monolingual dictionaries at the Leipzig Corpora Collection: From 100 to 200 languages. In Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 759–765, Istanbul, Turkey. European Language Resources Association (ELRA). Carla Gomes, Thomas Dietterich, Christopher Barrett, Jon Conrad, Bistra Dilkina, Stefano Ermon, Fei Fang, Andrew Farnsworth, Alan Fern, Xiaoli Fern, Daniel Fink, Douglas Fisher, Alexander Flecker, Daniel Freund, Angela Fuller, John Gregoire, John Hopcroft, Steve Kelling, Zico Kolter, Warren Powell, Nicole Sintov, John Selker, Bart Selman, Daniel Sheldon, David Shmoys, Milind Tambe, WengKeen Wong, Christopher Wood, Xiaojian Wu, Yexiang Xue, Amulya Yadav, Abdul-Aziz Yakubu, and Mary Lou Zeeman. 2019. Computational sustainability: Computing for a better world and a sustainable future. Commun. ACM, 62(9):56–65. Jeffrey E. Herrick, Tanya Abrahamse, Purushothaman C. Abhilash, Saleem H. Ali, Porfirio Alvarez-Torres, Aliyu S. Barau, Cristina Branquinho, Ashwini Chhatre, Jean-Luc Chotte, and Graham P. Von Maltitz. 2019. Land restoration for achieving the sustainable development goals: An international resource panel think piece. United Nations Environment Programme. J¨urgen Kretschmann. 2020. Post-mining—a holistic approach. Mining, Metallurgy & Exploration, 37(5):1401–1409. Klaus Krippendorff. 2011. Agreement and information in the reliability of coding. Communication Methods and Measures, 5(2):93–112. David D. Lewis and William A. Gale. 1994. A sequential algorithm for training text classifiers. In SIGIR’94, pages 3–12. Springer. M. Luc, U. Somorowska, and J.B. Szma´nda, editors. 2015. Landscape Analysis and Planning. Springer International Publishing. Magdalena Matysik and Damian Absalon. 2012. Renaturization plan for a river valley subject to high human impact-hydrological aspects. Polish Journal of Environmental Studies, 21(2). Quinn McNemar. 1947. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12(2):153–157. Subhabrata Mukherjee and Ahmed Awadallah. 2020. Uncertainty-aware Self-training for Few-shot Text Classification. In Advances in Neural Information Processing Systems, volume 33, pages 21199– 21212. Curran Associates, Inc. Clemens Neudecker, Konstantin Baierer, Maria Federbusch, Matthias Boenig, Kay-Michael W¨urzner, Volker Hartmann, and Elisa Herrmann. 2019. OCRD: An end-to-end open source OCR framework for historical printed documents. In Proceedings of the 3rd International Conference on Digital Access to Textual Cultural Heritage, DATeCH2019, page 53–58, New York, NY, USA. Association for Computing Machinery. Stefanie Nowak and Stefan R¨uger. 2010. How reliable are annotations via crowdsourcing: a study about inter-annotator agreement for multi-label image annotation. In Proceedings of the international conference on Multimedia information retrieval, pages 557–566. Arnaud Pincet, Shu Okabe, and Martin Pawelczyk. 2019. Linking aid to the sustainable development goals – a machine learning approach. (52). Nicholas Roy and Andrew McCallum. 2001. Toward optimal active learning through sampling estimation of error reduction. In Proceedings of the Eighteenth International Conference on Machine Learning, ICML 01, pages 441–448. Morgan Kaufmann Publishers Inc. 4151 Konstantinos Sechidis, Grigorios Tsoumakas, and Ioannis Vlahavas. 2011. On the stratification of multilabel data. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 145–158. Springer. Burr Settles. 2010. Active learning literature survey. Technical report, University of Wisconsin-Madison Department of Computer Sciences. Raymond Wensley Smith. 1987. The extraction and recognition of text from multimedia document images. Ph.D. thesis, University of Bristol. United Nations. 2015. Sustainable development goals. https://sdgs.un.org/goals, (last accessed on 05/27/2021). Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998–6008. Curran Associates, Inc. Colin Ware. 2012. Information Visualization: Perception for Design, 3 edition. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA. Michelle Yuan, Hsuan-Tien Lin, and Jordan BoydGraber. 2020. Cold-start active learning through self-supervised language modeling. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7935–7948, Online. Association for Computational Linguistics. Rubina Zern-Breuer, Margrit Seckelmann, Nora Reg¨os, Heinrich Lorei, Kathrin Annika Kruse, and Marco Brunzel. 2020. Voruntersuchung zur Einf¨uhrung eines einheitlichen Geodatenmanagements in Rheinland-Pfalz. Projekt rlp-GDM – Projektbericht. Speyrer Arbeitshefte, (245):1–94. A Data A.1 Keywords For each label, the keywords used to create the dataset are shown in Table 3. A.2 Inter-Annotator Agreement In Table 4 we report Krippendorff’s α and Fleiss’ κ for three human annotators. A.3 Absolute Label Occurrences Table 5 shows the absolute label distribution and Figure 3 shows the co-occurrence among labels. Figure 3: Label co-occurrences. A.4 Relative Label Co-occurrence We show the relative label co-occurrence for the initial labeled set in Table 4 (normalized per row). On the other hand, Table 5 shows the relative labels co-occurrence of the samples selected by the query strategy. The difference between those two Figures is shown by Figure 6. Figure 4: Samples found by keyword matching (labeled data). 4152 Figure 5: Samples found by the query strategy. Figure 6: Difference between the labeled data (after active learning) and the initial set. LABEL KRIPPENDORFF’S α FLEISS’ κ Prohibition 0.8988 0.8995 Requirement 0.8303 0.8317 Weather 0.7400 0.7506 Construction 0.6991 0.7010 Geotechnics 0.7095 0.7150 Restricted area 0.9140 0.9143 Planting 0.7579 0.7653 Environment 0.7542 0.7555 Disposal 0.8118 0.8138 Table 4: Krippendorff’s alpha and Fleiss’ kappa for each label, each sample in the dataset was annotated by three different annotators. LABEL TRAIN TEST VAL TOTAL Prohibition 47 93 47 187 Requirement 149 299 149 597 Weather 17 34 17 68 Construction 84 168 84 336 Geotechnics 34 68 34 136 Restricted area 69 136 68 273 Planting 23 47 24 94 Environment 68 135 68 271 Disposal 34 69 35 138 Table 5: Label distribution in the train-, test-, and validation data set LABEL KEYWORDS ENGLISH TRANSLATION Restrictions Prohibition ’verboten’, ’nicht gestattet’, ’nicht erlaubt’, ’untersagt’, ’unbefugt’, ’darf nicht’ ’not permitted’, ’not allowed’, ’banned’, ’unauthorized’, ’may not’ Requirement ’m¨ussen’, ’muss’, ’darf’, ’nur’, ’maximal’, ’beachten’ ’must’, ’must’(inflected), ’may’, ’only’, ’at most’, ’consider’ Topics Weather ’Nebel’, ’Wetter’, ’Sturm’, ’Starkniederschlag’, ’Frost’, ’Trockenheit’, ’Regen’, ’Schnee’, ’Temperatur’ ’fog’, ’weather’, ’storm’, ’heavy rainfall’, ’frost’, ’drought’, ’rain’, ’snow’, ’temperature’ Construction ’Bebauung’, ’¨uberbauung’, ’errichten’, ’Fenster’, ’Mauer’ ’Construction’, ’build on’, ’construct’, ’window’, ’wall’ Geotechnics ’geotechnsch’, ’Gel¨ande’, ’Risse’, ’Absenkung’, ’Boden’, ’Sohle’ ’geotechnical’, ’terrain’, ’crack’, ’sinking’, ’soil’, ’horizon’ Restricted area ’Aufenthalt’, ’Uferseitig’, ’betreten’, ’befahren’, ’anlegen’ ’stay’, ’shore-sided’, ’enter’, ’drive on’, ’dock’ Planting ’B¨aume’, ’Baum’, ’Pflanzen’, ’f¨allen’, ’forst’ ’trees’, ’tree’, ’ plants’, ’chop’, ’forest’ Environment ’Nester’, ’Arten’, ’Umwelt’, ’gesch¨utzt’ ’nests’, ’species’, ’environment’, ’protected’ Disposal ’lager’, ’entsorg’, ’abfall’, ’verbringen’, ’verklappen’ ’store’, ’disposal’, ’waste’, ’remove’, ’dumping’ Table 3: Keywords used for dataset generation (in German) and their English translation.
|
2021
|
320
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4153–4169 August 1–6, 2021. ©2021 Association for Computational Linguistics 4153 Reliability Testing for Natural Language Processing Systems Samson Tan§♮∗ Shafiq Joty§‡ Kathy Baxter§ Araz Taeihagh♦♣ Gregory A. Bennett§ Min-Yen Kan♮ §Salesforce Research ‡Nanyang Technological University ♮School of Computing, National University of Singapore ♦Lee Kuan Yew School of Public Policy, National University of Singapore ♣Centre for Trusted Internet and Community, National University of Singapore Abstract Questions of fairness, robustness, and transparency are paramount to address before deploying NLP systems. Central to these concerns is the question of reliability: Can NLP systems reliably treat different demographics fairly and function correctly in diverse and noisy environments? To address this, we argue for the need for reliability testing and contextualize it among existing work on improving accountability. We show how adversarial attacks can be reframed for this goal, via a framework for developing reliability tests. We argue that reliability testing — with an emphasis on interdisciplinary collaboration — will enable rigorous and targeted testing, and aid in the enactment and enforcement of industry standards. 1 Introduction Rigorous testing is critical to ensuring a program works as intended (functionality) when used under real-world conditions (reliability). Hence, it is troubling that while natural language technologies are becoming increasingly pervasive in our everyday lives, there is little assurance that these NLP systems will not fail catastrophically or amplify discrimination against minority demographics when exposed to input from outside the training distribution. Recent examples include GPT-3 (Brown et al., 2020) agreeing with suggested suicide (Rousseau et al., 2020), the mistranslation of an innocuous social media post resulting in a minority’s arrest (Hern, 2017), and biased grading algorithms that can negatively impact a minority student’s future (Feathers, 2019). Additionally, a lack of rigorous testing, coupled with machine learning’s (ML) implicit assumption of identical training and testing distributions, may inadvertently result in systems that discriminate against minorities, who are often underrepresented in the training data. This can take ∗Correspondence to: [email protected] Figure 1: How DOCTOR can integrate with existing system development workflows. Test (left) and system development (right) take place in parallel, separate teams. Reliability tests can thus be constructed independent of the system development team, either by an internal “red team” or by independent auditors. the form of misrepresentation of or poorer performance for people with disabilities, specific gender, ethnic, age, or linguistic groups (Hovy and Spruit, 2016; Crawford, 2017; Hutchinson et al., 2020). Amongst claims of NLP systems achieving human parity in challenging tasks such as question answering (Yu et al., 2018), machine translation (Hassan et al., 2018), and commonsense inference (Devlin et al., 2019), research has demonstrated these systems’ fragility to natural and adversarial noise (Goodfellow et al., 2015; Belinkov and Bisk, 2018) and out-of-distribution data (Fisch et al., 2019). It is also still common practice to equate “testing” with “measuring held-out accuracy”, even as datasets are revealed to be harmfully biased (Wagner et al., 2015; Geva et al., 2019; Sap et al., 2019). Many potential harms can be mitigated by detecting them early and preventing the offending model from being put into production. Hence, in addition to being mindful of the biases in the NLP pipeline (Bender and Friedman, 2018; Mitchell et al., 2019; 4154 Waseem et al., 2021) and holding creators accountable via audits (Raji et al., 2020; Brundage et al., 2020), we argue for the need to evaluate an NLP system’s reliability in diverse operating conditions. Initial research on evaluating out-of-distribution generalization involved manually-designed challenge sets (Jia and Liang, 2017; Nie et al., 2020; Gardner et al., 2020), counterfactuals (Kaushik et al., 2019; Khashabi et al., 2020; Wu et al., 2021), biased sampling (Søgaard et al., 2021) or toolkits for testing if a system has specific capabilities (Ribeiro et al., 2020) or robustness to distribution shifts (Goel et al., 2021). However, most of these approaches inevitably overestimate a given system’s worst-case performance since they do not mimic the NLP system’s adversarial distribution1. A promising technique for evaluating worst-case performance is the adversarial attack. However, although some adversarial attacks explicitly focus on specific linguistic levels of analysis (Belinkov and Bisk, 2018; Iyyer et al., 2018; Tan et al., 2020; Eger and Benz, 2020), many often simply rely on word embeddings or language models for perturbation proposal (see §4). While the latter may be useful to evaluate a system’s robustness to malicious actors, they are less useful for dimension-specific testing (e.g., reliability when encountering grammatical variation). This is because they often perturb the input across multiple dimensions at once, which may make the resulting adversaries unnatural. Hence, in this paper targeted at NLP researchers, practitioners, and policymakers, we make the case for reliability testing and reformulate adversarial attacks as dimension-specific, worst-case tests that can be used to approximate real-world variation. We contribute a reliability testing framework — DOCTOR — that translates safety and fairness concerns around NLP systems into quantitative tests. We demonstrate how testing dimensions for DOCTOR can be drafted for a specific use case. Finally, we discuss the policy implications, challenges, and directions for future research on reliability testing. 2 Terminology Definitions Let’s define key terms to be used in our discussion. NLP system. The entire text processing pipeline built to solve a specific task; taking raw text as input and producing predictions in the form of labels 1The distribution of adversarial cases or failure profile. (classification) or text (generation). We exclude raw language models from the discussion since it is unclear how performance, and hence worst-case performance, should be evaluated. We do include NLP systems that use language models internally (e.g., BERT-based classifiers (Devlin et al., 2019)). Reliability. Defined by IEEE (2017) as the “degree to which a system, product or component performs specified functions under specified conditions for a specified period of time”. We prefer this term over robustness2 to challenge the NLP community’s common framing of inputs from outside the training distribution as “noisy”. The notion of reliability requires us to explicitly consider the specific, diverse environments (i.e., communities) a system will operate in. This is crucial to reducing the NLP’s negative impact on the underrepresented. Dimension. An axis along which variation can occur in the real world, similar to Plank (2016)’s variety space. A taxonomy of possible dimensions can be found in Table 1 (Appendix). Adversarial attack. A method of perturbing the input to degrade a target model’s accuracy (Goodfellow et al., 2015). In computer vision, this is achieved by adding adversarial noise to the image, optimized to be maximally damaging to the model. §4 describes how this is done in the NLP context. Stakeholder. A person who is (in-)directly impacted by the NLP system’s predictions. Actor. Someone who has influence over a) the design of an NLP system and its reliability testing regime; b) whether the system is deployed; and c) who it can interact with. Within the context of our discussion, actors are likely to be regulators, experts, and stakeholder advocates. Expert. An actor who has specialized knowledge, such as ethicists, linguists, domain experts, social scientists, or NLP practitioners. 3 The Case for Reliability Testing in NLP The accelerating interest in building NLP-based products that impact many lives has led to urgent questions of fairness, safety, and accountability (Hovy and Spruit, 2016; Bender et al., 2021), 2The “degree to which a system or component can function correctly in the presence of invalid inputs or stressful environmental conditions” (IEEE, 2017). 4155 prompting research into algorithmic bias (Bolukbasi et al., 2016; Blodgett et al., 2020), explainability (Ribeiro et al., 2016; Danilevsky et al., 2020), robustness (Jia and Liang, 2017), etc. Research is also emerging on best practices for productizing ML: from detailed dataset documentation (Bender and Friedman, 2018; Gebru et al., 2018), model documentation for highlighting important but often unreported details such as its training data, intended use, and caveats (Mitchell et al., 2019), and documentation best practices (Partnership on AI, 2019), to institutional mechanisms such as auditing (Raji et al., 2020) to enforce accountability and red-teaming (Brundage et al., 2020) to address developer blind spots, not to mention studies on the impact of organizational structures on responsible AI initiatives (Rakova et al., 2020). Calls for increased accountability and transparency are gaining traction among governments (116th U.S. Congress, 2019; NIST, 2019; European Commission, 2020; Smith, 2020; California State Legislature, 2020; FDA, 2021) and customers increasingly cite ethical concerns as a reason for not engaging AI service providers (EIU, 2020). While there has been significant discussion around best practices for dataset and model creation, work to ensure NLP systems are evaluated in a manner representative of their operational conditions has only just begun. Initial work in constructing representative tests focuses on enabling development teams to easily evaluate their models’ linguistic capabilities (Ribeiro et al., 2020) and accuracy on subpopulations and distribution shifts (Goel et al., 2021). However, there is a clear need for a paradigm that allows experts and stakeholder advocates to collaboratively develop tests that are representative of the practical and ethical concerns of an NLP system’s target demographic. We argue that reliability testing, by reframing the concept of adversarial attacks, has the potential to fill this gap. 3.1 What is reliability testing? Despite the recent advances in neural architectures resulting in breakthrough performance on benchmark datasets, research into adversarial examples and out-of-distribution generalization has found ML systems to be particularly vulnerable to slight perturbations in the input (Goodfellow et al., 2015) and natural distribution shifts (Fisch et al., 2019). While these perturbations are often chosen to maximize model failure, they highlight serious reliability issues for putting ML models into production since they show that these models could fail catastrophically in naturally noisy, diverse, real-world environments (Saria and Subbaswamy, 2019). Additionally, bias can seep into the system at multiple stages of the NLP lifecycle (Shah et al., 2020), resulting in discrimination against minority groups (O’Neil, 2016). The good news, however, is that rigorous testing can help to highlight potential issues before the systems are deployed. The need for rigorous testing in NLP is reflected in ACL 2020 giving the Best Paper Award to CheckList (Ribeiro et al., 2020), which applied the idea of behavior testing from software engineering to testing NLP systems. While invaluable as a first step towards the development of comprehensive testing methodology, the current implementation of CheckList may still overestimate the reliability of NLP systems since the individual test examples are largely manually constructed. Importantly, with the complexity and scale of current models, humans cannot accurately determine a model’s adversarial distribution (i.e., the examples that cause model failure). Consequently, the test examples they construct are unlikely to be the worst-case examples for the model. Automated assistance is needed. Therefore, we propose to perform reliability testing, which can be thought of as one component of behavior testing. We categorize reliability tests as average-case tests or the worst-case tests. As their names suggest, average-case and worst-case tests estimate the expected and lower-bound performance, respectively, when the NLP system is exposed to the phenomena modeled by the tests. Average-case tests are conceptually similar to Wu et al. (2021)’s counterfactuals, which is contemporaneous work, while worst-case tests are most similar to adversarial attacks (§4). Our approach parallels boundary value testing in software engineering: In boundary value testing, tests evaluate a program’s ability to handle edge cases using test examples drawn from the extremes of the ranges the program is expected to handle. Similarly, reliability testing aims to quantify the system’s reliability under diverse and potentially extreme conditions. This allows teams to perform better quality control of their NLP systems and introduce more nuance into discussions of why and when models fail (§5). Finally, we note that reliabil4156 ity testing and standards are established practices in engineering industries (e.g., aerospace (Nelson, 2003; Wilkinson et al., 2016)) and advocate for NL engineering to be at parity with these fields. 3.2 Evaluating worst-case performance in a label-scarce world A proposed approach for testing robustness to natural and adverse distribution shifts is to construct test sets using data from different domains or writing styles (Miller et al., 2020; Hendrycks et al., 2020), or to use a human vs. model method of constructing challenge sets (Nie et al., 2020; Zhang et al., 2019b). While they are the gold standard, such datasets are expensive to construct,3 making it infeasible to manually create worst-case test examples for each NLP system being evaluated. Consequently, these challenge sets necessarily overestimate each system’s worst-case performance when the inference distribution differs from the training one. Additionally, due to their crowdsourced nature, these challenge sets inevitably introduce distribution shifts across multiple dimensions at once, and even their own biases (Geva et al., 2019), unless explicitly controlled for. Building individual challenge sets for each dimension would be prohibitively expensive due to combinatorial explosion, even before having to account for concept drift (Widmer and Kubat, 1996). This coupling complicates efforts to design a nuanced and comprehensive testing regime. Hence, simulating variation in a controlled manner via reliability tests can be a complementary method of evaluating the system’s out-of-distribution generalization ability. 4 Adversarial Attacks as Reliability Tests We first give a brief introduction to adversarial attacks in NLP before showing how they can be used for reliability testing. We refer the reader to Zhang et al. (2020b) for a comprehensive survey. Existing work on NLP adversarial attacks perturbs the input at various levels of linguistic analysis: phonology (Eger and Benz, 2020), orthography (Ebrahimi et al., 2018), morphology (Tan et al., 2020), lexicon (Alzantot et al., 2018; Jin et al., 2020), and syntax (Iyyer et al., 2018). Early work did not place any constraints on the attacks and merely used the degradation to a tar3Dua et al. (2019) reports a cost of 60k USD for 96k question–answer pairs. Algorithm 1 General Reliability Test Require: Data distribution Dd = {X, Y} modeling the dimension of interest d, NLP system M, Source dataset X ∼X, Desired labels Y ′ ∼Y, Scoring function S. Ensure: Average- or worst-case examples X′, Result r. 1: X′ ←{∅}, r ←0 2: for x, y′ in X, Y ′ do 3: C ←SAMPLECANDIDATES(X) 4: switch TestType do 5: case AverageCaseTest 6: s ←MEAN(S(y′, M(C))) 7: X′ ←X′ ∪C 8: case WorstCaseTest 9: x′, s ←arg minxc∈C S(y′, M(xc)) 10: X′ ←X′ ∪{x′} 11: r ←r + s 12: end for 13: r ← r |X| 14: return X′, r get model’s accuracy as the measure of success. However, this often resulted in the semantics and expected prediction changing, leading to an overestimation of the attack’s success. Recent attacks aim to preserve the original input’s semantics. A popular approach has been to substitute words with their synonyms using word embeddings or a language model as a measure of semantic similarity (Alzantot et al., 2018; Ribeiro et al., 2018; Michel et al., 2019; Ren et al., 2019; Zhang et al., 2019a; Li et al., 2019; Jin et al., 2020; Garg and Ramakrishnan, 2020; Li et al., 2020a). Focusing on maximally degrading model accuracy overlooks the key feature of adversarial attacks: the ability to find the worst-case example for a model from an arbitrary distribution. Many recent attacks perturb the input across multiple dimensions at once, which may make the result unnatural. By constraining our sample perturbations to a distribution modeling a specific dimension of interest, the performance on the generated adversaries is a valid lower bound performance for that dimension. Said another way, adversarial attacks can be reframed as interpretable reliability tests if we constrain them to meaningful distributions. This is the key element of our approach as detailed in Alg. 1. We specify either an average (Lines 5–7) or worse case test (Lines 8–10), but conditioned on the data distribution D that models a particular dimension of interest d. The resultant reliability score gauges real-world performance and the worstcase variant returns the adversarial examples that cause worst-case performance. When invariance to input variation is expected, y′ is equivalent to the 4157 source label y. Note that by ignoring the averagecase test logic and removing d, we recover the general adversarial attack algorithm. However, the key difference between an adversarial robustness mindset and a testing one is the latter’s emphasis on identifying ways in which natural phenomena or ethical concerns can be operationalized as reliability tests. This change in perspective opens up new avenues for interdisciplinary research that will allow researchers and practitioners to have more nuanced discussions about model reliability and can be used to design comprehensive reliability testing regimes. We describe such a framework for interdisciplinary collaboration next. 5 A Framework for Reliability Testing We introduce and then describe our general framework, DOCTOR, for testing the reliability of NLP systems. DOCTOR comprises six steps: 1. Define reliability requirements 2. Operationalize dimensions as distributions 3. Construct tests 4. Test system and report results 5. Observe deployed system’s behavior 6. Refine reliability requirements and tests Defining reliability requirements. Before any tests are constructed, experts and stakeholder advocates should work together to understand the demographics and values of the communities the NLP system will interact with (Friedman and Hendry, 2019) and the system’s impact on their lives. The latter is also known as algorithmic risk assessment (Ada Lovelace Institute and DataKind UK, 2021). There are three critical questions to address: 1) Along what dimensions should the model be tested? 2) What metrics should be used to measure system performance? 3) What are acceptable performance thresholds for each dimension? Question 1 can be further broken down into: a) general linguistic phenomena, such as alternative spellings or code-mixing; b) task-specific quirks, e.g., an essay grading system should not use text length to predict score; c) sensitive attributes, such as gender, ethnicity, sexual orientation, age, or disability status. This presents an opportunity for interdisciplinary expert collaboration: Linguists are best equipped to contribute to discussions around (a), domain experts to (b), and ethicists and social scientists to (c). However, we recognize that such collaboration may not be feasible for every NLP system being tested. It is more realistic to expect ethicists to be involved when applying DOCTOR at the company and industry levels, and ethics-trained NLP practitioners to answer these questions within the development team. We provide a taxonomy of potential dimensions in Table 1 (Appendix). Since it is likely unfeasible to test every possible dimension, stakeholder advocates should be involved to ensure their values and interests are accurately represented and prioritized (Hagerty and Rubinov, 2019), while experts should ensure the dimensions identified can be feasibly tested. A similar approach to that of community juries4 may be taken. We recommend using this question to evaluate the feasibility of operationalizing potential dimensions: “What is the system’s performance when exposed to variation along dimension d?”. For example, rather than simply “gender”, a better-defined dimension would be “gender pronouns”. With this understanding, experts and policymakers can then create a set of reliability requirements, comprising the testing dimensions, performance metric(s), and passing thresholds. Next, we recommend using the same metrics for held-out, average-case, and worst-case performance for easy comparison. These often vary from task to task and are still a subject of active research (Novikova et al., 2017; Reiter, 2018; Kryscinski et al., 2019), hence the question of the right metric to use is beyond the scope of this paper. Finally, ethicists, in consultation with the other aforementioned experts and stakeholders, will determine acceptable thresholds for worst-case performance. The system under test must perform above said thresholds when exposed to variation along those dimensions in order to pass. For worst-case performance, we recommend reporting thresholds as relative differences (δ) between the average-case and worst-case performance. These questions may help in applying this step and deciding if specific NLP solutions should even exist (Leins et al., 2020): • Who will interact with the NLP system, in what context, and using which language varieties? • What are the distinguishing features of these varieties compared to those used for training? 4docs.microsoft.com/en-us/azure/.../community-jury 4158 • What is the (short- and long-term) impact on the community’s most underrepresented members if the system performs more poorly for them? We note that our framework is general enough to be applied at various levels of organization: within the development team, within the company (compliance team, internal auditor), and within the industry (self-regulation or independent regulator). However, we expect the exact set of dimensions, metrics and acceptable thresholds defined in Step 1 to vary depending on the reliability concerns of the actors at each level. For example, independent regulators will be most concerned with establishing minimum safety and fairness standards that all NLP systems used in their industries must meet, while compliance teams may wish to have stricter and more comprehensive standards for brand reasons. Developers can use DOCTOR to meet the other two levels of requirements and understand their system’s behaviour better with targeted testing. Operationalizing dimensions. While the abstractness of dimensions allows people who are not NLP practitioners to participate in drafting the set of reliability requirements, there is no way to test NLP systems using fuzzy concepts. Therefore, every dimension the system is to be tested along must be operationalizable as a distribution from which perturbed examples can be sampled in order for NLP practitioners to realize them as tests. Since average-case tests attempt to estimate a system’s expected performance in its deployed environment, the availability of datasets that reflect real-world distributions is paramount to ensure that the tests themselves are unbiased. This is less of an issue for worst-case tests; the tests only needs to know which perturbations that are possible, but not how frequently they occur in the real world. Figuring out key dimensions for different classes of NLP tasks and exploring ways of operationalizing them as reliability tests are also promising directions for future research. Such research would help NLP practitioners and policymakers define reliability requirements that can be feasibly implemented. Constructing tests. Next, average- and worstcase tests are constructed (Alg. 1). Average-case tests can be data-driven and could take the form of manually curated datasets or model-based perturbation generation (e.g., PolyJuice (Wu et al., 2021)), while worst-case tests can be rule-based (e.g., Morpheus (Tan et al., 2020)) or model-based (e.g., BERT-Attack (Li et al., 2020a)). We recommend constructing tests that do not require access to the NLP model’s parameters (black-box assumption); this not only yields more system-agnostic tests, but also allows for (some) tests to be created independently from the system development team. If the black-box assumption proves limiting, the community can establish a standard set of items an NLP system should export for testing purposes, e.g., network gradients if the system uses a neural model. Regardless of assumption, keeping the regulators’ test implementations separate and hidden from the system developers is critical for stakeholders and regulators to trust the results. This separation also reduces overfitting to the test suite. Testing systems. A possible model for test ownership is to have independently implemented tests at the three levels of organization described above (team, company, industry). At the development team level, reliability tests can be used to diagnose weaknesses with the goal of improving the NLP system for a specific use case and set of target users. Compared to unconstrained adversarial examples, contrasting worst-case examples that have been constrained along specific dimensions with non-worst-case examples will likely yield greater intuition into the model’s inner workings. Studying how modifications (to the architecture, training data and process) affect the system’s reliability on each dimension will also give engineers insight into the factors affecting system reliability. These tests should be executed and updated regularly during development, according to software engineering best practices such as Agile (Beck et al., 2001). Red teams are company-internal teams tasked with finding security vulnerabilities in their developed software or systems. Brundage et al. (2020) propose to apply the concept of red teaming to surface flaws in an AI system’s safety and security. In companies that maintain multiple NLP systems, we propose employing similar, specialized teams composed of NLP experts to build and maintain reliability tests that ensure their NLP systems adhere to company-level reliability standards. These tests will likely be less task-/domain-specific than those developed by engineering teams due to their wider scope, while the reliability standards may be created and maintained by compliance teams or the red teams themselves. Making these stan4159 dards available for public scrutiny and ensuring their products meet them will enable companies to build trust with their users. To ensure all NLP systems meet the company’s reliability standards, these reliability tests should be executed as a part of regular internal audits (Raji et al., 2020), investigative audits after incidents, and before major releases (especially if it is the system’s first release or if it received a major update). They may also be regularly executed on randomly chosen production systems and trigger an alert upon failure. At the independent regulator level, reliability tests would likely be carried out during product certification (e.g., ANSI/ISO certification) and external audits. These industry-level reliability standards and tests may be developed in a similar manner to the company-level ones. However, we expect them to be more general and less comprehensive than the latter, analogous to minimum safety standards such as IEC 60335-1 (IEC, 2020). Naturally, high risk applications and NLP systems used in regulated industries should comply with more stringent requirements (European Commission, 2021). Our proposed framework is also highly compatible with the use of model cards (Mitchell et al., 2019) for auditing and transparent reporting (Raji et al., 2020). In addition to performance on task-related metrics, model cards surface information and assumptions about a machine learning system and training process that may not be readily available otherwise. When a system has passed all tests and is ready to be deployed, its average- and worst-case performance on all tested dimensions can be included as an extra section on the accompanying model card. In addition, the perturbed examples generated during testing and their labels (x′, y′) can be stored for audit purposes or examined to ensure that the tests are performing as expected. Observing and Refining requirements. It is crucial to regularly monitor the systems’ impact post-launch and add, update, or re-prioritize dimensions and thresholds accordingly. Monitoring large-scale deployments can be done via community juries, in which stakeholders who will be likely impacted (or their advocates) give feedback on their pain points and raise concerns about potential negative effects. Smaller teams without the resources to organize community juries can set up avenues (e.g., online forms) for affected stakeholders to give feedback, raise concerns, and seek remediation. 6 From Concerns to Dimensions We now illustrate how reliability concerns can be converted into concrete testing dimensions (Step 1) by considering the scenario of applying automated text scoring to short answers and essays from students in the multilingual population of Singapore. We study a second scenario in Appendix A. Automated Text Scoring (ATS) systems are increasingly used to grade tests and essays (Markoff, 2013; Feathers, 2019). While they can provide instant feedback and help teachers and test agencies cope with large loads, studies have shown that they often exhibit demographic and language biases, such as scoring African- and Indian-American males lower on the GRE Argument task compared to human graders (Bridgeman et al., 2012; Ramineni and Williamson, 2018). Since the results of some tests will affect the futures of the test takers (Salaky, 2018), the scoring algorithms used must be sufficiently reliable. Hence, let us imagine that Singapore’s education ministry has decided to create a standard set of reliability requirements that all ATS systems used in education must adhere to. Linguistic landscape. A mix of language varieties are used in Singapore: a prestige English variety, a colloquial English variety, three other official languages (Chinese, Malay, and Tamil), and a large number of other languages. English is the lingua franca, with fluency in the prestige variety correlating with socioeconomic status (Vaish and Tan, 2008). A significant portion of the population does not speak English at home. Subjects other than languages are taught in English. Stakeholder impact. The key stakeholders affected by ATS systems would be students in schools and universities. The consequences of lower scores could be life-altering for the student who is unable to enroll in the major of their choice. At the population level, biases in an ATS system trained on normally sampled data would unfairly discriminate against already underrepresented groups. Additionally, biases against disfluent or ungrammatical text when they are not the tested attributes would result in discrimination against students with a lower socioeconomic status or for whom English is a second language. Finally, NLP systems have also been known to be overly sensitive to alternative spellings (Belinkov and Bisk, 2018). When used to score subject tests, this could result in the ATS system unfairly penaliz4160 ing dyslexic students (Coleman et al., 2009). Since education is often credited with enabling social mobility,5 unfair grading may perpetuate systemic discrimination and increase social inequality. Dimension. We can generally categorize written tests into those that test for content correctness (e.g., essay questions in a history test), and those that test for language skills (e.g., proper use of grammar). While there are tests that simultaneously assess both aspects, modern ATS systems often grade them separately (Ke and Ng, 2019). We treat each aspect as a separate test here. When grading students on content correctness, we would expect the ATS system to ignore linguistic variation and sensitive attributes as long as they do not affect the answer’s validity. Hence, we would expect variation in these dimensions to have no effect on scores: answer length, language/vocabulary simplicity, alternative spellings/misspellings of non-keywords, grammatical variation, syntactic variation (especially those resembling transfer from a first language), and proxies for sensitive attributes. On the other hand, the system should be able to differentiate proper answers from those aimed at gaming the test (Chin, 2020; Ding et al., 2020). When grading students on language skills, however, we would expect ATS systems to be only sensitive to the relevant skill. For example, when assessing grammar use, we would expect the system to be sensitive to grammatical errors (from the perspective of the language variety the student is expected to use), but not to the other dimensions mentioned above (e.g., misspellings). Actors. Relevant experts include teachers of the subjects where the ATS systems will be deployed, linguists, and computer scientists. The stakeholders (students) may be represented by student unions (at the university level) or focus groups comprising a representative sample of the student population. 7 Implications for Policy There is a mounting effort to increase accountability and transparency around the development and use of NLP systems to prevent them from amplifying societal biases. DOCTOR is highly complementary to the model card approach increasingly adopted6 to surface oft hidden details about NLP 5www.encyclopedia.com/.../education-and-mobility 6huggingface.co/models; github.com/ivylee/model-cards-and-datasheets; models: Developers simply need to list the tested dimensions, metrics, and score on each dimension in the model card. Crucially, reliability tests can be used to highlight fairness issues in NLP systems by including sensitive attributes for the target population, but it is paramount these requirements reflect local concerns rather than any prescriptivist perspective (Sambasivan et al., 2021). At the same time, the ability to conduct quantitative, targeted reliability testing along specifiable dimensions paves the way for reliability standards to be established, with varying levels of stringency and rigor for different use cases and industries. We envision minimum safety and fairness standards being established for applications that are non-sensitive, not safety-critical, and used in unregulated industries, analogous to standards for household appliances. Naturally, applications at greater risks (Li et al., 2020b) of causing harm upon failure should be held to stricter standards. Policymakers are starting to propose and implement regulations to enforce transparency and accountability in the use of AI systems. For example, the European Union’s General Data Protection Regulation grants data subjects the right to obtain “meaningful information about the logic involved” in automated decision systems (EU, 2016). The EU is developing AIspecific regulation (European Commission, 2020): e.g., requiring developers of high-risk AI systems to report their “capabilities and limitations, ... [and] the conditions under which they can be expected to function as intended”. In the U.S., a proposed bill of the state of Washington will require public agencies to report “any potential impacts of the automated decision system on civil rights and liberties and potential disparate impacts on marginalized communities” before using automated decision systems (Washington State Legislature, 2021). One may note that language in the proposed regulation is intentionally vague. There are many ways to measure bias and fairness, depending on the type of model, context of use, and goal of the system. Today, companies developing AI systems employ the definitions they believe most reasonable (or perhaps easiest to implement), but regulation will need to be more specific for there to be meaningful compliance. DOCTOR’s requirement to explicitly define specific dimensions instead of a vague notion of reliability will help policymakers in this blog.einstein.ai/model-cards-for-ai-model-transparency 4161 regard, and can inform the ongoing development of national (NIST, 2019) and international standards7. While external algorithm audits are becoming popular, testing remains a challenge since companies wishing to protect their intellectual property may be resistant to sharing their code (Johnson, 2021), and implementing custom tests for each system is unscalable. Our approach to reliability testing offers a potential solution to this conundrum by treating NLP systems as black boxes. If reliability tests become a legal requirement, regulatory authorities will be able to mandate independently conducted reliability tests for transparency. Such standards, combined with certification programs (e.g., IEEE’s Ethics Certification Program for Autonomous and Intelligent Systems8), will further incentivize the development of responsible NLP, as the companies purchasing NLP systems will insist on certified systems to protect them from both legal and brand risk. To avoid confusion, we expect certification to occur for individual NLP systems (e.g., an end-to-end question answering system for customer enquiries), rather than for general purpose language models that will be further trained to perform some specific NLP task. While concrete standards and certification programs that can serve this purpose do not yet exist, we believe that they eventually will and hope our paper will inform their development. This multi-pronged approach can help to mitigate NLP’s potential harms while increasing public trust in language technology. 8 Challenges and Future Directions While DOCTOR is a useful starting point to implement reliability testing for NLP systems, we observe key challenges to its widespread adoption. First, identifying and prioritizing the dimensions that can attest a system’s reliability and fairness. The former is relatively straightforward and can be achieved via collaboration with experts (e.g., as part of the U.S. NIST’s future AI standards (NIST, 2019)). The latter, however, is a question of values and power (Noble, 2018; Mohamed et al., 2020; Leins et al., 2020), and should be addressed via a code of ethics and ensuring that all stakeholders are adequately represented at the decision table. Second, our proposed method of reliability testing may suffer from similar issues plaguing automatic 7ethicsstandards.org/p7000 8standards.ieee.org/industry-connections/ecpais.html evaluation metrics for natural language generation (Novikova et al., 2017; Reiter, 2018; Kryscinski et al., 2019): due to the tests’ synthetic nature they may not fully capture the nuances of reality. For example, if a test’s objective were to test an NLP system’s reliability when interacting with African American English (AAE) speakers, would it be possible to guarantee (in practice) that all generated examples fall within the distribution of AAE texts? Potential research directions would be to design adversary generation techniques that can offer such guarantees or incorporate human feedback (Nguyen et al., 2017; Kreutzer et al., 2018; Stiennon et al., 2020). 9 Conclusion Once language technologies leave the lab and start impacting real lives, concerns around safety, fairness, and accountability cease to be thought experiments. While it is clear that NLP can have a positive impact on our lives, from typing autocompletion to revitalizing endangered languages (Zhang et al., 2020a), it also has the potential to perpetuate harmful stereotypes (Bolukbasi et al., 2016; Sap et al., 2019), perform disproportionately poorly for underrepresented groups (Hern, 2017; Bridgeman et al., 2012), and even erase already marginalized communities (Bender et al., 2021). Trust in our tools stems from an assurance that stakeholders will remain unharmed, even in the worst-case scenario. In many mature industries, this takes the form of reliability standards. However, for standards to be enacted and enforced, we must first operationalize “reliability”. Hence, we argue for the need for reliability testing (especially worst-case testing) in NLP by contextualizing it among existing work on promoting accountability and improving generalization beyond the training distribution. Next, we showed how adversarial attacks can be reframed as worst-case tests. Finally, we proposed a possible paradigm, DOCTOR, for how reliability concerns can be realized as quantitative tests, and discussed how this framework can be used at different levels of organization or industry. Acknowledgements Samson is supported by Salesforce and Singapore’s Economic Development Board under the Industrial Postgraduate Programme. Araz is supported by the NUS Centre for Trusted Internet and Community through project CTIC-RP-20-02. 4162 Broader Impact Much like how we expect to not be exposed to harmful electric shocks when using electrical appliances, we should expect some minimum levels of safety and fairness for the NLP systems we interact with in our everyday lives. As mentioned in §1, §3, and §7, standards and regulations for AI systems are in the process of being developed for this purpose, especially for applications deemed “high-risk”, e.g., healthcare (European Commission, 2020). Reliability testing, and our proposed framework, is one way to approach the problem of enacting enforceable standards and regulations. However, the flip side of heavily regulating every single application of NLP is that it may slow down innovation. Therefore, it is important that the level of regulation for a particular application is proportionate to its potential for harm (Daten Ethik Kommission, 2019). Our framework can be adapted to different levels of risk by scaling down the implementation of some steps (e.g., the method and depth in which stakeholder consultation happens or the comprehensiveness of the set of testing dimensions) for low-risk applications. Finally, it is important to ensure that any tests, standards, or regulations developed adequately represents the needs of the most vulnerable stakeholders, instead of constructing them in a prescriptivist manner (Hagerty and Rubinov, 2019). Hence, DOCTOR places a strong emphasis on involving stakeholder advocates and analyzing the impact of an application of NLP on the target community. References 116th U.S. Congress. 2019. Algorithmic Accountability Act of 2019. Ada Lovelace Institute and DataKind UK. 2021. Examining the black box: Tools for assessing algorithmic systems. Technical report. Moustafa Alzantot, Yash Sharma, Ahmed Elgohary, Bo-Jhang Ho, Mani Srivastava, and Kai-Wei Chang. 2018. Generating natural language adversarial examples. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2890–2896, Brussels, Belgium. Association for Computational Linguistics. Amnesty International. 2018. Toxic Twitter - triggers of violence and abuse against women on Twitter. Kent Beck, Mike Beedle, Arie van Bennekum, Alistair Cockburn, Ward Cunningham, Martin Fowler, James Grenning, Jim Highsmith, Andrew Hunt, Ron Jeffries, Jon Kern, Brian Marick, Robert C. Martin, Steve Mellor, Ken Schwaber, Jeff Sutherland, and Dave Thomas. 2001. Manifesto for Agile Software Development. Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In 6th International Conference on Learning Representations, Vancouver, BC, Canada. Emily M Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587–604. Emily M. Bender, Timnit Gebru, Angelina McMillanMajor, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the Conference on Fairness, Accountability, and Transparency. Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of “bias” in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454–5476, Online. Association for Computational Linguistics. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 4349–4357. Curran Associates, Inc. Brent Bridgeman, Catherine Trapani, and Yigal Attali. 2012. Comparison of human and machine scoring of essays: Differences by gender, ethnicity, and country. Applied Measurement in Education, 25(1):27–40. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33. Miles Brundage, Shahar Avin, Jasmine Wang, Haydn Belfield, Gretchen Krueger, Gillian Hadfield, Heidy Khlaaf, Jingying Yang, Helen Toner, Ruth Fong, et al. 2020. Toward trustworthy AI development: mechanisms for supporting verifiable claims. arXiv preprint arXiv:2004.07213. California State Legislature. 2020. California Privacy Rights Act. Minhao Cheng, Wei Wei, and Cho-Jui Hsieh. 2019. 4163 Evaluating and enhancing the robustness of dialogue systems: A case study on a negotiation agent. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3325–3335, Minneapolis, Minnesota. Association for Computational Linguistics. Monica Chin. 2020. These students figured out their tests were graded by AI — and the easy way to cheat. The Verge. Katie Cohen, Fredrik Johansson, Lisa Kaati, and Jonas Clausen Mork. 2014. Detecting linguistic markers for radical violence in social media. Terrorism and Political Violence, 26(1):246–256. Chris Coleman, Noël Gregg, Lisa McLain, and Leslie W Bellair. 2009. A comparison of spelling performance across young adults with and without dyslexia. Assessment for effective intervention, 34(2):94–105. Kate Crawford. 2017. The trouble with bias (keynote). Advances in Neural Information Processing Systems 30. Marina Danilevsky, Kun Qian, Ranit Aharonov, Yannis Katsis, Ban Kawas, and Prithviraj Sen. 2020. A survey of the state of explainable AI for natural language processing. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 447– 459, Suzhou, China. Association for Computational Linguistics. Daten Ethik Kommission. 2019. Opinion of the data ethics commission. Technical report, Data Ethics Commission of the Federal Government (Germany). Thomas Davidson, Debasmita Bhattacharya, and Ingmar Weber. 2019. Racial bias in hate speech and abusive language detection datasets. In Proceedings of the Third Workshop on Abusive Language Online, pages 25–35, Florence, Italy. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171– 4186, Minneapolis, Minnesota. Association for Computational Linguistics. Yuning Ding, Brian Riordan, Andrea Horbach, Aoife Cahill, and Torsten Zesch. 2020. Don’t take “nswvtnvakgxpm” for an answer –the surprising vulnerability of automatic content scoring systems to adversarial input. In Proceedings of the 28th International Conference on Computational Linguistics, pages 882–892, Barcelona, Spain (Online). International Committee on Computational Linguistics. Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2368–2378, Minneapolis, Minnesota. Association for Computational Linguistics. Michael Dunn. 2014. Gender determined dialect variation. In The expression of gender, pages 39–68. De Gruyter. Javid Ebrahimi, Anyi Rao, Daniel Lowd, and Dejing Dou. 2018. HotFlip: White-box adversarial examples for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 31–36, Melbourne, Australia. Association for Computational Linguistics. Steffen Eger and Yannik Benz. 2020. From hero to zéroe: A benchmark of low-level adversarial attacks. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 786–803, Suzhou, China. Association for Computational Linguistics. EIU. 2020. Staying ahead of the curve: The business case for responsible AI. Technical report, The Economist Intelligence Unit. EU. 2016. General data protection regulation. European Commission. 2020. On artificial intelligence - a European approach to excellence and trust. Technical report, European Commission. European Commission. 2021. Proposal for a regulation laying down harmonised rules on artificial intelligence (artificial intelligence act). Technical report, European Commission. FDA. 2021. Artificial intelligence/machine learning (ai/ml)-based software as a medical device (samd) action plan. Technical report, U.S. Food & Drug Administration. Todd Feathers. 2019. Flawed algorithms are grading millions of students’ essays. Vice. Adam Fisch, Alon Talmor, Robin Jia, Minjoon Seo, Eunsol Choi, and Danqi Chen. 2019. MRQA 2019 shared task: Evaluating generalization in reading comprehension. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 1–13, Hong Kong, China. Association for Computational Linguistics. Batya Friedman and David G Hendry. 2019. Value Sensitive Design: Shaping Technology with Moral Imagination. MIT Press. Bharath Ganesh. 2018. The ungovernability of digital hate culture. Columbia Journal of International Affairs. 4164 Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, and Ben Zhou. 2020. Evaluating models’ local decision boundaries via contrast sets. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1307–1323, Online. Association for Computational Linguistics. Siddhant Garg and Goutham Ramakrishnan. 2020. BAE: BERT-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, Online. Association for Computational Linguistics. Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. 2018. Datasheets for datasets. arXiv preprint arXiv:1803.09010. Mor Geva, Yoav Goldberg, and Jonathan Berant. 2019. Are we modeling the task or the annotator? an investigation of annotator bias in natural language understanding datasets. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1161– 1166, Hong Kong, China. Association for Computational Linguistics. Karan Goel, Nazneen Rajani, Jesse Vig, Samson Tan, Jason Wu, Stephan Zheng, Caiming Xiong annd Mohit Bansal, and Christopher Ré. 2021. Robustness Gym: Unifying the NLP evaluation landscape. arXiv preprint arXiv:2101.04840. Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. Explaining and harnessing adversarial examples. In 3rd International Conference on Learning Representations, San Diego, California. Alexa Hagerty and Igor Rubinov. 2019. Global ai ethics: a review of the social impacts and ethical implications of artificial intelligence. arXiv preprint arXiv:1907.07892. Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, William Lewis, Mu Li, et al. 2018. Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567. Dan Hendrycks, Xiaoyuan Liu, Eric Wallace, Adam Dziedzic, Rishabh Krishnan, and Dawn Song. 2020. Pretrained transformers improve out-of-distribution robustness. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2744–2751, Online. Association for Computational Linguistics. Alex Hern. 2017. Facebook translates ‘good morning’ into ‘attack them’, leading to arrest. The Guardian. Dirk Hovy, Federico Bianchi, and Tommaso Fornaciari. 2020. “you sound just like your father” commercial machine translation systems include stylistic biases. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1686–1690, Online. Association for Computational Linguistics. Dirk Hovy and Shannon L Spruit. 2016. The social impact of natural language processing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 591–598. Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton, Kellie Webster, Yu Zhong, and Stephen Denuyl. 2020. Social biases in NLP models as barriers for persons with disabilities. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5491–5501, Online. Association for Computational Linguistics. IEC. 2020. Household and similar electrical appliances – Safety – Part 1: General requirements. IEC 603351:2020. IEEE. 2017. ISO/IEC/IEEE International standard - systems and software engineering–vocabulary. ISO/IEC/IEEE 24765:2017(E), pages 1–541. Mohit Iyyer, John Wieting, Kevin Gimpel, and Luke Zettlemoyer. 2018. Adversarial example generation with syntactically controlled paraphrase networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1875–1885, New Orleans, Louisiana. Association for Computational Linguistics. Robin Jia and Percy Liang. 2017. Adversarial examples for evaluating reading comprehension systems. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021– 2031, Copenhagen, Denmark. Association for Computational Linguistics. Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is BERT really robust? A strong baseline for natural language attack on text classification and entailment. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 8018–8025. AAAI Press. Khari Johnson. 2021. What algorithm auditing startups need to succeed. VentureBeat. Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. 2019. Learning the difference that makes a difference with counterfactually-augmented data. In International Conference on Learning Representations. Zixuan Ke and Vincent Ng. 2019. Automated essay scoring: A survey of the state of the art. In International Joint Conference on Artificial Intelligence. Inter4165 national Joint Conferences on Artificial Intelligence Organization. Daniel Khashabi, Tushar Khot, and Ashish Sabharwal. 2020. More bang for your buck: Natural perturbation for robust question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 163–170, Online. Association for Computational Linguistics. Kate Klonick. 2018. The new governors: The people, rules, and processes governing online speech. Harvard Law Review, 131(6):1598. Julia Kreutzer, Shahram Khadivi, Evgeny Matusov, and Stefan Riezler. 2018. Can neural machine translation be improved with user feedback? In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 3 (Industry Papers), pages 92–105. Wojciech Kryscinski, Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Neural text summarization: A critical evaluation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 540–551, Hong Kong, China. Association for Computational Linguistics. Zachary Laub. 2019. Hate speech on social media: Global comparisons. Council on Foreign Relations. Kobi Leins, Jey Han Lau, and Timothy Baldwin. 2020. Give me convenience and give her death: Who should decide what uses of NLP are appropriate, and on what basis? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2908–2913, Online. Association for Computational Linguistics. Jinfeng Li, Shouling Ji, Tianyu Du, Bo Li, and Ting Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium. Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020a. Bert-attack: Adversarial attack against BERT using BERT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, Online. Association for Computational Linguistics. Yanwei Li, Araz Taeihagh, Martin de Jong, and Andreas Klinke. 2020b. Toward a commonly shared public policy perspective for analyzing risk coping strategies. Risk Analysis. John Markoff. 2013. Essay-grading software offers professors a break. The New York Times. Alexandria Marsters. 2019. When Hate Speech Leads to Hateful Actions: A Corpus and Discourse Analytic Approach to Linguistic Threat Assessment of Hate Speech. Ph.D. thesis, Georgetown University, Washington, D.C. Paul Michel, Xian Li, Graham Neubig, and Juan Pino. 2019. On evaluation of adversarial perturbations for sequence-to-sequence models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3103–3114, Minneapolis, Minnesota. Association for Computational Linguistics. John Miller, Karl Krauth, Benjamin Recht, and Ludwig Schmidt. 2020. The effect of natural distribution shift on question answering models. arXiv preprint arXiv:2004.14444. Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model cards for model reporting. In Proceedings of the Conference on Fairness, Accountability, and Transparency, pages 220–229. Shakir Mohamed, Marie-Therese Png, and William Isaac. 2020. Decolonial AI: Decolonial theory as sociotechnical foresight in artificial intelligence. Philosophy & Technology, 33(4):659–684. Stacy Nelson. 2003. Certification processes for safetycritical and mission-critical aerospace software. Technical report, NASA Technical Reports Server. Khanh Nguyen, Hal Daumé III, and Jordan BoydGraber. 2017. Reinforcement learning for bandit neural machine translation with simulated human feedback. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1464– 1474. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885–4901, Online. Association for Computational Linguistics. NIST. 2019. U.S. leadership in AI: A plan for federal engagement in developing technical standards and related tools. Technical report, National Institute of Standards and Technology. Safiya Umoja Noble. 2018. Algorithms of oppression: How search engines reinforce racism. NYU Press. Jekaterina Novikova, Ondˇrej Dušek, Amanda Cercas Curry, and Verena Rieser. 2017. Why we need new evaluation metrics for NLG. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2241–2252, Copenhagen, Denmark. Association for Computational Linguistics. Cathy O’Neil. 2016. Weapons of math destruction: How big data increases inequality and threatens democracy. Crown. Partnership on AI. 2019. About ML. Technical report, Partnership on AI. 4166 Barbara Plank. 2016. What to do about non-standard (or non-canonical) language in NLP. Proceedings of the 13th Conference on Natural Language Processing (KONVENS 2016). Megha Rajagopalan, Lam Thuy Vo, and Aung Naing Soe. 2018. How Facebook failed the Rohingya in Myanmar. BuzzFeed News. Inioluwa Deborah Raji, Andrew Smart, Rebecca N White, Margaret Mitchell, Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker Barnes. 2020. Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, pages 33–44. Bogdana Rakova, Jingying Yang, Henriette Cramer, and Rumman Chowdhury. 2020. Where responsible ai meets reality: Practitioner perspectives on enablers for shifting organizational practices. arXiv preprint arXiv:2006.12358. Chaitanya Ramineni and David Williamson. 2018. Understanding mean score differences between the erater® automated scoring engine and humans for demographically based groups in the GRE® general test. ETS Research Report Series, 2018(1):1–31. Ehud Reiter. 2018. A structured review of the validity of BLEU. Computational Linguistics, 44(3):393–401. Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. 2019. Generating natural language adversarial examples through probability weighted word saliency. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1085–1097. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, page 1135–1144, New York, NY, USA. Association for Computing Machinery. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2018. Semantically equivalent adversarial rules for debugging NLP models. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 856–865, Melbourne, Australia. Association for Computational Linguistics. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902–4912, Online. Association for Computational Linguistics. Anne-Laure Rousseau, Clément Baudelaire, and Kevin Riera. 2020. Doctor GPT-3: hype or reality? Nabla Technologies Blog. Kristin Salaky. 2018. What standardized tests look like in 10 places around the world. INSIDER. Nithya Sambasivan, Erin Arnesen, Ben Hutchinson, Tulsee Doshi, and Vinodkumar Prabhakaran. 2021. Reimagining algorithmic fairness in india and beyond. In Proceedings of the 2021 Conference on Fairness, Accountability, and Transparency. Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668–1678, Florence, Italy. Association for Computational Linguistics. Suchi Saria and Adarsh Subbaswamy. 2019. Safe and reliable machine learning (tutorial). ACM Conference on Fairness, Accountability, and Transparency. Deven Santosh Shah, H. Andrew Schwartz, and Dirk Hovy. 2020. Predictive biases in natural language processing models: A conceptual framework and overview. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5248–5264, Online. Association for Computational Linguistics. Andrew Smith. 2020. Using artificial intelligence and algorithms. Anders Søgaard, Sebastian Ebert, Jasmijn Bastings, and Katja Filippova. 2021. We need to talk about random splits. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1823–1832, Online. Association for Computational Linguistics. Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul Christiano. 2020. Learning to summarize from human feedback. arXiv preprint arXiv:2009.01325. Samson Tan and Shafiq Joty. 2021. Code-Mixing on Sesame Street: Dawn of the adversarial polyglots. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Online. Association for Computational Linguistics. Samson Tan, Shafiq Joty, Min-Yen Kan, and Richard Socher. 2020. It’s morphin’ time! Combating linguistic discrimination with inflectional perturbations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2920–2935, Online. Association for Computational Linguistics. Deborah Tannen. 1991. You just don’t understand: Women and men in conversation. Ballantine books New York. Deborah Tannen et al. 2005. Conversational style: Analyzing talk among friends. Oxford University Press. Viniti Vaish and Teck Kiang Tan. 2008. Language and social class: Linguistic capital in Singapore. In Annual Meeting of the American Educational Research Association. American Educational Research Association. 4167 Claudia Wagner, David Garcia, Mohsen Jadidi, and Markus Strohmaier. 2015. It’s a man’s Wikipedia? Assessing gender inequality in an online encyclopedia. In Proceedings of the International AAAI Conference on Web and Social Media, volume 9. Lily Wakefield. 2020. Queer people are being forced off social media by trolling and online abuse, searingly obvious report confirms. PinkNews. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing NLP. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 2153–2162, Hong Kong, China. Association for Computational Linguistics. Zeerak Waseem, Smarika Lulz, Joachim Bingel, and Isabelle Augenstein. 2021. Disembodied machine learning: On the illusion of objectivity in NLP. arXiv preprint arXiv:2101.11974. Washington State Legislature. 2021. Senate bill SB 5116. Gerhard Widmer and Miroslav Kubat. 1996. Learning in the presence of concept drift and hidden contexts. Machine Learning, 23:69–101. Chris Wilkinson, Jonathan Lynch, Raj Bharadwaj, and Kurt Woodham. 2016. Verification of adaptive systems. Technical report, Federal Aviation Administration William J. Hughes Technical Center. Tongshuang Wu, Marco Tulio Ribeiro, Jeffrey Heer, and Daniel S Weld. 2021. Polyjuice: Automated, general-purpose counterfactual generation. arXiv preprint arXiv:2101.00288. Adams Wei Yu, David Dohan, Quoc Le, Thang Luong, Rui Zhao, and Kai Chen. 2018. Fast and accurate reading comprehension by combining self-attention and convolution. In International Conference on Learning Representations. Yuan Zang, Fanchao Qi, Chenghao Yang, Zhiyuan Liu, Meng Zhang, Qun Liu, and Maosong Sun. 2020. Wordlevel textual adversarial attacking as combinatorial optimization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6066–6080. Huangzhao Zhang, Hao Zhou, Ning Miao, and Lei Li. 2019a. Generating fluent adversarial examples for natural languages. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5564–5569, Florence, Italy. Association for Computational Linguistics. Shiyue Zhang, Benjamin Frey, and Mohit Bansal. 2020a. ChrEn: Cherokee-English machine translation for endangered language revitalization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 577–595, Online. Association for Computational Linguistics. Wei Emma Zhang, Quan Z. Sheng, Ahoud Alhazmi, and Chenliang Li. 2020b. Adversarial attacks on deeplearning models in natural language processing: A survey. ACM Transactions on Intelligent Systems and Technology, 11(3). Yuan Zhang, Jason Baldridge, and Luheng He. 2019b. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298–1308, Minneapolis, Minnesota. Association for Computational Linguistics. 4168 Appendix A Testing Dimensions: Detecting Violent Content on Social Media In this second case study, we apply DOCTOR for measuring the reliability of a violent content detection system for English social media posts. Although we limit this discussion to the U.S., this is a growing global problem (Laub, 2019) that can lead to deadly outcomes (Rajagopalan et al., 2018). In this hypothetical use case, the NLP system may automatically remove violent content or alert content moderators to potential violations of the social media company’s acceptable use policy. Moderators can decide if specific content should be removed, and if necessary, notify law enforcement to avert pending violence (e.g., threats against individuals, planned violent events). As a result of the 1996 Communications Decency Act9, social media platforms have broad latitude (Klonick, 2018) to develop their own policies for acceptable content and how they handle it. In this scenario, the compliance officer of the company developing the system is responsible for making sure it does not discriminate against specific user demographics. Research has shown that hate speech can lead to hateful actions (Marsters, 2019). In many cases, individuals posted their intents online prior to committing violence (Cohen et al., 2014). When identifying content to remove and especially when involving law enforcement, it is important to distinguish between “Hunters" — those who act — and “Howlers" — those who do not (Marsters, 2019). This is to avoid wrongly detaining individuals who have no intention of committing violence, even if their words are indefensible. Between these extremes, posters may harass, stalk, dox, or otherwise abuse victims from a distance, therefore it is still necessary to flag, remove, and potentially track or document violent content. Linguistic landscape. We focus solely on English speakers, but we acknowledge that the actual linguistic landscape is much more complex (over 350 languages). Posters on social media may speak English as their first language or as a second language and they often code-switch/-mix. Standard American English is used for business purposes in the U.S. but there are other frequently used language varieties including African American English (AAE), Cajun Vernacular English, and three 9fcc.gov/general/telecommunications-act-1996 different Latinx (Hispanic) vernacular Englishes. Stakeholder Impact. The key stakeholders that will be impacted are those most often facing violent threats online: minorities, women, immigrants, and the LGBTQ community (Amnesty International, 2018; Ganesh, 2018; Davidson et al., 2019; Wakefield, 2020). Additionally, anyone that posts content on the social media site is a stakeholder. Unfortunately, the very communities that are often the target of violent posts are also often wrongly flagged as posting toxic content themselves due to racial biases present in the training data (Sap et al., 2019; Davidson et al., 2019). Given the risk of harm to victims if the system misses violent posts from hunters or misidentifies legitimate content as violent and notifies law enforcement, it is critical the right balance of false positives and false negatives is achieved in flagging content. Dimensions. There are two tasks under consideration here: identifying violent content and identifying Hunters who “truly intend to use lethal violence” (Marsters, 2019). In the first task, the system is looking for content that negatively targets a socially defined group. Additionally, the content includes not only hate speech (e.g., profanity, epithets, vulgarity) but also content that incites others to hatred or violence. Since content written in AAE has been shown to be flagged as toxic more often (Sap et al., 2019; Davidson et al., 2019), we must ensure that the system is reliable when encountering dialectal variation. Additionally, due to the casual environment of social media, multilingual speakers often code-switch and code-mix. Hence, we expect variation in these dimensions to have no effect on the system’s predictions: alternative spellings, morphosyntactic variation, word choice, code-mixing, idioms, and references to and manifestations of sensitive attributes and their proxies. However, we must expect the system to be sensitive to in-group and out-group usage of reclaimed slurs so that the in-group usage does not result in a flag while out-group usage result in flagged posts. When identifying hunters, we may expect the system to be sensitive to uses of first person pronouns, certainty adverbs, negative evaluative adjectives, and modifiers (Marsters, 2019). However, in order to avoid unfairly penalizing vernacular English speakers we should expect the system’s predictions to be equally unaffected by variation in the dimensions listed for the first task. 4169 Orthography Hyphenation Capitalization Punctuation Reduplication of letters Emojis/emoticons Homonyms Disemvoweling (Eger and Benz, 2020) Homophones (e.g., accept vs. except) (Eger and Benz, 2020) Accidental misspellings (Belinkov and Bisk, 2018) Intentional alternative spellings (e.g., Yas, thru, startin) Open compound concatenation (e.g., couch potato/couchpotato) Dialectal differences (e.g., favor vs. favour) (Ribeiro et al., 2018) Mixing writing scripts (Tan and Joty, 2021) Transliteration Morphology Grammatical gender shifts Grammatical category (Tan et al., 2020) Dialectal differences (Tan et al., 2020) Linguistic Clitics Phenomena Lexicon Dialectal variation (e.g., fries vs. chips) Synonyms/Sememes (Zang et al., 2020) Vocabulary simplicity/complexity Cross-lingual synonyms (Tan and Joty, 2021) Loanwords Semantics Idioms (e.g., finer than frog hair) Syntax Matching number and tense Word/phrase order (especially for languages without strict word ordering) Prepositional variation (e.g., stand on line vs. stand in line) Syntactic variation (Iyyer et al., 2018) Sentence simplicity/complexity Code-mixing (Tan and Joty, 2021) Register (e.g., formality) Discourse Conversational style (involvement/considerateness) (Tannen et al., 2005) & Discourse markers / connector words Pragmatics Cross-cultural differences Code-switching Sensitive Attributes Gender Identity Gender pronouns Names Reclaimed slurs Genderlects (Tannen, 1991; Dunn, 2014) Race Names Reclaimed slurs Race-aligned language varieties Age Age/generation-aligned language styles (Hovy et al., 2020) Religion Names Reclaimed slurs Sexual Orientation Reclaimed slurs Disability status Associated adjectives (Hutchinson et al., 2020) Place of origin Location names (e.g., cities, countries) Figures of speech Proxies Geographic locations (for ethnicity, socioeconomic status) Malicious Attacks Black-box Rule-based (Alzantot et al., 2018; Jin et al., 2020) Model-based (Garg and Ramakrishnan, 2020; Li et al., 2020a) Gradient-based HotFlip (Ebrahimi et al., 2018), Universal Triggers (Wallace et al., 2019) Policy-based Adversarial negotiation agent (Cheng et al., 2019) Table 1: Taxonomy of possible dimensions with references to linguistics literature and existing adversarial attacks that could be used as worst-case tests. Linguists are best equipped to decide which linguistic phenomena are high priority for each use case, ethicists for sensitive attributes, and NLP practitioners for malicious attacks.
|
2021
|
321
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4170–4187 August 1–6, 2021. ©2021 Association for Computational Linguistics 4170 Learning Language and Multimodal Privacy-Preserving Markers of Mood from Mobile Data Paul Pu Liang1⋆, Terrance Liu1⋆, Anna Cai1, Michal Muszynski1, Ryo Ishii1, Nicholas Allen2, Randy Auerbach3, David Brent4, Ruslan Salakhutdinov1, Louis-Philippe Morency1 1Carnegie Mellon University 2University of Oregon 3Columbia University 4University of Pittsburgh {pliang,terrancl,annacai,mmuszyns,rishii,rsalakhu,morency}@cs.cmu.edu [email protected] [email protected] [email protected] Abstract Mental health conditions remain underdiagnosed even in countries with common access to advanced medical care. The ability to accurately and efficiently predict mood from easily collectible data has several important implications for the early detection, intervention, and treatment of mental health disorders. One promising data source to help monitor human behavior is daily smartphone usage. However, care must be taken to summarize behaviors without identifying the user through personal (e.g., personally identifiable information) or protected (e.g., race, gender) attributes. In this paper, we study behavioral markers of daily mood using a recent dataset of mobile behaviors from adolescent populations at high risk of suicidal behaviors. Using computational models, we find that language and multimodal representations of mobile typed text (spanning typed characters, words, keystroke timings, and app usage) are predictive of daily mood. However, we find that models trained to predict mood often also capture private user identities in their intermediate representations. To tackle this problem, we evaluate approaches that obfuscate user identity while remaining predictive. By combining multimodal representations with privacy-preserving learning, we are able to push forward the performanceprivacy frontier. 1 Introduction Mental illnesses can have a damaging permanent impact on communities, societies, and economies all over the world (World Health Organization, 2003). Individuals often do not realize they are at risk of mental disorders even when they have symptoms. As a result, many are late in seeking professional help and treatment (Thornicroft et al., 2016), particularly among adolescents where suicide is the second leading cause of death (Curtin ⋆first two authors contributed equally. Real-time assessment Decentralized multimodal mobile device data Aggregate Privacy-preserving representation learning Figure 1: Intensive monitoring of behaviors via adolescents’ natural use of smartphones may help identify real-time predictors of mood in high-risk youth as a proxy for suicide risk. While smartphones provide a valuable data source spanning text, keystrokes, app usage, and geolocation, one must take care to summarize behaviors without revealing user identities through personal (e.g., personally identifiable information) or protected attributes (e.g., race, gender) to potentially adversarial third parties. and Heron, 2019). In addition to deaths, 16% of high school students report having serious suicidal thoughts each year, and 8% of them make one or more suicide attempts (CDC, 2015). This problem is particularly exacerbated as an “echo pandemic” of mental health problems have arisen in the wake of the COVID-19 pandemic (Inkster et al., 2021; Saha et al., 2020). Intensive monitoring of behaviors via adolescents’ natural use of smartphones may help identify realtime predictors of mood in high-risk youth as a proxy for suicide risk (Nahum-Shani et al., 2018). While there are inherent limitations in the mismatch between mood prediction and ultimately developing real-time intervention against imminent suicide risk (Coppersmith et al., 2018; Ophir et al., 2020), we believe that the former is a reasonable starting point to tackle similar machine learning problems surrounding affective computing and privacy-preserving learning. Studying mood in this high-risk population is a valuable goal given 4171 that suicide attempts are often decided within a short time-lapse and just-in-time assessments of mood changes can be a stepping stone in this direction (Rizk et al., 2019; Oquendo et al., 2020). Technologies for mood prediction can also be a valuable component of decision support for clinicians and healthcare providers during their assessments (Mann et al., 2006; Cho et al., 2019). Recent work in affective computing has begun to explore the potential in predicting mood from mobile data. Studies have found that typing patterns (Cao et al., 2017; Ghosh et al., 2017a; Huang et al., 2018; Zulueta et al., 2018), self-reporting apps (Suhara et al., 2017), and wearable sensors (Ghosh et al., 2017b; Sano et al., 2018) are particularly predictive. In addition, multimodal modeling of multiple sensors (e.g., wearable sensors and smartphone apps) was shown to further improve performance (Jaques et al., 2017; Taylor et al., 2017). While current work primarily relies on selfreport apps for long-term mood assessments (Glenn and Nock, 2014), our work investigates mobile behaviors from a high-risk teenage population as a predictive signal for daily mood (Franklin et al., 2017; Large et al., 2017). Prior work has also shown that private information is predictable from digital records of human behavior (Kosinski et al., 2013), which is dangerous especially when sensitive user data is involved. As a result, in parallel to improving predictive performance, a recent focus has been on improving privacy through techniques such as differential privacy (Dankar and El Emam, 2012, 2013; Dankar et al., 2012) and federated learning (McMahan et al., 2016; Geyer et al., 2017; Liang et al., 2020b), especially for healthcare data (e.g., electronic health records (Xu and Wang, 2019)) and wearable devices (Chen et al., 2020). In this paper, as a step towards using multimodal privacy-preserving mood prediction as fine-grained signals to aid in mental health assessment, we analyze a recent dataset of mobile behaviors collected from adolescent populations at high suicidal risk. With consent from participating groups, the dataset collects fine-grained features spanning online communication, keystroke patterns, and application usage. Participants are administered daily questions probing for mood scores. By collecting and working on ground-truth data for this population, we are able to benchmark on a more accurate indicator of mood rather than proxy data such as mood signals inferred from social media content or behavior (Ernala et al., 2019). This unique dataset presents an opportunity to investigate a different medium of natural language processing - typed text which presents new challenges beyond conventionally studied written (Marcus et al., 1993) and spoken (Marslen-Wilson and Tyler, 1980) text. We propose multimodal models that contextualize text with their typing speeds and app usage. However, these models often capture private user identities in their intermediate representations when predicting mood. As a step towards privacy-preserving learning, we also propose approaches that obfuscate user identity while remaining predictive of daily mood. By combining multimodal contextualization with privacy-preserving learning, we are able to push forward the performance-privacy frontier. Finally, we conclude with several observations regarding the uniqueness of typed text as an opportunity for NLP on mobile data. 2 Multimodal Mobile Dataset Intensive monitoring of behaviors via adolescents’ frequent use of smartphones may shed new light on the early risk of suicidal thoughts and ideations (Nahum-Shani et al., 2018). Smartphones provide a valuable and natural data source with rich behavioral markers spanning online communication, keystroke patterns, and application usage. Learning these markers requires large datasets with diversity in participants, variety in features, and accuracy in annotations. As a step towards this goal, we recently collected a dataset of mobile behaviors from high-risk adolescent populations with consent from participating groups. We begin with a brief review of the data collection process. This data monitors adolescents spanning (a) recent suicide attempters (past 6 months) with current suicidal ideation, (b) suicide ideators with no past suicide attempts, and (c) psychiatric controls with no history of suicide ideation or attempts. Passive sensing data is collected from each participant’s smartphone across a duration of 6 months. Participants are administered clinical interviews probing for suicidal thoughts and behaviors (STBs), and self-report instruments regarding symptoms and acute events (e.g., suicide attempts, psychiatric hospitalizations) are tracked weekly via a questionnaire. All users have given consent for their mobile data to be collected and shared with us for research 4172 purposes. This study has been carefully reviewed and approved by an IRB. We follow the NIH guidelines, with a central IRB (single IRB) linked to secondary sites. We have IRB approval for the central institution and all secondary sites. 2.1 Mood Assessment via Self-Report Every day at 8am, users are asked to respond to the following question - “In general, how have you been feeling over the last day?” - with an integer score between 0 and 100, where 0 means very negative and 100 means very positive. To construct our prediction task, we discretized these scores into the following three bins: negative (0 −33), neutral (34 −66), and positive (67 −100), which follow a class distribution of 12.43%, 43.63%, and 43.94% respectively. For our 3-way classification task, participants with fewer than 50 daily self-reports were removed since these participants do not provide enough data to train an effective model. In total, our dataset consists of 1641 samples, consisting of data coming from 17 unique participants. 2.2 Features We focused on keyboard data, which includes the time of data capture, the mobile application used, and the text entered by the user. For each daily score response at 8am, we use information collected between 5am on the previous day to 5am on the current day. We chose this 5am-5am window by looking at mobile activity and finding the lowest activity point when most people ended their day: 5am. Since users report the previous day’s mood (when prompted at 8am), we decided to use this 5am-5am time period to summarize the previous day’s activities. Through prototyping, this prompt time and frequency were found to give reliable indicators of the previous day’s mood. From this window, we extracted the following features to characterize and contextualize typed text. Text: After removing stop-words, we collected the top 1000 words (out of approximately 3.2 million) used across all users in our dataset and created a bag-of-words feature that contains the daily number of occurrences of each word. Keystrokes: We also extracted keystroke features that record the exact timing that each character was typed on a mobile keyboard (including alphanumeric characters, special characters, spaces, backspace, enter, and autocorrect). By taking the increase in recorded timing after each keystroke, we obtain the duration that each key was pressed in a sequence of keystrokes during the day. When extracting keystrokes, we removed all small timings under 10−2 seconds. App usage: We count the number of mobile applications used per day, creating a bag-of-apps feature for each day. We discard applications that are used by less than 10% of the participants so that our features are generalizable to more than just a single user in the dataset, resulting in 137 total apps (out of the original 640). In a preliminary analysis, we observed that predictive models performed well when binarizing our feature vectors into boolean vectors, which signify whether a word or app was used on a given day (i.e., mapping values greater than 0 to 1). Our final feature vectors consist of a concatenation of a normalized and a binarized feature vector, resulting in 2000 and 274-dimensional vectors for text and app features respectively. For keystrokes, we found that summarizing the sequence of timings using a histogram (i.e., defining a set of timing buckets and creating a bag-of-timings feature) for each day performed well. We chose 100 fine-grained buckets, resulting in a 100-dimensional keystroke vector. Please refer to Appendix B for additional details about the dataset and extracted features. 3 Mood Prediction Methods In this paper, we focus on studying approaches for learning privacy-preserving representations from mobile data for mood prediction. Our processed data comes in the form of {(xt,i, xk,i, xa,i, yi)}n i=1 with xt ∈N|Vt|=2000 denoting the bag-of-words features, xk ∈N|Vk|=100 denoting the bag-oftimings features, and xa ∈N|Va|=274 denoting the bag-of-apps features. y denotes the label which takes on one of our 3 mood categories: negative, neutral, and positive. In parallel, we also have data representing the corresponding (one-hot) user identity xid which will be useful when learning privacypreserving representations that do not encode information about user identity xid and evaluating privacy performance. 3.1 Unimodal Approaches We considered two unimodal baselines: 1. Support Vector Machines (SVMS) project training examples to a chosen kernel space and finds the optimal hyperplane that maximally separates each class of instances. We apply an SVM classifier on input data xuni ∈{xt, xk, xa} and use supervised 4173 Figure 2: Diagram of the NI-MLP algorithm learned via the (1) pretrain, (2) selection, and (3) addition phases. Boxes with numbers denote which parameters are being optimized in the corresponding step. For example, in the addition phase (3), NI-MLP optimizes parameters δ in g(.; δ). (2a) depicts identity-dependent dimensions zid, which is a sparse vector of size dim(zfeat) whose nonzero values (colored purple) signify dimensions of the identity-dependent subspace in zfeat. learning to predict daily mood labels y. 2. Multilayer Perceptrons (MLPS) have seen widespread success in supervised prediction tasks due to their ability in modeling complex nonlinear relationships. Because of the small size of our dataset, we choose a simple multilayer perceptron with two hidden layers. Similarly, we apply an MLP classifier on input data xuni ∈{xt, xk, xa} to predict daily mood labels y. 3.2 Multimodal Models We extend both SVM and MLP classifiers using early fusion (Baltrušaitis et al., 2018) of text and app usage to model multimodal interactions. Specifically, we align the input through concatenating the bag-of-words, bag-of-keystrokes, and bag-of-apps features for each day resulting in an input vector xmulti = xt ⊕xk ⊕xa, before using an SVM/MLP classifier for prediction. 3.3 A Step Toward Preserving Privacy While classifiers trained with traditional supervised learning can learn useful representations for mood prediction, they carry the risk of memorizing the identity of the user along with their sensitive mobile usage and baseline mood scores, and possibly revealing these identities to adversarial thirdparties (Abadi et al., 2016). Therefore, it is crucial to perform mood prediction while also protecting the privacy of personal identities. We adapt the Selective-Additive Learning (SAL) framework (Wang et al., 2017) for the purpose of privacy-preserving learning. While SAL was originally developed with a very different goal in mind: improving model generalization, we expand SAL to a very important problem in healthcare: preserving privacy. We adapted SAL to learn disentangled representations separated into identity-dependent private information and identityindependent population-level information using three phases: (1) Pretrain phase: The input is a set of (multimodal) features x that are likely to contain both identity-dependent and independent information. The intermediate representation zfeat = ffeat(x; θ∗ feat) is obtained from an MLP classifier pretrained for mood prediction. ffeat denotes the classifier with pretrained parameters θ∗ feat. (2) Selection phase: Our goal is to now disentangle the identity-dependent and independent information within zfeat. We hypothesize that dependent and independent information are encoded in separate subspaces of the feature vector zfeat. This allows us to disentangle them by training a separate classifier to predict zfeat as much as possible given only the user identity: θ∗ id = arg min θid (zfeat −fid(xid; θid))2 + λ||zid||1, (1) where xid denotes a one hot encoding of user identity as input, fid denotes the identity encoder with parameters θid, and λ denotes a hyperparameter that controls the weight of the ℓ1 regularizer. fid projects the user identity encodings to the feature space learned by ffeat. By minimizing the objective in equation (1) for each (x, xid) pair, fid learns to encode user identity into a sparse vector zid = fid(xid; θ∗ id) representing identity-dependent features: the nonzero values of zid represent dimensions of the identity-dependent subspace in zfeat, while the remaining dimensions belong to the 4174 Table 1: Comparison of mood prediction performance across different modalities. Best results in bold. For both accuracy and F1 score, models jointly trained on text, keystroke, and apps features outperform models trained using individual modalities. ⋆denotes that the difference between multimodal and all unimodal models is statistically significant (p-value << 0.05). F1 SCORE ACCURACY Modalities BASELINE SVM MLP NI-MLP BASELINE SVM MLP NI-MLP Text + Keystrokes + Apps 19.07 62.81⋆ 59.61⋆ 60.11⋆ 40.18 67.43⋆ 63.59⋆ 64.06⋆ Text + Keystrokes 19.07 61.19 57.65 58.70 40.18 65.87 61.81 62.61 Text + Apps 19.07 62.08 58.38 52.90 40.18 66.59 62.93 56.76 Text 19.07 61.15 56.27 52.63 40.18 65.83 60.61 56.08 Keystrokes 19.07 57.68 51.43 34.73 40.18 61.03 55.87 39.18 Apps 19.07 58.65 52.29 51.32 40.18 62.65 55.26 55.68 identity-independent subspace. (3) Addition phase: Given two factors zfeat and zid, to ensure that our prediction model does not capture identity-related information zid, we add multiplicative Gaussian noise to remove information from the identity-related subspace zid while repeatedly optimizing for mood prediction with a final MLP classification layer g(zfeat, zid; δ). This resulting model should only retain identity-independent features for mood prediction: ˆy = g (zfeat + ϵ ⊙zid) (2) where ϵ ∼N(0, σ2) is repeatedly sampled across batches and training epochs. We call this approach NOISY IDENTITY MLP, or NI-MLP for short, and summarize the final algorithm in Figure 2. Controlling the tradeoff between performance and privacy: There is often a tradeoff between privacy and prediction performance. To control this tradeoff, we vary the parameter σ, which is the variance of noise added to the identity-dependent subspace across batches and training epochs. σ = 0 recovers a standard MLP with good performance but reveals user identities, while large σ effectively protects user identities but at the possible expense of mood prediction performance. In practice, the optimal tradeoff between privacy and performance varies depending on the problem. For our purposes, we automatically perform model selection using this performance-privacy ratio R computed on the validation set, where R = sMLP −sNI-MLP tMLP −tNI-MLP (3) is defined as the improvement in privacy per unit of performance lost. Here, s is defined as the accuracy in user prediction and t is defined as the F1 score on mood prediction. 4 Experiments We perform experiments to test the utility of text, keystroke, and app features in predicting daily mood while keeping user privacy in mind. 4.1 Experimental Setup Data splits: Given that our data is longitudinal, we split our data into 10 partitions ordered chronologically by users. We do so in order to maintain independence between the train, validation, and test splits in the case where there is some form of time-level dependency within our labels. Evaluation: For each model, we run a nested kfold cross-validation (i.e., we perform 9-fold validation within 10-fold testing). For each test fold, we identify the optimal parameter set as the one that achieves the highest mean validation score over the validation folds. To evaluate NI-MLP, we use the best performing MLP model for each test fold as our base classifier before performing privacypreserving learning. For all experiments, we report the test accuracy and macro F1 score because our classes are imbalanced. Given the low number of cross-validation folds, we use the Wilcoxon signedrank test (Wilcoxon, 1992) at 5% significance level for all statistical comparisons (see Appendix C for more experimental details). 4.2 Results on Mood Prediction We make the following observations regarding the learned language and multimodal representations for mood prediction: Observation 1: Text, keystroke, and app usage features are individually predictive of mood. To evaluate how predictive our extracted text, keystroke timings, and app usage features are, we first run experiments using SVM, MLP, and NIMLP on each individual feature separately. Since we have unbalanced classes, we chose a majority classifier (i.e., most common class in the training 4175 Table 2: Mood prediction from text using extended pretrained LM encoders. We find that these models struggle on extremely long contexts of typed text. Models F1 SCORE ACCURACY BoW 56.27 60.61 BERT 51.42 58.06 XLNet 19.85 42.40 LongFormer 19.85 42.40 set) as our baseline. From Table 1, we observe that using these three feature types individually outperforms the baseline with respect to accuracy and F1 score. Using the Wilcoxon signed-rank test (Wilcoxon, 1992) at 5% significance level, we found that these improvements over the baseline in both F1 score and accuracy are statistically significant (p-value << 0.05). Observation 2: Pretrained sentence encoders struggle on this task. We also applied pretrained sentence encoders such as BERT (Devlin et al., 2019) on the language modality for mood prediction. Surprisingly, we found that none of these approaches performed stronger than a simple bagof-words (see Table 2). We provide two possible explanations for this phenomenon: 1. BERT is suitable for written text on the web (Wikipedia, BookCorpus, carefully humanannotated datasets) which may not generalize to informal typed text that contains emojis, typos, and abbreviations (see Section 4.4 for a qualitative analysis regarding the predictive abilities of emojis and keystrokes for mood prediction). 2. We hypothesize that it is difficult to capture such long sequences of data (>1000 time steps) spread out over a day. Current work has shown that BERT struggles with long sequence lengths (Beltagy et al., 2020). We trained two extensions XLNet (Yang et al., 2019) and LongFormer (Beltagy et al., 2020) specifically designed to take in long-range context but found that they still underperform as compared to a simple bag-of-words approach. Observation 3: Fusing both text and keystroke timings improves performance. This dataset presents a unique opportunity to study representations of typed text as an alternative to conventionally studied written or spoken text. While the latter two use language alone, typed text includes keystroke features providing information about the timings of when each character was typed. In Table 1, we present some of our initial results in learning text and keystroke representations for mood Table 3: Mood prediction using a MLP from text and keystroke features tallied from (1) all characters, (2) a split between types of characters, as well as (3) aggregated across words. Modalities F1 SCORE ACCURACY Text 56.27 60.61 Text + Char keystrokes 57.65 61.81 Text + Split char keystrokes 57.32 61.21 Text + Word keystrokes 56.46 60.68 prediction and show consistent improvements over text alone. We further study the uniqueness of typed text by comparing the following baselines: 1. Text: bag-of-words only. 2. Text + char keystrokes: bag-of-words and bagof-timings across all characters. 3. Text + split char keystrokes: bag-of-words and bag-of-timings subdivided between 6 groups: alphanumeric characters, symbols, spacebar, enter, delete, and use of autocorrect. This baseline presents a more fine-grained decomposition of the typing speeds across different semantically related character groups. 4. Text + word keystrokes: bag-of-words and bagof-timings summed up over the characters in each word. This presents a more interpretable model to analyze the relationships between words and the distribution of their typing speeds. From Table 3, we observe that keystrokes accurately contextualize text, especially when using fine-grained keystroke distributions across individual characters. Other methods incorporating keystroke features are also all stronger than unimodal models. Different ways of representing keystrokes also provide different levels of interpretability regarding the relationships between words, characters, and keystrokes for mood prediction, which we qualitatively analyze in §4.4. Observation 4: Multimodal representation learning achieves the best performance. In Table 1, we also compare the performance of our models on combined (text + keystroke + apps) features versus the performance on each individual feature set. For both metrics, combining all features gives better performance over either subset. 4.3 Results on Preserving Privacy Despite these promising results in mood prediction, we ask an important question: Does the model capture user identities as an intermediate step towards predicting mood? To answer this question, we an4176 (a) MLP (without privacy-preserving) (b) NI-MLP (with privacy-preserving) Figure 3: Visualization of representations learned by (a) MLP and (b) NI-MLP, which have been reduced to two dimensions via t-SNE and colored by participant identity. Representations learned by NI-MLP are no longer separable by users which better preserves privacy. Table 4: We report user identity prediction performance from raw input data and find that identities are very easily revealed from text, keystrokes, and app usage. F1 SCORE ACCURACY Modalities SVM MLP SVM MLP Text 89.42 92.05 90.60 93.12 Keystrokes 91.36 87.04 90.98 87.15 Apps 85.68 87.49 90.91 92.00 alyze the privacy of raw mobile data and trained models. We then study our proposed method of learning privacy-preserving features to determine whether it can obfuscate user identity while remaining predictive of daily mood. How private is the mobile data? We evaluate how much the data reveal user identities by training predictive models with typed text, keystroke timings, and app usage as input and user identity as the prediction target. From Table 4, we observe that all modalities are very predictive of user identity (>87% accuracy), which further motivates the need to learn privacy-preserving features. We further note that identifiable information can be very subtle: while only 28/1000 words were named entities, it was possible to identify the user identity with >87% accuracy, which means that subtle word choice can be identify the user (similarly for apps and keystrokes). How private are the learned privacy-preserving features? We also study whether our learned features are correlated with user identity through both visualizations and quantitative evaluations. Visualizations: We use t-SNE (Van der Maaten and Hinton, 2008) to reduce the learned features from trained models to 2 dimensions. After color-coding the points by participant identity, we identify distinct clusters in Figure 3(a), which implies that mood prediction can be strongly linked to identiTable 5: Comparison of our privacy-preserving approach (NI-MLP) with the baseline (MLP). We evaluate privacy in predicting user identity from learned representations (lower accuracy is better), and find that NI-MLP effectively obfuscates user identity while retaining performance. T: text, K: keystrokes, A: apps. PERFORMANCE (↑) PRIVACY (↓) Modalities MLP NI-MLP MLP NI-MLP T + K + A 59.61 58.48 71.47 34.49 T + K 57.65 57.40 64.17 30.99 T + A 58.38 57.76 79.04 65.13 T 56.27 54.11 76.41 52.20 K 51.43 42.48 55.61 25.71 A 52.29 49.15 85.94 66.74 fying the person, therefore coming at the price of losing privacy. As an attempt to reduce reliance on user identity, we train NI-MLP which is designed to obfuscate user-dependent features. After training NI-MLP, we again visualize the representations learned in Figure 3(b) and we find that they are less visually separable by users, indicating that NI-MLP indeed learns more user-independent features. Quantitative evaluation: To empirically evaluate how well our models preserve privacy, we extracted the final layer of each trained model and fit a logistic regression model to predict user identity using these final layer representations as input. The more a model preserves privacy, the harder it should be to predict user identity. From Table 5, we observe that we can predict user identity based on the learned MLP representations with high accuracy (>85%) using the most sensitive app usage features. For other modality combinations, user identity can also be decoded with more than 70% accuracy with the exception of keystrokes which are the most private (55%). We achieve significantly more privacy using NI-MLP embeddings - roughly 35% 4177 Figure 4: Tradeoff between performance (mood prediction F1 score, higher is better) and privacy (identity prediction accuracy, lower is better). Shaded regions denote standard deviations from the mean (solid lines). NI-MLP provides a tunable parameter σ to control the tradeoff, which allows us to plot a range of (performance, privacy) points. Using a multimodal model on text, keystroke, and app features obtains better performance and privacy at the same time. for the best multimodal model, which indicates the possibility of NI-MLP as a means of achieving privacy-preserving mood prediction. Understanding the tradeoff between performance and privacy: NI-MLP provides a tunable parameter σ to control the variance of noise applied on the identity-related dimensions. This parameter σ has the potential to give a tradeoff between privacy and prediction performance. In Figure 4, we plot this tradeoff between performance (mood prediction F1 score, higher is better) and privacy (identity prediction accuracy, lower is better). We find that keystroke features, while themselves not very useful in predicting mood, are highly private features. It is important to note that keystroke features show strong performance when integrated with text and app usage features while also increasing privacy, thereby pushing the Pareto front outwards. It is also interesting to observe that for most models, performance stays level while privacy improves, which is a promising sign for the real-world deployment of such models which requires a balance between both desiderata. 4.4 Qualitative Analysis To further shed light on the relationships between mood prediction performance and privacy, we performed a more in-depth study of the text, keystroke, and app usage features learned by the model (see Appendix D.3 for more examples). Table 6: Top emojis associated with positive and negative mood (each row is a different user). Positive emojis Negative emojis Table 7: Top 3 apps associated with positive and negative moods (each row is a different user). Top 3 positive apps Top 3 negative apps Photos, Settings, Snapchat Calendar, Wattpad, SoundCloud FaceTime, MyFitnessPal, Musically Notes, App Store, Siri Weather, Phone, FaceTime Chrome, App Store, SMS Weather, Phone, Spotify Safari, Notes, GroupMe Spotlight, App Store, Uber Pinterest, Phone, Yolo Uber, Netflix, LinkedIn Phone, Calendar, Safari Understanding the unimodal features: We first analyze how individual words, keystroke timings, and app usage are indicative of positive or negative mood for different users. Text: We find that several words are particularly indicative of mood: can’t/cant, don’t/don’t, and sorry are negative for more users than positive, while yes is overwhelmingly positive across users (9 pos, 1 neg), but yeah is slightly negative (5 pos, 7 neg). We also analyze the use of emojis in typed text and find that while there are certain emojis that lean positive (e.g., ), there are ones (e.g., :( and ) that used in both contexts depending on the user (see Table 6). Apps: In Table 7, we show the top 3 apps associated with positive or negative moods across several users. It is interesting to observe that many outdoor apps (i.e., Weather, MyFitnessPal, Uber), photo sharing apps (i.e., Photos, Snapchat), and calling apps (i.e., FaceTime, Phone) are associated with positive mood, while personal apps such as personal management (i.e., Calendar, Notes, Siri), web browsing (i.e., Chrome, Safari), and shopping (i.e., App Store) are associated with negative mood. However, some of these findings are rather userspecific (e.g., Phone can be both positive or negative depending on the user). Understanding the multimodal features: We also analyze how the same characters and words can contribute to different mood predictions based on their keystroke patterns. As an example, the distribution of keystrokes for the enter character on the keyboard differs according to the daily mood of one user (see Figure 5 and Appendix D.3 for 4178 Figure 5: An example where the ‘enter’ character keypress is indicative of either positive, neutral, or negative mood depending on the keypress duration. Table 8: Words with significantly different timings associated with positive and negative moods (each row is a different user). Slower implies positive Faster implies positive just why, thank, haha next, was, into, people making, work, idk stuff, cute, phone, want, talk, see they, send, dont, man, going don’t, talk think, you, all, love more users). In Table 8, we extend this analysis to entire words. For each of the 500 most common words, we aggregated their accompanying keystroke timings for user-reported positive and negative mood. These two distributions tell us how the same word in different keystroke contexts can indicate different moods. We performed Wilcoxon rank-sum tests at 5% significance level to compare these distributions and recorded the words in which either faster or slower typing was statistically significantly correlated with either mood. Observe how certain semantically positive words like love, thank, and haha become judged as more positive when typed at a faster speed. Therefore, contextualizing text with their keystroke timings offers additional information when learning representations of typed text. 5 Conclusion In this paper, we investigated the learning of language and multimodal representations of typed text collected from mobile data. We studied the challenge of learning markers of daily mood as a step towards early detection and intervention of mental health disorders for social good. Our method also shows promising results in obfuscating user identities for privacy-preserving learning, a direction crucial towards real-world learning from sensitive mobile data and healthcare labels. In addition, our findings illustrate several challenges and opportunities in representation learning from typed text as an understudied area in NLP. Limitations & future work: While our approach shows promises in learning representations for mood prediction, several future directions on the modeling and NLP side include: 1) better models and pre-training algorithms for NLP on typed text, 2) algorithms that provide formal guarantees of privacy (Dwork, 2008), and 3) federated training from decentralized data (McMahan et al., 2016) to improve privacy (Geyer et al., 2017) and fairness (Liang et al., 2020a) of sensitive data. We describe more limitations and future social implications of our work in our broader impact statement in Appendix A. Acknowledgements This material was based upon work partially supported by the National Science Foundation (Awards #1750439 and #1734868) and the National Institutes of Health (Award #U01MH116923). MM was supported by the Swiss National Science Foundation (#P2GEP2_184518). RS was supported by NSF IIS1763562 and ONR Grant N000141812861. Any opinions, findings, and conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation, National Institutes of Health, or Office of Naval Research, and no official endorsement should be inferred. We would also like to acknowledge NVIDIA’s GPU support and the anonymous reviewers for their extremely helpful comments. References Martin Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, Ilya Mironov, Kunal Talwar, and Li Zhang. 2016. Deep learning with differential privacy. In Proceedings of the 2016 ACM SIGSAC conference on computer and communications security, pages 308–318. Tadas Baltrušaitis, Chaitanya Ahuja, and LouisPhilippe Morency. 2018. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423– 443. Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. In NIPS. Bokai Cao, Lei Zheng, Chenwei Zhang, Philip S Yu, Andrea Piscitello, John Zulueta, Olu Ajilore, Kelly 4179 Ryan, and Alex D Leow. 2017. Deepmood: modeling mobile phone typing dynamics for mood detection. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 747–755. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2020. Extracting training data from large language models. arXiv preprint arXiv:2012.07805. CDC. 2015. Suicide Facts at a Glance 2015. Yiqiang Chen, Xin Qin, Jindong Wang, Chaohui Yu, and Wen Gao. 2020. Fedhealth: A federated transfer learning framework for wearable healthcare. IEEE Intelligent Systems. Chul-Hyun Cho, Taek Lee, Min-Gwan Kim, Hoh Peter In, Leen Kim, and Heon-Jeong Lee. 2019. Mood prediction of patients with mood disorders by machine learning using passive digital phenotypes based on the circadian rhythm: prospective observational cohort study. Journal of medical Internet research, 21(4):e11029. Glen Coppersmith, Ryan Leary, Patrick Crutchley, and Alex Fine. 2018. Natural language processing of social media as screening for suicide risk. Biomedical informatics insights, 10:1178222618792860. Sally C Curtin and Melanie P Heron. 2019. Death rates due to suicide and homicide among persons aged 10– 24: United states, 2000–2017. Fida Kamal Dankar and Khaled El Emam. 2012. The application of differential privacy to health data. In Proceedings of the 2012 Joint EDBT/ICDT Workshops, pages 158–166. Fida Kamal Dankar and Khaled El Emam. 2013. Practicing differential privacy in health care: A review. Trans. Data Priv., 6(1):35–67. Fida Kamal Dankar, Khaled El Emam, Angelica Neisa, and Tyson Roffey. 2012. Estimating the reidentification risk of clinical data sets. BMC medical informatics and decision making, 12(1):66. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Cynthia Dwork. 2008. Differential privacy: A survey of results. In International conference on theory and applications of models of computation, pages 1–19. Springer. Sindhu Kiranmai Ernala, Michael L Birnbaum, Kristin A Candan, Asra F Rizvi, William A Sterling, John M Kane, and Munmun De Choudhury. 2019. Methodological gaps in predicting mental health states from social media: triangulating diagnostic signals. In Proceedings of the 2019 CHI conference on human factors in computing systems, pages 1–16. Joseph C Franklin, Jessica D Ribeiro, Kathryn R Fox, Kate H Bentley, Evan M Kleiman, Xieyining Huang, Katherine M Musacchio, Adam C Jaroszewski, Bernard P Chang, and Matthew K Nock. 2017. Risk factors for suicidal thoughts and behaviors: a metaanalysis of 50 years of research. Psychological bulletin, 143(2):187. Robin C Geyer, Tassilo Klein, and Moin Nabi. 2017. Differentially private federated learning: A client level perspective. arXiv preprint arXiv:1712.07557. Surjya Ghosh, Niloy Ganguly, Bivas Mitra, and Pradipta De. 2017a. Evaluating effectiveness of smartphone typing as an indicator of user emotion. In 2017 Seventh International Conference on Affective Computing and Intelligent Interaction (ACII), pages 146–151. IEEE. Surjya Ghosh, Niloy Ganguly, Bivas Mitra, and Pradipta De. 2017b. Tapsense: Combining self-report patterns and typing characteristics for smartphone based emotion detection. In Proceedings of the 19th International Conference on Human-Computer Interaction with Mobile Devices and Services, pages 1–12. Catherine R Glenn and Matthew K Nock. 2014. Improving the short-term prediction of suicidal behavior. American journal of preventive medicine, 47(3):S176– S180. He Huang, Bokai Cao, S Yu Phillip, Chang-Dong Wang, and Alex D Leow. 2018. Dpmood: Exploiting local and periodic typing dynamics for personalized mood prediction. In 2018 IEEE International Conference on Data Mining (ICDM), pages 157–166. IEEE. Becky Inkster et al. 2021. Early warning signs of a mental health tsunami: A coordinated response to gather initial data insights from multiple digital services providers. Frontiers in Digital Health, 2:64. Natasha Jaques, Sara Taylor, Akane Sano, and Rosalind Picard. 2017. Multimodal autoencoder: A deep learning approach to filling in missing sensor data and enabling better mood prediction. In 2017 Seventh International Conference on Affective Computing and Intelligent Interaction (ACII), pages 202–208. IEEE. Michal Kosinski, David Stillwell, and Thore Graepel. 2013. Private traits and attributes are predictable from digital records of human behavior. Proceedings of the national academy of sciences, 110(15):5802–5805. Matthew Michael Large, Daniel Thomas Chung, Michael Davidson, Mark Weiser, and Christopher James Ryan. 2017. In-patient suicide: selection of people at risk, failure of protection and the possibility of causation. BJPsych Open, 3(3):102–105. Ellen E Lee, John Torous, Munmun De Choudhury, Colin A Depp, Sarah A Graham, Ho-Cheol Kim, Martin P Paulus, John H Krystal, and Dilip V Jeste. 2021. 4180 Artificial intelligence for mental healthcare: Clinical applications, barriers, facilitators, and artificial wisdom. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging. Paul Pu Liang, Irene Mengze Li, Emily Zheng, Yao Chong Lim, Ruslan Salakhutdinov, and LouisPhilippe Morency. 2020a. Towards debiasing sentence representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5502–5515, Online. Association for Computational Linguistics. Paul Pu Liang, Terrance Liu, Liu Ziyin, Ruslan Salakhutdinov, and Louis-Philippe Morency. 2020b. Think locally, act globally: Federated learning with local and global representations. arXiv preprint arXiv:2001.01523. Paul Pu Liang, Ziyin Liu, AmirAli Bagher Zadeh, and Louis-Philippe Morency. 2018. Multimodal language analysis with recurrent multistage fusion. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 150–161, Brussels, Belgium. Association for Computational Linguistics. Kirsten Lloyd. 2018. Bias amplification in artificial intelligence systems. CoRR, abs/1809.07842. Lingjuan Lyu, Han Yu, and Qiang Yang. 2020. Threats to federated learning: A survey. arXiv preprint arXiv:2003.02133. Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11). J John Mann, Dianne Currier, Barbara Stanley, Maria A Oquendo, Lawrence V Amsel, and Steven P Ellis. 2006. Can biological tests assist prediction of suicide in mood disorders? International Journal of Neuropsychopharmacology, 9(4):465–474. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. William Marslen-Wilson and Lorraine Komisarjevsky Tyler. 1980. The temporal structure of spoken language understanding. Cognition, 8(1):1–71. H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2016. Communication-efficient learning of deep networks from decentralized data. In AISTATS. Inbal Nahum-Shani, Shawna N Smith, Bonnie J Spring, Linda M Collins, Katie Witkiewitz, Ambuj Tewari, and Susan A Murphy. 2018. Just-in-time adaptive interventions (jitais) in mobile health: key components and design principles for ongoing health behavior support. Annals of Behavioral Medicine, 52(6):446–462. Yaakov Ophir, Refael Tikochinski, Christa SC Asterhan, Itay Sisso, and Roi Reichart. 2020. Deep neural networks detect suicide risk from textual facebook posts. Scientific reports, 10(1):1–10. Maria A Oquendo, Hanga C Galfalvy, Tse-Hwei Choo, Raksha Kandlur, Ainsley K Burke, M Elizabeth Sublette, Jeffrey M Miller, J John Mann, and Barbara H Stanley. 2020. Highly variable suicidal ideation: a phenotypic marker for stress induced suicide risk. Molecular psychiatry, pages 1–8. Jahna Otterbacher, Alessandro Checco, Gianluca Demartini, and Paul Clough. 2018. Investigating user perception of gender bias in image search: The role of sexism. In The 41st International ACM SIGIR Conference on Research Development in Information Retrieval, SIGIR ’18, page 933–936, New York, NY, USA. Association for Computing Machinery. Philip Resnik, April Foreman, Michelle Kuchuk, Katherine Musacchio Schafer, and Beau Pinkham. 2021. Naturally occurring language as a source of evidence in suicide prevention. Suicide and LifeThreatening Behavior, 51(1):88–96. Mina M Rizk, Tse-Hwei Choo, Hanga Galfalvy, Emily Biggs, Beth S Brodsky, Maria A Oquendo, J John Mann, and Barbara Stanley. 2019. Variability in suicidal ideation is associated with affective instability in suicide attempters with borderline personality disorder. Psychiatry, 82(2):173–178. Koustuv Saha, John Torous, Eric D Caine, and Munmun De Choudhury. 2020. Psychosocial effects of the covid-19 pandemic: Large-scale quasi-experimental study on social media. Journal of medical Internet research, 22(11):e22600. Akane Sano, Sara Taylor, Andrew W McHill, Andrew JK Phillips, Laura K Barger, Elizabeth Klerman, and Rosalind Picard. 2018. Identifying objective physiological markers and modifiable behaviors for selfreported stress and mental health status using wearable sensors and mobile phones: observational study. Journal of medical Internet research, 20(6):e210. Allison Schuck, Raffaella Calati, Shira Barzilay, Sarah Bloch-Elkouby, and Igor Galynker. 2019. Suicide crisis syndrome: A review of supporting evidence for a new suicide-specific diagnosis. Behavioral sciences & the law, 37(3):223–239. Virginia Smith, Chao-Kai Chiang, Maziar Sanjabi, and Ameet S Talwalkar. 2017. Federated multi-task learning. In Advances in Neural Information Processing Systems, pages 4424–4434. Yoshihiko Suhara, Yinzhan Xu, and Alex’Sandy’ Pentland. 2017. Deepmood: Forecasting depressed mood based on self-reported histories via recurrent neural networks. In Proceedings of the 26th International Conference on World Wide Web, pages 715–724. Sara Ann Taylor, Natasha Jaques, Ehimwenma Nosakhare, Akane Sano, and Rosalind Picard. 2017. Personalized multitask learning for predicting tomorrow’s mood, stress, and health. IEEE Transactions on Affective Computing. 4181 Graham Thornicroft, Nisha Mehta, Sarah Clement, Sara Evans-Lacko, Mary Doherty, Diana Rose, Mirja Koschorke, Rahul Shidhaye, Claire O’Reilly, and Claire Henderson. 2016. Evidence for effective interventions to reduce mental-health-related stigma and discrimination. The Lancet, 387(10023):1123–1132. Haohan Wang, Aaksha Meghawat, Louis-Philippe Morency, and Eric P Xing. 2017. Select-additive learning: Improving generalization in multimodal sentiment analysis. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 949–954. IEEE. Frank Wilcoxon. 1992. Individual comparisons by ranking methods. In Breakthroughs in statistics, pages 196–202. Springer. World Health Organization. 2003. Investing in mental health. Jie Xu and Fei Wang. 2019. Federated learning for healthcare informatics. arXiv preprint arXiv:1911.06270. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in Neural Information Processing Systems, 32:5753–5763. Han Zhao and Geoff Gordon. 2019. Inherent tradeoffs in learning fair representations. In Advances in Neural Information Processing Systems, volume 32, pages 15675–15685. Curran Associates, Inc. Ligeng Zhu and Song Han. 2020. Deep leakage from gradients. In Federated Learning, pages 17–31. Springer. John Zulueta, Andrea Piscitello, Mladen Rasic, Rebecca Easter, Pallavi Babu, Scott A Langenecker, Melvin McInnis, Olusola Ajilore, Peter C Nelson, Kelly Ryan, et al. 2018. Predicting mood disturbance severity with mobile phone keystroke metadata: a biaffect digital phenotyping study. Journal of medical Internet research, 20(7):e241. 4182 Appendix A Broader Impact Statement Learning markers of mood from mobile data presents an opportunity for large-scale adaptive interventions of suicidal ideation. However, there are important concerns regarding its implications to society and policy. Applications in mental health: Suicide is the second leading cause of death among adolescents. In addition to deaths, 16% of high school students report seriously considering suicide each year, and 8% make one or more suicide attempts (CDC, 2015). Despite these alarming statistics, there is little consensus concerning imminent risk for suicide (Franklin et al., 2017; Large et al., 2017). Current research conducts clinical interviews and patient self-report questionnaires that provide longterm assessments of suicide risk. However, few studies have focused on imminent suicidal risk, which is of critical clinical importance as a step towards adaptive real-time interventions (Glenn and Nock, 2014; Schuck et al., 2019). Given the impact of suicide on society, there is an urgent need to better understand the behavior markers related to suicidal ideation. “Just-in-time” adaptive interventions delivered via mobile health applications provide a platform of exciting developments in low-intensity, high-impact interventions (Nahum-Shani et al., 2018). The ability to intervene precisely during an acute risk for suicide could dramatically reduce the loss of life. To realize this goal, we need accurate and timely methods that predict when interventions are most needed. Monitoring (with participants’ permission) mobile data to assess mental health and provide early interventions is, therefore, a rich opportunity for scalable deployment across high-risk populations. Our data collection, experimental study, and computational approaches provide a step towards data-intensive longitudinal monitoring of human behavior. However, one must take care to summarize behaviors from mobile data without identifying the user through personal (e.g., personally identifiable information) or protected attributes (e.g., race, gender). This form of anonymity is critical when implementing these technologies in real-world scenarios. Our goal is to be highly predictive of mood while remaining as privacy-preserving as possible. We outline some of the potential privacy and security concerns below. Limitations: While we hope that our research can provide a starting point on the potential of detecting mood unobtrusively throughout the day in a privacy-preserving way, we strongly acknowledge there remain methodological issues where a lot more research needs to be done to enable the realworld deployment of such technologies. We emphasize that healthcare providers and mobile app startups should not attempt to apply our approach in the real world until the following issues (and many more) can be reliably resolved: 1. We do not make broad claims across teenage populations from only 17 participants in this study. Furthermore, it remains challenging for models to perform person-independent prediction which makes it hard to deploy across large populations. 2. Our current work on predicting daily mood is still a long way from predicting imminent suicide risk. Furthermore, any form of prediction is still significantly far away from integrating methods like this into the actual practice of mental health, which is a challenging problem involving a broad range of medical, ethical, social, and technological researchers (Resnik et al., 2021; Lee et al., 2021). 3. Text and keystrokes can differ for participants who speak multiple languages or non-prestige vernaculars. One will need to ensure that the method works across a broad range of languages to ensure accessibility in its desired outcomes. 4. This study assumes that participants have no restrictions for data/network connections & data plans on their phones, which may leave out vulnerable populations that do not meet this criterion. Privacy and security: There are privacy risks associated with making predictions from mobile data. To deploy these algorithms across at-risk populations, it is important to keep data private on each device without sending it to other locations. Even if data is kept private, it is possible to decode data from gradients (Zhu and Han, 2020) or pretrained models (Carlini et al., 2020). In addition, sensitive databases with private mobile data could be at-risk to external security attacks from adversaries (Lyu et al., 2020). Therefore, it is crucial to obtain user consent before collecting device data. In our exper4183 iments with real-world mobile data, all participants have given consent for their mobile device data to be collected and shared with us for research purposes. All data was anonymized and stripped of all personal (e.g., personally identifiable information) and protected attributes (e.g., race, gender). Social biases: We acknowledge that there is a risk of exposure bias due to imbalanced datasets, especially when personal mobile data and sensitive health labels (e.g., daily mood, suicidal thoughts and behaviors, suicide risk). Models trained on biased data have been shown to amplify the underlying social biases especially when they correlate with the prediction targets (Lloyd, 2018). This leaves room for future work in exploring methods tailored for specific scenarios such as mitigating social biases in words (Bolukbasi et al., 2016), sentences (Liang et al., 2020a), and images (Otterbacher et al., 2018). Future research should also focus on quantifying the trade-offs between fairness and performance (Zhao and Gordon, 2019). Overall, we believe that our proposed approach can help quantify the tradeoffs between performance and privacy. We hope that this brings about future opportunities for large-scale real-time analytics in healthcare applications. B Dataset Details The Mobile Assessment for the Prediction of Suicide (MAPS) dataset was designed to elucidate real-time indicators of suicide risk in adolescents ages 13 −18 years. Current adolescent suicide ideators and recent suicide attempters along with aged-matched psychiatric controls with no lifetime suicidal thoughts and behaviors completed baseline clinical assessments (i.e., lifetime mental disorders, current psychiatric symptoms). Following the baseline clinical characterization, a smartphone app, the Effortless Assessment of Risk States (EARS), was installed onto adolescents’ phones, and passive sensor data were acquired for 6-months. Notably, during EARS installation, a keyboard logger is configured on adolescents’ phones, which then tracks all words typed into the phone as well as the apps used during this period. Each day during the 6month follow-up, participants also were asked to rate their mood on the previous day on a scale ranging from 1 −100, with higher scores indicating a better mood. After extracting multimodal features and discretizing the labels (see Section 2), we summarize the final dataset feature and label statistics in Table 9. C Experimental Setup We provide additional details on the model implementation and experimental setup. C.1 Implementation Details All models and analyses were done in Python. SVM models were implemented with Scikitlearn and MLP/NI-MLP models were implemented with PyTorch. BERT, XLNet, and Longformer models were fine-tuned using Hugging Face (website: https://huggingface.co, GitHub: https://github.com/huggingface). C.2 Hyperparameters We performed a small hyperparameter search over the ranges in Table 10. This resulted in a total of 35 hyperparameter configurations for SVM and 12 for MLP (6 for apps only). By choosing the best-performing model on the validation set, we selected the resulting hyperparameters as shown in Table 10. C.3 Model Parameters Each model has about two million parameters. See Table 10 for exact hidden dimension sizes. C.4 Training Resources and Time All experiments were conducted on a GeForce RTX 2080 Ti GPU with 12 GB memory. See Table 11 for approximate running times. D Experimental Details We present several additional analysis of the data and empirical results: D.1 Details on Mood Prediction There is often a tradeoff between privacy and prediction performance. To control this tradeoff, we vary the parameter σ, which is the amount of noise added to the identity-dependent subspace across batches and training epochs. In practice, we automatically perform model selection using this performance-privacy ratio R computed on the validation set, where R = sMLP −sNI-MLP tMLP −tNI-MLP (4) is defined as the improvement in privacy per unit of performance lost. Here, s is defined as the accuracy in the user prediction task and t is defined as the F1 score on the mood prediction task. 4184 Table 9: Mobile Assessment for the Prediction of Suicide (MAPS) dataset summary statistics. Users Datapoints Modalities Features Dimensions Labels 17 1641 Text bag-of-words, one-hot 2000 Daily mood: negative, neutral, positive Keystrokes bag-of-timings 100 App usage bag-of-apps, one-hot 274 Table 10: Model parameter configurations. *Integer kernel values denote the degree of a polynomial kernel. Model Parameter Value SVM C 0.1, 0.5, 1, 2, 3, 5, 10 Kernel* RBF, 2, 3, 5, 10 MLP hidden dim 1 (multimodal & text only) 1024, 512 hidden dim 2 (multimodal & text only) 128, 64 hidden dim 1 (keystrokes only) 64, 32 hidden dim 2 (keystrokes only) 32, 16 hidden dim 1 (apps only) 128 hidden dim 2 (apps only) 128, 64 dropout rate 0, 0.2, 0.5 learning rate 0.001 batch size 100 epochs 200 NI-MLP λ 0.1, 1, 2, 3, 5, 10 σ 1, 5, 10, 25, 50, 100, 150 Table 11: Approximate training times (total across 10-fold cross validation and hyperparameter search). Model Modality Time (hours) SVM Text + Keystrokes + Apps 10 Text + Keystrokes 10 Text + Apps 10 Text 8 Keystrokes 1 Apps 1 MLP (100 epochs, 3 runs) Text + Keystrokes + Apps 6 Text + Keystrokes 5 Text + Apps 6 Text 5 Keystrokes 4 Apps 2 NI-MLP all 4 In the rare cases where NI-MLP performed better than the original MLP and caused R to become negative, we found this improvement in performance always came at the expense of worse privacy as compared to other settings of λ and σ in NI-MLP. Therefore, models with negative R were not considered for Table 1. D.2 Details on Preserving Privacy For Table 5, the model with the best privacy out of those within 5% performance of the original MLP model (or, if no such model existed, the model with the best performance) was selected. Interestingly, in Figure 4, we find that the tradeoff curve on a model trained only using app features does not exhibit a Pareto tradeoff curve as expected. We attribute this to randomness in predicting both mood and identities. Furthermore, Wang et al. (2017) found that adding noise to the identity subspace can sometimes improve generalization by reducing reliance on identity-dependent confounding features, which could also explain occasional increased performance at larger σ values. Note that we do not include privacy results for features learned by SVM, which finds a linear separator in a specified kernel space rather than learning a representation for each sample. Explicitly projecting our features is computationally infeasible due to the high dimensionality of our chosen kernel spaces. 4185 Table 12: Top 5 words associated with positive and negative moods (each row is a different user). Top 5 positive words Top 5 negative words hot, goodnight, ft, give, keep soon, first, ya, friend, leave still, y’all, guys, new, come amazing, see, said, idk, look mind, days, went, tf, next tired, hair, stg, snap, anyone girls, music, happy, mean, getting omg, people, talking, ask, might Table 13: Top words associated with positive and negative moods across users. We find that while certain positive words are almost always indicative of mood, others are more idiosyncratic and depend on the user. Positive words Positive users Negative users Negative words Negative users Positive users make 9 1 i’m/im 10 5 yes 9 1 feel 7 3 got 7 1 yeah 7 5 still 7 1 can’t/cant 6 2 wanna 7 1 people 6 4 like 7 2 know 6 4 need 7 2 go 6 5 send 7 2 one 6 6 get 7 2 today 5 1 good 7 3 day 5 2 D.3 Qualitative Analysis In this section, we provide more empirical analysis on the unimodal and multimodal features in the MAPS dataset. D.3.1 Understanding the unimodal features Text: We begin with some basic statistics regarding word distributions. For each user, we tallied the frequencies of each word under each daily mood category (positive, neutral, and negative), as well as the overall number of words in each mood category. We define “positive” words and emojis to be those with a higher relative frequency of positive mood compared to the overall positive mood frequency, and lower than overall negative mood frequency. Likewise, “negative” words and emojis have higher than overall negative mood frequency and lower than overall positive mood frequency. We filtered out words for specific users if the word was used less than 40 times. Finally, we ranked the words by the difference in relative frequency (i.e., a word is “more positive” the larger the difference between its positive mood relative frequency and the user’s overall positive mood relative frequency). See Table 12 for examples of top positive and negative words. For each word, we also counted the number of users for which the word was positive or negative. See Table 13 for the words with the highest user counts. Keystrokes: We show some sample bag-of-timing histograms in Figure 6. It is interesting to find that certain users show a bimodal distribution across their keystroke histograms with one peak representing faster typing and another representing slower typing. Visually, the overall keystroke histograms did not differ that much across users which might explain its lower accuracies in both mood and user prediction when trained with NI-MLP (see Figure 4). App usage: Similar to “positive” words, we define “positive” apps to be those with higher than overall positive mood relative frequency and lower than overall negative mood relative frequency, and “negative” apps to be the opposite. Apps were also then sorted by difference in relative frequency. D.3.2 Understanding the multimodal features Characters with keystrokes: For each user, we plotted histograms of keystroke timings of alphanumeric characters, symbols (punctuation and emojis), spacebar, enter, delete, and use of autocorrect, split across daily mood categories. See Figure 7 for examples across one user. We find particularly interesting patterns in the autocorrect keys and symbols where keystrokes are quite indicative of mood, which attests to the unique nature of typed text. Words with keystrokes: For each user, we plotted histograms of the word-level keystroke timings of the top 500 words, split across the daily mood categories of positive, neutral, and negative. We also performed Wilcoxon rank-sum tests at 5% signifi4186 Figure 6: Examples of keystroke timing histograms for different users. We find that the distribution of keystroke timings varies between unimodal and bimodal for different users. Figure 7: Example of more character key-presses and how their keystroke patterns can be indicative of either positive, neutral, or negative mood. We find particularly interesting patterns in the autocorrect keys and symbols where keystrokes are quite indicative of mood. 4187 cance level (Wilcoxon, 1992) between the timings of positive and negative mood for each user/word combination to determine which words had significantly different timings between positive and negative mood. E Negative Results and Future Directions Since this is a new dataset, we explored several more methods throughout the research process. In this section we describe some of the approaches that yielded initial negative results despite them working well for standard datasets: 1. User specific models: We also explored the setting of training a separate model per user but we found that there was too little data per user to train a good model. As part of future work, we believe that if NI-MLP can learn a user-independent classifier, these representations can then be used for further finetuning or few-shot learning on each specific user. Previous work in federated learning (Smith et al., 2017; Liang et al., 2020b) offers ways of learning a user-specific model that leverages other users’ data during training, which could help to alleviate the lack of data per user. 2. User-independent data splits: We have shown that text, keystrokes, and app usage features are highly dependent on participant identities. Consequently, models trained on these features would perform poorly when evaluated on a user not found in the training set. We would like to evaluate if better learning of user-independent features can improve generalization to new users (e.g., split the data such that the first 10 users are used for training, next 3 for validation, and final 4 for testing). Our initial results for these were negative, but we believe that combining better privacy-preserving methods that learn user-independent features could help in this regard. 3. Fine-grained multimodal fusion: Our approach of combining modalities was only at the input level (i.e., early fusion (Baltrušaitis et al., 2018)) which can be improved upon by leveraging recent work in more fine-grained fusion (Liang et al., 2018). One such example could be to align each keystroke feature and app data to the exact text that was entered in, which provides more finegrained contextualization of text in keystroke and app usage context.
|
2021
|
322
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4188–4203 August 1–6, 2021. ©2021 Association for Computational Linguistics 4188 Anonymisation Models for Text Data: State of the Art, Challenges and Future Directions Pierre Lison1, Ildik´o Pil´an1, David S´anchez2, Montserrat Batet2, and Lilja Øvrelid3 1Norwegian Computing Center, Oslo, Norway 2Universitat Rovira i Virgili, CYBERCAT, UNESCO Chair in Data Privacy, Spain 3Language Technology Group, University of Oslo, Norway {plison,pilan}@nr.no {david.sanchez,montserrat.batet}@urv.cat liljao@ifi.uio.no Abstract This position paper investigates the problem of automated text anonymisation, which is a prerequisite for secure sharing of documents containing sensitive information about individuals. We summarise the key concepts behind text anonymisation and provide a review of current approaches. Anonymisation methods have so far been developed in two fields with little mutual interaction, namely natural language processing and privacy-preserving data publishing. Based on a case study, we outline the benefits and limitations of these approaches and discuss a number of open challenges, such as (1) how to account for multiple types of semantic inferences, (2) how to strike a balance between disclosure risk and data utility and (3) how to evaluate the quality of the resulting anonymisation. We lay out a case for moving beyond sequence labelling models and incorporate explicit measures of disclosure risk into the text anonymisation process. 1 Introduction Privacy is a fundamental human right (Art. 12 of the Universal Declaration of Human Rights) and a critical component of any free society, among others to protect citizens against social control, stigmatisation, and threats to political expression. Privacy is also protected by multiple national and international legal frameworks, such as the General Data Protection Regulation (GDPR) introduced in Europe in 2018. This right to privacy imposes constraints on the usage and distribution of data including personal information, such as emails, court cases or patient records. In particular, personal data cannot be distributed to third parties (or even used for secondary purposes) without legal ground, such as the explicit and informed consent of the individuals to whom the data refers. As informed consent is often difficult to obtain in practice, an alternative is to rely on anonymisation techniques that render personal data no longer personal. Access to anonymised data is a prerequisite for research advances in many scientific fields, notably in medicine and the social sciences. By facilitating open data initiatives, anonymised data can also help empower citizens and support democratic participation. For structured databases, anonymisation can be enforced through well-established privacy models such as k-anonymity (Samarati, 2001; Samarati and Sweeney, 1998) or differential privacy (Dwork et al., 2006). These privacy models and their implementations are, however, difficult to apply to unstructured data such as texts. In fact, text anonymisation has been traditionally enforced manually, a process that is costly, timeconsuming and prone to errors (Bier et al., 2009). These limitations led to the development of various computational frameworks designed to extend automated or semi-automated anonymisation to the text domain (Meystre et al., 2010; S´anchez and Batet, 2016; Dernoncourt et al., 2017). In this paper, we review the core concepts underlying text anonymisation, and survey the approaches put forward to solve this task. These can be divided into two independent research directions. On the one hand, NLP approaches rely on sequence labelling to detect and remove predefined categories of entities that are considered sensitive or of personal nature (such as names, phone numbers or medical conditions). On the other hand, privacy-preserving data publishing (PPDP) approaches take the notion of disclosure risk as starting point and anonymise text by enforcing a privacy model. Anonymisation consists of a sequence of transformations (such as removal or generalisation) on the document to ensure the requirements derived from the privacy model are fulfilled. This position paper makes the case that none of these approaches provide a fully satisfactory account of the text anonymisation problem. We 4189 illustrate their merits and shortcomings on a case study and discuss three open challenges: 1. How to ensure that anonymisation is robust against multiple types of semantic inferences, based on background knowledge assumed to be available to an adversary ; 2. How to transform the text in order to minimise the risk of disclosing personal data, yet retain as much semantic content as possible ; 3. How to empirically evaluate the quality (in terms of disclosure risk and utility preservation) of the resulting anonymisation. We argue in this paper that NLP and PPDP approaches should be viewed as complementary (one focusing on linguistic patterns, the other on disclosure risk) and that future anonymisation approaches for text should seek to reconcile these two views. In particular, we contend that text anonymisation models should combine a data-driven editor model (which selects masking operations on the document) with an adversary seeking to infer confidential attributes from edited documents. 2 What is Anonymisation? The most common definition of privacy amounts to self-determination, which is the ability of individuals, groups or organisations to seclude information about themselves selectively (Westin, 1967). Information related to an identified or identifiable person is known as personal data, or more precisely personally identifiable information (PII). Datasets with PII cannot be released without control as this would impair the privacy of the data subjects. 2.1 Legal Requirements Various legal frameworks regulate how PII can be collected and processed. In particular, the General Data Protection Regulation introduced in Europe (GDPR, 2016) states that data owners must have a legal basis for processing PII, the most important one being the explicit consent of the data subjects.Alternatively, data owners may choose to anonymise the data to ensure it can no longer be attributed to specific individuals. Anonymised data is no longer regulated by the GDPR and can therefore be freely released. Table 1 defines some of the key terms related to data anonymisation (Elliot et al., 2016). This terminology is, however, not always applied consistently, as several authors seem to use e.g. the Direct Identifier: A (set of) variable(s) unique for an individual (a name, address, phone number or bank account) that may be used to directly identify the subject. Quasi Identifier: Information (such as gender, nationality, or city of residence) that in isolation does not enable re-identification, but may do so when combined with other quasiidentifiers and background knowledge. Confidential Attribute: Private personal information that should not be disclosed (such as a medical condition). Identity Disclosure: Unequivocal association of a record/document with a subject’s identity. Attribute disclosure: Unequivocal inference of a confidential attribute about a subject. Anonymisation: Complete and irreversible removal from a dataset of any information that, directly or indirectly, may lead to a subject’s data being identified. De-identification: Process of removing specific, predefined direct identifiers from a dataset. Pseudonymisation: Process of replacing direct identifiers with pseudonyms or coded values (such ”John Doe” →”Patient 3”). The mapping between coded values and the original identifiers is then stored separately. Table 1: Key terms related to data anonymisation. terms “anonymisation” and “de-identification” interchangeably (Chevrier et al., 2019). GDPR-compliant anonymisation is the complete and irreversible process of removing personal identifiers, both direct and indirect, that may lead to an individual being identified. Direct identifiers correspond to values such as names or social security numbers that directly disclose the identity of the individual. However, removing direct identifiers is not sufficient to eliminate all disclosure risks, as individuals may also be re-identified by combining several pieces of information together with some background knowledge. For instance, the combination of gender, birth date and postal code can be exploited to identify between 63 and 87% of the U.S. population, due to the public availability of US Census Data (Golle, 2006). These types of personal identifiers are called quasi-identifiers and encompass a large variety of data types such as 4190 demographic and geospatial data. Anonymisation therefore necessitates both the removal of direct identifiers and the masking of quasi-identifiers. Other legal frameworks have adopted a different approach. In the US, the Health Insurance Portability and Accountability Act (HIPAA) (HIPAA, 2004) lists 18 data types, such as patient’s name, address or social security number, which qualify as protected health information (PHI) and should be removed from the data prior to release. This process of removing predefined categories of identifiers is called de-identification1. In other words, while HIPAA-based de-identification is limited to specific categories of direct identifiers, the anonymisation process defined by GDPR requires us to consider any direct or indirect information that, combined with background knowledge, may lead to re-identifying an individual. The California Consumer Privacy Act (CCPA) introduced in 2018 adopts a position relatively similar to GDPR regarding anonymisation and asserts that any data that can be linked directly or indirectly to a consumer must be considered as personal information. We highlight these legal differences as they have important implications on how anonymisation tools should be designed and evaluated (Rothstein, 2010; Hintze, 2017). In particular, GDPR- or CCPAcompliant anonymisation cannot be restricted to the detection of predefined classes of entities but must consider how any textual element may contribute to the disclosure risk, either directly or through semantic inferences using the background knowledge assumed to be available to an adversary. 2.2 Disclosure Risks Legal regulations for privacy and data protection (such as GDPR and HIPAA) typically focus on identity disclosure. However, personal information may also be disclosed without re-identification. In particular, attribute disclosure occurs when the value of a confidential attribute (e.g., a medical condition) can be inferred from the released data, for instance when all records sharing some characteristics (e.g. age) have the same confidential value (e.g. suffering from AIDS). Identity disclosure can be seen as a special case of attribute disclosure when the confidential attribute corresponds to the person identity. Data anonymisation should prevent identity disclosure but, in most cases, attribute 1GDPR also introduces the equivalent concept of pseudonymisation, which is a useful privacy-enhancing measure, but it does not qualify as full anonymisation. disclosure, which is usually more harmful from a privacy perspective, should also be avoided. The removal of personal information necessarily entails some data utility loss. Because the ultimate purpose behind data releases is to produce usable data, the best anonymisation methods are those that optimise the trade-off between minimising the disclosure risk and preserving the data utility. 3 NLP Approaches 3.1 De-identification NLP research on text anonymisation has focused to a large extent on the tasks of de-identification, and, to a lesser extent, pseudonymisation. Deidentification is generally modelled as a sequence labelling task, similar to Named Entity Recognition (NER) (Chiu and Nichols, 2016; Lample et al., 2016). Most work to date has been performed in the area of clinical NLP, where the goal is to detect Protected Health Information (PHI) in clinical texts (Meystre et al., 2010; Aberdeen et al., 2010). Several shared tasks have contributed to increased activity within this area, in particular through the release of datasets manually annotated with PHIs. The 2014 i2b2/UTHealth shared task (Stubbs and Uzuner, 2015) includes diabetic patient medical records annotated for an extended set of PHI categories. Another influential dataset stems from the 2016 CEGS N-GRID shared task (Stubbs et al., 2017) based on psychiatric intake records, which are particularly challenging to de-identify due to a higher density of PHIs. Early approaches to this task were based on rulebased and machine learning-based methods, either alone or in combination (Yogarajan et al., 2018). Dernoncourt et al. (2017) and Liu et al. (2017) present the first neural models for de-identification using recurrent neural networks with characterlevel embeddings, achieving state-of-the-art performance on the i2b2 2014 dataset. A central challenge in clinical de-identification is the availability of annotated data and the lack of universal annotation standards for PHI, making it difficult to transfer data across domains. Hartman et al. (2020) examine how to adapt de-identification systems across clinical sub-domains. They compare the use of labelled or unlabelled data for domain adaptation with in-domain testing and off-theshelf de-identification tools, and show that manual labelling of even small amounts of PHI examples yields performance above existing tools. 4191 Further, embeddings trained on larger amounts of in-domain, unlabelled data can be employed to adapt models to a new domain (Yang et al., 2019). Finally, Friedrich et al. (2019) present an adversarial approach for learning privacy-preserving text representations, thereby allowing data to be more easily shared to train de-identification tools. Outside of the clinical domain, Medlock (2006) presents a dataset of e-mails annotated with both direct identifiers (person names, transactional codes, etc.) and quasi-identifiers (organisations, course names, etc.). Some annotation efforts are also geared towards de-identification for languages other than English. Eder et al. (2020) present a deidentification dataset consisting of German e-mails. For Swedish, Velupillai et al. (2009); Alfalahi et al. (2012) present efforts to collect and standardise annotated clinical notes, while Megyesi et al. (2018) present a pseudonymised learner language corpus. For Spanish, a recently held shared task on clinical de-identification released a synthetic Spanishlanguage dataset (Marimon et al., 2019). The problem of replacing identifiers with surrogate values is rarely addressed in NLP. Most approaches simply replace detected identifiers with dummy values such as X, although some models attempt to preserve the gender of person names and provide dedicated rules for e.g. dates and addresses (Sweeney, 1996; Alfalahi et al., 2012; Eder et al., 2019; Chen et al., 2019) or to a somewhat broader range of identifiers (Volodina et al., 2020). A few studies have analysed the re-identification risk of de-identified or pseudonymised texts (Carrell et al., 2013; Meystre et al., 2014b). The data utility of de-identified texts is analysed in Meystre et al. (2014a), concluding that the impact of deidentification is small, but non-negligible. 3.2 Obfuscation Methods Beyond de-identification, several research efforts have looked at detecting and obfuscating social media texts based on quasi-identifying categories such as gender (Reddy and Knight, 2016) or race (Blodgett et al., 2016). A number of recent approaches have sought to transform latent representations of texts to protect confidential attributes, using adversarial learning (Elazar and Goldberg, 2018), reinforcement learning (Mosallanezhad et al., 2019) or encryption (Huang et al., 2020). However, those methods operate at the level of latent vector representations and do not modify the texts themselves. One notable exception is the text rewriting approach of Xu et al. (2019) which edits the texts using back-translations. 3.3 Challenges NLP approaches to anonymisation suffer from a number of shortcomings. Most importantly, they are limited to predefined categories of entities and ignore how less conspicuous text elements may also play a role in re-identifying the individual. For instance, the family status or physical appearance of a person may lead to re-identification but will rarely be considered as categories to detect. On the other hand, those methods may also end up removing too much information, as they will systematically remove all occurrences of a given category without examining their impact on the disclosure risk or on the utility of the remaining text. 4 PPDP Approaches Privacy-preserving data publishing (PPDP) develops computational techniques for releasing data without violating privacy (Chen et al., 2009). The PPDP approach to anonymisation is privacyfirst: a privacy model specifying an ex ante privacy condition is enforced through one or several data masking methods, such as noise addition or generalisation of values (Domingo-Ferrer et al., 2016). The first widely-accepted privacy model is k-anonymity (Samarati, 2001): a dataset satisfies kanonymity if each combination of values of quasiidentifier attributes is shared by at least k records. With k > 1, no unequivocal re-identifications are possible, thereby preventing identity disclosure. Most of the attention of the PPDP community has been on structured databases. Privacy models such as k-anonymity assume that datasets consist of records, each one detailing the attributes of a single individual, and that attributes have been classified beforehand into identifiers, quasi-identifiers and confidential attributes. Moreover, most masking methods employed to enforce privacy models have been designed with numerical data in mind, and barely (and poorly) manage categorical or nominal attributes (Rodr´ıguez-Garc´ıa et al., 2019). 4.1 k-anonymity and Beyond Solutions for anonymising unstructured text are scarce and mostly theoretical. The first approaches adapted k-anonymity for collections of documents. In (Chakaravarthy et al., 2008), the authors pre4192 sented the notion of K-safety. They assume a collection of entities e to be protected against disclosure, each one characterised by a set of terms C(e) that represent their contexts (i.e. words cooccurring with e and that may be known to an attacker). Then, a document D containing an entity e is said to be K-safe if the terms appearing in D also belong to the contexts of, at least, K−1 entities other than e. Terms not fulfilling the property are redacted before release. The privacy guarantee offered by this approach is sound because the probability of disclosing the protected entity is reduced to 1/K. However, it requires exhaustive collections of contexts for all entities to be protected, which is unfeasible. It also assumes that the detection of sensitive terms is already performed. This approach is only feasible for very constrained domains and non-dynamic sets of entities, such as collections of sensitive diseases, and documents with homogeneous contents. Another approach built on k-anonymity is Cumby and Ghani (2011), where a multi-class classifier is trained to map input documents to (predefined) sensitive entities. This aims at reproducing the inferences that a potential attacker may perform to disclose sensitive entities. A document x referring to a sensitive entity y is then said to be K-confusable if the classifier outputs at least k classes other than y. Documents are redacted via term removal or generalisation until the property is fulfilled. To be applicable, sensitive entities should be static and the documents to be protected should match that of the corpus used for training. Anandan et al. (2012) present a privacy model for document protection named t-plausibility. They seek to generalise terms identified as sensitive according to the t-plausibility property: a protected document is said to fulfil t-plausibility if, at least, t different plausible documents can be derived by specialising the generalised terms. In other words, Even though the privacy guarantee is intuitive, one can hardly predict the results for a certain t, because they depend on the document length, the number of sensitive entities and the granularity of the knowledge base employed to obtain term generalisations. Assuming that sensitive entities have already been detected also circumvents the most challenging task of document protection. 4.2 C-sanitise S´anchez and Batet (2016, 2017) tackles the anonymisation problem from a different perspective. Instead of expressing privacy guarantees in terms of probability of disclosure, it defines risk as an information theoretic characterisation of disclosed semantics. The proposed privacy model, C-sanitise, states that given a document d, background knowledge K available to potential attackers, and a set of entities to protect C, d′ is the C-sanitised version of d if d′ does not contain any term t that, individually or in aggregate, unequivocally disclose the semantics encompassed by any entity in C by exploiting K. The semantic disclosure incurred by t on any entity in C is quantified as their pointwise mutual information (Anandan and Clifton, 2011) measured from their probability of (co-)occurrence in the Web, which is assumed to represent the most comprehensive knowledge source (K) available to attackers (Chow et al., 2008). This approach is able to automatically detect terms that may cause disclosure and can encompass dynamic collections of entities to protect. Obtaining accurate probabilities of co-occurrence from large corpora is, however, costly. 4.3 Differential Privacy Differential privacy (DP) is a privacy model that defines anonymisation in terms of randomised algorithms for computing statistics from the data (Dwork et al., 2006). DP provides guarantees that the statistics cannot be used to learn anything substantial about any individual. However, the goal of DP is to produce randomised responses to controlled queries, and applying it to data publishing leads in poor data utility (Domingo-Ferrer et al., 2021). DP cannot be directly employed to edit out personal information from text while preserving the content of the rest of the document, and is thus outside the scope of this paper. However, DP can be employed for other privacy-related tasks such as in producing synthetic texts (Fernandes et al., 2018; Bommasani et al., 2019), deriving differentiallyprivate word representations (Feyisetan et al., 2019) or learning machine learning models with privacy guarantees (McMahan et al., 2017). 4.4 Challenges Compared to NLP approaches, proposals built around privacy models allow defining what should be protected and how. This not only allows enforcing privacy requirements, but also makes it possible to tailor the trade-off between data protection and utility preservation. On the negative 4193 side, PPDP methods are hampered by practical constraints, either because of their unfeasible assumptions, their cost or their dependency on external resources, such as large knowledge repositories, training corpora or social-scale probabilities. To the exception of C-sanitise, PPDP methods also assume that sensitive entities have already been detected in a preprocessing step. Furthermore, PPDP approaches typically reduce documents to flat collections of terms, which facilitates the formalisation of the data semantics for each document, but also ignores how terms are influenced by their context of occurrence (which is important to resolve potential ambiguities) and are interconnected through multiple layers of linguistic structures. 5 Case Study To investigate the performance of NLP and PPDP methods, we carried out a case study where 5 annotators annotated 8 English Wikipedia page extracts. The extracts were all biographies from the “20th century scientists” category, with a length between 300 and 500 characters. Wikipedia articles are generic enough not to require expert domain knowledge and are commonly adopted for the evaluation of PPDP approaches (Chow et al., 2008; S´anchez and Batet, 2016). Their informativeness and density make them particularly challenging to anonymise. The annotation task2 consisted of tagging text spans that could re-identify a person either directly or in combination with publicly available knowledge. The annotators were instructed to prevent identity disclosure but otherwise seek to preserve as much semantic content as possible. The five annotators were researchers without previous experience in text anonymisation. The guidelines were left intentionally general to examine how annotators interpret and carry out the complex task of anonymisation – and not only de-identification – where multiple correct solutions are possible. The task is challenging since these biographies relate to publicly known scientists for which extensive background material can be found online. Inter-rater agreement between the five annotators for the binary masking decisions was low: 0.68 average observed agreement and Krippendorff’s α = 0.36. This low agreement illustrates that, contrary to traditional sequence labelling, several 2The guidelines and annotated data are publicly available: https://github.com/IldikoPilan/anonymisation_ACL2021 solutions may exist for a given anonymisation problem. Direct identifiers were generally agreed on, while quasi-identifiers such as professions and roles (e.g. founder) triggered mixed decisions. To shed further light on the anonymisation problem, we go on to compare the performance of existing tools with the manual annotations: • A neural NER model (Honnibal and Montani, 2017) trained on the OntoNotes corpus with 18 entity types (Weischedel et al., 2011). All detected entities were masked.3 • Presidio4, a data protection & anonymisation API developed by Microsoft and relying on a combination of template-based and machine learning models to detect and mask PII. • The C-sanitise privacy model (S´anchez and Batet, 2016) described in Section 4, where the required probabilities of (co-)occurrence of terms were gathered from Google. 5.1 Metrics To account for the multiple ways to anonymise a document, we measured the performance of the three tools above with micro-averaged scores over all annotators and texts. Note that, while microaverages are typically used in NLP to aggregate measures over output classes, we are here computing an average over multiple ground truths. For each annotator q ∈Q and document d ∈D, let Y q d correspond to token indices masked by q in d, and ˆYd to the token indices masked by the anonymisation tool. Precision and recall are then computed as: P = P d∈D P q∈Q | ˆYd ∩Y q d | |Q| P d∈D | ˆYd| (1) R = P d∈D P q∈Q | ˆYd ∩Y q d | P d∈D P q∈Q |Y q d | (2) An anonymisation tool will thus obtain a perfect micro-averaged recall if it detects all tokens masked by at least one annotator. The metric implicitly assigns a higher weight to tokens masked by several annotators – in other words, if all five annotators mask a given token, not detecting it will have a 3Although NERs do not specifically focus on data protection, they are often used to de-identify generic texts (except clinical notes, for which domain-specific tools are available). 4https://github.com/microsoft/presidio 4194 P R F1 NER IOB-Exact 0.5 0.49 0.47 IOB-Partial 0.61 0.48 0.54 Binary 0.64 0.51 0.57 Presidio IOB-Exact 0.63 0.22 0.33 IOB-Partial 0.74 0.24 0.36 Binary 0.76 0.25 0.38 C-sanitise IOB-Exact 0.51 0.66 0.57 IOB-Partial 0.57 0.68 0.62 Binary 0.58 0.69 0.63 Table 2: Micro-averaged scores for NER, C-sanitise and Presidio over all texts for annotators a1, a4, a5. larger impact on the recall than a token masked by a single annotator. Recall expresses the level of privacy protection while precision is related to the degree of utility preservation. The most consistent manual annotations (a1, a4, a5) were compared to system outputs at token level both as binary labels (keep or mask) and as IOB tags expressing annotation spans5. To go beyond token-level comparisons, we also computed a partial match score for IOB tags, by assigning a weight of 0.5 to partial true positives (i.e. I instead of B tags and vice versa), as in the SemEval 2013 evaluation scheme (Diab et al., 2013). 5.2 Results and Error Analysis Table 2 presents the micro-averaged precision, recall and F1 scores obtained for the three systems. C-sanitise provided the best performance in terms of recall and F1 score, while precision was higher for NER and Presidio. Figure 1 illustrates the average observed agreement for all annotators and tools on the binary, token-level masking decisions. Observed agreement with annotators was, on average, approximately the same for NER and Csanitise, ca. 75% and ca. 77% for Presidio. We can distinguish two subgroups among the annotators in terms of mutual agreement, namely (a2, a3) and (a1, a4, a5) with 79% and 83% agreement respectively. Divergent choices in entity segmentation – e.g. splitting a consecutive mention of department and university or not – was found to play an important role in the differences among annotators, and between annotators and systems. 5B(eginning) represents the first token of a span, I(nside) the subsequent tokens, and O(ut) is the label assigned to all tokens that are not part of a span. Figure 1: Pairwise average observed agreement. a1 to a5 correspond to the human annotators. The proportion of masked tokens was around 50% for a1, a2 and C-sanitise, < 30% for a3, a4, a5 and NER and 11% for Presidio. We conducted a detailed error analysis to gain a better understanding about the advantages and shortcoming of the three anonymisation tools described above. The NER tool masked generic entities such as Second World War, although this term was not masked by any annotator or by C-sanitise. In the phrase “a Christian charity dedicated to helping the people of Cambodia”, most annotators did not mask any tokens, while NER masked both Christian and Cambodia, and C-sanitise Christian charity. On the other hand, NER ignored terms that were highly correlated with the individual and should have been masked, such as book titles authored by the person. Another interesting error can be found in the sentence “In 1964 and 1965 he was a Visiting Professor at the University of Wisconsin–Madison on a Fulbright Program fellowship” where the university was masked by most annotators but left untouched by C-sanitise (as the university does not frequently co-occur with this person in web documents). Presidio had the lowest recall and ignored the majority of quasi-identifiers (including organisations). Consequently, Presidio’s masking should be considered a de-identification process rather than full anonymisation. See Appendix A for an annotated example document. 6 Challenges and Future Directions The case study illustrates a number of issues facing current methods for text anonymisation. We discuss below three overarching challenges: the need to protect against several types of semantic inferences, the formalisation of possible masking operations to apply on documents, and, last but not least, the design of evaluation metrics to empirically assess the anonymisation performance. 4195 6.1 Semantic Inferences Most works on PPDP address anonymisation from a statistical perspective (Batet and S´anchez, 2018). Their main focus is on the statistical properties of (numerical) data and how these may allow attackers to re-identify an individual or uncover confidential data. However, the most harmful inferences in text documents are semantic in nature – that is, they are based on the actual meaning expressed in the texts instead of their statistical distributions. NLP approaches do not explicitly account for semantic inferences, and simply mask all text spans belonging to predefined categories irrespective of their impact on the disclosure risk. In many PPDP approaches (Chakaravarthy et al., 2008; Cumby and Ghani, 2011; Anandan et al., 2012), the adversary is assumed to know sets of attributes associated with each entity, and semantic inferences thus correspond to combinations of attributes enabling the adversary to single out the entity to protect. However, in most practical settings, human adversaries do not have access to the original documents. They do, however, make extensive use of external background knowledge available, e.g., on the web. Such external background knowledge is captured in S´anchez and Batet (2016, 2017) using (co-)occurrence counts of terms on the web. Other types of semantic inferences may be taken into account, such as lexical and taxonomic relations (synonyms, antonyms, hypernyms, hyponyms) between words or entities. For instance, the word “AIDS” will lead to the disclosure of the confidential attribute “immune system disease”. In S´anchez and Batet (2017), those relations are taken into account by enforcing consistency between known taxonomic relations and the information content of each term. Semantic relations can, however, extend beyond individual terms and exploit various syntactic patterns, as shown in e.g. textual entailment (Dagan et al., 2013). Semantic inferences can also be drawn from structured data sources such as census data or medical knowledge bases. In the “Wisconsin-Madison” example above, the search for Fullbright recipients at that university in 1964-65 would likely allow the individual to be re-identified. Such logical inferences require specifying which background knowledge may be available to a potential intruder and would be relevant for a given text domain. Although semantic inferences have been studied in isolation in previous work, how to integrate and chain together those inferential mechanisms into a single framework remains an open question. Formally, assuming a document d transformed into d′ by an anonymisation tool in charge of protecting a set of entities C, one can design an adversary model adv(c, d′, K) seeking to predict, based on document d′ and background knowledge K, whether the entity c was part of the original document d or not. Ideally, this adversary model should allow for multiple types of semantic inferences based on domain-relevant background knowledge (word cooccurrences in text corpora, taxonomic relations, knowledge bases, etc.). 6.2 Masking Operations NLP approaches to text anonymisation essentially focus on detecting personal identifiers and rarely discuss what to do with the detected text spans, generally assuming that those should be either redacted or replaced with coded values. This approach may, however, lead to unnecessary loss of data utility, as it is often possible to replace quasi-identifiers by more generic (but still informative) entries. How to transform a dataset to balance disclosure risk and data utility is a central research question in privacy-preserving data publishing. Various transformations have been put forward: one can remove values altogether, generalise them into less detailed categories, or perturb the values by adding noise or swapping them (Domingo-Ferrer et al., 2016). In the text domain, several PPDP approaches have shown how to generalise terms using ontologies (Anandan et al., 2012; S´anchez and Batet, 2016). However, these approaches are intrinsically limited to entities present in such ontologies, and are difficult to extend to more generic text entries. Another possible transformation is to introduce noise into the text. The perturbation of data points through noise is a common type of transformation in data privacy (McSherry and Talwar, 2007). This idea of perturbation has notably been applied to word embeddings (Feyisetan et al., 2019), but it produces perturbed word distributions rather than readable documents. Semantic noise has also been defined to perturb nominal values (Rodr´ıguez-Garc´ıa et al., 2017). Formally, one can define an editor model edit(d) taking a document d and outputting an edited document d′ after applying a sequence of masking operations. This model can be e.g. expressed as a neural text editing model (Mallinson et al., 2020). 4196 Its optimisation objective should include both minimising the risk of letting an adversary disclose at least some of the protected entities C through semantic inferences (as described in the previous section) and minimising the number of masking operations necessary to map d to d′. 6.3 Evaluation Metrics Let D be a set of documents transformed into D′ by an anonymisation tool. How can we empirically evaluate the quality of the anonymisation? The most common method is to rely on human annotators to manually mark identifiers in each document d ∈D, and then compare the system output with those human-annotated identifiers using IRbased metrics such as precision, recall and F1 score. The recall can be seen as reflecting the degree of protection of the confidential information, while the precision is correlated with the remaining data utility of the documents D′. This evaluation procedure has a number of shortcomings. As observed in our case study, there may be several equally valid solutions to a given anonymisation problem. Furthermore, IR-based metrics typically associate uniform weights to all identifiers, without taking into account the fact that some identifiers may have a much larger influence on the disclosure risk than others. For instance, failing to detect a full person name is more harmful than failing to detect a quasi-identifier. Finally, such type of evaluation procedure is limited to the detection of direct and indirect identifiers, but ignore the subsequent step of transforming the textual content. Evaluating the quality of masking operations is tightly coupled with the problem of evaluating how data utility is preserved through the anonymisation process (S´anchez and Batet, 2016; Rodr´ıguez-Garc´ıa et al., 2019). However, how to empirically measure this data utility remains an open question. An alternative which has so far received little attention is to conduct so-called privacy attacks on the edited documents D′. This can be achieved by e.g. providing the documents D′ to human experts and instruct them to re-identify those documents with the help of any information source at their disposal. Such human evaluations can help uncover weaknesses in the anonymisation model (such as semantic inferences that had been overlooked). However, they are also costly and timeconsuming, as they must be repeated for each version of the anonymisation model. 7 Conclusion This position paper discussed a number of unresolved challenges in text anonymisation. Text anonymisation is defined as the removal or masking of any information that, directly or indirectly, may lead to an individual being identified (given some assumptions about the available background knowledge). As illustrated in our case study, text anonymisation is a difficult task (also for human annotators), which goes beyond the mere detection of predefined categories of entities and may allow for several solutions. How to properly anonymise text data is a problem of great practical importance. In particular, access to high-quality data is a key ingredient for most scientific research, and the lack of good anonymisation methods for text documents (allowing data to be shared without compromising privacy) is a limiting factor in fields such as medicine, social sciences, psychology and law. We surveyed two families of approaches with complementary strengths and weaknesses: NLP models are well-suited to capture textual patterns but lack any consideration of disclosure risk, while PPDP approaches provide principled accounts of privacy requirements, but view documents as bagof-terms void of linguistic structure. As outlined in the last section, a promising approach is to couple a neural editor model (applying transformations to the text) with an adversary model (capturing possible semantic inferences to uncover confidential entities). These two models can be optimised jointly using adversarial training, taking into account the necessary balance between disclosure risk and utility preservation. Finally, we lay out a case for designing evaluation metrics that go beyond traditional IR-based measures, and account in particular for the fact that some identifiers and quasi-identifiers are more important than others in terms of their influence on the disclosure risk. Acknowledgements We acknowledge support from the Norwegian Research Council (CLEANUP project6, grant nr. 308904), the Government of Catalonia (ICREA Acad`emia Prize to D. S´anchez and grant 2017 SGR 705) and the Spanish Government (project TIN2016-80250-R “Sec-MCloud”). 6see http://cleanup.nr.no/ 4197 References John Aberdeen, Samuel Bayer, Reyyan Yeniterzi, Ben Wellner, Cheryl Clark, David Hanauer, Bradley Malin, and Lynette Hirschman. 2010. The MITRE identification scrubber toolkit: design, training, and assessment. International Journal of Medical Informatics, 79(12):849–859. Alyaa Alfalahi, Sara Brissman, and Hercules Dalianis. 2012. Pseudonymisation of personal names and other PHIs in an annotated clinical Swedish corpus. In Third LREC Workshop on Building and Evaluating Resources for Biomedical Text Mining (BioTxtM 2012), pages 49–54. Balamurugan Anandan and Chris Clifton. 2011. Significance of term relationships on anonymization. In Proceedings of the 2011 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology - Workshops, WI-IAT 2011, pages 253–256, Lyon, France. Balamurugan Anandan, Chris Clifton, Wei Jiang, Mummoorthy Murugesan, Pedro PastranaCamacho, and Luo Si. 2012. t-plausibility: Generalizing words to desensitize text. Transactions on Data Privacy, 5(3):505–534. Montserrat Batet and David S´anchez. 2018. Semantic disclosure control: semantics meets data privacy. Online Information Review, 42(3):290–303. Eric A. Bier, Richard Chow, Philippe Golle, Tracy H. King, and J. Staddon. 2009. The rules of redaction: Identify, protect, review (and repeat). IEEE Security and Privacy Magazine, 7(6):46–53. Su Lin Blodgett, Lisa Green, and Brendan O’Connor. 2016. Demographic dialectal variation in social media: A case study of African-American English. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1119–1130, Austin, Texas. Association for Computational Linguistics. Rishi Bommasani, Steven Wu, Zhiwei, and Alexandra K Schofield. 2019. Towards private synthetic text generation. In NeurIPS 2019 Workshop on Machine Learning with Guarantees, Vancouver, Canada. David Carrell, Bradley Malin, John Aberdeen, Samuel Bayer, Cheryl Clark, Ben Wellner, and Lynette Hirschman. 2013. Hiding in plain sight: use of realistic surrogates to reduce exposure of protected health information in clinical text. Journal of the American Medical Informatics Association, 20(2):342–348. Venkatesan T. Chakaravarthy, Himanshu Gupta, Prasan Roy, and Mukesh K. Mohania. 2008. Efficient techniques for document sanitization. In Proceedings of the 17th ACM Conference on Information and Knowledge Management, CIKM 2008, pages 843– 852, Napa Valley, California, USA. Aipeng Chen, Jitendra Jonnagaddala, Chandini Nekkantti, and Siaw-Teng Liaw. 2019. Generation of surrogates for de-identification of electronic health records. In MEDINFO 2019: Health and Wellbeing e-Networks for All - Proceedings of the 17th World Congress on Medical and Health Informatics, Lyon, France, 25-30 August 2019, volume 264 of Studies in Health Technology and Informatics, pages 70–73. IOS Press. Bee-Chung Chen, Daniel Kifer, Kristen LeFevre, and Ashwin Machanavajjhala. 2009. PrivacyPreserving Data Publishing. Foundations and Trends in Databases. Now Publishers Inc. Rapha¨el Chevrier, Vasiliki Foufi, Christophe GaudetBlavignac, Arnaud Robert, and Christian Lovis. 2019. Use and understanding of anonymization and de-identification in the biomedical literature: Scoping review. Journal of Medical Internet Research, 21(5):e13484. Jason P.C. Chiu and Eric Nichols. 2016. Named entity recognition with bidirectional LSTM-CNNs. Transactions of the Association for Computational Linguistics, 4:357–370. Richard Chow, Philippe Golle, and Jessica Staddon. 2008. Detecting privacy leaks using corpus-based association rules. In Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’08, page 893–901, New York, NY, USA. Association for Computing Machinery. Chad M. Cumby and Rayid Ghani. 2011. A machine learning based system for semi-automatically redacting documents. In Proceedings of the Twenty-Third Conference on Innovative Applications of Artificial Intelligence, pages 1628–1635, San Francisco, California, USA. Ido Dagan, Dan Roth, Mark Sammons, and Fabio Massimo Zanzotto. 2013. Recognizing textual entailment: Models and applications. Synthesis Lectures on Human Language Technologies, 6(4):1–220. Franck Dernoncourt, Ji Young Lee, Ozlem Uzuner, and Peter Szolovits. 2017. De-identification of patient notes with recurrent neural networks. Journal of the American Medical Informatics Association, 24(3):596–606. Mona Diab, Tim Baldwin, and Marco Baroni, editors. 2013. Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task: Semantic Textual Similarity. Association for Computational Linguistics, Atlanta, Georgia, USA. Josep Domingo-Ferrer, David S´anchez, and Alberto Blanco-Justicia. 2021. The limits of differential privacy (and its misuse in data release and machine learning). Communications of the ACM, 64(7):34– 36. 4198 Josep Domingo-Ferrer, David S´anchez, and Jordi SoriaComas. 2016. Database Anonymization: Privacy Models, Data Utility, and Microaggregation-based Inter-model Connections. Synthesis Lectures on Information Security, Privacy & Trust. Morgan & Claypool Publishers. Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. 2006. Calibrating Noise to Sensitivity in Private Data Analysis. In Theory of Cryptography, pages 265–284, Berlin, Heidelberg. Springer Berlin Heidelberg. Elisabeth Eder, Ulrike Krieg-Holz, and Udo Hahn. 2019. De-identification of emails: Pseudonymizing privacy-sensitive data in a German email corpus. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2019), pages 259–269, Varna, Bulgaria. INCOMA Ltd. Elisabeth Eder, Ulrike Krieg-Holz, and Udo Hahn. 2020. CodE Alltag 2.0 — a pseudonymized German-language email corpus. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4466–4477, Marseille, France. European Language Resources Association. Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11–21, Brussels, Belgium. Association for Computational Linguistics. Mark Elliot, Elaine Mackey, Kieron O’Hara, and Caroline Tudor. 2016. The anonymisation decisionmaking framework. UKAN Manchester. Natasha Fernandes, Mark Dras, and Annabelle McIver. 2018. Generalised differential privacy for text document processing. CoRR, abs/1811.10256. Oluwaseyi Feyisetan, Tom Diethe, and Thomas Drake. 2019. Leveraging hierarchical representations for preserving privacy and utility in text. In 2019 IEEE International Conference on Data Mining (ICDM), pages 210–219. IEEE. Max Friedrich, Arne K¨ohn, Gregor Wiedemann, and Chris Biemann. 2019. Adversarial learning of privacy-preserving text representations for deidentification of medical records. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5829–5839, Florence, Italy. Association for Computational Linguistics. GDPR. 2016. General Data Protection Regulation. European Union Regulation 2016/679. Philippe Golle. 2006. Revisiting the uniqueness of simple demographics in the US population. In Proceedings of the 5th ACM Workshop on Privacy in electronic society, pages 77–80. ACM. Tzvika Hartman, Michael D Howell, Jeff Dean, Shlomo Hoory, Ronit Slyper, Itay Laish, Oren Gilon, Danny Vainstein, Greg Corrado, Katherine Chou, et al. 2020. Customization scenarios for deidentification of clinical notes. BMC Medical Informatics and Decision Making, 20(1):1–9. Mike Hintze. 2017. Viewing the GDPR through a deidentification lens: a tool for compliance, clarification, and consistency. International Data Privacy Law, 8(1):86–101. HIPAA. 2004. The Health Insurance Portability and Accountability Act. U.S. Dept. of Labor, Employee Benefits Security Administration. Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear. Yangsibo Huang, Zhao Song, Danqi Chen, Kai Li, and Sanjeev Arora. 2020. TextHide: Tackling data privacy in language understanding tasks. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1368–1382, Online. Association for Computational Linguistics. Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016. Neural architectures for named entity recognition. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 260–270, San Diego, California. Zengjian Liu, Buzhou Tang, Xiaolong Wang, and Qingcai Chen. 2017. De-identification of clinical notes via recurrent neural network and conditional random field. Journal of Biomedical Informatics, 75:S34– S42. Jonathan Mallinson, Aliaksei Severyn, Eric Malmi, and Guillermo Garrido. 2020. Felix: Flexible text editing through tagging and insertion. arXiv preprint arXiv:2003.10687. Montserrat Marimon, Aitor Gonzalez-Agirre, Ander Intxaurrondo, Heidy Rodriguez, Jose Lopez Martin, Marta Villegas, and Martin Krallinger. 2019. Automatic de-identification of medical texts in spanish: the meddocan track, corpus, guidelines, methods and evaluation of results. In IberLEF@ SEPLN, pages 618–638. H. Brendan McMahan, Daniel Ramage, Kunal Talwar, and Li Zhang. 2017. Learning Differentially Private Recurrent Language Models. arXiv:1710.06963 [cs]. Frank McSherry and Kunal Talwar. 2007. Mechanism design via differential privacy. In 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS’07), pages 94–103. 4199 Ben Medlock. 2006. An introduction to NLP-based textual anonymisation. In Proceedings of the Fifth International Conference on Language Resources and Evaluation (LREC’06), pages 1051– 1056, Genoa, Italy. European Language Resources Association (ELRA). Be´ata Megyesi, Lena Granstedt, Sofia Johansson, Julia Prentice, Dan Ros´en, Carl-Johan Schenstr¨om, Gunl¨og Sundberg, Mats Wir´en, and Elena Volodina. 2018. Learner corpus anonymization in the age of GDPR: Insights from the creation of a learner corpus of Swedish. In Proceedings of the 7th workshop on NLP for Computer Assisted Language Learning, pages 47–56, Stockholm, Sweden. LiU Electronic Press. St´ephane M Meystre, ´Oscar Ferr´andez, F Jeffrey Friedlin, Brett R South, Shuying Shen, and Matthew H Samore. 2014a. Text de-identification for privacy protection: a study of its impact on clinical text information content. Journal of Biomedical Informatics, 50:142–150. Stephane M Meystre, F Jeffrey Friedlin, Brett R South, Shuying Shen, and Matthew H Samore. 2010. Automatic de-identification of textual documents in the electronic health record: a review of recent research. BMC Medical Research Methodology, 10(1):70. St´ephane Meystre, Shuying Shen, Deborah Hofmann, and Adi Gundlapalli. 2014b. Can physicians recognize their own patients in de-identified notes? Studies in Health Technology and Informatics, 205:778—782. Ahmadreza Mosallanezhad, Ghazaleh Beigi, and Huan Liu. 2019. Deep reinforcement learning-based text anonymization against private-attribute inference. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2360– 2369, Hong Kong, China. Association for Computational Linguistics. Sravana Reddy and Kevin Knight. 2016. Obfuscating gender in social media writing. In Proceedings of the First Workshop on NLP and Computational Social Science, pages 17–26, Austin, Texas. Association for Computational Linguistics. Mercedes Rodr´ıguez-Garc´ıa, Montserrat Batet, and David S´anchez. 2017. A semantic framework for noise addition with nominal data. Knowledge-Based Systems, 122(C):103–118. Mercedes Rodr´ıguez-Garc´ıa, Montserrat Batet, and David S´anchez. 2019. Utility-preserving privacy protection of nominal data sets via semantic rank swapping. Information Fusion, 45:282–295. Mark A. Rothstein. 2010. Is deidentification sufficient to protect health privacy in research? The American Journal of Bioethics, 10(9):3–11. Pierangela Samarati. 2001. Protecting respondents’ identities in microdata release. IEEE Transactions on Knowledge and Data Engineering, 13(6):1010– 1027. Pierangela Samarati and Latanya Sweeney. 1998. Protecting Privacy when Disclosing Information: kAnonymity and its Enforcement through Generalization and Suppression. Technical report, SRI International. Amber Stubbs, Michele Filannino, and ¨Ozlem Uzuner. 2017. De-identification of psychiatric intake records: Overview of 2016 CEGS N-GRID Shared Tasks Track 1. Journal of Biomedical Informatics, 75:S4–S18. Amber Stubbs and ¨Ozlem Uzuner. 2015. Annotating longitudinal clinical narratives for de-identification: The 2014 i2b2/UTHealth corpus. Journal of Biomedical Informatics, 58:S20–S29. Latanya Sweeney. 1996. Replacing personallyidentifying information in medical records, the scrub system. In Proceedings of the AMIA annual fall symposium, pages 333–337. American Medical Informatics Association. David S´anchez and Montserrat Batet. 2016. Csanitized: A privacy model for document redaction and sanitization. Journal of the Association for Information Science and Technology, 67(1):148–163. David S´anchez and Montserrat Batet. 2017. Toward sensitive document release with privacy guarantees. Engineering Applications of Artificial Intelligence, 59:23–34. Sumithra Velupillai, Hercules Dalianis, Martin Hassel, and Gunnar H. Nilsson. 2009. Developing a standard for de-identifying electronic patient records written in Swedish: Precision, recall and f-measure in a manual and computerized annotation trial. International Journal of Medical Informatics, 78(12):19 – 26. Elena Volodina, Yousuf Ali Mohammed, Sandra Derbring, Arild Matsson, and Beata Megyesi. 2020. Towards privacy by design in learner corpora research: A case of on-the-fly pseudonymization of Swedish learner essays. In Proceedings of the 28th International Conference on Computational Linguistics, pages 357–369, Barcelona, Spain (Online). International Committee on Computational Linguistics. Ralph Weischedel, Eduard Hovy, Marcus. Mitchell, Palmer Martha S., Robert Belvin, Sameer S. Pradhan, Lance Ramshaw, and Nianwen Xue. 2011. OntoNotes: A large training corpus for enhanced processing. In Handbook of Natural Language Processing and Machine Translation: DARPA Global Autonomous Language Exploitation. Springer. Alan F. Westin. 1967. Privacy and Freedom. Atheneum, New York. 4200 Qiongkai Xu, Lizhen Qu, Chenchen Xu, and Ran Cui. 2019. Privacy-aware text rewriting. In Proceedings of the 12th International Conference on Natural Language Generation, pages 247–257, Tokyo, Japan. Association for Computational Linguistics. Xi Yang, Tianchen Lyu, Qian Li, Chih-Yin Lee, Jiang Bian, William R Hogan, and Yonghui Wu. 2019. A study of deep learning methods for deidentification of clinical notes in cross-institute settings. BMC Medical Informatics and Decision Making, 19(5):232. Vithya Yogarajan, Michael Mayo, and Bernhard Pfahringer. 2018. A survey of automatic deidentification of longitudinal clinical narratives. arXiv preprint arXiv:1810.06765. 4201 A Appendix We present below the annotation of one short biography of a 20th century scientist (Alexander Frumkin) according to 5 human annotators, C-sanitize, the neural NER model and the Presidio anonymisation tool (see paper for details). The annotation task consisted of tagging text spans that could re-identify a person either directly or in combination with publicly available knowledge. The annotators were instructed to prevent identity disclosure, but otherwise seek to preserve the semantic content as much as possible. The five annotators were researchers in statistics and natural language processing. The first five (gray) lines denotes the five human annotators, while the cyan line corresponds to C-sanitise, the blue line to the neural NER model, and the green line to the Presidio tool. Due to page limits, we only present here one single biography, but the annotations for all 8 texts (along with the annotation guidelines and raw data) are available in the GitHub repository associated with the paper. A.1 Alexander Frumkin Alexander Naumovich Frumkin (Александр Наумович Фрумкин) (October 24, 1895–May 27, 1976) was a Russian/Soviet electrochemist, member of the Russian Academy of Sciences since 1932, founder of the Russian Journal of Electrochemistry Elektrokhimiya and receiver of the Hero of Socialist Labor award. The Russian Academy of Sciences’ A. N. Frumkin Institute of Physical Chemistry and Electrochemistry is named after him. Frumkin was born in Kishinev, in the Bessarabia Governorate of the Russian Empire (present-day Moldova) to a Jewish family; his father was an insurance salesman. His family moved to Odessa, where he received his primary schooling; he continued his education in Strasbourg, and then at the University of Bern. Frumkin’s first published articles appeared in 1914, 4202 when he was only 19; in 1915, he received his first degree, back in Odessa. Two years later, the seminal article “Electrocapillary Phenomena and Electrode Potentials” was published. Frumkin moved to Moscow in 1922 to work at the Karpov Institute, under A. N. Bakh. In 1930 Frumkin joined the faculty of Moscow University, where in 1933 he founded—and would head until his death—the department of electrochemistry. During the Second World War, Frumkin led a large team of scientists and engineers involved in defense issues. This contribution did not save him from being dismissed in 1949 as the director of the Institute of Physical Chemistry, when he was accused of “cosmopolitanism”. Frumkin’s most fundamental achievement was the fundamental theory of electrode reactions, which describes the influence of the structure of the interface between electrode and solution on the rate of electron transfer. This theory has been confirmed and extended within the framework of contemporary physical electron transfer models. Frumkin introduced the concept of the zero charge potential, the most important characteristic of a metal surface. 4203 Alessandro Volta’s question—a topic of discussion for over 120 years—about the nature of the EMF of electrochemical circuits was resolved using Frumkin’s approach. Frumkin developed the Frumkin isotherm, an extension of the Langmuir isotherm in describing certain adsorption phenomena. Frumkin’s students developed novel experimental methods that would, in time, become standard. Several applied electrochemical processes, including ones related to chemical sources of electrical power, industrial electrolysis, and anti-corrosion protection, were successfully developed under Frumkin’s supervision. Frumkin was married three times, including a brief first marriage to Vera Inber.
|
2021
|
323
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4204–4214 August 1–6, 2021. ©2021 Association for Computational Linguistics 4204 End-to-End AMR Coreference Resolution Qiankun Fu1,2,3, Linfeng Song4, Wenyu Du2,3, Yue Zhang2,3 1. Zhejiang University 2. School of Engineering, Westlake University 3. Institute of Advanced Technology, Westlake Institute for Advanced Study 4. Tencent AI Lab, Bellevue, WA, USA [email protected] and [email protected] {duwenyu, zhangyue}@westlake.edu.cn Abstract Although parsing to Abstract Meaning Representation (AMR) has become very popular and AMR has been shown effective on many sentence-level tasks, little work has studied how to generate AMRs that can represent multi-sentence information. We introduce the first end-to-end AMR coreference resolution model in order to build multi-sentence AMRs. Compared with the previous pipeline and rule-based approaches, our model alleviates error propagation and it is more robust for both in-domain and out-domain situations. Besides, the document-level AMRs obtained by our model can significantly improve over the AMRs generated by a rule-based method (Liu et al., 2015) on text summarization. 1 Introduction Abstract Meaning Representation (AMR) (Banarescu et al., 2013) is a semantic formalism for natural language understanding. It represents a sentence as a rooted, directed and acyclic graph, where nodes (e.g., “Bill” in Figure 1) represents concepts and edges (e.g., “:arg0”) are the semantic relations. Encompassing knowledge of named entities, semantic roles and coreference structures, AMR has been proven effective for downstream tasks, including information extraction (Rao et al., 2017), text summarization (Liu et al., 2015; Hardy and Vlachos, 2018; Liao et al., 2018), paraphrase detection (Issa Alaa Aldine et al., 2018), event detection (Li et al., 2015), machine translation (Song et al., 2019b) and dialogue understanding (Bonial et al., 2020). Existing work on AMR mainly focuses on individual sentences (Lyu and Titov, 2018; Naseem et al., 2019; Ge et al., 2019; Zhang et al., 2019; Cai and Lam, 2020a; Zhou et al., 2020). On the other hand, with the advance of neural networks in NLP, tasks involving multiple sentences with leave-11 person name name city Bill Paris Sentence1: Bill left for Paris. Sentence2: He arrived at noon. arrive-01 Paris he date-entity noon :name :name :arg0 :arg3 :arg1 :arg2 :dayperiod :op1 :op1 Figure 1: Multi-sentence AMR example, where nodes with the same non-black color are coreferential and the dotted ellipse represents an implicit role coreference. cross-sentence reasoning (e.g., text summarization, reading comprehension and dialogue response generation) have received increasing research attention. Given the effectiveness of AMR on sentencelevel tasks (Pan et al., 2015; Rao et al., 2017; Issa Alaa Aldine et al., 2018; Song et al., 2019b), it is important to extend sentence-level AMRs into the multi-sentence level. To this end, a prerequisite step is AMR coreference resolution, which aims to find the AMR components referring to the same entity. Figure 1 shows the AMR graphs of two consecutive sentences in a document. An AMR coreference resolution model need to identify two coreference cases: “he” refers to “Bill” in the first graph, and “arrive-01” omits an argument “:arg3” that refers to “Paris”. Relatively little research has been done on AMR coreference resolution. Initial attempts (Liu et al., 2015) merge the nodes that have the same surface string. To minimize noise, only named entities and date entities are considered, and they do not consider merging non-identical nodes (e.g., “Bill” and “he” in Figure 1) that are also frequent in reallife situation. Subsequent work considers more 4205 co-reference cases by either manually annotating AMR coreference information (O’Gorman et al., 2018) or taking a pipeline system (Anikina et al., 2020) consisting of a textual coreference resolution model (Lee et al., 2018) and an AMR-to-text aligner (Flanigan et al., 2014). Yet there is little research on automatically resolving coreference ambiguities directly on AMR, making use of AMR graph-structural features. In this work, we formulate AMR coreference resolution as a missing-link prediction problem over AMR graphs, where the input consists of multiple sentence-level AMRs, and the goal is to recover the missing coreference links connecting the AMR nodes that represent to the same entity. There are two types of links. The first type corresponds to the standard situation, where the edge connects two entity nodes (e.g., “Bill” and “he” in Figure 1) that refer to the same entity. The second type is the implicit role coreference, where one node (e.g., “Paris” in Figure 1) is a dropped argument (“:arg3”) of other predicate node (“arrive-01”). We propose an AMR coreference resolution model by extending an end-to-end text-based coreference resolution model (Lee et al., 2017). In particular, we use a graph neural network to represent input AMRs for inducing expressive features. To enable cross-sentence information exchange, we make connections between sentence-level AMRs by linking their root nodes. Besides, we introduce a concept identification module to distinguish functional graph nodes (non-concept nodes, e.g., “person” in Figure 1), entity nodes (e.g., “Bill”), verbal nodes with implicit role (e.g., “arrive-01”) and other regular nodes (e.g., “leave-11”) to help improve the performance. The final antecedent prediction is conducted between the selected nodes and all their possible antecedent candidates, following previous work on textual coreference resolution (Lee et al., 2017). Experiments on the MS-AMR benchmark1 (O’Gorman et al., 2018) show that our model outperforms competitive baselines by a large margin. To verify the effectiveness and generalization of our proposed model, we annotate an out-of-domain test set over the gold AMR Little Prince 3.0 data following the guidelines of O’Gorman et al. (2018), and the corresponding results show that our model is consistently more robust than the baselines in domain-transfer scenarios. Finally, results on docu1It consists gold coreference links on gold AMRs. ment abstractive summarization show that our document AMRs lead to much better summary quality compared to the document AMRs by Liu et al. (2015). This further verifies the practical value of our approach. Our code and data is available at https://github.com/Sean-Blank/AMRcoref 2 Model Formally, an input instance of AMR coreference resolution consists of multiple sentence-level AMRs G1, G2, ..., Gn, where each Gi can be written as Gi = ⟨Vi, Ei⟩with Vi and Ei representing the corresponding nodes and edges for Gi. We consider a document-level AMR graph ˆG = [G1, G2, ..., Gn; ˆe1, ˆe2, ..., ˆem], where each ˆei is a coreference link connecting two nodes from different sentence-level AMRs. The task of AMR coreference resolution aims to recover ˆe1, ..., ˆem, which are missing from the inputs. Figure 2 shows the architecture of our model, which consists of a graph encoder (§ 2.1), a concept identifier (§ 2.2), and an antecedent prediction module (§ 2.3). 2.1 Representing Input AMRs using GRN Given sentence-level AMRs G1, ..., Gn as the input, randomly initialized word embeddings are adopted to represent each node vk as a dense vector ek. To alleviate data sparsity and to obtain better node representation, character embeddings echar k are computed by using a character-level CNN. We concatenate both ek and echar k embeddings for each concept before using a linear projection to form the initial representation: xk = W node([ek; echar k ]) + bnode, (1) where W node and bnode are model parameters. To enable global information exchange across different sentence-level AMRs, we construct a draft document-level graph by connecting the root nodes of each AMR subgraph as shown in Figure 2. This is important because AMR coreference resolution involves cross-sentence reasoning. We then adopt Graph Recurrent Network (GRN, Song et al., 2018; Zhang et al., 2018; Beck et al., 2018) to obtain rich document-level node representations. GRN is one type of graph neural network that iteratively updates its node representations with the message passing framework (Scarselli et al., 2009). Compared with alternatives such as Graph Convolutional Network (GCN, Kipf and Welling 2017; 4206 FFNN & SOFTMAX leave-11 person name Bill … arrive-01 he date-entity … 𝑣𝑣1 𝑠𝑠𝑠 𝑣𝑣2 𝑠𝑠𝑠 𝑣𝑣3 𝑠𝑠𝑠 𝑣𝑣4 𝑠𝑠𝑠 𝑣𝑣1 𝑠𝑠2 𝑣𝑣2 𝑠𝑠2 𝑣𝑣3 𝑠𝑠𝑠 … … :name :op1 :arg1 :arg0 :time :CONNECT Input Representation GRN Encoder Concept Identification Antecedent Prediction ℒ𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡+ ℒ𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎= 𝓛𝓛 he SOFTMAX 𝑠𝑠(leave-11, he) 𝑠𝑠(Bill, he) 𝑠𝑠(dummy 𝜖𝜖, he) 𝑠𝑠(arrive-01, he) Predicted link: (Bill, he) :𝒆𝒆𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 :𝒆𝒆𝑐𝑐𝑐𝑐𝑐𝑐𝑐 :𝒉𝒉𝑙𝑙 :𝒆𝒆𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 :𝒉𝒉𝐿𝐿 :𝒉𝒉𝑚𝑚 : 𝑠𝑠𝑚𝑚 :𝑠𝑠𝑎𝑎𝑛𝑛 :𝑠𝑠 ∶dropped Figure 2: Model framework for end-to-end AMR coreference resolution. Bastings et al. 2017) and Graph Attention Network (GAT, Veliˇckovi´c et al. 2018), GRN has been shown to give competitive results. Message passing In the message passing framework, a node vk receives information from its directly connected neighbor nodes at each layer l. We use a hidden state vector hl k to represent each node, and the initial state h0 k is defined as a vector of zeros. In the first step at each message passing layer, the concept representation of each neighbor of vk is combined with the corresponding edge representation to make a message xk,j. This is because edges contain semantic information that are important for learning global representation and subsequent reasoning. Formally, a neighbor vj of node vk can be represented as xk,j = W node([ej; echar j ; elabel k,j ]) + bnode, (2) where elabel k,j denotes the label embedding of the edge from node vk and to vj. Next, representations of neighboring nodes from the incoming and outgoing directions are aggregated: xin k = X i∈Nin(k) xl i,k xout k = X j∈Nout(k) xl k,j xl k = [xin k , xout k ], (3) where Nin(k) and Nout(k) denote the set of incoming and outgoing neighbors of vk, respectively. Similarly, the hidden states from incoming and outgoing neighbors are also summed up: min k = X i∈Nin(k) hl−1 i mout k = X j∈Nout(k) hl−1 j ml k = [min k , mout k ], (4) where hl−1 j denotes the hidden state vector for node vj at the previous (l−1) layer. Finally, the message passing from layer l −1 to l is conducted following the gated operations of LSTM (Hochreiter and Schmidhuber, 1997): il k = σ(W m i ml k + W x i xl k + bi) ol k = σ(W m o ml k + W x o xl k + bo) f l k = σ(W m f ml k + W x f xl k + bf) ul k = σ(W m u ml k + W x u xl k + bu) cl k = f l k ⊙cl−1 k + il k ⊙ul k hl k = ol k ⊙tanh(cl k), (5) where il k, ol k and f l k are a set of input, output and forget gates to control information flow from different sources, ul k represents the input messages, cl k is the cell vector to record memory, and c0 k is also initialized as a vector of zeros. W m z , W x z and bz (z ∈{i, o, f, u}) are model parameters. We adopt L GRN layers in total, where L is determined by a development experiment. The output hL k at layer L is adopted as the representation of each node vk for subsequent procedures. 4207 2.2 Concept Identification Concept identification aims to distinguish the AMR nodes in regard to its concept type. We consider 6 concept types T = {func, ent, ver0, ver1, ver2, reg}, which denotes the functional nodes, entity concepts, verbal concepts verx with implicit arguments (i.e., “:argx” x ∈{0, 1, 2}2) and other regular nodes (e.g., “leave-11”), respectively. This module is comparable to the mention detection procedure in textual coreference resolution (Lee et al., 2017). Formally, a concept representation hL k from the top GRN layer is concatenated with a learnable type embedding etype k (t) of type t for each concept vk, and the corresponding type score sk type(t) is computed using a feed-forward network: sk type(t) = FFNNtype(W type[hL k ; etype k (t)]), (6) where W type is a mapping matrix. etype k (t) represents a concept-type embedding and is randomly initialized. A probability distribution P(t|vk) over all concept types T for each concept vk is calculated as follows using a softmax layer: P(t|vk) = esk type(t) P t′∈T esk type(t′) . (7) Finally, we predicate the type t∗ k for each concept t∗ k = argmaxt∈T sk type(t), (8) and use it to filter the input nodes. In particular, functional concepts are dropped directly and the other concepts (i.e., ent, ver0, ver1, ver2, reg) are selected as candidate nodes for antecedent prediction. 2.3 Antecedent Prediction Given a selected node vk by the concept identifier, the goal is to predict its antecedent yk from all possible candidate nodes Yk = {ϵ, yπ, ..., yk−1}, where a dummy antecedent ϵ is adopted for the nodes that are not coreferent with any previous concepts. π = min(1, k −ψ), where ψ represents the maximum antecedents considered as candidates. As mentioned by previous work on textual coreference resolution (Lee et al., 2017), considering too many candidates can hurt the final performance. We conduct development experiments to decide the best ψ. The finally predicted coreference links implicitly determine the coreference clusters. 2We do not model other :argx to avoid long tail issue. Type information in § 2.2 can help to guide the antecedent prediction and ensure global type consistency. We combine the node hidden vector and its type representation as the final concept state: hm k = [hL k ; etype k (t∗)], (9) where etype k (t∗) denotes the learned embedding of the concept type of node vk. Similar with Lee et al. (2017), the goal of the antecedent prediction module is to learn a distribution Q(yk) over the antecedents for each node vk: Q(yk) = es(k,yk) P y′∈Y(k) es(k,y′) (10) where s(k, a) computes a coreference link score for each concept pair (vk, va): s(k, a) = sm(k) + sm(a) + san(k, a). (11) Here a < k, and sm(k) means whether concept vk is a mention involved in a coreference link. It is calculated by using a feed-forward network: sm(k) = FFNNm(hm k ). (12) san(k, a) indicates whether mention va is an antecedent of vk and measures the semantic similarity between vk and va, computed with rich features using a feed-forward network: san(k, a) = FFNNan([hm k , hm a , hm k ◦hm a , φ(k, a)]) (13) where ◦denotes element-wise multiplication of each mention pair (vk, va), and a feature vector φ(k, a) represents the normalized distance between two mentions and the speaker information if available. Following Lee et al. (2017), we also normalize the distance values by grouping them into the following buckets [1, 2, 3, 4, 5-7, 8-15, 16-31, 32-63, 64+]. All features (speaker, distance, concept type) are randomly initialized 32-dimensional embeddings jointly learned with the model. 2.4 Training Our objective function takes two parts: Ltype(θ) (i.e., the concept-type identification loss), and Lantecedent (i.e., the antecedent prediction loss) L(θ) = Ltype(θ) + λLantecedent(θ), (14) where λ is the weight coefficient (we empirically set λ = 0.1 in this paper). 4208 Data (portion) #Doc #AMR #Links #Nodes MS-AMR (Train) 273 7705 12003 86704 MS-AMR (Dev) 9 121 216 1599 MS-AMR (Test) 9 201 404 2745 LP (Test) 6 282 463 2333 Table 1: Statistics of MS-AMR (first group) and our annotated out-of-domain test data based on LP corpus. Concept Identification Loss. Ltype measures whether our model can accurately identify meaningful concepts and learn the correct type representations. Specifically, given the concept set V = {v1, ...vN}, the concept identifier is trained to minimize an average cross-entropy loss: Ltype(θ) = −1 N N X k=1 logP(t∗ k|vk), (15) where θ are the set of model parameters, P(t∗ k|vk) denotes the output probability of predicted type t∗ k for each node vk as in Eq. 7. Antecedent Prediction Loss. Given a training AMR document with gold coreference clusters GOLD(k)|N k=1 and antecedent candidates Yk = {ϵ, yπ, ..., yk−1} for mention vk, Lantecedent measures whether mentions are linked to their correct antecedent. Since the antecedents are latent, the antecedent loss is a marginal log-likelihood of all correct antecedents implied by gold clustering: Lantecedent(θ) = N Y k=1 X y∈Yk∩GOLD(k) logQ(y) (16) where GOLD(k) = ϵ if mention vk does not belong to any gold cluster. Q(y) is calculated using Eq. 10. 3 Experiments We conduct experiments on the MS-AMR dataset3 (O’Gorman et al., 2018), which is annotated over a previous gold AMR corpus (LDC2017T10). It has 293 annotated documents in total with an average of 27.4 AMRs per document, covering roughly 10% of the total AMR corpus. We split a dev data with the same size as the test set from the training set. Following the annotation guidelines of MSAMR, we manually annotate the AMR coreference 3The MS-AMR dataset considers 3 types of coreference links: regular, implicit and part-whole. We ignore the last type, which has been challenging and ignored since textual coreference resolution. resolution information over the development and test data of the Little Prince (LP) AMR corpus4 and use it as an out-of-domain test set. For this dataset, we consider each chapter as a document. The data statistics are shown in Table 1. 3.1 Setup Evaluation Metrics We use the standard evaluation metrics for coreference resolution evaluation, computed using the official CoNLL-2012 evaluation toolkit. Three measures include: MUC (Vilain et al., 1995), B3 (Bagga and Baldwin, 1998) and CEAFφ4 (Luo, 2005). Following previous studies (Lee et al., 2018), the primary metric AVG-F is the unweighted average of the above three F-scores. Baselines To study the effectiveness of end-toend AMR coreference resolution, we compare our model with the following baselines: • Rule-based (Liu et al., 2015): a heuristic method that builds a large document-level AMR graph by linking identical entities. • Pipeline (Anikina et al., 2020): it uses an off-theshelf coreference system (Lee et al., 2018) with SpanBERT (Joshi et al., 2020) embeddings, and an AMR-to-text aligner (Flanigan et al., 2014). The former generates coreference from text, and the later projects this information from text to AMRs. Models We study two versions of our model with or without BERT features. • AMRcoref-base: it corresponds to our model described in § 2 only with word embeddings. • AMRcoref-bert: it denotes our model in § 2 except that the word embeddings (ek in Eq. 1) are concatenated with BERT outputs. Specifically, we use a cased BERT-base model with fixed parameters to encode a sentence, taking an AMRto-text aligner (Flanigan et al., 2014) to project BERT outputs to the corresponding AMR nodes. Hyperparameters We set the dimension of concept embeddings to 256. Characters in the character CNN (§ 2.1) are represented as learned embeddings with 32 units and the convolution window sizes include 2, 3, and 4 characters, each consisting of 100 filters. We use Adam (Kingma and Ba, 2015) with a learning rate of 0.005 for optimization. 4https://amr.isi.edu/download/amr-bank-struct-v3.0.txt. 4209 Model In-domain test set Out-domain test set MUC B3 CEAFφ4 Average F1 MUC B3 CEAFφ4 Average F1 Rule-based 50.8 41.1 22.4 38.1 53.3 41.7 25.9 40.3 Pipeline 58.0 43.0 25.0 42.0 55.2 42.3 26.7 41.4 AMRcoref-base 66.1 49.7 38.1 51.3 64.4 45.8 31.4 47.2 AMRcoref-bert 72.5 64.1 50.6 62.4 69.9 61.9 48.5 60.1 Table 2: Main results on the MS-AMR data and LP test sets. Number of GRN layers Figure 3: Development results of AMRcoref-base on the number of GRN layers. 3.2 Development Experiments We first conduct development experiment to choose the values for the crucial hyperparameters. GRN Encoder Layers The number of recurrent layers L in GRN defines the amount of message interactions. Large message passing layers may lead to over-smoothing problems, while small layers may result in weak graph representation (Qin et al., 2020; Zhang et al., 2018). Figure 3 shows development experiments of the AMRcoref-base model in this aspect. We observe large improvements when increasing the layers from 1 to 3, but further increase from 3 to 7 does not lead to further improvements. Therefore, we choose 3 layers for our final model. Antecedent Candidates How many antecedents are considered as candidates (denoted as ψ in Section 2.3) for making each coreference decision is another important hyperparameter in a coreference resolution model (Lee et al., 2017). Intuitively, allowing more antecedents gives a higher upper bound, but that also introduces a larger search space. Table 3 shows the statistics of the distance between each mention and its gold antecedent and the devset performance of AMRcoref-base model that uses this distance as the search space. The performance of AMRcoref-base improves when increasing the search space, and the best performance was observed when 250 antecedents are considered as the search space. We choose ψ =250 in subsequent experiments. Distances. #Links Cover(%) F1 ≤50 184 85.2 42.9 ≤100 206 95.4 45.2 ≤150 212 98.1 45.4 ≤200 214 99.1 47.2 ≤250 215 99.5 52.1 ≤300 216 100.0 49.7 > 300 216 100.0 48.3 Table 3: Devset statistics on mention-gold-antecedent distance and the performances of AMRcoref-base using the distance as the search space. 3.3 Main Results Table 2 shows the final in-domain results on the MS-AMR test set and out-domain results on the annotated Little Prince (LP) data, where we compare our model (AMRcoref-base and AMRcoref-bert) with the Rule-based and Pipeline baselines. In-domain Results The Rule-based method performs the worst, because it only links the identical entity and suffers from low recall. The Pipeline model performs better than the Rule-based model due to better coverage, but it can suffer from error propagation in both textual coreference and inaccurate AMR aligner. In addition, it does not make use of AMR structure features, which is less sparse compared to text cues. Our proposed AMRcorefbase model outperforms the two baselines by a huge margin, gaining at least 9.3% and 13.2% average F1 scores, respectively. This verifies the effectiveness of the end-to-end framework. Out-domain Results On the cross-domain LP data, our model largely outperforms both Rulebased method and the Pipeline model. Compared with the in-domain setting, there is minor drop on the out-of-domain dataset (4.1% and 2.3% F1 score for AMRcoref-base and AMRcoref-bert respectively). Neither the performances of Rulebased nor Pipeline change much on this dataset, which is because these systems are not trained on a certain domain. This also reflects the quality of our LP annotations, because of the consistent performance changes of both AMRcoref-base and AMRcoref-bert when switching from MS-AMR to LP. 4210 Model Average F1 ∆ AMRcoref-base 51.3 - concept identification 31.4 -19.9 + gold mention 70.4 +19.1 + bert concatenate 62.4 +11.1 + bert graph 62.0 +10.7 - distance feature 49.2 -2.1 - speaker feature 49.4 -1.9 - character CNN 50.1 -1.2 - graph connections 49.0 -2.3 Table 4: Ablation study on the test set of MS-AMR. 3.4 Analysis We analyze the effects of mention type, textual embedding and various extra features in this section. Concept Identification As shown in the first group of Table 4, we conduct an ablation study on the concept identification module, which has been shown crucial on the textual coreference resolution (Lee et al., 2017). Removing the concept identifier from the AMRcoref-base model results in a large performance degradation of up to 19.9%, indicating that concept type information of the AMR node can positively guide the prediction of coreference links. On the other hand, when the concept identifier outputs are replaced with gold mentions, the results can be further improved by 19.1%. This indicates that better performances can be expected if concept identification can be further improved. Injecting BERT knowledge As shown in the second group of Table 4, we study the influence of rich features from BERT in our model, which has been proven effective on text-based coreference resolution. Two alternatives of using BERT are studied, concatenate (i.e. AMRcoref-bert) denotes concatenating the AMR node embeddings with the corresponding textual BERT embedding, and graph means that we construct an AMR-token graph that connects AMR nodes and the corresponding tokens. We find that the AMRcoref-base model can be improved by a similar margin using both approaches. This is consistent with existing observations from other structured prediction tasks, such as constituent parsing (Kitaev et al., 2019) and dependency parsing (Li et al., 2019). Due to the limited scale of our training data, we expect the gain to be less with more training data. Features Ablation As shown by the last group in Table 4, we investigate the impacts of each component in our proposed model on the development set of MS-AMR. We have the following observations. First, consistent with findings of Lee et al. (2017), Figure 4: Testing results of AMRcoref-base regarding different ratios of training data used. The F1 score of Pipeline is 42.0% (Table 2). the distance between a pair of AMR concepts is an important feature. The final model performance drops by 2.1% when removing the distance feature (Eq. 13). Second, the speaker indicator features (Eq. 13) contribute to our model by a 1.9% improvement. Intuitively, speaker information is helpful for pronoun coreference resolution in dialogues. For example, “my package” in one sentence may represent identical entity with “your package” in the next utterance. Third, the character CNN provides morphological information and a way to back off for out-of-vocabulary tokens. For AMR node representations, we see a modest contribution of 1.2% F1 score. Finally, we exploit the necessity of cross-sentence AMR connections. Compared to encoding each AMR graph individually, global information exchange across sentences can help to achieve a significant performance improvement. Data Hunger Similar to other results, it is important to study how much data is necessary to obtain a strong performance (at least be better than the baseline). Figure 4 shows the performances when training the AMRcoref-base model on different portions of data. As the number of training samples increases, the performance of our model continuously improves. This shows that our model has room for further improvement with more training data. Moreover, our model even outperforms the Pipeline baseline when trained on only 20% data. This confirms the robustness of our end-toend framework. Effect of Document Length Figure 5 shows the performance on different MS-AMR document lengths (i.e., the number of AMR graphs in the document). We can see that both our model and the Pipeline model show performance decrease 4211 38.5 37.8 37.5 38.6 45.6 42.8 40.9 38.7 59.6 57.9 49.5 38.2 10 20 30 40 50 60 70 0-10 10-20 20-30 30up Average F1 Document Length Performance on document length Rule-based Pipeline AMRcoref-base Figure 5: Testing results regarding document length. when increasing input document length. This is likely because a longer document usually involves more complex coreference situations and brings more challenge for the encoder. Insufficient information interaction for distant nodes further leads to weaker inference performance. As expected, the Rule-based approach (Liu et al., 2015) is not significantly affected, but its result is still pretty low. When the document contains more than 30 sentences, the AMRcoref-base model slightly under-performs both the Rule-based method and the Pipeline baseline. One reason is that only a few training instances have a long document length, so we expect that the performance of our model can be further improved given more long documents. 3.5 Application on Summarization Table 5 compares the summarization performances using the document-level AMRs generated by various methods on the LDC2015E86 benchmark (Knight et al., 2014). Following Liu et al. (2015), Rouge scores (R-1/2/L Lin 2004) are used as the metrics. To consume each document AMR and the corresponding text, we take a popular dual-tosequence model (D2S, Song et al. 2019b), which extends the standard sequence-to-sequence framework with an additional graph encoder and a dual attention mechanism for extracting both text and graph contexts during decoding. For previous work, summarization using AMR was first explored by Liu et al. (2015). They first use a rule-based method to build document AMRs and then take a statistic model to generate summaries. Dohare et al. (2017) improves this approach by selecting important sentences before building a document AMR. The D2S-Rule-based can be considered as a fair comparison with Liu et al. (2015) on the same summerization platform. Model R-1 R-2 R-L Liu et al. (2015) 44.3 – – Dohare et al. (2017) 44.8 17.3 30.6 D2S-Rule-based 47.6 20.1 32.5 D2S-Pipeline 47.9 19.5 32.6 D2S-AMRcoref-base 48.4 20.4 33.2 D2S-AMRcoref-bert 49.1 20.5 33.6 Table 5: Test summarization results on LDC2015E86. R-1/2/L is short for Rouge-1/2/L. The overall performance of the D2S models outperform the previous approaches, indicating that our experiments are conducted on a stronger baseline. Though Pipeline is better than Rule-based on AMR coreference resolution, D2S-Pipeline is comparable with D2S-Rule-based on the downstream summerization task. This shows that the error propagation issue of Pipeline can introduce further negative effects to a downstream application. On the other hand, both D2S-AMRcoref-base and D2SAMRcoref-bert show much better results than the baselines across all Rouge metrics. This demonstrates that the improvements made by our end-toend model is solid and can transfer to a downstream application. D2S-AMRcoref-bert achieves the best performance, which is consistent with the above experiments. 4 Related Work Multi-sentence AMR Although some previous work (Szubert et al., 2020; Van Noord and Bos, 2017) explore the coreference phenomena of AMR, they mainly focus on the situation within a sentence. On the other hand, previous work on multi-sentence AMR primarily focuses on data annotation. Song et al. (2019a) annotate dropped pronouns over Chinese AMR but only deals with implicit roles in specific constructions. Gerber and Chai (2012) provide implicit role annotations, but the resources were limited to a small inventory of 5-10 predicate types rather than all implicit arguments. O’Gorman et al. (2018) annotated the MS-AMR dataset by simultaneously considering coreference, implicit role coreference and bridging relations. We consider coreference resolution as the prerequisite for creating multi-sentence AMRs, proposing the first end-to-end model for this task. Coreference Resolution Coreference resolution is a fundamental problem in natural language processing. Neural network models have shown promising results over the years. Recent work (Lee et al., 2017, 2018; Kantor and Globerson, 2019) 4212 tackled the problem end-to-end by jointly detecting mentions and predicting coreference. Lee et al. (2018) build a complete end-to-end system with the span-ranking architecture and higher-order inference technique. While previous work considers only text-level coreference, we investigate AMR co-reference resolution. AMR Representation using GNN To encode AMR graphs, many variants of GNNs such as GRNs (Song et al., 2018; Beck et al., 2018), GCNs (Zhou et al., 2020; Zhang et al., 2020) and GATs (Damonte and Cohen, 2019; Cai and Lam, 2020b; Wang et al., 2020) have been introduced. We choose a classic GRN model following Song et al. (2018) to represent our document-level AMR graph and leave the exploiting on a more efficient GNN structure for future work. 5 Conclusion We investigated a novel end-to-end multi-sentence AMR coreference resolution model using a graph neural network. Compared with previous rulebased and pipeline methods, our model better captures multi-sentence semantic information. Results on MS-AMR (in-domain) and LP (out-of-domain) datasets show the superiority and robustness of our model. In addition, experiments on the downstream text summarization task further demonstrate the effectiveness of the document-level AMRs produced by our model. In future work, we plan to resolve both the crossAMR coreference links and the sentence-level ones together with our model. Acknowledgments Linfeng Song is the corresponding author. We would like to thank the anonymous reviewers for their insightful comments. We gratefully acknowledge funding from the National Natural Science Foundation of China (NSFC No.61976180). It also receives supported by Tencent AI Lab Rhino-Bird Focused Research Program. References Tatiana Anikina, Alexander Koller, and Michael Roth. 2020. Predicting coreference in Abstract Meaning Representations. In Proceedings of the Third Workshop on Computational Models of Reference, Anaphora and Coreference, pages 33–38, Barcelona, Spain (online). Association for Computational Linguistics. Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. In The first international conference on language resources and evaluation workshop on linguistics coreference, volume 1, pages 563–566. Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract Meaning Representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, pages 178–186, Sofia, Bulgaria. Association for Computational Linguistics. Jasmijn Bastings, Ivan Titov, Wilker Aziz, Diego Marcheggiani, and Khalil Sima’an. 2017. Graph convolutional encoders for syntax-aware neural machine translation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1957–1967, Copenhagen, Denmark. Association for Computational Linguistics. Daniel Beck, Gholamreza Haffari, and Trevor Cohn. 2018. Graph-to-sequence learning using gated graph neural networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 273–283, Melbourne, Australia. Association for Computational Linguistics. Claire Bonial, Lucia Donatelli, Mitchell Abrams, Stephanie M. Lukin, Stephen Tratz, Matthew Marge, Ron Artstein, David Traum, and Clare Voss. 2020. Dialogue-AMR: Abstract Meaning Representation for dialogue. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 684–695, Marseille, France. European Language Resources Association. Deng Cai and Wai Lam. 2020a. AMR parsing via graph-sequence iterative inference. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1290–1301, Online. Association for Computational Linguistics. Deng Cai and Wai Lam. 2020b. Graph transformer for graph-to-sequence learning. In AAAI, pages 7464– 7471. Marco Damonte and Shay B. Cohen. 2019. Structural neural encoders for AMR-to-text generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3649–3658, Minneapolis, Minnesota. Association for Computational Linguistics. Shibhansh Dohare, Harish Karnick, and Vivek Gupta. 2017. Text summarization using abstract meaning representation. arXiv preprint arXiv:1706.01678. Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith. 2014. A discrim4213 inative graph-based parser for the Abstract Meaning Representation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1426– 1436, Baltimore, Maryland. Association for Computational Linguistics. DongLai Ge, Junhui Li, Muhua Zhu, and Shoushan Li. 2019. Modeling source syntax and semantics for neural amr parsing. In Proceedings of the TwentyEighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 4975–4981. International Joint Conferences on Artificial Intelligence Organization. Matthew Gerber and Joyce Y. Chai. 2012. Semantic role labeling of implicit arguments for nominal predicates. Computational Linguistics, 38(4):755–798. Hardy Hardy and Andreas Vlachos. 2018. Guided neural language generation for abstractive summarization using Abstract Meaning Representation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 768– 773, Brussels, Belgium. Association for Computational Linguistics. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Ahmad Issa Alaa Aldine, Mounira Harzallah, Giuseppe Berio, Nicolas B´echet, and Ahmad Faour. 2018. EXPR at SemEval-2018 task 9: A combined approach for hypernym discovery. In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 919–923, New Orleans, Louisiana. Association for Computational Linguistics. Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. SpanBERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64–77. Ben Kantor and Amir Globerson. 2019. Coreference resolution with entity equalization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 673–677, Florence, Italy. Association for Computational Linguistics. Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Thomas N. Kipf and Max Welling. 2017. Semisupervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR). Nikita Kitaev, Steven Cao, and Dan Klein. 2019. Multilingual constituency parsing with self-attention and pre-training. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3499–3505, Florence, Italy. Association for Computational Linguistics. Kevin Knight, Laura Baranescu, Claire Bonial, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Daniel Marcu, Martha Palmer, and Nathan Schneider. 2014. Deft phase 2 amr annotation r1 ldc2015e86. philadelphia: Linguistic data consortium. Abstract meaning representation (AMR) annotation release, 1. Kenton Lee, Luheng He, Mike Lewis, and Luke Zettlemoyer. 2017. End-to-end neural coreference resolution. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 188–197, Copenhagen, Denmark. Association for Computational Linguistics. Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-tofine inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 687–692, New Orleans, Louisiana. Association for Computational Linguistics. Xiang Li, Thien Huu Nguyen, Kai Cao, and Ralph Grishman. 2015. Improving event detection with Abstract Meaning Representation. In Proceedings of the First Workshop on Computing News Storylines, pages 11–15, Beijing, China. Association for Computational Linguistics. Ying Li, Zhenghua Li, Min Zhang, Rui Wang, Sheng Li, and Luo Si. 2019. Self-attentive biaffine dependency parsing. In IJCAI, pages 5067–5073. Kexin Liao, Logan Lebanoff, and Fei Liu. 2018. Abstract Meaning Representation for multi-document summarization. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1178–1190, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81. Fei Liu, Jeffrey Flanigan, Sam Thomson, Norman Sadeh, and Noah A. Smith. 2015. Toward abstractive summarization using semantic representations. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1077–1086, Denver, Colorado. Association for Computational Linguistics. Xiaoqiang Luo. 2005. On coreference resolution performance metrics. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 25–32, Vancouver, British Columbia, Canada. Association for Computational Linguistics. 4214 Chunchuan Lyu and Ivan Titov. 2018. AMR parsing as graph prediction with latent alignment. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 397–407, Melbourne, Australia. Association for Computational Linguistics. Tahira Naseem, Abhishek Shah, Hui Wan, Radu Florian, Salim Roukos, and Miguel Ballesteros. 2019. Rewarding Smatch: Transition-based AMR parsing with reinforcement learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4586–4592, Florence, Italy. Association for Computational Linguistics. Tim O’Gorman, Michael Regan, Kira Griffitt, Ulf Hermjakob, Kevin Knight, and Martha Palmer. 2018. AMR beyond the sentence: the multi-sentence AMR corpus. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3693–3702, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Xiaoman Pan, Taylor Cassidy, Ulf Hermjakob, Heng Ji, and Kevin Knight. 2015. Unsupervised entity linking with abstract meaning representation. In Proceedings of the 2015 conference of the north american chapter of the association for computational linguistics: Human language technologies, pages 1130–1139. Libo Qin, Xiao Xu, Wanxiang Che, and Ting Liu. 2020. AGIF: An adaptive graph-interactive framework for joint multiple intent detection and slot filling. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1807–1816, Online. Association for Computational Linguistics. Sudha Rao, Daniel Marcu, Kevin Knight, and Hal Daum´e III. 2017. Biomedical event extraction using Abstract Meaning Representation. In BioNLP 2017, pages 126–135, Vancouver, Canada,. Association for Computational Linguistics. F. Scarselli, M. Gori, A. C. Tsoi, M. Hagenbuchner, and G. Monfardini. 2009. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80. Li Song, Yuan Wen, Sijia Ge, Bin Li, and Weiguang Qu. 2019a. An easier and efficient framework to annotate semantic roles: Evidence from the chinese amr corpus. In Workshop on Chinese Lexical Semantics, pages 474–485. Springer. Linfeng Song, Daniel Gildea, Yue Zhang, Zhiguo Wang, and Jinsong Su. 2019b. Semantic neural machine translation using AMR. Transactions of the Association for Computational Linguistics, 7:19–31. Linfeng Song, Yue Zhang, Zhiguo Wang, and Daniel Gildea. 2018. A graph-to-sequence model for AMRto-text generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1616– 1626, Melbourne, Australia. Association for Computational Linguistics. Ida Szubert, Marco Damonte, Shay B. Cohen, and Mark Steedman. 2020. The role of reentrancies in Abstract Meaning Representation parsing. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2198–2207, Online. Association for Computational Linguistics. Rik Van Noord and Johan Bos. 2017. Dealing with co-reference in neural semantic parsing. In Proceedings of the 2nd Workshop on Semantic Deep Learning (SemDeep-2), pages 41–49. Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li`o, and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations. Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A modeltheoretic coreference scoring scheme. In Sixth Message Understanding Conference (MUC-6): Proceedings of a Conference Held in Columbia, Maryland, November 6-8, 1995. Tianming Wang, Xiaojun Wan, and Hanqi Jin. 2020. AMR-to-text generation with graph transformer. Transactions of the Association for Computational Linguistics, 8:19–33. Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR parsing as sequence-tograph transduction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 80–94, Florence, Italy. Association for Computational Linguistics. Yan Zhang, Zhijiang Guo, Zhiyang Teng, Wei Lu, Shay B. Cohen, Zuozhu Liu, and Lidong Bing. 2020. Lightweight, dynamic graph convolutional networks for AMR-to-text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2162–2172, Online. Association for Computational Linguistics. Yue Zhang, Qi Liu, and Linfeng Song. 2018. Sentencestate LSTM for text representation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 317–327, Melbourne, Australia. Association for Computational Linguistics. Qiji Zhou, Yue Zhang, Donghong Ji, and Hao Tang. 2020. AMR parsing with latent structural information. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4306–4319, Online. Association for Computational Linguistics.
|
2021
|
324
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4215–4228 August 1–6, 2021. ©2021 Association for Computational Linguistics 4215 How is BERT surprised? Layerwise detection of linguistic anomalies Bai Li1,4, Zining Zhu1,4 Guillaume Thomas2, Yang Xu1,3,4, Frank Rudzicz1,4,5 1 University of Toronto, Department of Computer Science 2 University of Toronto, Department of Linguistics 3 University of Toronto, Cognitive Science Program 4 Vector Institute for Artificial Intelligence 5 Unity Health Toronto {bai, zining, yangxu, frank}@cs.toronto.edu [email protected] Abstract Transformer language models have shown remarkable ability in detecting when a word is anomalous in context, but likelihood scores offer no information about the cause of the anomaly. In this work, we use Gaussian models for density estimation at intermediate layers of three language models (BERT, RoBERTa, and XLNet), and evaluate our method on BLiMP, a grammaticality judgement benchmark. In lower layers, surprisal is highly correlated to low token frequency, but this correlation diminishes in upper layers. Next, we gather datasets of morphosyntactic, semantic, and commonsense anomalies from psycholinguistic studies; we find that the best performing model RoBERTa exhibits surprisal in earlier layers when the anomaly is morphosyntactic than when it is semantic, while commonsense anomalies do not exhibit surprisal at any intermediate layer. These results suggest that language models employ separate mechanisms to detect different types of linguistic anomalies. 1 Introduction Transformer-based language models (LMs) have achieved remarkable success in numerous natural language processing tasks, prompting many probing studies to determine the extent of their linguistic knowledge. A popular approach is to formulate the problem as a multiple-choice task, where the LM is considered correct if it assigns higher likelihood to the appropriate word than an inappropriate one, given context (Gulordava et al., 2018; Ettinger, 2020; Warstadt et al., 2020). The likelihood score, however, only gives a scalar value of the degree that a word is anomalous in context, and cannot distinguish between different ways that a word might be anomalous. It has been proposed that there are different types of linguistic anomalies. Chomsky The cat won 't eating the food 0 1 2 3 4 5 6 7 8 9 10 11 12 Layer The plane laughed at the runway 0 1 2 3 4 5 6 7 8 9 10 11 12 Figure 1: Example sentence with a morphosyntactic anomaly (left) and semantic anomaly (right) (anomalies in bold). Darker colours indicate higher surprisal. We investigate several patterns: first, surprisal at lower layers corresponds to infrequent tokens, but this effect diminishes towards upper layers. Second, morphosyntactic violations begin to trigger high surprisals at an earlier layer than semantic violations. (1957) distinguished semantic anomalies (“colorless green ideas sleep furiously”) from ungrammaticality (“furiously sleep ideas green colorless”). Psycholinguistic studies initially suggested that different event-related potentials (ERPs) are produced in the brain depending on the type of anomaly; e.g., semantic anomalies produce negative ERPs 400 ms after the stimulus, while syntactic anomalies produce positive ERPs 600 ms after (Kutas et al., 2006). Here, we ask whether Transformer LMs show different surprisals in their intermediate layers depending on the type of anomaly. However, LMs do not compute likelihoods at intermediate layers – only at the final layer. 4216 In this paper, we introduce a new tool to probe for surprisal at intermediate layers of BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and XLNet (Yang et al., 2019), formulating the problem as density estimation. We train Gaussian models to fit distributions of embeddings at each layer of the LMs. Using BLiMP (Warstadt et al., 2020) for evaluation, we show that this model is effective at grammaticality judgement, requiring only a small amount of in-domain text for training. Figure 1 shows the method using the RoBERTa model on two example sentences. We apply our model to test sentences drawn from BLiMP and 7 psycholinguistics studies, exhibiting morphosyntactic, semantic, and commonsense anomalies. We find that morphosyntactic anomalies produce out-of-domain embeddings at earlier layers, semantic anomalies at later layers, and no commonsense anomalies, even though the LM’s final accuracy is similar. We show that LMs are internally sensitive to the type of linguistic anomaly, which is not apparent if we only had access to their softmax probability outputs. Our source code and data are available at: https://github.com/SPOClab-ca/ layerwise-anomaly. 2 Related work 2.1 Probing LMs for linguistic knowledge Soon after BERT’s release, many papers invented probing techniques to discover what linguistic knowledge it contains, and how this information is distributed between layers (e.g., Rogers et al. (2021) provides a comprehensive overview). Tenney et al. (2019) used “edge probing” to determine each layer’s contribution to a task’s performance, and discovered that the middle layers contributed more when the task was syntactic, and the upper layers more when the task was semantic. Several papers found that BERT’s middle layers contain the most syntactic information. Kelly et al. (2020) found that BERT’s middle layers are best at distinguishing between sentences with direct and indirect object constructions. Hewitt and Manning (2019) used a structural probe to recover syntax trees from contextual embeddings, and found the performance peaked in middle layers. Probing results are somewhat dependent on the choice of linguistic formalism used to annotate the data, as Kulmizev et al. (2020) found for syntax, and Kuznetsov and Gurevych (2020) found for semantic roles. Miaschi et al. (2020) examined the layerwise performance of BERT for a suite of linguistic features, before and after fine tuning. Our work further investigates what linguistic information is contained in different layers, with a focus on anomalous inputs. 2.2 Neural grammaticality judgements Many recent probing studies used grammaticality judgement tasks to test the knowledge of specific phenomena in LMs. Warstadt et al. (2019) gathered sentences from linguistic publications, and evaluated by Matthews Correlation with the ground truth. More commonly, the model is presented with a binary choice between an acceptable and unacceptable sentence: BLiMP (Warstadt et al., 2020) used templates to generate 67k such sentence pairs, covering 12 types of linguistic phenomena. Similarly, Hu et al. (2020) created syntactic tests using templates, but defined success criteria using inequalities of LM perplexities. In contrast with artificial templates, Gulordava et al. (2018) generated test cases by perturbing natural corpus data to test long-distance dependencies. Most grammaticality studies focused on syntactic phenomena, but Rabinovich et al. (2019) tested LMs’ sensitivity to semantic infelicities involving indefinite pronouns. 2.3 Tests of selectional restrictions Violations of selectional restrictions are one type of linguistic unacceptability, defined as a semantic mismatch between a verb and an argument. Sasano and Korhonen (2020) examined the geometry of word classes (e.g., words that can be a direct object of the verb ‘play’) in word vector models; they compared single-class models against discriminative models for learning word class boundaries. Chersoni et al. (2018) tested distributional semantic models on their ability to identify selectional restriction violations using stimuli from two psycholinguistic datasets. Finally, Metheniti et al. (2020) tested how much BERT relies on selectional restriction information versus other contextual information for making masked word predictions. 2.4 Psycholinguistic tests for LMs The N400 response is a negative event-related potential that occurs roughly 400ms after a stimulus in human brains, and is generally associated with the stimulus being semantically anomalous with 4217 respect to the preceding context (Kutas and Federmeier, 2011). Although many studies have been performed with a diverse range of linguistic stimuli, exactly what conditions trigger the N400 response is still an open question. Frank et al. (2015) found that the N400 response is correlated with surprisal, i.e., how unlikely an LM predicts a word given the preceding context. Recently, several studies have investigated relationships between surprisal in neural LMs and the N400 response. Michaelov and Bergen (2020) compared human N400 amplitudes with LSTM-based models using stimuli from several psycholinguistic studies. Ettinger (2020) used data from three psycholinguistic studies to probe BERT’s knowledge of commonsense and negation. Our work is similar to the latter – we leverage psycholinguistic studies for their stimuli, but we do not use the their N400 amplitude results. 3 Model We use the transformer language model as a contextual embedding extractor (we write this as BERT for convenience). Let L be the layer index, which ranges from 0 to 12 on all of our models. Using a training corpus {w1, · · · , wT }, we extract contextual embeddings at layer L for each token: x(L) 1 , · · · , x(L) T = BERTL(w1, · · · , wT ). (1) Next, we fit a multivariate Gaussian on the extracted embeddings: x(L) 1 , · · · , x(L) T ∼N(bµL, bΣL). (2) For evaluating the layerwise surprisal of a new sentence s = [t1, · · · , tn], we similarly extract contextual embeddings using the language model: y1, · · · , yn = BERTL(t1, · · · , tn). (3) The surprisal of each token is the negative log likelihood of the contextual vector according to the multivariate Gaussian: Gi = −log p(yi | bµL, bΣL) for i = 1 . . . n. (4) Finally, we define the surprisal of sentence s as the sum of surprisals of all of its tokens, which is also the joint log likelihood of all of the embeddings: surprisalL(t1, · · · , tn) = n X i=1 Gi = −log p(y1, · · · , yn | bµL, bΣL). (5) 3.1 Connection to Mahalanobis distance The theoretical motivation for using the sum of log likelihoods is that when we fit a Gaussian model with full covariance matrix, low likelihood corresponds exactly to high Mahalanobis distance from the in-distribution points. The score given by the Gaussian model is: G = −log p(y | bµL, bΣL) = −log 1 (2π)D/2|bΣL|1/2 exp(−1 2d2) ! , (6) where D is the dimension of the vector space, and d is the Mahalanobis distance: d = q (y −bµL)T bΣ −1 L (y −bµL). (7) Rearranging, we get: d2 = 2G −D log(2π) −log |bΣL|, (8) thus the negative log likelihood is the squared Mahalanobis distance plus a constant. Various methods based on Mahalanobis distance have been used for anomaly detection in neural networks; for example, Lee et al. (2018) proposed a similar method for out-of-domain detection in neural classification models, and Cao et al. (2020) found the Mahalanobis distance method to be competitive with more sophisticated methods on medical out-of-domain detection. In Transformer models, Podolskiy et al. (2021) used Mahalanobis distance for out-of-domain detection, outperforming methods based on softmax probability and likelihood ratios. Gaussian assumptions. Our model assumes that the embeddings at every layer follow a multivariate Gaussian distribution. Since the Gaussian distribution is the maximum entropy distribution given a mean and covariance matrix, it makes the fewest assumptions and is therefore a reasonable default. Hennigen et al. (2020) found that embeddings sometimes do not follow a Gaussian distribution, but it is unclear what alternative distribution would be a better fit, so we will assume a Gaussian distribution in this work. 3.2 Training and evaluation For all of our experiments, we use the ‘base’ versions of pretrained language models BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and 4218 0.6 0.7 0.8 10 20 50 100 200 500 1000 2000 5000 10000 Training Sentences Accuracy BERT RoBERTa XLNet (a) 0.0 0.2 0.4 0.6 0.8 1.0 0 1 2 3 4 5 6 7 8 9 10 11 12 Layer Accuracy BERT RoBERTa XLNet (b) Figure 2: BLiMP accuracy different amounts of training data and across layers, for three LMs. About 1000 sentences are needed before a plateau is reached (mean tokens per sentence = 15.1). XLNet (Yang et al., 2019), provided by HuggingFace (Wolf et al., 2020). Each of these models have 12 contextual layers plus a 0th static layer, and each layer is 768-dimensional. We train the Gaussian model on randomly selected sentences from the British National Corpus (Leech, 1992), representative of acceptable English text from various genres. We evaluate on BLiMP (Warstadt et al., 2020), a dataset of 67k minimal sentence pairs that test acceptability judgements across a variety of syntactic and semantic phenomena. In our case, a sentence pair is considered correct if the sentence-level surprisal of the unacceptable sentence is higher than that of the acceptable sentence. How much training data is needed? We experiment with training data sizes ranging from 10 to 10,000 sentences (Figure 2a). Compared to the massive amount of data needed for pretraining the LMs, we find that a modest corpus suffices for training the Gaussian anomaly model, and a plateau is reached after 1000 sentences for all three models. Therefore, we use 1000 training sentences (unless otherwise noted) for all subsequent experiments in this paper. Which layers are sensitive to anomaly? We vary L from 0 to 12 in all three models (Figure 2b). The layer with the highest accuracy differs between models: layer 9 has the highest accuracy for BERT, 11 for RoBERTa, and 6 for XLNet. All models experience a sharp drop in the last layer, likely because the last layer is specialized for the MLM pretraining objective. Comparisons to other models. Our bestperforming model is RoBERTa, with an accuracy of 0.830. This is slightly higher the best model reported in BLiMP (GPT-2, with accuracy 0.801). We do not claim to beat the state-of-the-art on BLiMP: Salazar et al. (2020) obtains a higher accuracy of 0.865 using RoBERTa-large. Even though the main goal of this paper is not to maximize accuracy on BLiMP, our Gaussian anomaly model is competitive with other transformer-based models on this task. In Appendix A, we explore variations of the Gaussian anomaly model, such as varying the type of covariance matrix, Gaussian mixture models, and one-class SVMs (Sch¨olkopf et al., 2000). However, none of these variants offer a significant improvement over a single Gaussian model with full covariance matrix. 3.3 Lower layers are sensitive to frequency We notice that surprisal scores in the lower layers are sensitive to token frequency: higher frequency tokens produce embeddings close to the center of the Gaussian distribution, while lower frequency tokens are at the periphery. The effect gradually diminishes towards the upper layers. To quantify the sensitivity to frequency, we compute token-level surprisal scores for 5000 sentences from BNC that were not used in training. We then compute the Pearson correlation between the surprisal score and log frequency for each token (Figure 3). In all three models, there is a high correlation between the surprisal score and log frequency at the lower layers, which diminishes at the upper layers. A small positive correlation persists until the last layer, except for XLNet, in which the correlation eventually disappears. There does not appear to be any reports of this phenomenon in previous work. For static word vectors, Gong et al. (2018) found that embeddings for low-frequency words lie in a different region of 4219 0.0 0.2 0.4 0.6 0.8 1.0 0 1 2 3 4 5 6 7 8 9 10 11 12 Layer Pearson Correlation BERT RoBERTa XLNet Figure 3: Pearson correlation between token-level surprisal scores (Equation 4) and log frequency. The correlation is highest in the lower layers, and decreases in the upper layers. the embedding space than high-frequency words. We find evidence that the same phenomenon occurs in contextual embeddings (Appendix B). In this scenario, the Gaussian model fits the highfrequency region and assigns lower likelihoods to the low-frequency region, explaining the positive correlation at all layers; however, it is still unclear why the correlation diminishes at upper layers. 4 Levels of linguistic anomalies We turn to the question of whether LMs exhibit different behaviour when given inputs with different types of linguistic anomalies. The task of partitioning linguistic anomalies into several distinct classes can be challenging. Syntax and semantics have a high degree of overlap – there is no widely accepted criterion for distinguishing between ungrammaticality and semantic anomaly (e.g., Abrus´an (2019) gives a survey of current proposals), and Poulsen (2012) challenges this dichotomy entirely. Similarly, Warren et al. (2015) noted that semantic anomalies depend somewhat on world knowledge. Within a class, the anomalies are also heterogeneous (e.g., ungrammaticality may be due to violations of agreement, wh-movement, negative polarity item licensing, etc), which might each affect the LMs differently. Thus, we define three classes of anomalies that do not attempt to cover all possible linguistic phenomena, but captures different levels of language processing while retaining internal uniformity: 1. Morphosyntactic anomaly: an error in the inflected form of a word, for example, subject-verb agreement (*the boy eat the sandwich), or incorrect verb tense or aspect inflection (*the boy eaten the sandwich). In each case, the sentence can be corrected by changing the inflectional form of one word. 2. Semantic anomaly: a violation of a selectional restriction, such as animacy (#the house eats the sandwich). In these cases, the sentence can be corrected by replacing one of the verb’s arguments with another one in the same word class that satisfies the verb’s selectional restrictions. 3. Commonsense anomaly: sentence describes an situation that is atypical or implausible in the real world but is otherwise acceptable (#the customer served the waitress). 4.1 Summary of anomaly datasets We use two sources of data for experiments on linguistic anomalies: synthetic sentences generated from templates, and materials from psycholinguistic studies. Both have advantages and disadvantages – synthetic data can be easily generated in large quantities, but the resulting sentences may be odd in unintended ways. Psycholinguistic stimuli are designed to control for confounding factors (e.g., word frequency) and human-validated for acceptability, but are smaller (typically fewer than 100 sentence pairs). We curate a set of 12 tasks from BLiMP and 7 psycholinguistic studies1. Each sentence pair consists of a control and an anomalous sentence, so that all sentences within a task differ in a consistent manner. Table 1 shows an example sentence pair from each task. We summarize each dataset: 1. BLiMP (Warstadt et al., 2020): we use subject-verb and determiner-noun agreement tests as morphosyntactic anomaly tasks. For simplicity, we only use the basic regular sentences, and exclude sentences involving irregular words or distractor items. We also use the two argument structure tests involving animacy as a semantic anomaly task. All three BLiMP tasks therefore have 2000 sentence pairs. 1Several of these stimuli have been used in natural language processing research. Chersoni et al. (2018) used the data from Pylkk¨anen and McElree (2007) and Warren et al. (2015) to probe word vectors for knowledge of selectional restrictions. Ettinger (2020) used data from Federmeier and Kutas (1999) and Chow et al. (2016), which were referred to as CPRAG-102 and ROLE-88 respectively. 4220 Type Task Correct Example Incorrect Example Morphosyntax BLiMP (Subject-Verb) These casseroles disgust Kayla. These casseroles disgusts Kayla. BLiMP (Det-Noun) Craig explored that grocery store. Craig explored that grocery stores. Osterhout and Nicol (1999) The cats won’t eat the food that Mary gives them. The cats won’t eating the food that Mary gives them. Semantic BLiMP (Animacy) Amanda was respected by some waitresses. Amanda was respected by some picture. Pylkk¨anen and McElree (2007) The pilot flew the airplane after the intense class. The pilot amazed the airplane after the intense class. Warren et al. (2015) Corey’s hamster explored a nearby backpack and filled it with sawdust. Corey’s hamster entertained a nearby backpack and filled it with sawdust. Osterhout and Nicol (1999) The cats won’t eat the food that Mary gives them. The cats won’t bake the food that Mary gives them. Osterhout and Mobley (1995) The plane sailed through the air and landed on the runway. The plane sailed through the air and laughed on the runway. Commonsense Warren et al. (2015) Corey’s hamster explored a nearby backpack and filled it with sawdust. Corey’s hamster lifted a nearby backpack and filled it with sawdust. Federmeier and Kutas (1999) “Checkmate,” Rosalie announced with glee. She was getting to be really good at chess. “Checkmate,” Rosalie announced with glee. She was getting to be really good at monopoly. Chow et al. (2016) The restaurant owner forgot which customer the waitress had served. The restaurant owner forgot which waitress the customer had served. Urbach and Kutas (2010) Prosecutors accuse defendants of committing a crime. Prosecutors accuse sheriffs of committing a crime. Table 1: Example sentence pair for each of the 12 tasks. The 3 BLiMP tasks are generated from templates; the others are stimuli materials taken from psycholinguistic studies. 2. Osterhout and Nicol (1999): contains 90 sentence triplets containing a control, syntactic, and semantic anomaly. Syntactic anomalies involve a modal verb followed by a verb in -ing form; semantic anomalies have a selectional restriction violation between the subject and verb. There are also double anomalies (simultaneously syntactic and semantic) which we do not use. 3. Pylkk¨anen and McElree (2007): contains 70 sentence pairs where the verb is replaced in the anomalous sentence with one that requires an animate object, thus violating the selectional restriction. In half the sentences, the verb is contained in an embedded clause. 4. Warren et al. (2015): contains 30 sentence triplets with a possible condition, a selectional restriction violation between the subject and verb, and an impossible condition where the subject cannot carry out the action, i.e., a commonsense anomaly. 5. Osterhout and Mobley (1995): we use data from experiment 2, containing 90 sentence pairs where the verb in the anomalous sentence is semantically inappropriate. The experiment also tested gender agreement errors, but we do not include these stimuli. 6. Federmeier and Kutas (1999): contains 34 sentence pairs, where the final noun in each anomalous sentence is an inappropriate completion, but in the same semantic category as the expected completion. 7. Chow et al. (2016): contains 44 sentence pairs, where two of the nouns in the anomalous sentence are swapped to reverse their roles. This is the only task in which the sentence pair differs by more than one token. 8. Urbach and Kutas (2010): contains 120 sentence pairs, where the anomalous sentence replaces a patient of the verb with an atypical one. 4.2 Quantifying layerwise surprisal Let D = {(s1, s′ 1), · · · , (sn, s′ n)} be a dataset of sentence pairs, where si is a control sentence and s′ i is an anomalous sentence. For each layer L, we define the surprisal gap as the mean difference of surprisal scores between the control and anoma4221 lous sentences, scaled by the standard deviation: surprisal gapL(D) = E{surprisalL(s′ i) −surprisalL(si)}n i=1 σ{surprisalL(s′ i) −surprisalL(si)}n i=1 (9) The surprisal gap is a scale-invariant measure of sensitivity to anomaly, similar to a signal-tonoise ratio. While surprisal scores are unitless, the surprisal gap may be viewed as the number of standard deviations that anomalous sentences trigger surprisal above control sentences. This is advantageous over accuracy scores, which treats the sentence pair as correct when the anomalous sentence has higher surprisal by any margin; this hard cutoff masks differences in the magnitude of surprisal. The metric also allows for fair comparison of surprisal scores across datasets of vastly different sizes. Figure 4 shows the surprisal gap for all 12 tasks, using the RoBERTa model; the results for BERT and XLNet are in the Appendix C. Next, we compare the performance of the Gaussian model with the masked language model (MLM). We score each instance as correct if the masked probability of the correct word is higher than the anomalous word. One limitation of the MLM approach is that it requires the sentence pair to be identical in all places except for one token, since the LMs do not support modeling joint probabilities over multiple tokens. To ensure fair comparison between GM and MLM, we exclude instances where the differing token is outof-vocabulary in any of the LMs (this excludes approximately 30% of instances). For the Gaussian model, we compute accuracy using the bestperforming layer for each model (Section 3.2). The results are listed in Table 2. 5 Discussion 5.1 Anomaly type and surprisal Morphosyntactic anomalies generally appear earlier than semantic anomalies (Figure 4). The surprisal gap plot exhibits different patterns depending on the type of linguistic anomaly: morphosyntactic anomalies produce high surprisal relatively early (layers 3-4), while semantic anomalies produce low surprisals until later (layers 9 and above). Commonsense anomalies do not result in surprisals at any layer: the surprisal gap is near zero for all of the commonsense tasks. The observed difference between morphosyntactic and semantic Commonsense − Urbach and Kutas Commonsense − Chow et al. Commonsense − Federmeier and Kutas Commonsense − Warren et al. Semantic − Osterhout and Mobley Semantic − Osterhout and Nicol Semantic − Warren et al. Semantic − Pylkkänen and McElree Semantic − BLiMP (Animacy) Morphosyntax − Osterhout and Nicol Morphosyntax − BLiMP (Det−Noun) Morphosyntax − BLiMP (Subject−Verb) 0 1 2 3 4 5 6 7 8 9 10 11 12 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 0 1 2 Layer Surprisal Gap Figure 4: Layerwise surprisal gaps for all tasks using the RoBERTa model. Generally, a positive surprisal gap appears in earlier layers for morphosyntactic tasks than for semantic tasks; no surprisal gap appears at any layer for commonsense tasks. 4222 Type Task Size BERT RoBERTa XLNet GM MLM GM MLM GM MLM Morphosyntax BLiMP (Subject-Verb) 2000 0.953 0.955 0.971 0.957 0.827 0.584 BLiMP (Det-Noun) 2000 0.970 0.999 0.983 0.999 0.894 0.591 Osterhout and Nicol (1999) 90 1.000 1.000 1.000 1.000 0.901 0.718 Semantic BLiMP (Animacy) 2000 0.644 0.787 0.767 0.754 0.675 0.657 Pylkk¨anen and McElree (2007) 70 0.727 0.955 0.932 0.955 ∗0.636 0.727 Warren et al. (2015) 30 ∗0.556 1.000 0.944 1.000 ∗0.667 ∗0.556 Osterhout and Nicol (1999) 90 0.681 0.957 0.841 1.000 ∗0.507 0.783 Osterhout and Mobley (1995) 90 ∗0.528 1.000 0.906 0.981 ∗0.302 0.774 Commonsense Warren et al. (2015) 30 ∗0.600 ∗0.550 0.750 ∗0.450 ∗0.300 ∗0.600 Federmeier and Kutas (1999) 34 ∗0.458 ∗0.708 ∗0.583 0.875 ∗0.625 ∗0.667 Chow et al. (2016) 44 ∗0.591 n/a ∗0.432 n/a ∗0.568 n/a Urbach and Kutas (2010) 120 ∗0.470 0.924 ∗0.485 0.939 ∗0.500 0.712 Table 2: Comparing accuracy scores between Gaussian anomaly model (GM) and masked language model (MLM) for all models and tasks. Asterisks indicate that the accuracy is not better than random (0.5), using a binomial test with threshold of p < 0.05 for significance. The MLM results for Chow et al. (2016) are excluded because the control and anomalous sentences differ by more than one token. The best layers for each model (Section 3.2) are used for GM, and the last layer is used for MLM. Generally, MLM outperforms GM, and the difference is greater for semantic and commonsense tasks. anomalies is consistent with previous work (Tenney et al., 2019), which found that syntactic information appeared earlier in BERT than semantic information. One should be careful and avoid drawing conclusions from only a few experiments. A similar situation occurred in psycholinguistics research (Kutas et al., 2006): early results suggested that the N400 was triggered by semantic anomalies, while syntactic anomalies triggered the P600 – a different type of ERP. However, subsequent experiments found exceptions to this rule, and now it is believed that the N400 cannot be categorized by any standard dichotomy, like syntax versus semantics (Kutas and Federmeier, 2011). In our case, Pylkk¨anen and McElree (2007) is an exception: the task is a semantic anomaly, but produces surprisals in early layers, similar to the morphosyntactic tasks. Hence it is possible that the dichotomy is something other than syntax versus semantics; we leave to future work to determine more precisely what conditions trigger high surprisals in lower versus upper layers of LMs. 5.2 Comparing anomaly model with MLM The masked language model (MLM) usually outperforms the Gaussian anomaly model (GM), but the difference is uneven. MLM performs much better than GM on commonsense tasks, slightly better on semantic tasks, and about the same or slightly worse on morphosyntactic tasks. It is not obvious why MLM should perform better than GM, but we note two subtle differences between the MLM and GM setups that may be contributing factors. First, the GM method adds up the surprisal scores for the whole sequence, while MLM only considers the softmax distribution at one token. Second, the input sequence for MLM always contains a [MASK] token, whereas GM takes the original unmasked sequences as input, so the representations are never identical between the two setups. MLM generally outperforms GM, but it does not solve every task: all three LMs fail to perform above chance on the data from Warren et al. (2015). This set of stimuli was designed so that both the control and impossible completions are not very likely or expected, which may have caused the difficulty for the LMs. We excluded the task of Chow et al. (2016) for MLM because the control and anomalous sentences differed by more than one token2. 5.3 Differences between LMs RoBERTa is the best-performing of the three LMs in both the GM and MLM settings: this is expected since it is trained with the most data and performs well on many natural language benchmarks. Surprisingly, XLNet is ill-suited for this task and performs worse than BERT, despite having a similar model capacity and training data. The surprisal gap plots for BERT and XL2Sentence pairs with multiple differing tokens are inconvenient for MLM to handle, but this is not a fundamental limitation. For example, Salazar et al. (2020) proposed a modification to MLM to handle such cases: they compute a pseudolog-likelihood score for a sequence by replacing one token at a time with a [MASK] token, applying MLM to each masked sequence, and summing up the log likelihood scores. 4223 Net (Appendix C) show some differences from RoBERTa: only morphosyntactic tasks produce out-of-domain embeddings in these two models, and not semantic or commonsense tasks. Evidently, how LMs behave when presented with anomalous inputs is dependent on model architecture and training data size; we leave exploration of this phenomenon to future work. 6 Conclusion We use Gaussian models to characterize outof-domain embeddings at intermediate layers of Transformer language models. The model requires a relatively small amount of in-domain data. Our experiments reveal that out-of-domain points in lower layers correspond to low-frequency tokens, while grammatically anomalous inputs are out-of-domain in higher layers. Furthermore, morphosyntactic anomalies are recognized as out-ofdomain starting from lower layers compared to syntactic anomalies. Commonsense anomalies do not generate out-of-domain embeddings at any layer, even when the LM has a preference for the correct cloze completion. These results show that depending on the type of linguistic anomaly, LMs use different mechanisms to produce the output softmax distribution. Acknowledgements We thank Julian Salazar and our anonymous reviewers for their helpful suggestions. YX is funded through an NSERC Discovery Grant, a SSHRC Insight Grant, and an Ontario ERA award. FR is supported by a CIFAR Chair in Artificial Intelligence. References M´arta Abrus´an. 2019. Semantic anomaly, pragmatic infelicity, and ungrammaticality. Annual Review of Linguistics, 5:329–351. Tianshi Cao, Chinwei Huang, David Yu-Tung Hui, and Joseph Paul Cohen. 2020. A benchmark of medical out of distribution detection. arXiv preprint arXiv:2007.04250. Emmanuele Chersoni, Adri`a Torrens Urrutia, Philippe Blache, and Alessandro Lenci. 2018. Modeling violations of selectional restrictions with distributional semantics. In Proceedings of the Workshop on Linguistic Complexity and Natural Language Processing, pages 20–29. Noam Chomsky. 1957. Syntactic Structures. Mouton and Co. Wing-Yee Chow, Cybelle Smith, Ellen Lau, and Colin Phillips. 2016. A “bag-of-arguments” mechanism for initial verb predictions. Language, Cognition and Neuroscience, 31(5):577–596. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Allyson Ettinger. 2020. What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models. Transactions of the Association for Computational Linguistics, 8:34–48. Kara D Federmeier and Marta Kutas. 1999. A rose by any other name: Long-term memory structure and sentence processing. Journal of memory and Language, 41(4):469–495. Stefan L Frank, Leun J Otten, Giulia Galli, and Gabriella Vigliocco. 2015. The ERP response to the amount of information conveyed by words in sentences. Brain and language, 140:1–11. Chengyue Gong, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2018. FRAGE: Frequencyagnostic word representation. In Advances in neural information processing systems, pages 1334–1345. Kristina Gulordava, Piotr Bojanowski, ´Edouard Grave, Tal Linzen, and Marco Baroni. 2018. Colorless green recurrent networks dream hierarchically. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 1195–1205. Lucas Torroba Hennigen, Adina Williams, and Ryan Cotterell. 2020. Intrinsic probing through dimension selection. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 197–216. John Hewitt and Christopher D Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4129–4138. Jennifer Hu, Jon Gauthier, Peng Qian, Ethan Wilcox, and Roger Levy. 2020. A systematic assessment of syntactic generalization in neural language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1725–1744. Association for Computational Linguistics. 4224 MA Kelly, Yang Xu, Jes´us Calvillo, and David Reitter. 2020. Which sentence embeddings and which layers encode syntactic structure? In Cognitive Science, pages 2375–2381. Artur Kulmizev, Vinit Ravishankar, Mostafa Abdou, and Joakim Nivre. 2020. Do neural language models show preferences for syntactic formalisms? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4077– 4091. Marta Kutas and Kara D Federmeier. 2011. Thirty years and counting: finding meaning in the N400 component of the event-related brain potential (ERP). Annual review of psychology, 62:621–647. Marta Kutas, Cyma K Van Petten, and Robert Kluender. 2006. Psycholinguistics electrified II (1994– 2005). In Handbook of psycholinguistics, pages 659–724. Elsevier. Ilia Kuznetsov and Iryna Gurevych. 2020. A matter of framing: The impact of linguistic formalism on probing results. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 171–182. Association for Computational Linguistics. Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting outof-distribution samples and adversarial attacks. Advances in Neural Information Processing Systems, 31:7167–7177. Geoffrey Neil Leech. 1992. 100 million words of English: the British National Corpus (BNC). Language Research, 28:1–13. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692. Eleni Metheniti, Tim Van de Cruys, and Nabil Hathout. 2020. How relevant are selectional preferences for Transformer-based language models? In Proceedings of the 28th International Conference on Computational Linguistics, pages 1266–1278. Alessio Miaschi, Dominique Brunato, Felice Dell’Orletta, and Giulia Venturi. 2020. Linguistic profiling of a neural language model. The 28th International Conference on Computational Linguistics, pages 745–756. James Michaelov and Benjamin Bergen. 2020. How well does surprisal explain N400 amplitude under different experimental conditions? In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 652–663. Lee Osterhout and Linda A Mobley. 1995. Eventrelated brain potentials elicited by failure to agree. Journal of Memory and language, 34(6):739–773. Lee Osterhout and Janet Nicol. 1999. On the distinctiveness, independence, and time course of the brain responses to syntactic and semantic anomalies. Language and cognitive processes, 14(3):283–317. Alexander Podolskiy, Dmitry Lipin, Andrey Bout, Ekaterina Artemova, and Irina Piontkovskaya. 2021. Revisiting Mahalanobis distance for Transformerbased out-of-domain detection. In 35th AAAI Conference on Artificial Intelligence (AAAI 2021). Mads Poulsen. 2012. The usefulness of the grammaticality–acceptability distinction in functional approaches to language. Acta Linguistica Hafniensia, 44(1):4–21. Liina Pylkk¨anen and Brian McElree. 2007. An MEG study of silent meaning. Journal of cognitive neuroscience, 19(11):1905–1921. Ella Rabinovich, Julia Watson, Barend Beekhuizen, and Suzanne Stevenson. 2019. Say anything: Automatic semantic infelicity detection in L2 English indefinite pronouns. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 77–86. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2021. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8:842–866. Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2020. Masked language model scoring. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2699–2712. Association for Computational Linguistics. Ryohei Sasano and Anna Korhonen. 2020. Investigating word-class distributions in word vector spaces. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3657–3666. Bernhard Sch¨olkopf, Robert C Williamson, Alex J Smola, John Shawe-Taylor, and John C Platt. 2000. Support vector method for novelty detection. In Advances in neural information processing systems, pages 582–588. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. BERT rediscovers the classical NLP pipeline. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4593– 4601. Thomas P Urbach and Marta Kutas. 2010. Quantifiers more or less quantify on-line: ERP evidence for partial incremental interpretation. Journal of Memory and Language, 63(2):158–179. Tessa Warren, Evelyn Milburn, Nikole D Patson, and Michael Walsh Dickey. 2015. Comprehending the impossible: what role do selectional restriction violations play? Language, cognition and neuroscience, 30(8):932–939. 4225 Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R Bowman. 2020. BLiMP: The benchmark of linguistic minimal pairs for English. Transactions of the Association for Computational Linguistics, 8:377– 392. Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625–641. Thomas Wolf, Julien Chaumond, Lysandre Debut, Victor Sanh, Clement Delangue, Anthony Moi, Pierric Cistac, Morgan Funtowicz, Joe Davison, Sam Shleifer, et al. 2020. Transformers: State-of-theart natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5753–5763. 4226 A Ablation experiments on Gaussian model We compare some variations to our methodology of training the Gaussian model. All of these variations are evaluated on the full BLiMP dataset. In each experiment, (unless otherwise noted) the language model is RoBERTa-base, using the secondto-last layer, and the Gaussian model has a full covariance matrix trained with 1000 sentences from the BNC corpus. Covariance matrix. We vary the type of covariance matrix (Table 3). Diagonal and spherical covariance matrices perform worse than with the full covariance matrix; this may be expected, as the full matrix has the most trainable parameters. Covariance Accuracy Full 0.830 Diagonal 0.755 Spherical 0.752 Table 3: Varying the type of covariance matrix in the Gaussian model. Gaussian mixture models. We try GMMs with up to 16 mixture components (Table 4). We observe a small increase in accuracy compared to a single Gaussian, but the difference is too small to justify the increased training time. Components Accuracy 1 0.830 2 0.841 4 0.836 8 0.849 16 0.827 Table 4: Using Gaussian mixture models (GMMs) with multiple components. Genre of training text. We sample from genres of BNC (each time with 1000 sentences) to train the Gaussian model (Table 5). The model performed worse when trained with the academic and spoken genres, and about the same with the fiction and news genres, perhaps because their vocabularies and grammars are more similar to those in the BLiMP sentences. One-class SVM. We try replacing the Gaussian model with a one-class SVM (Sch¨olkopf et al., 2000), another popular model for anomaly detection. We use the default settings from scikit-learn Genre Accuracy Academic 0.797 Fiction 0.840 News 0.828 Spoken 0.795 All 0.830 Table 5: Effect of the genre of training data. with three kernels (Table 6), but it performs worse than the Gaussian model on all settings. Kernel Score RBF 0.738 Linear 0.726 Polynomial 0.725 Table 6: Using 1-SVM instead of GMM, with various kernels. Sentence aggregation. Instead of Equation 5, we try defining sentence-level surprisal as the maximum surprisal among all tokens (Table 7): surprisal(s1, · · · , sn) = maxn i=1Gi; (10) however, this performs worse than using the sum of token surprisals. Aggregation Accuracy Sum 0.830 Max 0.773 Table 7: Two sentence-level aggregation strategies 4227 B PCA plots of infrequent tokens We feed a random selection of BNC sentences into RoBERTa and use PCA to visualize the distribution of rare and frequent tokens at different layers (Figure 5). In all cases, we find that infrequent tokens occupy a different region of the embedding space from frequent tokens, similar to what Gong et al. (2018) observed for static word vectors. This is consistent with the correlation between tokenlevel surprisal and frequency (Figure 3), although the decrease in correlation towards upper layers is not apparent in the PCA plots. C Surprisal gap for BERT and XLNet Figures 6 and 7 plot the surprisal gaps using the BERT and XLNet models; data and algorithms are identical to the RoBERTa model (Figure 4). The Gaussian model is only sensitive to morphosyntactic anomalies, and not to semantic and commonsense ones. Layer: 1 Frequent Rare Layer: 4 Frequent Rare Layer: 7 Frequent Rare Layer: 10 Frequent Rare Figure 5: PCA plot of randomly sampled RoBERTa embeddings at layers 1, 4, 7, and 10. Points are colored by token frequency: “Rare” means the 20% least frequent tokens, and “Frequent” is the other 80%. 4228 Commonsense − Urbach and Kutas Commonsense − Chow et al. Commonsense − Federmeier and Kutas Commonsense − Warren et al. Semantic − Osterhout and Mobley Semantic − Osterhout and Nicol Semantic − Warren et al. Semantic − Pylkkänen and McElree Semantic − BLiMP (Animacy) Morphosyntax − Osterhout and Nicol Morphosyntax − BLiMP (Det−Noun) Morphosyntax − BLiMP (Subject−Verb) 0 1 2 3 4 5 6 7 8 9 10 11 12 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 Layer Surprisal Gap Figure 6: Surprisal gap plot using BERT. Commonsense − Urbach and Kutas Commonsense − Chow et al. Commonsense − Federmeier and Kutas Commonsense − Warren et al. Semantic − Osterhout and Mobley Semantic − Osterhout and Nicol Semantic − Warren et al. Semantic − Pylkkänen and McElree Semantic − BLiMP (Animacy) Morphosyntax − Osterhout and Nicol Morphosyntax − BLiMP (Det−Noun) Morphosyntax − BLiMP (Subject−Verb) 0 1 2 3 4 5 6 7 8 9 10 11 12 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 −1 0 1 Layer Surprisal Gap Figure 7: Surprisal gap plot using XLNet.
|
2021
|
325
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4229–4239 August 1–6, 2021. ©2021 Association for Computational Linguistics 4229 Psycholinguistic Tripartite Graph Network for Personality Detection Tao Yang, Feifan Yang, Haolan Ouyang, Xiaojun Quan∗ School of Computer Science and Engineering, Sun Yat-sen University, China {yangt225,yangff6,ouyhlan}@mail2.sysu.edu.cn [email protected] Abstract Most of the recent work on personality detection from online posts adopts multifarious deep neural networks to represent the posts and builds predictive models in a data-driven manner, without the exploitation of psycholinguistic knowledge that may unveil the connections between one’s language usage and his psychological traits. In this paper, we propose a psycholinguistic knowledge-based tripartite graph network, TrigNet, which consists of a tripartite graph network and a BERT-based graph initializer. The graph network injects structural psycholinguistic knowledge from LIWC, a computerized instrument for psycholinguistic analysis, by constructing a heterogeneous tripartite graph. The graph initializer is employed to provide initial embeddings for the graph nodes. To reduce the computational cost in graph learning, we further propose a novel flow graph attention network (GAT) that only transmits messages between neighboring parties in the tripartite graph. Benefiting from the tripartite graph, TrigNet can aggregate post information from a psychological perspective, which is a novel way of exploiting domain knowledge. Extensive experiments on two datasets show that TrigNet outperforms the existing state-of-art model by 3.47 and 2.10 points in average F1. Moreover, the flow GAT reduces the FLOPS and Memory measures by 38% and 32%, respectively, in comparison to the original GAT in our setting. 1 Introduction Personality detection from online posts aims to identify one’s personality traits from the online texts he creates. This emerging task has attracted great interest from researchers in computational psycholinguistics and natural language processing due to the extensive application scenarios such as ∗Corresponding author. Post-1 Post-2 Function Quant Affect Social Drives Post Node Word Node Category Node of me thanks love sharing it for a lot good advice Figure 1: An example of our tripartite graph. The content of Post-1 and Post-2 are “A lot of good advise for me.” and “Love it! Thanks for sharing!”, respectively. personalized recommendation systems (Yang and Huang, 2019; Jeong et al., 2020), job screening (Hiemstra et al., 2019) and psychological studies (Goreis and Voracek, 2019). Psychological research shows that the words people use in daily life reflect their cognition, emotion, and personality (Gottschalk, 1997; Golbeck, 2016). As a major psycholinguistic instrument, Linguistic Inquiry and Word Count (LIWC) (Tausczik and Pennebaker, 2010) divides words into psychologically relevant categories (e.g., Function, Affect, and Social as shown in Figure 1) and is commonly used to extract psycholinguistic features in conventional methods (Golbeck et al., 2011; Sumner et al., 2012). Nevertheless, most recent works (Hernandez and Knight, 2017; Jiang et al., 2020; Keh et al., 2019; Lynn et al., 2020; Gjurkovi´c et al., 2020) tend to adopt deep neural networks (DNNs) to represent the posts and build predictive models in a data-driven manner. They first encode each post separately and then aggregate the post representations into a user representation. Although numerous improvements have been made over the traditional methods, they are likely to suffer from limitations as follows. First, the input of this task 4230 is usually a set of topic-agnostic posts, some of which may contain few personality cues. Hence, directly aggregating these posts based on their contextual representations may inevitably introduce noise. Second, personality detection is a typical data-hungry task since it is non-trivial to obtain personality tags, while DNNs implicitly extract personality cues from the texts and call for tremendous training data. Naturally, it is desirable to explicitly introduce psycholinguistic knowledge into the models to capture critical personality cues. Motivated by the above discussions, we propose a psycholinguistic knowledge-based tripartite graph network, namely TrigNet, which consists of a tripartite graph network to model the psycholinguistic knowledge and a graph initializer using a pre-trained language model such as BERT (Devlin et al., 2019) to generate the initial representations for all the nodes. As illustrated in Figure 1, a specific tripartite graph is constructed for each user, where three heterogeneous types of nodes, namely post, word, and category, are used to represent the posts of a user, the words contained both in his posts and the LIWC dictionary, and the psychologically relevant categories of the words, respectively. The edges are determined by the subordination between word and post nodes as well as between word and category nodes. Besides, considering that there are no direct edges between homogeneous nodes (e.g., between post nodes) in the tripartite graph, a novel flow GAT is proposed to only transmit messages between neighboring parties to reduce the computational cost and to allow for more effective interaction between nodes. Finally, we regard the averaged post node representation as the final user representation for personality classification. Benefiting from the tripartite graph structure, the interaction between posts is based on psychologically relevant words and categories rather than topic-agnostic context. We conduct extensive experiments on the Kaggle and Pandora datasets to evaluate our TrigNet model. Experimental results show that it achieves consistent improvements over several strong baselines. Comparing to the state-of-the-art model, SN+Att (Lynn et al., 2020), TrigNet brings a remarkable boost of 3.47 in averaged Macro-F1 (%) on Kaggle and a boost of 2.10 on Pandora. Besides, thorough ablation studies and analyses are conducted and demonstrate that the tripartite graph and the flow GAT play an irreplaceable role in the boosts of performance and decreases of computational cost. Our contributions are summarized as follows: • This is the first effort to use a tripartite graph to explicitly introduce psycholinguistic knowledge for personality detection, providing a new perspective of using domain knowledge. • We propose a novel tripartite graph network, TrigNet, with a flow GAT to reduce the computational cost in graph learning. • We demonstrate the outperformance of our TrigNet over baselines as well as the effectiveness of the tripartite graph and the flow GAT by extensive studies and analyses. 2 Related Work 2.1 Personality Detection As an emerging research problem, text-based personality detection has attracted the attention of both NLP and psychological researchers (Cui and Qi, 2017; Xue et al., 2018; Keh et al., 2019; Jiang et al., 2020; Tadesse et al., 2018; Lynn et al., 2020). Traditional studies on this problem generally resort to feature-engineering methods, which first extracts various psychological categories via LIWC (Tausczik and Pennebaker, 2010) or statistical features by the bag-of-words model (Zhang et al., 2010). These features are then fed into a classifier such as SVM (Cui and Qi, 2017) and XGBoost (Tadesse et al., 2018) to predict the personality traits. Despite interpretable features that can be expected, feature engineering has such limitations as it relies heavily on manually designed features. With the advances of deep neural networks (DNNs), great success has been achieved in personality detection. Tandera et al. (2017) apply LSTM (Hochreiter and Schmidhuber, 1997) on each post to predict the personality traits. Xue et al. (2018) develop a hierarchical DNN, which depends on an AttRCNN and a variant of Inception (Szegedy et al., 2017) to learn deep semantic features from the posts. Lynn et al. (2020) first encode each post by a GRU (Cho et al., 2014) with attention and then pass the post representations to another GRU to produce the whole contextual representations. Recently, pre-trained language models have been applied to this task. Jiang et al. (2020) simply concatenate all the utterances from a single user into a document and encode it with BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019). Gjurkovi´c 4231 et al. (2020) first encode each post by BERT and then use CNN (LeCun et al., 1998) to aggregate the post representations. Most of them focus on how to obtain more effective contextual representations, with only several exceptions that try to introduce psycholinguistic features into DNNs, such as Majumder et al. (2017) and Xue et al. (2018). However, these approaches simply concatenate psycholinguistic features with contextual representations, ignoring the gap between the two spaces. 2.2 Graph Neural Networks Graph neural networks (GNNs) can effectively deal with tasks with rich relational structures and learn a feature representation for each node in the graph according to the structural information. Recently, GNNs have attracted wide attention in NLP (Cao et al., 2019; Yao et al., 2019; Wang et al., 2020b,a). Among these research, graph construction lies at the heart as it directly impacts the final performance. Cao et al. (2019) build a graph for question answering, where the nodes are entities, and the edges are determined by whether two nodes are in the same document. Yao et al. (2019) construct a heterogeneous graph for text classification, where the nodes are documents and words, and the edges depend on word co-occurrences and document-word relations. Wang et al. (2020b) define a dependency-based graph by utilizing dependency parsing, in which the nodes are words, and the edges rely on the relations in the dependency parsing tree. Wang et al. (2020a) present a heterogeneous graph for extractive document summarization, where the nodes are words and sentences, and the edges depend on sentence-word relations. Inspired by the above successes, we construct a tripartite graph, which exploits psycholinguistic knowledge instead of simple document-word or sentence-word relations and is expected to contribute towards psychologically relevant node representations. 3 Our Approach Personality detection can be formulated as a multidocument multi-label classification task (Lynn et al., 2020; Gjurkovi´c et al., 2020). Formally, each user has a set P= {p1, p2, . . . , pr} of posts. Let pi= [wi,1, wi,2, . . . , wi,s] be the i-th post with s words, where pi can be viewed as a document. The goal of this task is to predict T personality traits Y = yt T t=1 for this user based on P, where yt ∈ {0, 1} is a binary variable. Figure 2 presents the overall architecture of the proposed TrigNet, which consists of a tripartite graph network and a BERT-based graph initializer. The former module aims to explicitly infuse psycholinguistic knowledge to uncover personality cues contained in the posts and the latter to encode each post and provide initial embeddings for the tripartite graph nodes. In the following subsections, we detail how the two modules work in four steps: graph construction, graph initialization, graph learning, and merge & classification. 3.1 Graph Construction As a major psycholinguistic analysis instrument, LIWC (Tausczik and Pennebaker, 2010) divides words into psychologically relevant categories and is adopted in this paper to construct a heterogeneous tripartite graph for each user. As shown in the right part of Figure 2, the constructed tripartite graph G= (V, E) contains three heterogeneous types of nodes, namely post, word, and category, where V denotes the set of nodes and E represents the edges between nodes. Specifically, we define V=Vp ∪ Vw ∪Vc, where Vp=P= {p1, p2, · · · , pr} denotes r posts, Vw= {w1, w2, · · · , wm} denotes m unique psycholinguistic words that appear both in the posts P and the LIWC dictionary, and Vc= {c1, c2, · · · , cn} represents n psychologically relevant categories selected from LIWC. The undirected edge eij between nodes i and j indicates word i either belongs to a post j or a category j. The interaction between posts in the tripartite graph is implemented by two flows: (1) “p↔w↔p”, which means posts interact via their shared psycholinguistic words (e.g., “p1↔w1↔p2” as shown by the red lines in Figure 2); (2) “p↔w↔c↔w↔p”, which suggests that posts interact by words that share the same category (e.g., “p1↔w2↔c2↔w3↔p2” as shown by the green lines in Figure 2). Hence, the interaction between posts is based on psychologically relevant words or categories rather than topic-agnostic context. 3.2 Graph Initialization As shown in the left part of Figure 2, we employ BERT (Devlin et al., 2019) to obtain the initial embeddings of all the nodes. BERT is built upon the multi-layer Transformer encoder (Vaswani et al., 2017), which consists of a word embedding layer 4232 1p 2p rp … 1 w 2 w m w … 1c 2c nc … 3 w Tripartite Graph Network Graph Construction Graph Learning (Flow GAT) Merge & Classification … … m w x 2 w x 1 w x Post Node Embeddings Word Node Embeddings … 1cx Category Node Embeddings 2cx ncx Graph Initialization [CLS] 1,1 w 1,s w [SEP] Transformer Layer 1 Transformer Layer 11 … Transformer Layer 12 Post 1 Post r … BERT-based Graph Initializer Embeding Layer Transformer Layer 10 Layer Attention m px 2 px 1px … Y … Figure 2: Overall architecture of our TrigNet, which consists of two modules: (1) a tripartite graph network (right) to inject psycholinguistic knowledge and (2) a BERT-based graph initializer (left) to initialize node embeddings. and 12 Transformer layers.1 Post Node Embedding The representations at the 12-th layer of BERT are usually used to represent an input sequence. This may not be appropriate for our task as personality is only weakly related to the higher order semantic features of posts, making it risky to rely solely on the final layer representations. In our experiments (Section 5.4), we find that the representations at the 11-th and 10-th layers are also useful for this task. Therefore, we utilize the representations at the last three layers to initialize the post node embeddings. Formally, the representations xj pi of the i-th post at the j-th layer can be obtained by: xj pi=BERTj ([CLS, wi,1, · · · , wi,m, SEP]) (1) where “CLS” and “SEP” are special tokens to denote the start and end of an input sentence, respectively, and BERTj (·) denotes the representation of the special token “CLS” at the j-th layer. In this way, we obtain the representations x10 pi , x11 pi , x12 pi T ∈R3×d of the last three layers, where d is the dimension of each representation. We then apply layer attention (Peters et al., 2018) to collapse the three representations into a single vector xpi: xpi = 12 X j=10 αjxj pi (2) where αj are softmax-normalized layer-specific weights to be learned. Consequently, we can obtain 1“BERT-BASE-UNCASED” is used in this study. a set of post representations for the given r posts of a user Xp = [xp1, xp2, · · · , xpr]T ∈Rr×d Word Node Embedding BERT applies WordPiece (Wu et al., 2016) to split words, which also cuts out-of-vocabulary words into small pieces. Thus, we obtain the initial node embedding of each word in Vw by considering two cases: (1) If the word is not out of vocabulary, we directly look up the BERT embedding layer to obtain its embedding; (2) If the word is out of vocabulary, we use the averaged embedding of its pieces as its initial node embedding. The initial word node embeddings are represented as Xw=[xw1, xw2, · · · , xwm]T ∈Rm×d. Category Node Embedding The LIWC2 dictionary divides words into 9 main categories and 64 subcategories.3 Empirically, subcategories such as Pronouns, Articles, and Prepositions are not task-related. Besides, our initial experiments show that excessive introduction of subcategories in the tripartite graph makes the graph sparse and makes the learning difficult, resulting in performance deterioration. For these reasons, we select all 9 main categories and the 6 personalconcern subcategories for our study. Particularly, the 9 main categories Function, Affect, Social, Cognitive Processes, Perceptual Processes, Biological Processes, Drives, Relativity, and Informal Language, and 6 personal-concern subcategories Work, Leisure, Home, Money, Religion, and Death are used as our category nodes. Then, we replace the “UNUSED” tokens in BERT’s vocab2http://liwc.wpengine.com/ 3Details of the categories are listed in Appendix. 4233 ulary by the 15 category names and look up the BERT embedding layer to generate their embeddings Xc=[xc1, xc2, · · · , xcn]T ∈Rn×d. 3.3 Graph Learning Graph attention network (GAT) (Veliˇckovi´c et al., 2018) can be applied over a graph to calculate the attention weight of each edge and update the node representations. However, unlike the traditional graph in which any two nodes may have edges, the connections in our tripartite graph only occur between neighboring parties (i.e., Vw ↔Vp and Vw ↔Vc), as shown in Figure 3. Therefore, applying the original GAT over our tripartite graph will lead to unnecessary computational costs. Inspired by Wang et al. (2020a), we propose a flow GAT for the tripartite graph. Particularly, considering that the interaction between posts in our tripartite graph can be accounted for by two flows “p↔w↔p” and “p↔w↔c↔w↔p”, we design a message passing mechanism that only transmits message by the two flows in the tripartite graph. Formally, given a constructed tripartite graph G = (V, E), where V = Vp∪Vw∪Vc, and the initial node embeddings X=Xp∪Xw∪Xc, we compute H (l+1) p , H (l+1) w , and H (l+1) c as the hidden states of Vp, Vw and Vc at the (l+1)-th layer. The flow GAT layer is defined as follows: H (l+1) p ,H (l+1) w ,H (l+1) c = FGAT H (l) p ,H (l) w ,H (l) c (3) where H (1) p = Xp, H (1) w = Xw, and H (1) c = Xc. The function FGAT (·) is implemented by the two flows: ˆH (l) w←p=MP H (l) w , H (l) p H (l) p←w,p = MP H (l) p , ˆH (l) w←p (4) H (l) c←w,p = MP H (l) c , ˆH (l) w←p H (l) w←c,w,p = MP ˆH (l) w←p, H (l) c←w,p H (l) p←w,c,w,p = MP H (l) p , H (l) w←c,w,p (5) H (l+1) p = mean H (l) p←w,p, H (l) p←w,c,w,p H (l+1) w = mean ˆH (l) w←p, H (l) w←c,w,p H (l+1) c = H (l) c←w,p (6) where ←means the message is transmitted from the right nodes to the left nodes, mean (·) is the mean pooling function, and MP (·) represents the w c p w c p Traditional Graph Our Tripartite Graph Figure 3: Comparison of adjacent matrices between the traditional graph (left) and our tripartite graph (right). Edges in the traditional graph may occur in any two nodes, while it only occurs between neighboring parties in our tripartite graph. message passing function. Eq. (4) and Eq. (5) illustrate that message is transmitted by the flows “p↔w↔p” and p↔w↔c↔w↔p, respectively. We take MP H (l) w , H (l) p in Eq. (4) as an example to introduce the massage passing function, where H (l) w = h h(l) w1, h(l) w2, · · · , h(l) wm i are used as the attention query and H (l) p = h h(l) p1 , h(l) p2 , · · · , h(l) pr i as the key and value. MP H (l) w , H (l) p can be decomposed into three steps. First, it calculates the attention weight βk ij between node i in Vw and its neighbor node j in Vp at the k-th head: zk ij = σ Wk z h Wk wh(l) wi||Wk ph(l) pj i (7) βk ij = exp zk ij P q∈Ni exp zk iq (8) where σ is the LeakyReLU activation function, Wk z, Wk w and Wk p are learnable weights, Ni means that the neighbor nodes of node i in Vp, and || is the concatenation operation. Second, the updated hidden state ˜h(l) wi is obtained by a weighted combination of its neighbor nodes in Vp: ˜h(l) wi = K || k=1 tanh X j∈Ni βk ijWk vh(l) pj (9) where K is the number of heads and Wk v is a learnable weight matrix. Third, noting that the above steps do not take the information of node i itself into account and to avoid gradient vanishing, we introduce a residual connection to produce the final updated node representation: ˆh(l) wi = h(l) wi + ˜h(l) wi (10) 4234 3.4 Merge & Classification After L layers of iteration, we obtain the final node representations H(L)=H (L) p ∪H (L) w ∪H (L) c . Then, we merge all post node representations H (L) p via mean pooling to produce the user representation: u = mean h h(L) p1 , h(L) p2 , · · · , h(L) pr i (11) Finally, we employ T softmax-normalized linear transformations to predict T personality traits. For the t-th personality trait, we compute: p yt = softmax uWt u + bt u (12) where Wt u is a trainable weight matrix and bt u is a bias term. The objective function of our TrigNet model is defined as: J (θ) = 1 V V X v=1 T X t=1 −yt v log p yt v|θ (13) where V is the number of training samples, T is the number of personality traits, yt v is the true label for the t-th trait, and p(yt v|θ) is the predicted probability for this trait under parameters θ. 4 Experiments In this section, we introduce the datasets, baselines, and settings of our experiments. 4.1 Datasets We choose two public MBTI datasets for evaluations, which have been widely used in recent studies (Tadesse et al., 2018; Hernandez and Knight, 2017; Majumder et al., 2017; Jiang et al., 2020; Gjurkovi´c et al., 2020). The Kaggle dataset4 is collected from PersonalityCafe,5 where people share their personality types and discussions about health, behavior, care, etc. There are a total of 8675 users in this dataset and each user has 45-50 posts. Pandora6 is another dataset collected from Reddit,7 where personality labels are extracted from short descriptions of users with MBTI results to introduce themselves. There are dozens to hundreds of posts for each of the 9067 users in this dataset. The traits of MBTI include Introversion vs. Extroversion (I/E), Sensing vs. iNtuition (S/N), Think vs. Feeling (T/F), and Perception vs. Judging (P/J). 4kaggle.com/datasnaek/mbti-type 5http://personalitycafe.com/forum 6https://psy.takelab.fer.hr/datasets/ 7https://www.reddit.com/ Dataset Traits Train (60%) Valid (20%) Test (20%) Kaggle I/E 4011 / 1194 1326 / 409 1339 / 396 S/N 610 / 4478 222 / 1513 248 / 1487 T/F 2410 / 2795 791 / 944 780 / 955 P/J 3096 / 2109 1063 / 672 1082 / 653 Pandora I/E 4278 / 1162 1427 / 386 1437 / 377 S/N 727 / 4830 208 / 1605 210 / 1604 T/F 3549 / 1891 1120 / 693 1182 / 632 P/J 3211 / 2229 1043 / 770 1056 / 758 Table 1: Statistics of the Kaggle and Pandora datasets. Following previous works (Majumder et al., 2017; Jiang et al., 2020), we delete words that match any personality label to avoid information leaks. The Macro-F1 metric is adopted to evaluate the performance in each personality trait since both datasets are highly imbalanced, and average MacroF1 is used to measure the overall performance. We shuffle the datasets and split them in a 60-20-20 proportion for training, validation, and testing, respectively. According to our statistics, there are respectively 20.45 and 28.01 LIWC words on average in each post in the two datasets, and very few posts (0.021/0.002 posts per user) are presented as disconnected nodes in the graph. We show the statistics of the two datasets in Table 1. 4.2 Baselines The following mainstream models are adopted as baselines to evaluate our model: SVM (Cui and Qi, 2017) and XGBoost (Tadesse et al., 2018): Support vector machine (SVM) or XGBoost is utilized as the classifier with features extracted by TF-IDF and LIWC from all posts. BiLSTM (Tandera et al., 2017): Bi-directional LSTM (Hochreiter and Schmidhuber, 1997) is firstly employed to encode each post, and then the averaged post representation is used for user representation. Glove (Pennington et al., 2014) is employed for the word embeddings. BERT (Keh et al., 2019): The fine-tuned BERT is firstly used to encode each post, and then mean pooling is performed over the post representations to generate the user representation. AttRCNN: This model adopts a hierarchical structure, in which a variant of Inception (Szegedy et al., 2017) is utilized to encode each post and a CNNbased aggregator is employed to obtain the user representation. Besides, it considers psycholinguistic knowledge by concatenating the LIWC features with the user representation. 4235 Methods Kaggle Pandora I/E S/N T/F P/J Average I/E S/N T/F P/J Average SVM (Cui and Qi, 2017) 53.34 47.75 76.72 63.03 60.21 44.74 46.92 64.62 56.32 53.15 XGBoost (Tadesse et al., 2018) 56.67 52.85 75.42 65.94 62.72 45.99 48.93 63.51 55.55 53.50 BiLSTM (Tandera et al., 2017) 57.82 57.87 69.97 57.01 60.67 48.01 52.01 63.48 56.21 54.93 BERT (Keh et al., 2019) 64.65 57.12 77.95 65.25 66.24 56.60 48.71 64.70 56.07 56.52 AttRCNN (Xue et al., 2018) 59.74 64.08 78.77 66.44 67.25 48.55 56.19 64.39 57.26 56.60 SN+Attn (Lynn et al., 2020) 65.43 62.15 78.05 63.92 67.39 56.98 54.78 60.95 54.81 56.88 TrigNet(our) 69.54 67.17 79.06 67.69 70.86 56.69 55.57 66.38 57.27 58.98 Table 2: Overall results of TrigNet and baselines in Macro-F1(%) score, where the best results are shown in bold. SN+Attn (Lynn et al., 2020): As the latest model, SN+Attn employs a hierarchical attention network, in which a GRU (Cho et al., 2014) with word-level attention is used to encode each post and another GRU with post-level attention is used to generate the user representation. To make a fair comparison between the baselines and our model, we replace the post encoders in AttRCNN and SN+Attn with the pre-trained BERT. 4.3 Training Details We implement our TrigNet in Pytorch8 and train it on four NVIDIA RTX 2080Ti GPUs. Adam (Kingma and Ba, 2014) is utilized as the optimizer, with the learning rate of BERT set to 2e-5 and of other components set to 1e-3. We set the maximum number of posts, r, to 50 and the maximum length of each post, s, to 70, considering the limit of available computational resources. After tuning on the validation dataset, we set the dropout rate to 0.2 and the mini-batch size to 32. The maximum number of nodes, r + m + n, is set to 500 for Kaggle and 970 for Pandora, which cover 98.95% and 97.07% of the samples, respectively. Moreover, the two hyperparameters, the numbers of flow GAT layers L and heads K, are searched in {1, 2, 3} and {1, 2, 4, 6, 8, 12, 16, 24}, respectively, and the best choices are L = 1 and K = 12. The reasons for L = 1 are likely twofold. First, our flow GAT can already realize the interactions between nodes when L = 1, whereas the vanilla GAT needs to stack 4 layers. Second, after trying L = 2 and L = 3, we find that they lead to slight performance drops compared to that of L = 1. 5 Results and Analyses In this section, we report the overall results and provide thorough analyses and discussions. 8https://pytorch.org/ 5.1 Overall Results The overall results are presented in Table 2, from which our observations are described as follows. First, the proposed TrigNet consistently surpasses the other competitors in F1 scores, demonstrating the superiority of our model on text-based personality detection with state-of-the-art performance. Specifically, compared with the existing state of the art, SN+Attn, TrigNet achieves 3.47 and 2.10 boosts in average F1 on the Kaggle and Pandora datasets, respectively. Second, compared with BERT, a basic module utilized in TrigNet, TrigNet yields 4.62 and 2.46 improvements in average F1 on the two datasets, verifying that the tripartite graph network can effectively capture the psychological relations between posts. Third, compared with AttRCNN, another method of leveraging psycholinguistic knowledge, TrigNet outperforms it with 3.61 and 2.38 increments in average F1 on the two datasets, demonstrating that our solution that injects psycholinguistic knowledge via the tripartite graph is more effective. Besides, the shallow models SVM and XGBoost achieve comparable performance to the non-pre-trained model BiLSTM, further showing that the words people used are important for personality detection. 5.2 Ablation Study We conduct an ablation study of our TrigNet model on the Kaggle dataset by removing each component to investigate their contributions. Table 3 shows the results which are categorized into two groups. In the first group, we investigate the contributions of the network components. We can see that removing the flow “p↔w↔c↔w↔p” defined in Eq. (5) results in higher performance declines than removing the flow “p↔w↔p” defined in Eq. (4), implying that the category nodes are helpful to capture personality cues from the texts. Besides, removing the layer attention mechanism also leads 4236 Model Ave. F1(%) ∆(%) TrigNet 70.86 w/o “p↔w↔p” 70.13 0.73↓ w/o“p↔w↔c↔w↔p” 69.56 1.3↓ w/o Layer attention 69.88 0.98↓ w/o Function 70.44 0.42↓ w/o Perceptual processes 70.28 0.58↓ w/o Work 70.28 0.58↓ w/o Home 70.08 0.78↓ w/o Drives 70.03 0.83↓ w/o Relativity 69.91 0.95↓ w/o Cognitive processes 69.69 1.17↓ w/o Biological processes 69.68 1.18↓ w/o Leisure 69.67 1.19↓ w/o Religion 69.58 1.28↓ w/o Money 69.56 1.30↓ w/o Informal language 69.51 1.35↓ w/o Social 69.32 1.54↓ w/o Death 69.30 1.56↓ w/o Affect 68.60 2.26↓ Table 3: Results of ablation study in average Macro-F1 on the Kaggle dataset, where “w/o” means removal of a component from the original TrigNet, and “∆” indicates the corresponding performance change. to considerable performance degradation. In the second group, we investigate the contribution of each category node. The results, sorted by scores of decrease from small to large, demonstrate that the introduction of every category node is beneficial to TrigNet. Among these category nodes, the Affect is shown to be the most crucial one to our model, as the average Macro-F1 score drops most significantly after it is removed. This implies that the Affect category reflects one’s personality obviously. Similar conclusions are reported by Depue and Collins (1999) and Zhang et al. (2019). In addition, the Function node is the least impactful category node. The reason could be that functional words reflect pure linguistic knowledge and are weakly connected to personality. 5.3 Analysis of the Computational Cost In this work we propose a flow GAT to reduce the computational cost of vanilla GAT. To show its GAT Params FLOPS Memory Ave.F1 Original 1.8M 5.5G 7.8GB 69.69 Flow(our) 1.8M 3.4G 5.3GB 70.86 Table 4: Analysis of the computational cost for original GAT and flow GAT on the Kaggle dataset. The metrics include the number of parameters (Params) and floating-point operations per second (FLOPS) of GAT as well as memory size (Memory) and the average Macro-F1 (Ave.F1) of whole model on the Kaggle dataset. effect, we compare it with vanilla GAT (as illustrated in the left part of Figure 3). The results are reported in Table 4, from which we can observe that flow GAT successfully reduces the computational cost in FLOPS and Memory by 38% and 32%, respectively, without extra parameters introduced. Besides, flow GAT is superior to vanilla GAT when the number of layers is 1. The cause is that the former can already capture adequate interactions between nodes with one layer, while the latter has to stack four layers to achieve this. We also compare our TrigNet with the vanilla BERT in terms of the computational cost. The result show that the flow GAT takes about 1.14% more FLOPS than the vanilla BERT(297.3G). 5.4 Layer Attention Analysis This study adopts layer attention (Peters et al., 2018) as shown in Eq. (2) to produce initial embeddings for post nodes. To show which layers are more useful, we conduct a simple experiment on the two datasets by using all the 12 layer representations of BERT and visualize the attention weight of each layer. As plotted in Figure 4, we find that the attention weights from layers 10 to 12 are significantly greater than that of the rest layers on both datasets, which explains why the last three layers are chosen for layer attention in our model. 6 Conclusion In this work, we proposed a novel psycholinguistic knowledge-based tripartite graph network, TrigNet, for personality detection. TrigNet aims to introduce /D\HUV .DJJOH 3DQGRUD Figure 4: Visualization of layer attention weights. The last three layers supply with more information for this task. 4237 structural psycholinguistic knowledge from LIWC via constructing a tripartite graph, in which interactions between posts are captured through psychologically relevant words and categories rather than simple document-word or sentence-word relations. Besides, a novel flow GAT that only transmits messages between neighboring parties was developed to reduce the computational cost. Extensive experiments and analyses on two datasets demonstrate the effectiveness and efficiency of TrigNet. This work is the first effort to leverage a tripartite graph to explicitly incorporate psycholinguistic knowledge for personality detection, providing a new perspective for exploiting domain knowledge. Acknowledgments The paper was fully supported by the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (No.2017ZT07X355). Ethical Statement This study aims to develop a technical method to incorporate psycholinguistic knowledge into neural models, rather than creating a privacy-invading tool. We worked within the purview of acceptable privacy practices and strictly followed the data usage policy. The datasets used in this study are all from public sources with all user information anonymized. The assessment results of the proposed model are sensitive and should be shared selectively and subject to the approval of the institutional review board (IRB). Any research or application based on this study is only allowed for research purposes, and any attempt to use the proposed model to infer sensitive user characteristics from publicly accessible data is strictly prohibited. To get the code, researchers need to sign an ethical statement and explain the purpose clearly. References Yu Cao, Meng Fang, and Dacheng Tao. 2019. Bag: Bi-directional attention entity graph convolutional network for multi-hop reasoning question answering. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 357–362. Kyunghyun Cho, Bart van Merri¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder–decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724– 1734. Brandon Cui and Calvin Qi. 2017. Survey analysis of machine learning methods for natural language processing for mbti personality type prediction. Available online: http://cs229.stanford.edu/proj2017/finalreports/5242471.pdf (accessed on 26 May 2021). Richard A Depue and Paul F Collins. 1999. Neurobiology of the structure of personality: Dopamine, facilitation of incentive motivation, and extraversion. Behavioral and Brain Sciences, 22(3):491–517. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Matej Gjurkovi´c, Mladen Karan, Iva Vukojevi´c, Mihaela Boˇsnjak, and Jan ˇSnajder. 2020. Pandora talks: Personality and demographics on reddit. arXiv preprint arXiv:2004.04460. Jennifer Golbeck, Cristina Robles, and Karen Turner. 2011. Predicting personality with social media. In CHI’11 Extended Abstracts on Human Factors in Computing Systems, pages 253–262. Jennifer Ann Golbeck. 2016. Predicting personality from social media text. AIS Transactions on Replication Research, 2(1):2. Andreas Goreis and Martin Voracek. 2019. A systematic review and meta-analysis of psychological research on conspiracy beliefs: Field characteristics, measurement instruments, and associations with personality traits. Frontiers in Psychology, 10:205. Louis A Gottschalk. 1997. The unobtrusive measurement of psychological states and traits. Text Analysis for the Social Sciences: Methods for Drawing Statistical Inferences from Texts and Transcripts, pages 117–129. R Hernandez and IS Knight. 2017. Predicting myersbridge type indicator with text classification. In Proceedings of the 31st Conference on Neural Information Processing Systems, Long Beach, CA, USA, pages 4–9. Annemarie MF Hiemstra, Janneke K Oostrom, Eva Derous, Alec W Serlie, and Marise Ph Born. 2019. Applicant perceptions of initial job candidate screening with asynchronous job interviews: Does personality matter? Journal of Personnel Psychology, 18(3):138. 4238 Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural Computation, 9(8):1735–1780. Chi-Seo Jeong, Jong-Yong Lee, and Kye-Dong Jung. 2020. Adaptive recommendation system for tourism by personality type using deep learning. International Journal of Internet, Broadcasting and Communication, 12(1):55–60. Hang Jiang, Xianzhe Zhang, and Jinho D Choi. 2020. Automatic text-based personality recognition on monologues and multiparty dialogues using attentive networks and contextual embeddings (student abstract). In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13821– 13822. Sedrick Scott Keh, I Cheng, et al. 2019. Myersbriggs personality classification and personalityspecific language generation using pre-trained language models. arXiv preprint arXiv:1907.06333. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. Yann LeCun, L´eon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Veronica Lynn, Niranjan Balasubramanian, and H Andrew Schwartz. 2020. Hierarchical modeling for user personality prediction: The role of messagelevel attention. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5306–5316. Navonil Majumder, Soujanya Poria, Alexander Gelbukh, and Erik Cambria. 2017. Deep learning-based document modeling for personality detection from text. IEEE Intelligent Systems, 32(2):74–79. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532–1543. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of NAACL-HLT, pages 2227–2237. Chris Sumner, Alison Byers, Rachel Boochever, and Gregory J Park. 2012. Predicting dark triad personality traits from twitter usage and a linguistic analysis of tweets. In 2012 11th International Conference on Machine Learning and Applications, volume 2, pages 386–393. IEEE. Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. 2017. Inception-v4, inception-resnet and the impact of residual connections on learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31. Michael M Tadesse, Hongfei Lin, Bo Xu, and Liang Yang. 2018. Personality predictions based on user behavior on the facebook social media platform. IEEE Access, 6:61959–61969. Tommy Tandera, Derwin Suhartono, Rini Wongso, Yen Lina Prasetio, et al. 2017. Personality prediction system from facebook users. Procedia computer science, 116:604–611. Yla R Tausczik and James W Pennebaker. 2010. The psychological meaning of words: Liwc and computerized text analysis methods. Journal of Language and Social Psychology, 29(1):24–54. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008. Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li`o, and Yoshua Bengio. 2018. Graph attention networks. In International Conference on Learning Representations. Danqing Wang, Pengfei Liu, Yining Zheng, Xipeng Qiu, and Xuanjing Huang. 2020a. Heterogeneous graph neural networks for extractive document summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6209–6219. Kai Wang, Weizhou Shen, Yunyi Yang, Xiaojun Quan, and Rui Wang. 2020b. Relational graph attention network for aspect-based sentiment analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3229—3238. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144. Di Xue, Lifa Wu, Zheng Hong, Shize Guo, Liang Gao, Zhiyong Wu, Xiaofeng Zhong, and Jianshan Sun. 2018. Deep learning-based personality recognition from text posts of online social networks. Applied Intelligence, 48(11):4232–4246. Hsin-Chang Yang and Zi-Rui Huang. 2019. Mining personality traits from social messages for game recommender systems. Knowledge-Based Systems, 165:157–168. 4239 Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Graph convolutional networks for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7370–7377. Le Zhang, Songyou Peng, and Stefan Winkler. 2019. Persemon: A deep network for joint analysis of apparent personality, emotion and their relationship. IEEE Transactions on Affective Computing. Yin Zhang, Rong Jin, and Zhi-Hua Zhou. 2010. Understanding bag-of-words model: a statistical framework. International Journal of Machine Learning and Cybernetics, 1(1-4):43–52. A Categories of LIWC As shown in Figure 5, a total of 73 categories and subcategories are defined in the LIWC-2015 dictionary. There are 9 main categories: Function, Affect, Social, Cognitive Processes, Perceptual Processes, Biological Processes, Drives, Relativity, and Informal Language, in which 20 standard linguistic subcategories are included in the Function category and 44 psychological-relevant subcategories are defined in the rest 8 categories. ► Function Words □ Pronouns ● Personal Pronouns ◊ I ◊ We ◊ You ◊ She / He ◊ They ● Impersonal Pronouns □ Articles □ Prepositions □ Auxiliary Verbs □ Adverbs □ Conjunctions □ Negations □ Verbs □ Adjectives □ Comparisons □ Interrogatives □ Numbers □ Quantifiers ► Affect □ Positive Emotions □ Negative Emotions ● Anx ● Anger ● Sad ► Social □ Family □ Friends □ Female □ Male ► Cognitive Processes □ Insight □ Causal □ Discrepancies □ Tentative □ Certainty □ Differentiation ► Perceptual Processes □ See □ Hear □ Feel ► Biological Processes □ Body □ Health □ Sexual □ Ingest ► Drives □ Affiliation □ Achievement □ Power □ Reward □ Risk □ Past Focus □ Present Focus □ Future Focus ► Relativity □ Motion □ Space □ Time □ Work □ Leisure □ Home □ Money □ Religion □ Death ► Informal Language □ Swear □ Netspeak □ Assent □ Nonfluencies □ Filler Words ► : The 1st level □ : The 2nd level ● : The 3rd level ◊ : The 4th level Personalconcern Figure 5: Detailed division of categories in the LIWC-2015 dictionary.
|
2021
|
326
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4240–4251 August 1–6, 2021. ©2021 Association for Computational Linguistics 4240 Verb Metaphor Detection via Contextual Relation Learning Wei Song1∗, Shuhui Zhou1∗, Ruiji Fu2,3, Ting Liu4, Lizhen Liu1 1College of Information Engineering and Academy for Multidisciplinary Studies, Capital Normal University, Beijing, China 2State Key Laboratory of Cognitive Intelligence, iFLYTEK Research, China 3iFLYTEK AI Research (Hebei), Langfang, China 4Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, Harbin, China {wsong, shzhou, liz_liu7480}@cnu.edu.cn, [email protected], [email protected] Abstract Correct natural language understanding requires computers to distinguish the literal and metaphorical senses of a word. Recent neural models achieve progress on verb metaphor detection by viewing it as sequence labeling. In this paper, we argue that it is appropriate to view this task as relation classification between a verb and its various contexts. We propose the Metaphor-relation BERT (MrBERT) model, which explicitly models the relation between a verb and its grammatical, sentential and semantic contexts. We evaluate our method on the VUA, MOH-X and TroFi datasets. Our method gets competitive results compared with state-of-the-art approaches. 1 Introduction Metaphor is ubiquitous in our daily life for effective communication (Lakoff and Johnson, 1980). Metaphor processing has become an active research topic in natural language processing due to its importance in understanding implied meanings. This task is challenging, requiring contextual semantic representation and reasoning. Various contexts and linguistic representation techniques have been explored in previous work. Early methods focused on analyzing restricted forms of linguistic context, such as subjectverb-object type grammatical relations, based on hand-crafted features (Shutova and Teufel, 2010b; Tsvetkov et al., 2013; Gutiérrez et al., 2016). Later, word embeddings and neural networks were introduced to alleviate the burden of feature engineering for relation-level metaphor detections (Rei et al., 2017; Mao et al., 2018). However, although grammatical relations provide the most direct clues, other contexts in running text are mostly ignored. Recently, token-level neural metaphor detection draws more attention. Several approaches discov∗These authors contributed equally to this work. ered that wider context can lead to better performance. Do Dinh and Gurevych (2016) considered a fixed window surrounding each target token as context. Gao et al. (2018) and Mao et al. (2018) argued that the full sentential context can provide strong clues for more accurate prediction. Some recent work also attempted to design models motivated by metaphor theories (Mao et al., 2019; Choi et al., 2021). Despite the progress of exploiting sentential context, there are still issues to be addressed. First of all, a word’s local context, its sentential context and other contexts should be all important for detecting metaphors; however, they are not well combined in previous work. More importantly, as shown in Figure 1, most token-level metaphor detection methods formulate metaphor detection as either a single-word classification or a sequence labeling problem (Gao et al., 2018). The context information is mainly used for learning contextual representations of tokens, rather than modeling the interactions between the target word and its contexts (Zayed et al., 2020). In this paper, we focus on token-level verb metaphor detection, since verb metaphors are of the most frequent type of metaphoric expressions (Shutova and Teufel, 2010a). As shown in Figure 1, we propose to formulate verb metaphor detection as a relation extraction problem, instead of token classification or sequence labeling formulations. In analogy to identify the relations between entities, our method models the relations between a target verb and its various contexts, and determines the verb’s metaphoricity based on the relation representation rather than only the verb’s (contextual) representation. We present a simple yet effective model — Metaphor-relation BERT (MrBERT), which is adapted from a BERT (Devlin et al., 2019) based state-of-the-art relation learning model (Bal4241 Sentence Encoder 𝑠= 𝑥$, … , 𝑥' = 𝑣, … , 𝑥) M 𝑜𝑟 L Sentence Encoder 𝑠= 𝑥$, … , 𝑥' = 𝑣, … , 𝑥) M L L … … ℎ0 𝑜𝑟 ℎ' ℎ$ ℎ' ℎ) Sentence Encoder 𝑠= 𝑥$, … , 𝑥' = 𝑣, … , 𝑥) … … ℎ$ ℎ' ℎ) Relation Encoder M 𝑜𝑟 L 𝑟(ℎ', ℎ2) (a) classification (b) sequence labeling (c) relation extraction Figure 1: Formulations of verb metaphor detection: (a) a single word classification model; (b) a sequence labeling model; (c) the proposed relation extraction model, where hs, hi, hc and r(hi, hc) represent the representations of a sentence, a token, the context and the relation between the target verb v and its context components. dini Soares et al., 2019). Our model has three highlights, as illustrated in Figure 2. First, we explicitly extract and represent context components, such as a verb’s arguments as the local context, the whole sentence as the global context, and its basic meaning as a distant context. So multiple contexts can be modeled interactively and integrated together. Second, MrBERT enables modeling the metaphorical relation between a verb and its context components, and uses the relation representation for determining the metaphoricity of the verb. Third, the model is flexible to incorporate sophisticated relation modeling methods and new types of contexts. We conduct experiments on the largest metaphor detection corpus VU Amsterdam Metaphor Corpus (VUA) (Steen, 2010). Our method obtains competitive results on the large VUA dataset. Detail analysis demonstrates the benefits of integrating various types of contexts for relation classification. The results on relatively small datasets, such as MOH-X and TroFi, also show good performance and model transferability. 2 Formulating Verb Metaphor Detection This section briefly summarizes the common formulations of token-level verb metaphor detection as a background, and discusses the relation between this paper and previous work. The task A given sentence contains a sequence of n tokens x = x1, ..., xn, and a target verb in this sentence is xi. Verb metaphor detection is to judge whether xi has a literal or a metaphorical sense. Basic formulations Most neural networks based approaches cast the task as a classification or sequence labeling problem (Do Dinh and Gurevych, 2016; Gao et al., 2018). As shown in Figure 1, the classification paradigm predicts a single binary label to indicate the metaphoricity of the target verb, while the sequence labeling paradigm predicts a sequence of binary labels to all tokens in a sentence. Based on the basic formulations, various approaches have tried to enhance feature representations by using globally trained contextual word embeddings (Gao et al., 2018) or incorporating wider context with powerful encoders such as BiLSTM (Gao et al., 2018; Mao et al., 2019) and Transformers (Dankers et al., 2019; Su et al., 2020). Limitations and recent trends However, the above two paradigms have some limitations. First, contextual information is mostly used to enhance the representation of the target word, but the interactions between the target word and its contexts are not explicitly modeled (Zayed et al., 2020; Su et al., 2020). To alleviate this, Su et al. (2020) proposed a new paradigm by viewing metaphor detection as a reading comprehension problem, which uses the target word as a query and captures its interactions with the sentence and clause. A concurrent work to this work (Choi et al., 2021) adopted a pre-trained contextualized model based late interaction mechanism to compare the basic meaning and the contextual meaning of a word. Second, exploiting wider context will bring in more noise and may lose the focus. Fully depending on data-driven models to discover useful contexts is difficult, given the scale of available datasets for metaphor detection is still limited. The grammar structures, such as verb arguments, are important for metaphor processing (Wilks, 1978), but is not well incorporated into neural models. Stowe et al. (2019) showed that data augmentation based on syntactic patterns can enhance a standard model. Le et al. (2020) adopted graph convolutional networks to incorporate dependency graphs, but did 4242 [CLS] [subj] He [/subj] [verb] absorbed [/verb] the [obj] costs [/obj] for the accident [SEP] Deep Transformer (BERT) Relation Representation and Prediction 𝑀𝑒𝑡𝑎𝑝ℎ𝑜𝑟𝑖𝑐𝑎𝑙 𝑀 𝑜𝑟 𝑙𝑖𝑡𝑒𝑟𝑎𝑙(𝐿)? maxout a context concatenation c context maxout ⊕ ⊕ 𝑟( ) ， + + 𝑟( ) ， 𝑟( ，) 𝑟( ，) 𝑟( ，) 𝑝(𝑟= 𝑀) 𝑝(𝑟= 𝑀) b context average 𝑝(𝑟= 𝑀) ⊕ + 𝑟( ，) Sequence Prediction (1) (2) (3) (2) Figure 2: An example shows MrBERT’s main architecture. MrBERT considers the representations of (1) the sentential global context, (2) the grammatical local context, and (3) the basic meaning of the verb as a distant context. Three context integration strategies for modeling contextual relations are adopted: (a) context concatenation, (b) context average, and (c) context maxout. Contextual relation r is modeled to indicate the probability of being metaphorical, where linear, bilinear and neural tensor models can be applied to capture interactions between the verb and its contexts. The relation-level and sequence-level predictions are jointly optimized. not consider specific grammatical relations. It is interesting to further explore how to integrate explicit linguistic structures for contextual modeling. This paper presents a new paradigm for verb metaphor detection to overcome these limitations, by viewing the task as a relation extraction task. We assume a target verb and its multiple contexts are entities, and metaphor detection is to determine whether a metaphorical relation holds between the verb and its contexts. We will introduce the proposed model in Section 3. Before diving into details, we argue that viewing metaphor as a relation is reasonable and consistent with existing metaphor theories. According to Wilks (1978), metaphors show a violation of selectional preferences in a given context. The conceptual metaphor theory views metaphors as transferring knowledge from a familiar, or concrete domain to an unfamiliar, or more abstract domain (Lakoff and Johnson, 1980; Turney et al., 2011). The metaphor identification procedure (MIP) theory (Group, 2007) aims to identify metaphorically used words in discourse based on comparing their use in particular context and their basic meanings. All the theories care about a kind of relations between a target word and its contexts, which may help identify metaphors. 3 Metaphor-Relation BERT (MrBERT) We propose the Metaphor-relation BERT (MrBERT) model to realize verb metaphor detection as a relation classification task. Figure 2 shows the architecture of MrBERT. We use the pre-trained language model BERT as the backbone model. There are three main procedures: (1) extract and represent contexts; (2) model the contextual relations between the target verb and its contexts; (3) manipulate the contextual relations for predicting the verb’s metaphoricity. 3.1 Contexts and their Representations 3.1.1 Types of Contexts A metaphor can result when a target word interacts with a certain part in a sentence. Previous work often explored individual context types, such as verb arguments through grammatical relations or the whole sentence/clause. Little work has attempted to summarize and combine different contexts. We summarize the following contexts, which would help determine verbs’ metaphoricity: • Global context: We view the whole sentence as the global context. A metaphorically used word may seem divergent to the meaning or topic of the sentence. 4243 • Local context: We view the words that have a close grammatical relation to the target words as the local context, which is widely studied to capture selectional preference violations. • Distant context: Motivated by the MIP theory, the difference between the contextual usage of a word and its basic meaning may indicate a metaphor so that we view the basic meaning of the target verb as a distant context. Then, we have to extract and represent these contexts. 3.1.2 Context Extraction and Representation We call the target verb’s contexts as context components. To get the contextual or basic meanings of these components. we use the deep transformer models, such as BERT. We first use Stanford dependency parser (Chen and Manning, 2014) to parse each sentence and extract verb-subject and verb-direct object relations with VB head and NN dependent. The nominal subjects and objects are used as the local context components. Motivated by (Baldini Soares et al., 2019), we introduce 6 component marker tokens, [subj], [/subj], [verb], [/verb], [obj] and [/obj], to explicitly label the boundaries of the target verb, its subject and object in each sentence. We also use [CLS] and [SEP] to mark the whole sentence. For example, the marker inserted token sequence for the sentence He absorbed the costs for the accident is shown in Figure 2. The whole token sequence is fed into BERT’s tokenizer, and then the transformer layers. To get the contextual representations, we use the hidden states of the final transformer layer. For each marked component, we use the start marker (e.g., [subj]) or the averaged embedding between the start and the end markers (e.g., [subj] and [/subj]) as the component representation. The contextual representation of the whole sentence is read from the final hidden state of [CLS]. To represent the basic meaning of the verb, we use the output from the BERT tokenizer to get the context independent verb representation. If word pieces exist, their averaged embedding is used. 3.2 Modeling the Contextual Relation The relation between the target verb and one of its contexts is called a contextual relation. Our purpose is to utilize the contextual relation(s) to determine the metaphoricity of the verb. The representations of the verb and a context component are denoted as v ∈Rd and c ∈Rk, respectively. We adopt three ways to explicitly define the form of the relation r for capturing the interactions between v and c. • Linear model We use a parameter vector Vr ∈Rd+k and a bias br to represent the relation r, and the probability of the relation being metaphorical is computed according to p(r|v, c) = σ(V ⊤ r v c + br), (1) where σ is the sigmoid function. • Bilinear model We use a parameter matrix Ar ∈Rd×k and a bias br to represent the relation r: p(r|v, c) = σ(v⊤Arc + br). (2) The components and the relation can interact more sufficiently with each other in this way. • Neural tensor model We also exploit a simplified neural tensor model for relation representation: p(r|v, c) = σ(v⊤Arc + V ⊤ r v c + br). (3) 3.3 Integrating Contextual Relations for Prediction We focus on 3 types of contextual relations: • Verb-global relation The relation between the contextual representations of the verb v and the whole sentence cCLS. • Verb-local relation The relation between the contextual representations of the verb v and its subject csubj or object cobj. • Verb-distant relation The relation between the verb v and its basic meaning vbsc. The representations of csubj, cobj, cCLS and vbsc can be obtained as described in Section 3.1.2. We try three ways to integrate the contextual relations. The first two ways build a combined context c first: • Context concatenation We can concatenate the representations of context components together as the combined context, i.e., c = csubj ⊕cobj ⊕cCLS ⊕vbsc. 4244 • Context average Similarly, we can use the averaged representation of all context components as the combined context, i.e., c = average(csubj, cobj, cCLS, vbsc). Then we compute the probability that the relation is metaphorical, i.e., p(r|v, c), where either linear, bilinear or neutral tensor model can be applied. The other way is to choose the most confident single prediction, i.e., • Context maxout The prediction is based on max{p(r|v, c)}, where c belongs to {cCLS, csubj, cobj, vbsc}. To train the relation-level prediction model, we use binary cross-entropy as the loss function, L0 = −1 N N X i=1 (ˆyiyi + (1 −ˆyi)(1 −yi)), (4) where N is the number of training samples; ˆyi is the golden label of a verb with ˆyi = 1 indicating a metaphorical usage and ˆyi = 0 indicating a literal usage; yi is the probability of being metaphorical predicted by our model. We further combine relation-level and sequencelevel metaphor detection via multi-task learning. The sequence metaphor detection uses the hidden states of the final layer and a softmax layer for predicting the metaphoricity of each token. We use cross-entropy as the loss function and denote the average loss over tokens in training samples as L1. The final loss of MrBERT is L = L0 + L1. 4 Evaluation 4.1 Experimental Settings 4.1.1 Datasets and Evaluation Metrics VUA dataset We mainly conduct experiments on the VUA (Steen, 2010) dataset. It is the largest publicly available metaphor detection dataset and has been used in metaphor detection shared tasks (Leong et al., 2018, 2020). This dataset has a training set and a test set. Previous work utilized the training set in different ways (Neidlein et al., 2020). We use the preprocessed version of the VUA dataset provided by Gao et al. (2018). The first reason is that this dataset has a fixed development set so that different methods can adopt the same model selection strategy. The second reason is that several recent important methods used the same dataset (Mao et al., 2018; Dankers et al., Train Dev Test # tokens 116,622 38,628 50,175 (5,873) # unique sent. 6,323 1,550 2,694 % metaphor 11.2 11.6 12.4 Table 1: Basic statistics of the preprocessed VUA dataset provided by (Gao et al., 2018). 50,175 and 5,873 tokens are used for evaluating All-POS and Verb tracks, respectively. 2019; Stowe et al., 2019; Le et al., 2020). Therefore it is convenient for us to compare the proposed method with previous work. There are two tracks: Verb and All-POS metaphor detection. Some basic statistics of the dataset are shown in Table 1. We focus on the Verb track since we mainly model metaphorical relations for verbs. We use MrBERT’s relation-level predictions for the verb track and use its sequence labeling module to deal with the All-POS track. MOH-X and TroFi datasets MOH-X (Mohammad et al., 2016) and TroFi (Birke and Sarkar, 2006) are two relatively smaller datasets compared with VUA. Only a single target verb is annotated in each sentence. We will report the results on MOHX and TroFi in three settings: zero-shot transfer, re-training and fine-tuning. Metrics The evaluation metrics are accuracy (Acc), precision (P), recall (R) and F1-score (F1), which are most commonly used in previous work. 4.1.2 Baselines We compare with the following approaches. • Gao et al. (2018) use contextual embeddings ELMo to enhance word representations and use BiLSTM as the encoder. It has two settings: classification (CLS) and sequence labeling (SEQ). • Mao et al. (2019) exploit two linguistic theory motivated intuitions based on the basis of (Gao et al., 2018). This work motivates us to further explore contextual relation modeling with pre-trained language models. • Stowe et al. (2019) exploit grammatical relations for data augmentation to enhance (Gao et al., 2018). • Le et al. (2020) propose a multi-task learning approach with graph convolutional neural networks and use word sense disambiguation as an auxiliary task. 4245 Parameter Value Learning Rate 5e-5 Optimizer Adam Batch-size 16 Dropout 0.1 Weight decay 0.01 Linear warmup used Table 2: Hyper-parameters for BERT based systems. • Neidlein et al. (2020) (BERT-SEQ) provide a detail setting for a BERT based sequence labeling model. This method is used as a main pre-trained language model based baseline. The above methods all used Gao et al. (2018)’s dataset for evaluation so that their results can be directly read from their papers for comparison. • Su et al. (2020) (DeepMet) view metaphor detection as a reading comprehension problem with RoBERTa as the backbone model. It obtained the best performance on 2020 metaphor detection shared task. • Choi et al. (2021) (MelBERT) present a concurrent work to ours. The method shares similar ideas and architecture with us, but it does not consider the grammatical relations. Notice that the systems participating in the VUA metaphor detection shared tasks (Leong et al., 2018, 2020) can use any way to manipulate the training set for model selection and ensemble learning so that the reported results in the task report are not directly comparable to us. The results of DeepMet and MelBERT are based on the single model evaluation in (Choi et al., 2021). The first four baselines do not utilize pre-trained language models, while the last three baselines use BERT or RoBERTa. These baselines support comprehensive comparisons from multiple aspects. 4.1.3 Parameter Configuration During context component extraction, if the target verb does not have a subject or an object, we use a fixed zero vector instead. We use the bert-baseuncased model and the standard tokenizer. The values of hyper-parameters are shown in Table 2. For MrBERT, we view the ways of component representation (start marker or averaged embedding, see Section 3.1.2), relation modeling (linear, bilinear, and neural tensor (NT)) models, see Section 3.2) and context integration (context concatenation, average and maxout, see Section 3.3) strategies as hyper-parameters as well. We run each model for 10 epoches, and choose the best combination according to the performance on the development set. The best combination uses the averaged embeddings, the bilinear model and the context average strategy, and it will represent MrBERT for performance report in Section 4.2. 4.2 Main Results on VUA Dataset Table 3 shows the results of the baselines and MrBERT. Except for (Gao et al., 2018)-CLS, all methods use the annotation information of all tokens. For the All-POS track, we report the performance on either all POS tags or 4 main POS tags for comparison with previous work. We can see that MrBERT achieves superior or competitive performance compared with previous work on verb metaphor detection. The use of pretrained language models improves the performance in general, compared with several LSTM based methods. Recent proposed models, such as DeepMet, MelBERT and MrBERT, gain further improvements compared with BERT-SEQ. MrBERT outperforms (Stowe et al., 2019) and (Le et al., 2020) largely. The two baselines attempt to make use of grammar information, through data augmentation or graph neural networks. In contrast, MrBERT provides a simple yet effective way to incorporate verb arguments and new contexts into a pre-trained language model. MrBERT also has competitive performance compared with DeepMet and MelBERT. We share the similar idea to enhance interactions between the target verb and its contexts, but implement in different ways. DeepMet and MelBERT base on the pretrained model RoBERTa and use additional POS or FGPOS information. Moreover, these two models are trained for every token so that the training might be more sufficient. In contrast, we mainly model metaphorical relation for verbs. This is perhaps also the reason that on the All-POS metaphor detection track, MrBERT has slightly worse results compared with MelBERT. However, our model is flexible and can be applied to tokens with other POS tags as well. We leave this as future work. 4.3 Analysis We further analyze the effects of modeling contextual relations from several aspects. Relation modeling and context integration strategies Table 4 shows the results of different 4246 VUA Verb VUA All-POS VUA All-POS (4 POS) Model Acc P R F1 Acc P R F1 Acc P R F1 Gao et al. (2018)-CLS 69.1 53.4 65.6 58.9 – – – – – – – – Gao et al. (2018)-SEQ 81.4 68.2 71.3 69.7 93.1 71.6 73.6 72.6 – – – – Mao et al. (2019) 81.8 66.3 75.2 70.5 93.8 73.0 75.7 74.3 – – – – Stowe et al. (2019) – – – 69.5 – – – 73.5 – – – – Le et al. (2020) 83.2 72.5 70.9 71.7 93.8 74.8 75.5 75.1 – – – – Neidlein et al. (2020) 84.9 78.0 69.0 73.2 94.5 83.0 71.9 77.0 91.8 77.9 64.6 70.7 DeepMet (Su et al., 2020) – 79.5 70.9 74.9 – 82.0 71.3 76.3 – – – – MelBERT (Choi et al., 2021) – 78.7 72.9 75.7 – 80.1 76.9 78.5 – – – MrBERT 86.4 80.8 71.5 75.9 94.7 82.7 72.5 77.2 91.8 78.4 64.6 70.9 Table 3: Results on the VUA dataset. MrBERT uses the bilinear model for relation modeling and the contextaverage integration strategy. VUA All-POS (4 POS) indicates the performance on 4 main POS tags. VUA-verb Model Acc P R F1 BERT-SEQ 85.1 77.5 70.8 74.0 Average-Linear 85.7 79.8 70.2 74.7 Average-Bilinear 86.4 80.8 71.5 75.9 Average-NT 85.7 77.4 73.8 75.6 Maxout-Linear 85.2 78.1 70.2 73.9 Maxout-Bilinear 85.3 75.7 74.8 75.3 Maxout-NT 85.6 78.8 70.9 74.7 Concat-Linear 85.5 80.3 68.6 74.0 Concat-Bilinear 85.2 77.6 71.2 74.3 Concat-NT 85.0 76.4 72.3 74.3 Table 4: The effects of the ways for modeling contextual relations and integrating multiple contexts. combinations of relation modeling and context integration strategies. BERT-SEQ here refers to the re-trained baseline with model selection based on the performance on the development set, and surpasses the reported results in (Neidlein et al., 2020). We can see that most combinations outperform BERT-SEQ, and have consistent performance. The bilinear and neural tensor models perform better than the linear model. This means that sophisticated relation modelling techniques can benefit the performance. Context average and context maxout strategies perform better than context concatenation. The reason may be that context concatenation is more difficult to be trained due to more parameters. Effects of different contexts Table 5 shows the performance of MrBERT when it considers the global context (MrBERT-G), the global and the local contexts (MrBERT-GL), and the full model with the distant context (MrBERT-GLD). Each model is trained separately, with the same model selection procedure. We can see that integrating multiple contexts leads to better performance. VUA-verb Model Acc P R F1 MrBERT-G 85.2 77.3 71.9 74.5 MrBERT-GL 85.5 76.8 73.9 75.3 MrBERT-GLD 86.4 80.8 71.5 75.9 Table 5: The performance of MrBERT when considering different types of contexts: G, L and D indicate global, local and distant contexts, respectively. MrBERT explicitly incorporates verb arguments through grammatical relations as the local context, which differs from other methods. We are interested in the effect of such information. We analyze MrBERT-G and MrBERT-GL. Table 6 shows the distribution of auto-extracted verbsubject and verb-direct object relations in the VUA test dataset. ∆F1 values indicate the improvements of MrBERT-G compared with BERT-SEQ in F1. We can see that MrBERT-G outperforms BERTSEQ mainly when verb’s arguments are incomplete. For verbs with complete verb-subject and verb-direct object structures, little improvement is gained. Table 7 shows the corresponding performance of MrBERT-GL. Better performance is obtained for verbs with all status of grammatical relations. The improvement on verbs in the lower right corner is obvious. In these cases, the verbs are usually intransitive verbs or used as a noun or an adjective. The benefit of involving grammatical relations may be that it helps keep a dynamic and balanced focus between the global and local contexts according to the signals expressed by the grammatical structure. Intuitively, the effect of incorporating grammatical relations should be more obvious for metaphor detection in long sentences, since the local and global contexts are quite different. To verify this, we divide sentences in the test dataset into bins 4247 Verbsubject Verb-direct object Yes No total Yes 1,324 (36%) ∆F1=0.0 2,035 (23%) ∆F1= +0.57 3,359 No 1,201 (38%) ∆F1=+0.05 1,313 (27%) ∆F1= +1.51 2,514 total 2,525 3,348 Table 6: The distribution of available syntactic patterns in VUA-verb test dataset and the improved F1 score of MrBERT-G compared with BERT-SEQ. The figures in brackets are the percentage of metaphors. Verbsubject Verb-direct object Yes No total Yes 1,324 (36%) ∆F1=0.47 2,035 (23%) ∆F1= +0.65 3,359 No 1,201 (38%) ∆F1=0.93 1,313 (27%) ∆F1= +4.29 2,514 total 2,525 3,348 Table 7: Similar to Table 6, this table shows the improved F1 score of MrBERT-GL, instead of MrBERTG, compared with BERT-SEQ. according to the number of clauses. Figure 3 confirms our hypothesis that MrBERT obtains larger improvements on sentences with more clauses, indicating that incorporating grammatical relations can help filter noisy information. Finally, the use of distant context obtains a further improvement. This observation is consistent with the conclusion of (Choi et al., 2021). It also indicates that the BERT tokenizer’s embedding can be used to approximate the representation of the target verb’s basic meaning. 4.4 Results on MOH-X and TroFi Datasets Table 8 shows the results on the MOH-X and TroFi datasets. In the zero-shot transfer setting, MrBERT obtains better performance compared with DeepMet and MelBERT on both datasets. The performance of DeepMet and MelBERT is read from (Choi et al., 1 2 3 4 4+ Number of clauses 0.60 0.65 0.70 0.75 0.80 0.85 F1 BERT-SEQ MrBERT Figure 3: The F1 scores of MrBERT and BERT-SEQ for sentences with different number of clauses. MOH-X Model Acc P R F1 CV Gao et al. (2018) 78.5 75.3 84.3 79.1 Mao et al. (2019) 79.8 77.5 83.1 80.0 Le et al. (2020) 79.9 79.7 80.5 79.6 MrBERT 81.9 80.0 85.1 82.1 MrBERT-finetune 84.9 84.1 85.6 84.2 Trans. DeepMet 79.9 76.5 77.9 MelBERT 79.3 79.7 79.2 MrBERT 79.3 75.9 84.1 79.8 TroFi Model Acc P R F1 CV Gao et al. (2018) 74.6 70.7 71.6 71.1 Mao et al. (2019) 75.2 68.6 76.8 72.4 Le et al. (2020) 76.4 73.1 73.6 73.2 MrBERT 75.1 70.4 74.3 72.2 MrBERT-finetune 76.7 73.9 72.1 72.9 Trans. DeepMet 53.7 72.9 61.7 MelBERT 53.4 74.1 62.0 MrBERT 61.1 53.8 75.0 62.7 Table 8: The experimental results on MOH-X and TroFi, where CV indicates 10-fold cross-validation and Trans. indicates transferring the trained MrBERT on VUA to the target datasets. 2021). The results means MrBERT has good zeroshot transferability, although these datasets have quite different characteristics. In the 10-fold cross-validation setting, the retrained MrBERT can also obtain superior or competitive results compared with previous work. If we continue to fine-tune the pre-trained MrBERT on the target datasets, better performance can be obtained, especially on the MOH-X dataset. 5 Related Work Metaphor detection is a key task in metaphor processing (Veale et al., 2016). It is typically viewed as a classification problem. The early methods were based on rules (Fass, 1991; Narayanan, 1997), 4248 while most recent methods are data-driven. Next, we summarize data-driven methods from the perspective of context types that have been explored. Grammatical relation-level detection This line of work is to determine the metaphoricity of a given grammatical relation, such as verbsubject, verb-direct object or adjective-noun relations (Shutova et al., 2016). The key to this category of work is to represent semantics and capture the relation between the arguments. Feature-based methods are based on handcrafted linguistic features. Shutova and Teufel (2010b) proposed to cluster nouns and verbs to construct semantic domains. Turney et al. (2011) and Shutova and Sun (2013) considered the abstractness of concepts and context. Mohler et al. (2013) exploited Wikipedia and WordNet to build domain signatures. Tsvetkov et al. (2014) combined abstractness, imageability, supersenses, and cross-lingual features. Bulat et al. (2017) exploited attribute-based concept representations. The above handcrafted features heavily rely on linguistic resources and expertise. Recently, distributed representations are exploited for grammatical relation-level metaphor detection. Distributed word embeddings were used as features (Tsvetkov et al., 2014) or to measure semantic relatedness (Gutiérrez et al., 2016; Mao et al., 2018). Visual distributed representations were also proven to be useful (Shutova et al., 2016). Rei et al. (2017) designed a supervised similarity network to capture interactions between words. Song et al. (2020) modeled metaphors as attribute-dependent domain mappings and presented a knowledge graph embedding approach for modeling nominal metaphors. Zayed et al. (2020) identified verb-noun and adjective-noun phrasal metaphoric expressions by modeling phrase representations as a context. Token-level detection Another line of work formulates metaphor detection as a single token classification or sequence labeling problem (Do Dinh and Gurevych, 2016; Gao et al., 2018; Mao et al., 2019). These approaches are mostly based on neural network architectures and learn representations in an end-to-end fashion. These approaches depend on token-level human annotated datasets, such as the widely used VUA dataset (Steen, 2010). BiLSTM plus pre-trained word embeddings is one of the popular architectures for this task (Gao et al., 2018; Mao et al., 2019). Recently, Transformer based pre-trained language models become the most popular architecture in the metaphor detection shared task (Leong et al., 2020). Multitask learning (Dankers et al., 2019; Rohanian et al., 2020; Le et al., 2020; Chen et al., 2020) and discourse context (Dankers et al., 2020) have been exploited as well. Discussion The grammatical relation-level and token-level metaphor detection consider different aspects of information. Grammatical relations incorporate syntactic structures, which are well studied in selectional preferences (Wilks, 1975, 1978) and provide important clues for metaphor detection. However, sentential context is also useful but is ignored. In contrast, token-level metaphor detection explores wider context and gains improvements, but syntactic information is neglected and as discussed in (Zayed et al., 2020), interactions between metaphor components are not explicitly modeled. This paper aims to combine the grammatical relation-level, token-level and semantic-level information through pre-trained language model based contextual relation modeling. 6 Conclusion This paper presented the Metaphor-relation BERT (MrBERT) model for verb metaphor detection. We propose a new view to formulate the task as modeling the metaphorical relation between the target verb and its multiple context components, i.e., contextual relations. We propose and evaluate various ways to extract, model and integrate contextual relations for metaphoricity prediction. We conduct comprehensive experiments on the VUA dataset. The evaluation shows that MrBERT achieves superior or competitive performance compared with previous methods. We also observe that incorporating grammatical relations can help balance local and global contexts, and the basic meaning of the verb as a distant context is effective. Further experiments on small datasets MOH-X and TroFi also show good model transferability of MrBERT. Acknowledgments This work is supported by the National Natural Science Foundation of China (Nos. 61876113, 61876112), Beijing Natural Science Foundation (No. 4192017), Support Project of Highlevel Teachers in Beijing Municipal Universities in the Period of 13th Five-year Plan (CIT&TCD20170322). Lizhen Liu is the corresponding author. 4249 References Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the blanks: Distributional similarity for relation learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2895–2905, Florence, Italy. Association for Computational Linguistics. Julia Birke and Anoop Sarkar. 2006. A clustering approach for nearly unsupervised recognition of nonliteral language. In 11th Conference of the European Chapter of the Association for Computational Linguistics, Trento, Italy. Association for Computational Linguistics. Luana Bulat, Stephen Clark, and Ekaterina Shutova. 2017. Modelling metaphor with attribute-based semantics. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 523–528, Valencia, Spain. Association for Computational Linguistics. Danqi Chen and Christopher Manning. 2014. A fast and accurate dependency parser using neural networks. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 740–750, Doha, Qatar. Association for Computational Linguistics. Xianyang Chen, Chee Wee (Ben) Leong, Michael Flor, and Beata Beigman Klebanov. 2020. Go figure! multi-task transformer-based architecture for metaphor detection using idioms: ETS team in 2020 metaphor shared task. In Proceedings of the Second Workshop on Figurative Language Processing, pages 235–243, Online. Association for Computational Linguistics. Minjin Choi, Sunkyung Lee, Eunseong Choi, Heesoo Park, Junhyuk Lee, Dongwon Lee, and Jongwuk Lee. 2021. MelBERT: Metaphor detection via contextualized late interaction using metaphorical identification theories. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1763–1773, Online. Association for Computational Linguistics. Verna Dankers, Karan Malhotra, Gaurav Kudva, Volodymyr Medentsiy, and Ekaterina Shutova. 2020. Being neighbourly: Neural metaphor identification in discourse. In Proceedings of the Second Workshop on Figurative Language Processing, pages 227–234, Online. Association for Computational Linguistics. Verna Dankers, Marek Rei, Martha Lewis, and Ekaterina Shutova. 2019. Modelling the interplay of metaphor and emotion through multitask learning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2218–2229, Hong Kong, China. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Erik-Lân Do Dinh and Iryna Gurevych. 2016. Tokenlevel metaphor detection using neural networks. In Proceedings of the Fourth Workshop on Metaphor in NLP, pages 28–33, San Diego, California. Association for Computational Linguistics. Dan Fass. 1991. met*: A method for discriminating metonymy and metaphor by computer. Computational linguistics, 17(1):49–90. Ge Gao, Eunsol Choi, Yejin Choi, and Luke Zettlemoyer. 2018. Neural metaphor detection in context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 607–613, Brussels, Belgium. Association for Computational Linguistics. Pragglejaz Group. 2007. Mip: A method for identifying metaphorically used words in discourse. Metaphor and symbol, 22(1):1–39. E. Dario Gutiérrez, Ekaterina Shutova, Tyler Marghetis, and Benjamin Bergen. 2016. Literal and metaphorical senses in compositional distributional semantic models. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 183–193, Berlin, Germany. Association for Computational Linguistics. George Lakoff and Mark Johnson. 1980. Metaphors we live by. University of Chicago press. Duong Le, My Thai, and Thien Nguyen. 2020. Multitask learning for metaphor detection with graph convolutional neural networks and word sense disambiguation. In AAAI, pages 8139–8146. Chee Wee (Ben) Leong, Beata Beigman Klebanov, Chris Hamill, Egon Stemle, Rutuja Ubale, and Xianyang Chen. 2020. A report on the 2020 VUA and TOEFL metaphor detection shared task. In Proceedings of the Second Workshop on Figurative Language Processing, pages 18–29, Online. Association for Computational Linguistics. Chee Wee (Ben) Leong, Beata Beigman Klebanov, and Ekaterina Shutova. 2018. A report on the 2018 VUA metaphor detection shared task. In Proceedings of the Workshop on Figurative Language Processing, pages 56–66, New Orleans, Louisiana. Association for Computational Linguistics. 4250 Rui Mao, Chenghua Lin, and Frank Guerin. 2018. Word embedding and WordNet based metaphor identification and interpretation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1222–1231, Melbourne, Australia. Association for Computational Linguistics. Rui Mao, Chenghua Lin, and Frank Guerin. 2019. End-to-end sequential metaphor identification inspired by linguistic theories. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3888–3898, Florence, Italy. Association for Computational Linguistics. Saif Mohammad, Ekaterina Shutova, and Peter Turney. 2016. Metaphor as a medium for emotion: An empirical study. In Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics, pages 23–33. Michael Mohler, David Bracewell, Marc Tomlinson, and David Hinote. 2013. Semantic signatures for example-based linguistic metaphor detection. In Proceedings of the First Workshop on Metaphor in NLP, pages 27–35, Atlanta, Georgia. Association for Computational Linguistics. Srini Narayanan. 1997. Knowledge-based action representations for metaphor and aspect (KARMA). Ph.D. thesis, Ph. D. thesis, University of California at Berkeley. Arthur Neidlein, Philip Wiesenbach, and Katja Markert. 2020. An analysis of language models for metaphor recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3722–3736. Marek Rei, Luana Bulat, Douwe Kiela, and Ekaterina Shutova. 2017. Grasping the finer point: A supervised similarity network for metaphor detection. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1537–1546, Copenhagen, Denmark. Association for Computational Linguistics. Omid Rohanian, Marek Rei, Shiva Taslimipoor, and Le An Ha. 2020. Verbal multiword expressions for identification of metaphor. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2890–2895, Online. Association for Computational Linguistics. Ekaterina Shutova, Douwe Kiela, and Jean Maillard. 2016. Black holes and white rabbits: Metaphor identification with visual features. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 160–170, San Diego, California. Association for Computational Linguistics. Ekaterina Shutova and Lin Sun. 2013. Unsupervised metaphor identification using hierarchical graph factorization clustering. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 978–988, Atlanta, Georgia. Association for Computational Linguistics. Ekaterina Shutova and Simone Teufel. 2010a. Metaphor corpus annotated for source - target domain mappings. In Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC’10), Valletta, Malta. European Language Resources Association (ELRA). Ekaterina Shutova and Simone Teufel. 2010b. Metaphor corpus annotated for source-target domain mappings. In LREC, volume 2, pages 2–2. Citeseer. Wei Song, Jingjin Guo, Ruiji Fu, Ting Liu, and Lizhen Liu. 2020. A knowledge graph embedding approach for metaphor processing. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:406–420. Gerard Steen. 2010. A method for linguistic metaphor identification: From MIP to MIPVU, volume 14. John Benjamins Publishing. Kevin Stowe, Sarah Moeller, Laura Michaelis, and Martha Palmer. 2019. Linguistic analysis improves neural metaphor detection. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 362–371, Hong Kong, China. Association for Computational Linguistics. Chuandong Su, Fumiyo Fukumoto, Xiaoxi Huang, Jiyi Li, Rongbo Wang, and Zhiqun Chen. 2020. DeepMet: A reading comprehension paradigm for token-level metaphor detection. In Proceedings of the Second Workshop on Figurative Language Processing, pages 30–39, Online. Association for Computational Linguistics. Yulia Tsvetkov, Leonid Boytsov, Anatole Gershman, Eric Nyberg, and Chris Dyer. 2014. Metaphor detection with cross-lingual model transfer. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 248–258, Baltimore, Maryland. Association for Computational Linguistics. Yulia Tsvetkov, Elena Mukomel, and Anatole Gershman. 2013. Cross-lingual metaphor detection using common semantic features. In Proceedings of the First Workshop on Metaphor in NLP, pages 45– 51, Atlanta, Georgia. Association for Computational Linguistics. Peter Turney, Yair Neuman, Dan Assaf, and Yohai Cohen. 2011. Literal and metaphorical sense identification through concrete and abstract context. In Proceedings of the 2011 Conference on Empirical 4251 Methods in Natural Language Processing, pages 680–690, Edinburgh, Scotland, UK. Association for Computational Linguistics. Tony Veale, Ekaterina Shutova, and Beata Beigman Klebanov. 2016. Metaphor: A computational perspective. Synthesis Lectures on Human Language Technologies, 9(1):1–160. Yorick Wilks. 1975. A preferential, pattern-seeking, semantics for natural language inference. Artificial intelligence, 6(1):53–74. Yorick Wilks. 1978. Making preferences more active. Artificial intelligence, 11(3):197–223. Omnia Zayed, John P. McCrae, and Paul Buitelaar. 2020. Contextual modulation for relationlevel metaphor identification. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 388–406, Online. Association for Computational Linguistics.
|
2021
|
327
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4252–4261 August 1–6, 2021. ©2021 Association for Computational Linguistics 4252 Improving Speech Translation by Understanding and Learning from the Auxiliary Text Translation Task Yun Tang, Juan Pino, Xian Li, Changhan Wang, Dmitriy Genzel Facebook AI {yuntang,juancarabina,xianl,changhan,dgenzel}@fb.com Abstract Pretraining and multitask learning are widely used to improve the speech to text translation performance. In this study, we are interested in training a speech to text translation model along with an auxiliary text to text translation task. We conduct a detailed analysis to understand the impact of the auxiliary task on the primary task within the multitask learning framework. Our analysis confirms that multitask learning tends to generate similar decoder representations from different modalities and preserve more information from the pretrained text translation modules. We observe minimal negative transfer effect between the two tasks and sharing more parameters is helpful to transfer knowledge from the text task to the speech task. The analysis also reveals that the modality representation difference at the top decoder layers is still not negligible, and those layers are critical for the translation quality. Inspired by these findings, we propose three methods to improve translation quality. First, a parameter sharing and initialization strategy is proposed to enhance information sharing between the tasks. Second, a novel attention-based regularization is proposed for the encoders and pulls the representations from different modalities closer. Third, an online knowledge distillation is proposed to enhance the knowledge transfer from the text to the speech task. Our experiments show that the proposed approach improves translation performance by more than 2 BLEU over a strong baseline and achieves state-of-theart results on the MUST-C English-German, English-French and English-Spanish language pairs. 1 Introduction End-to-end methods have achieved significant progress in speech to text translation (ST) and even surpassed the traditional pipeline-based methods in some applications (Niehues et al., 2019; Salesky and Black, 2020). However, the success of endto-end methods relies on large amounts of training data, which is quite expensive to obtain and relatively small in practice. Building ST systems from pretrained models with multitask learning (MTL) is widely used to overcome the limited training data issue (Weiss et al., 2017; Anastasopoulos and Chiang, 2018; Bahar et al., 2019; Indurthi et al., 2020; Wang et al., 2020b; Li et al., 2020). Nevertheless, little prior work has been devoted to understanding the interactions between different tasks. Standley et al. (2020) conduct an empirical study on computer vision tasks for MTL. They find many “assumptions” for MTL may not be held for specific applications. For example, “similar” tasks do not necessarily train better together. In this study, we focus on training the ST model along with an auxiliary text to text machine translation (MT) task. We are interested in the task interactions with different modalities and in improving the primary ST task with the help from the auxiliary MT task. The model is initialized with pretrained modules from automatic speech recognition (ASR) and MT. Two types of analysis are conducted on the fine-tuned multitask learned models. The first focuses on the model variation by comparing fine-tuned models with pretrained models for different tasks. The second aims to measure internal representation differences due to different modalities. The analysis leads to three main findings. First, the analysis confirms that MTL tends to generate similar model representations for different input modalities and preserves more information from the pretrained MT modules. Second, we do not observe significant negative transfer effect from the MT task to the corresponding ST task. Sharing more parameters is helpful to transfer knowledge to the primary ST task. Finally, the top layers in the ST decoder are more critical to the translation 4253 performance and they are also more sensitive to the modality difference. The model representations from different modalities demonstrate larger difference for the top layers in our analysis. Inspired by these findings, we propose three techniques to enhance the performance of the primary ST task. First, we propose to maximize parameter sharing between the ST and MT tasks, i.e. the entire decoder and the top encoder layers. Those shared parameters are initialized with the corresponding MT models. Second, a cross-attentive regularization is introduced for the encoders. It minimizes the L2 distance between two reconstructed encoder output sequences and encourages the encoder outputs from different modalities to be closer to each other. Finally, an online knowledge distillation learning is introduced for MTL in order to enhance knowledge transfer from the MT to the ST task. Our contributions are summarized as follows: 1. A detailed analysis is conducted on the interaction between the primary ST task and the auxiliary MT task. 2. A parameter sharing and initialization strategy are proposed to encourage information sharing between tasks. 3. Cross-attentive regularization and online knowledge distillation are proposed to reduce the model representation difference between different modalities and enhance the knowledge transfer from the MT task to the ST task. 4. Our system achieves state of the art results on the MUST-C English-German (EN-DE), English-French (EN-FR) and English-Spanish (EN-ES) language pairs, with 2 or more BLEU gains over strong baselines. 2 Related Work Multitask learning aims to improve generalization by leveraging domain-specific information contained in the training signals of related tasks (Vandenhende et al., 2020). Compared with single task, MTL has many advantages, such as the potential to improve performance by sharing complementary information or acting as a regularizer. Many previous works focus on learning a good model for all tasks. Chen et al. (2018) study the gradients from different tasks and conduct task dependent gradient normalization to encourage different tasks to learn at similar speed. Maninis et al. Figure 1: Joint Training framework. The speech to text translation task is depicted as dark gray line, text to text translation task is illustrated as light gray line. The parameters in blue modules are shared between two tasks. (2019); Liu et al. (2019a); Pfeiffer et al. (2020) introduce task-dependent components to enhance individual task performance. Weiss et al. (2017) explore different multitask training strategies for ST, and they find the oneto-many strategy, in which an encoder is shared between the ST and ASR tasks, is more effective. Anastasopoulos and Chiang (2018) further extend it to a triangle structure by concatenating ASR and ST models. Bahar et al. (2019) compare different multitask strategies for the ST task, and they confirm many-to-one strategy, in which MT and ST are trained together and the decoder is shared between two tasks, is effective if extra bitext data is used. In this work, we carefully study the relation between co-trained tasks in the many-to-one strategy, and the analysis results guide us to propose three techniques to learn more from the auxiliary MT task and enhance the ST performance further. Model analysis Chatterji et al. (2020) propose criticality analysis to measure the importance of different modules from the trained model. Parameters 4254 in the selected module or layer are partially rolled back to the initial values, and the module criticality or importance is measured by the performance drop after modification. Larger performance drops indicate a more critical module. Inspired by their work, we extend it to the analysis on the jointly trained models with different pretrained modules and schemes. Raghu et al. (2017); Morcos et al. (2018) propose to employ canonical correlation to measure the similarity between different models given the same input. We extend their work to study a model with inputs from different modalities. 3 Methods The proposed ST system is co-trained with the MT task as depicted in Figure 1. The modules in the primary ST task are connected with dark gray lines and the auxiliary MT task is illustrated with light gray lines. The parameters in the blue modules are shared between the two tasks. During inference with speech input, only modules related to the ST task are used. The model has two encoders, a text encoder and a speech encoder, to take text and speech input respectively. The decoder is shared between the two tasks. To encourage knowledge sharing between the two tasks, the top encoder layers are also shared. The parameters of the shared modules are initialized with a pretrained MT model. A novel crossattentive regularization is proposed to reduce the distance between encoder outputs from different input modalities. We also introduce a novel online knowledge distillation method where the output from the auxiliary MT task is used to guide the ST model training. The cross-attentive regularization and online knowledge distillation are illustrated as orange modules in Figure 1 and the details are presented in the following two subsections. 3.1 Cross-Attentive Regularization The cross-attentive regularization (CAR) is proposed to increase the similarity between the text encoder outputs and their corresponding speech encoder outputs. Hence, the performance of the more difficult ST task can be improved by learning from the relatively easier MT task. Encoder output sequences from different modalities can not be compared directly since they have different lengths. In CAR, the two reconstructed sequences are calculated from the text output sequence via self-attention or the speech output sequence via cross attention over the text output sequence. The two reconstructed sequences have the same length and the distance is simply measured as the L2 distance between the two sequences. Formally, we denote a speech to text translation training sample as a triplet o = (Xs, xt, y). Xs ∈ Rds×N, xt ∈RM, and y ∈RK are the speech feature input, text token input and target text output respectively. N, M and K are the corresponding sequence lengths. Assume Hs = (hs 1, hs 2, · · ·, hs N) and Ht = (ht 1, ht 2, · · ·, ht M), hs n, ht m ∈Rdh are outputs from the speech encoder and text encoder respectively, where dh is the dimension of the output states. A similarity matrix S ∈RN×M is defined as the cosine distance between the tensors in the two sequences: si,j = (hs i)′ · ht j ||hs i||2||ht j||2 (1) where si,j is the ith row and jth column component in S. The text encoder outputs Ht are reconstructed through the speech encoder outputs Hs and similarity matrix S as below. Hs→t = Hs · softmax(S) (2) Ht→t, the reconstruction of Ht from itself, can be computed similarly via self-attention. CAR is defined as the L2 distance between the two reconstruction encoder outputs: LCAR(θs) = 1 M Hs→t −sg[Ht→t] 2 (3) where sg[·] is the stop-gradient operator and θs are the ST model parameters. By optimizing the model with CAR, the speech encoder is encouraged to learn from more accurate text encoder and generates similar encoder outputs after reconstruction. CAR is inspired by the attention mechanism between the encoder and decoder where the decoder states are reconstructed through encoder output states via the attention mechanism. 3.2 Online Knowledge Distillation Knowledge distillation (KD) is widely used for model compression (Hinton et al., 2015; Kim and Rush, 2016) where a smaller student network is trained to mimic the original teacher network by minimizing the loss between the student and teacher outputs. The ST task is considerably more difficult than the MT task since the speech input is noisier and more ambiguous than the text input. 4255 The accuracy of the MT model is usually much higher than the corresponding ST model. Knowledge distillation from a well trained MT model to a ST model has been proved to be an effective way to improve the ST performance (Liu et al., 2019b; Gaido et al., 2020). In this work, we extend knowledge distillation to the MTL framework where both ST and MT are fine-tuned simultaneously with shared parameters. Concretely, we assume an MTL model learns from a data set D with target vocabulary size |V |. The training criterion is to minimize negative log likelihood (NLL) for each example o = (Xs, xt, y) ∈D from the training data: LNLL(θs) = − D X o K X k=1 |V | X v=1 δ(yk = v) log p(yk = v|y<k, Xs, θs) (4) where δ(·) is the indicator function and p the distribution from the ST model (parameterized by θs). Assume the probability distribution for yk given text input xt and MT model θt is q(yk = v|y<k, xt, θt), the knowledge distillation loss is defined as minimizing the cross-entropy with the MT’s probability distribution LKD(θs) = − D X o K X k=1 |V | X v=1 q(yk = v|y<k, xt, θt) log p(yk = v|y<k, Xs, θs) (5) The overall loss is the combination of crossattentive regularization, knowledge distillation loss, negative log likelihood loss for both ST and MT, as follows: L(θs, θt) = αLNLL(θs) + (1 −α)LKD(θs) +λLCAR(θs) + LNLL(θt) (6) where α and λ are predefined hyper-parameters. 4 Experimental Setup Experiments are conducted on three MUSTC (Gangi et al., 2019a) language pairs: EN-DE, EN-ES and EN-FR. The models are developed and analyzed on the dev set and the final results are reported on the tst-COMMON set. We use WMT parallel data from different years, 2013 for Spanish, 2014 for German, and 2016 for French, as extra text training corpus for MTL. Case-sensitive detokenized BLEU is reported by SACREBLEU with default options (Post, 2018). We use the “T-Md” configuration from (Wang et al., 2020a) in all experiments. The speech encoder has 12 transformer layers while the decoder is with 6 transformer layers. For the MTL model, the text encoder has 6 transformer layers. The transformer layer has an input embedding size of 512 and middle layer dimension 2048. We share parameters of all 6 text encoder transformer layers with the top 6 transformer layers in the speech encoder, hence both encoders use the same modules to generate the encoder outputs. The Adam optimizer (Kingma and Ba, 2014) with a learning rate 0.002 is employed in the experiments. Label smoothing and dropout rate are both set to 0.1. We choose α = 0.8 and λ = 0.02 in Equation 6 through grid search ([0.1, 1.0] for α and [0.01, 0.05] for λ). Input speech is represented as 80D log melfilterbank coefficients computed every 10ms with a 25ms window. Global channel mean and variance normalization is applied. The SpecAugment (Park et al., 2019) data augmentation with the LB policy is applied in all experiments. The input text tokens are converted into their corresponding pronunciation form as phoneme sequences (Tang et al., 2021; Renduchintala et al., 2018). The grapheme to phoneme conversion is done through the “g2p en” python package (Lee and Kim, 2018). The leading phoneme in a word is appended with an extra “ ” to mark word boundaries. In total, the vocabulary size for the input phonemes is 134. The target vocabulary consists of 10k “unigram” subword units learned by SentencePiece (Kudo and Richardson, 2018) with full character coverage of all training text data. All ST or jointly trained models are initialized with pretrained ASR and MT modules. The ASR model is trained on the same English speech training data from MUST-C with the “T-Md” configuration too. The pretrained MT models are trained for each language pair with the aforementioned WMT data. The MT encoder and decoder configurations are the same as the text encoder and decoder in the MTL model mentioned above. The models are fine-tuned to 100 epochs using 8 V100 GPUs for approximate one day. The batch size is 10,000 frames for speech to text translation samples and 10,000 tokens for parallel text samples per GPU. The model parameters are updated every 4 batches. Speech training samples and text input samples are used to update the model alternatively. 4256 Model Encoder Configuration Speech Text Shared ST ASR None None JT ASR MT None JT-S-ASR ASR MT ASR JT-S-MT ASR MT MT Table 1: Model initialization schemes The models are trained with FAIRSEQ (Ott et al., 2019; Wang et al., 2020a). The last 10 checkpoints are averaged for inference with beam size 5. 1. 5 MTL Analysis 5.1 Model Variation We extend Chatterji et al. (2020)’s work to analyze a MTL model. We initialize models with different pretrained modules and fine-tune them for ST and MT tasks within the MTL framework. The pretrained modules come from ASR and MT tasks. Criticality analysis is conducted on the ST model after the MTL fine-tuning step. The parameters in the selected modules are interpolated with corresponding parameters in the pretrained modules. MUST-C EN-DE dev set is used for BLEU computation. With different interpolation ratios, we obtain different BLEU scores. The BLEU difference comes from two sources. The first one comes from the selected module itself. If the module is important and sensitive, very small perturbation could result in a nontrivial BLEU difference as (Chatterji et al., 2020). Another source of difference is that if the selected module changes significantly to adapt to the ST task, rewinding the parameters back to the initial task may lead to a substantial decrease in BLEU. We attempt to quantify the extent of the degradation from the second source, which can be indicative of the model variation from the pretrained task to the ST task. This is accomplished by comparing the BLEU differences for the same module but using different initialization and training schemes. Table 1 lists models initialized with different pretrained modules. “ST” designates a ST model trained with the single ST task, “JT” corresponds to a ST model trained with the primary ST task and auxiliary MT task together. “JT-S-ASR” and “JTS-MT” are another two jointly trained models but 1The source code will be released at https://github.com/pytorch/fairseq/tree/master/examples/speech text joint to text (a) ST Enc. (b) ST Dec. Figure 2: Criticality analysis for the “ST” model. with the top encoder layers shared as described in section 4. The difference between the two models is how we initialized the shared encoder layers, either from the pretrained ASR model for “JT-SASR” or from the pretrained MT model for “JT-SMT”. ST Figure 2 shows the analysis for the “ST” model. The x-axis is the interpolation ratio and “1.0” means the pretrained parameters are used. The y-axis is the relative change in BLEU compared with the well-trained ST model. It is clear that higher layers are more critical to the performance. Around 5 BLEU decrease is observed on the top encoder layer (11) and top decoder layer (5) during the criticality tests. The following analysis will compare with Figure 2 and we can separate the aforementioned second source from the first one. JT Figure 3 presents the analysis for the “JT” model. The jointly trained model shows smaller degradation compared with “ST” for the decoder layers. This indicates that training the ST and MT tasks together helps to preserve more information from the original MT decoder and partially remedies the catastrophic forgetting (McCloskey and Cohen, 1989) during the finetuning phase. On the other hand, after rolling parameters back to the initial ASR model, the jointly trained model shows a larger degradation for the encoder layers. This means that the speech encoder in the jointly trained model has deviated far away from the speech encoder in the initial ASR task. We conclude that the shared decoder is subject to more constraints since it is optimized toward both MT and ST tasks while the speech encoder has to undergo larger changes in order to align with the text encoder, although there is no parameter sharing between two encoders. JT-S-ASR and JT-S-MT Results for models with 4257 (a) JT Enc. (b) JT Dec. Figure 3: Criticality analysis for the “JT” model. (a) JT-S-ASR Enc. (b) JT-S-ASR Dec. Figure 4: Criticality analysis for the “JT-S-ASR” model. The shared encoder layers are initialized with the layers from the ASR encoder. the top encoder layers shared are presented in Figure 4 and 5. In “JT-S-MT”, the top 6 shared encoder layers are initialized with the pretrained MT encoder. We illustrate their BLEU difference trajectories with dotted lines in Figure 5 (a) so they can be easily distinguished from other layers initialized from the ASR encoder. The BLEU difference for the top encoder layer is down from 20.2 to 17.6 when the parameters are replaced with the ones in the pretrained ASR encoder. It is further reduced to 10.0 if the shared layers are initialized with MT encoder layers. The BLEU differences in the decoder layers are mixed. The performance of “JT-S-ASR” degrades quickly in the criticality test for the top decoder layer, while “JT-S-MT performs similarly in the test as “JT” decoder. We argue that the top layers in the finetuned ST encoder might be closer to the MT encoder than the ASR encoder. It preserves more information from the MT task by sharing more parameters between two tasks and initializing them with pretrained MT modules. This is a desirable property since we want to transfer more knowledge from the text corpus to the ST task. (a) JT-S-MT Enc. (b) JT-S-MT Dec. Figure 5: Criticality analysis for the “JT-S-MT” model. The shared encoder layers are initialized with the layers from the MT encoder. Figure 6: Comparison of decoder layers correlation coefficients between text and speech input (“JT-S-MT”). 5.2 Modality Variation The jointly trained model takes input from two modalities, i.e. text or speech, and we are interested in the model internal representation difference for paired inputs. Given text target y, we extract the decoder hidden state representations for the corresponding text input xt and speech input Xs. The decoder representation difference solely comes from different input modalities. The difference is quantified by the correlation coefficient over all samples evaluated between two input modalities: rs,t(l, d) = σst(l, d) σs(l, d)σt(l, d) (7) where σz(l, d), z ∈[s, t] is the standard deviations of decoder hidden states at layer l for component d in all samples, and σst(l, d) is the corresponding covariance. The layer-wise correlation coefficient is the average of all components: rs,t(l) = 1 D X d rs,t(l, d) (8) Figure 6 depicts the correlation coefficient between speech input and text input for each decoder layer in the model “JT-S-MT”. The x-axis is the number of training epochs and the y-axis represents the correlation coefficient for each layer. There 4258 Data corpus #pars(m) DE ES FR Gangi et al. (2019b) 30 17.7 20.9 26.5 Inaguma et al. (2020) 22.9 28.0 32.7 Pino et al. (2020) 435 25.2 34.5 ST 76 21.5 28.1 33.8 JT 76 24.1 29.0 35.1 JT Proposed 76 26.8 31.0 37.4 Table 2: BLEU on three language pairs in the MuST-C tst-COMMON datasets. are two observations. First, the correlation coefficients become larger and close to “1.0” as training converges. Second, the higher the layer, the smaller the correlation coefficient. We hypothesize that the inputs to the lower layers are dominated by the decoder text embeddings, which are the same for both modalities, and the inputs to the higher layers would contain more information from the encoder outputs, which result in the decoder internal representation differences. The analysis shows a well trained MTL decoder has similar representations for paired text and speech input. However, the top decoder layers still have nontrivial representation differences due to different modalities. 6 Experimental Results 6.1 Main Results The main ST results are presented in Table 2. The first three rows are results from the literature. “ST” and “JT” are models initialized as Table 1 and studied in section 5. The last row (“JT Proposed”) presents results from the proposed system, in which the top encoder layers and decoder are shared, and the models are optimized following Equation 6. The second column (“pars(m)”) lists the number of parameters used during inference. From Table 2, our “ST” baseline is comparable to the previously reported results except (Pino et al., 2020), who use a much larger model and additional weakly supervised speech training data. As expected, the vanilla joint training baseline (“JT”) outperforms the “ST” baseline with the help of extra bitext training data. Finally, the proposed joint training model (“JT Proposed”) achieves 2.0∼2.7 BLEU gains over the strong joint training baseline (“JT”). 6.2 Ablation Table 3 breaks down the performance gains into individual components/changes. Sharing encoder layers improves the quality for all three language pairs EN-DE EN-ES EN-FR JT 24.1 29.0 35.1 JT-S-ASR 24.4 29.4 35.4 JT-S-MT 24.7 29.7 35.3 + CAR 25.0 30.4 36.2 + CAR + KD 26.8 31.0 37.4 Table 3: Ablation study. (a) JT Proposed Enc. (b) JT Proposed Dec. Figure 7: Criticality analysis for “JT Proposed”. (“JT” v.s. “JT-S-ASR”). Initializing the shared encoder layers with pretrained MT modules leads to BLEU increase for two of the three evaluated translation pairs (“JT-S-ASR” v.s. “JT-S-MT”). For EN-FR, the degradation is minimal (-0.1 BLEU). Overall, sharing top encoder layers can increase BLEU by 0.2∼0.7 (“JT-S-MT” v.s. “JT”). CAR further improves the translation by another 0.3∼0.9 BLEU. The best results are achieved by applying the shared top encoder layers, CAR and online KD together. They are about 2.9+ BLEU better than the single task based system (“ST”) and achieve 2+ BLEU increase on top of the strong vanilla joint training system(“JT”). Figure 7 demonstrates the model variation for the proposed system on the MUST-C EN-DE dev set. Compared with Figure 5, the decoder shows less degradation during the criticality test and it shows CAR and online KD help to preserve more information from the MT task. Figure 8 shows the corresponding correlation coefficients between paired text and speech input from the top decoder Figure 8: Correlation coefficient for the top decoder layers (epoch 100). 4259 Model BLEU JT-S-MT 24.7 JT-S-MT + Adapter 24.7 JT-S-MT + Dedicated Attention 24.2 Table 4: BLEU score for models with task dependent components layer from different model configurations. It also confirms that the proposed methods, i.e., shared top encoder layers, CAR and online KD, all reduce the modality difference substantially. 6.3 Task Dependent Components In MLT, many works (Maninis et al., 2019; Liu et al., 2019a; Zhang et al., 2020; Pfeiffer et al., 2020) employ task-dependent components to alleviate the negative transfer effect. In Table 4, we compare the “JT-S-MT” model with two variants using different task-dependent components. The first one (“JT-S-MT + Adapter”) (Bapna et al., 2019) adds an extra adapter module on the top of the speech encoder. Hence, the speech encoder outputs, which are generated from shared encoder layers, are further processed to reduce the difference between speech input and text input. The adapter module consists of a linear layer and layer normalization layer. The second variant (“JT-S-MT + Dedicated Attention”) (Blackwood et al., 2018) introduces dedicated decoder modules for different tasks. Attention layers between encoder and decoder, and the layer normalization modules are not shared between the ST and MT tasks. It gives the decoder more flexibility to handle information from different modalities. The results show the extra adapter layer doesn’t bring gain while the task dependent attention module actually makes the performance worse. It indicates that the negative transfer effect is not significant in this study and adding extra task-dependent components might not be necessary. 6.4 Impact on the MT Task As shown in Table 2, training ST models with an auxiliary MT task improves the translation quality substantially. It may be interesting to examine the impact on the auxiliary task itself. We evaluate the MT model jointly trained with the ST task. Results are shown in Table 5. “ST (JT Proposed)” in the first row corresponds to the best results obtained for the ST task. The detailed experimental setup is described in Appendix A. For reference, we also EN-DE EN-ES EN-FR ST (JT Proposed) 26.8 31.0 37.4 MT (Gangi et al., 2019a) 28.1 34.2 42.2 MT 25.4 27.7 33.5 MT (Tuned) 29.6 34.3 41.4 MT (JT) 28.9 33.9 41.6 MT (JT Proposed) 30.5 34.7 42.3 Table 5: Comparison between ST and MT. include the MT evaluation results from MUSTC (Gangi et al., 2019a) in the second row. All MT models (in the last 4 rows) take phoneme sequences as input instead of SentencePiece. “MT” (row 3) shows the results from pretrained MT models on WMT. In the “MT (Tuned)” row, the MT models pretrained on WMT are fine-tuned on the MUST-C datasets. The large improvements clearly show a domain mismatch between WMT and MUST-C. The MT models trained with WMT data are improved after fine-tuning, and they are comparable with the ones reported in (Gangi et al., 2019a), though the input token is in pronunciation form, which is more ambiguous than the corresponding SentencePiece unit. “MT (JT)” and “MT (JT Proposed)” are results from the co-trained MT models in “JT” and “JT Proposed” respectively. After fine-tuning using both MuST-C (speech and text) and WMT (text only) training data, the auxiliary MT models perform better than the corresponding ST models. The proposed techniques further improve the co-trained MT models by 0.7∼1.6 BLEU. While this is a surprising result, we note that the dedicated MT models may be improved with better hyperparameter tuning. In conclusion, the results show the proposed methods are effective to unify two tasks into one model with minimal negative transfer effect. 7 Conclusions In this study, we focus on understanding the interactions between the ST and MT tasks under the MTL framework, and on boosting the performance of the primary ST model with the auxiliary MT task. Two types of analysis on model variation and modality variation, are conducted on the MTL models. The analysis demonstrates MTL helps to preserve information from the MT task and generates similar model representations for different modalities. We observe a minimal negative transfer effect between the two tasks. Sharing more parameters can further boost the information transfer from 4260 the MT task to the ST model. The analysis also reveals that the model representation difference due to modality difference is nontrivial, especially for the top decoder layers, which are critical for the translation performance. Inspired by the findings, we propose three techniques to increase knowledge transfer from the MT task to the ST task. These techniques include parameter sharing and initialization strategy to improve the information sharing between tasks, CAR and online KD to encourage the ST system to learn more from the auxiliary MT task and then generate similar model representations from different modalities. Our results show that the proposed methods improve translation performance and achieve state-of–the-art results on three MUST-C language pairs. References Antonios Anastasopoulos and David Chiang. 2018. Tied multitask learning for neural speech translation. In NAACL-HLT. Parnia Bahar, Tobias Bieschke, and Hermann Ney. 2019. A comparative study on end-to-end speech to text translation. In ASRU. Ankur Bapna, N. Arivazhagan, and Orhan Firat. 2019. Simple, scalable adaptation for neural machine translation. In EMNLP/IJCNLP. G. Blackwood, Miguel Ballesteros, and T. Ward. 2018. Multilingual neural machine translation with taskspecific attention. In COLING. Niladri S. Chatterji, Behnam Neyshabur, and H. Sedghi. 2020. The intriguing role of module criticality in the generalization of deep networks. In ICLR. Z. Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. 2018. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML. Marco Gaido, Mattia Antonino Di Gangi, Matteo Negri, and Marco Turchi. 2020. End-toend speech-translation with knowledge distillation: Fbk@iwslt2020. Mattia Antonino Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019a. MuST-C: a multilingual speech translation corpus. In NAACL-HLT. Mattia Antonino Di Gangi, Matteo Negri, and Marco Turchi. 2019b. One-to-many multilingual end-toend speech translation. In ASRU. Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531. H. Inaguma, S. Kiyono, K. Duh, S. Karita, N. Soplin, T. Hayashi, and S. Watanabe. 2020. Espnet-st: Allin-one speech translation toolkit. In ACL. Sathish Reddy Indurthi, HouJeung Han, Nikhil Kumar Lakumarapu, Beom seok Lee, Insoo Chung, Sang-Ha Kim, and Chanwoo Kim. 2020. Endend speech-to-text translation with modality agnostic meta-learning. In ICASSP. Yoon Kim and Alexander M. Rush. 2016. Sequencelevel knowledge distillation. In EMNLP. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. In ICLR. T. Kudo and J. Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In EMNLP. Y. Lee and T. Kim. 2018. Learning pronunciation from a foreign language in speech synthesis networks. ArXiv. Xian Li, Changhan Wang, Yun Tang, Chau Tran, Yuqing Tang, Juan Pino, Alexei Baevski, Alexis Conneau, and Michael Auli. 2020. Multilingual speech translation with efficient finetuning of pretrained models. arXiv: Computation and Language. Shikun Liu, Edward Johns, and A. Davison. 2019a. End-to-end multi-task learning with attention. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1871–1880. Yuchen Liu, Hao Xiong, Zhongjun He, Jiajun Zhang, Hua Wu, Haifeng Wang, and Chengqing Zong. 2019b. End-to-end speech translation with knowledge distillation. In Interspeech. K. Maninis, Ilija Radosavovic, and I. Kokkinos. 2019. Attentive single-tasking of multiple tasks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1851–1860. M. McCloskey and N. J. Cohen. 1989. Catastrophic interference in connectionist networks: The sequential learning problem. Psychology of Learning and Motivation, 24:109–165. Ari S. Morcos, M. Raghu, and S. Bengio. 2018. Insights on representational similarity in neural networks with canonical correlation. In NeurIPS. Jan Niehues, R. Cattoni, Sebastian St¨uker, Matteo Negri, Marco Turchi, Elizabeth Salesky, Ramon Sanabria, Lo¨ıc Barrault, Lucia Specia, and Marcello Federico. 2019. The IWSLT 2019 evaluation campaign. Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, S. Gross, Nathan Ng, David Grangier, and M. Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL. 4261 D. Park, W. Chan, Y. Zhang, C. Chiu, B. Zoph, E. Cubuk, and Q. Le. 2019. Specaugment: A simple data augmentation method for automatic speech recognition. Interspeech. Jonas Pfeiffer, Ivan Vulic, Iryna Gurevych, and Sebastian Ruder. 2020. MAD-X: An adapter-based framework for multi-task cross-lingual transfer. In EMNLP. J. Pino, Q. Xu, X. Ma, M. Dousti, and Y. Tang. 2020. Self-training for end-to-end speech translation. In Interspeech. Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186– 191, Brussels, Belgium. Association for Computational Linguistics. M. Raghu, J. Gilmer, J. Yosinski, and Jascha SohlDickstein. 2017. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. In NIPS. A. Renduchintala, S. Ding, M. Wiesner, and S. Watanabe. 2018. Multi-modal data augmentation for endto-end asr. In INTERSPEECH. Elizabeth Salesky and Alan W Black. 2020. Phone features improve speech translation. In ACL. T. Standley, A. Zamir, Dawn Chen, L. Guibas, Jitendra Malik, and S. Savarese. 2020. Which tasks should be learned together in multi-task learning? In ICML. Yun Tang, Juan Pino, Changhan Wang, Xutai Ma, and Dmitriy Genzel. 2021. A general multi-task learning framework to leverage text data for speech to text tasks. In ICASSP. Simon Vandenhende, S. Georgoulis, Wouter Van Gansbeke, M. Proesmans, Dengxin Dai, and L. Gool. 2020. Multi-task learning for dense prediction tasks: A survey. arXiv: Computer Vision and Pattern Recognition. C. Wang, Y. Tang, X. Ma, A. Wu, D. Okhonko, and J. Pino. 2020a. fairseq s2t: Fast speech-to-text modeling with fairseq. In AACL (demo). Chengyi Wang, Yu Wu, Shujie Liu, Zhenglu Yang, and Ming Zhou. 2020b. Bridging the gap between pretraining and fine-tuning for end-to-end speech translation. In AAAI. Ron J. Weiss, Jan Chorowski, Navdeep Jaitly, Yonghui Wu, and Zhifeng Chen. 2017. Sequence-tosequence models can directly translate foreign speech. In INTERSPEECH. Biao Zhang, Philip Williams, Ivan Titov, and Rico Sennrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In ACL. A Appendix The detailed experimental setup for “MT” and “MT(Tuned)” in Table 5 are described as below. We trained MT models for each language pair in “EN-DE”, “EN-ES”, and “EN-FR”. The training data is from WMT from different years, 2013 for “EN-ES”, 2014 for “EN-DE” and 2016 for “ENFR”. We use “transformer wmt en de” architecture from Fairseq. The models are with embedding size 512 and feed-forward layer dimension 2048. Both encoder and decoder are with 6 transformer layers. The input is phoneme sequence and output is SentencePiece sequence. The vocabularies are shared with the corresponding speech to text translation models. The models are optimized with Adam with learning rate equal to 0.001. Beside experiments in Table 5, the trained MT models are used to initialize the jointly trained models. We further fine-tuned the “MT” models trained from WMT data to MUST-C data sets using source transcripts and target translation labels. No speech data is used. Similar to the “MT” models, Adam optimizer with learning rate equal to 0.001 is used. The models are fine-tuned on the corresponding MUST-C data sets for 15 epochs and the checkpoints from the last 5 epochs are averaged for evaluation.
|
2021
|
328
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4262–4274 August 1–6, 2021. ©2021 Association for Computational Linguistics 4262 Probing Toxic Content in Large Pre-Trained Language Models Nedjma Ousidhoum, Xinran Zhao, Tianqing Fang, Yangqiu Song, Dit-Yan Yeung Department of Computer Science and Engineering The Hong Kong University of Science and Technology [email protected], [email protected], [email protected], [email protected], [email protected] Abstract Large pre-trained language models (PTLMs) have been shown to carry biases towards different social groups which leads to the reproduction of stereotypical and toxic content by major NLP systems. We propose a method based on logistic regression classifiers to probe English, French, and Arabic PTLMs and quantify the potentially harmful content that they convey with respect to a set of templates. The templates are prompted by a name of a social group followed by a cause-effect relation. We use PTLMs to predict masked tokens at the end of a sentence in order to examine how likely they enable toxicity towards specific communities. We shed the light on how such negative content can be triggered within unrelated and benign contexts based on evidence from a large-scale study, then we explain how to take advantage of our methodology to assess and mitigate the toxicity transmitted by PTLMs. 1 Introduction The recent gain in size of pre-trained language models (PTLMs) has had a large impact on state-of-theart NLP models. Although their efficiency and usefulness in different NLP tasks is incontestable, their shortcomings such as their learning and reproduction of harmful biases cannot be overlooked and ought to be addressed. Present work on evaluating the sensitivity of language models towards stereotypical content involves the construction of assessment benchmarks (Nadeem et al., 2020; Tay et al., 2020; Gehman et al., 2020) in addition to the study of the potential risks associated with the use and deployment of PTLMs (Bender et al., 2021). Previous work on probing PTLMs focuses on their syntactic and semantic limitations (Hewitt and Manning, 2019; Marvin and Linzen, 2018), lack of domainspecific knowledge (Jin et al., 2019), and absence of commonsense (Petroni et al., 2019; Lin et al., 2020). However, except for a recent evaluation process of hurtful sentence completion (Nozza et al., 2021), we notice a lack of large-scale probing experiments for quantifying toxic content in PTLMs or systemic methodologies to measure the extent to which they generate harmful content about different social groups. In this paper, we present an extensive study which examines the generation of harmful content by PTLMs. First, we create cloze statements which are prompted by explicit names of social groups followed by benign and simple actions from the ATOMIC cause-effect knowledge graph patterns (Sap et al., 2019b). Then, we use a PTLM to predict possible reasons for these actions. We look into how BERT (Devlin et al., 2019), RoBERTa (Liu et al., 2019), and GPT-2 (Radford et al., 2019) associate unrelated and detrimental causes to basic everyday actions and examine how frequently the predicted words relate to specific social groups. Moreover, we study the same phenomenon in two other languages by translating more than 700 ATOMIC commonsense actions to Arabic and French, along with names of social groups, then run the same experiments using the French PTLM CamemBERT (Martin et al., 2020), and the Arabic AraBERT (Antoun et al., 2020). We find that, overall, the predicted content can also be irrelevant and offensive especially when the subject of the sentence is part of a marginalized community in the predominant culture of the language. In order to gauge the generated toxicity by different language models, we train simple toxicity classifiers based on logistic regression using available hate speech and offensive language datasets. We reduce the classification bias using a two-step approach to first, filter out examples with identity words which typically lead classifiers to predict a toxic label, then perform a second classification step on the remaining examples. We further con4263 Masked LM Top 10 predictions Jewish, Muslim, hungry, poor, gay, rich, afraid, Arab, blind, sick An arab calls everyone because he is [MASK] Input sentence x Jewish sick p([MASK]|x) 0.26 0.06 0.02 0.01 Muslim blind Figure 1: An example of generated content using BERT. Intuitively, one would think that adjectives would be prioritized over than ethnic/religious affiliations in a cause-effect cloze statement which appears not to be the case. Stereotypical and ethnic/religious terms are highlighted in bold font. duct a human evaluation step on 100 automatically labeled examples in each language. Experimental results show that distinct PTLMs demonstrate different percentages of generated toxic content based on the patterns that we use. The human annotations confirm the existence of toxicity in English, French, and Arabic PTLMS and show that, despite their imperfections, the classifiers can be used as toxicity pointers. Our main contributions can be summarized in the following. • We perform a large-scale extensible study on toxic content in PTLMs without relying on datasets which are specific to such a task. • We quantify common misconceptions and wrongly attributed designations to people from different communities. This assessment can be taken into account when using a PTLM for toxic language classification, and when adopting a mitigation strategy in NLP experiments. • We develop a large dataset based on structured patterns that can later be used for the evaluation of toxic language classification and harmful content within PTLMs. We make our data resources publicly available to the community. 1 The rest of the paper is organized as follows. We first introduce our methodology in Section 2. 1The link to code and data is https://github.com/ HKUST-KnowComp/Probing_toxicity_in_PTLMs. PATTERN PersonX ACTION because he [MASK]. PersonX ACTION because of his [MASK]. ManX ACTION because he [MASK]. ManX ACTION because of his [MASK]. WomanX ACTION because she [MASK]. WomanX ACTION because of her [MASK]. Table 1: Patterns used with the ATOMIC actions. Given the nature of PTLMs and for the sake of our multilingual study, we use the pronouns he and she even for the non-gendered PersonX. ManX and WomanX refer to a man and a woman from specific social groups such as a Black man and an Asian woman, respectively. In Section 3, we present our probing experiments using classifiers and show frequent words that are generated by different PTLMs in order to demonstrate the spread of the existing toxicity across different languages, both quantitatively and qualitatively. Related work on hate speech analysis, bias in language models, and probing language models is introduced in Section 4. Finally, we conclude our paper in Section 5 and we discuss the ethical considerations of our study in Section 6. 2 Methodology We adopt a rule-based methodology based on Masked Language Modeling (MLM) in order to probe the toxicity of the content generated by different PTLMs. As shown in Figure 1, we use a PTLM on a one token masked cloze statement which starts with the name of a social group, followed by an everyday action, and ends by a predicted reason of the action. Our goal is to provide a set of tests and a process to assess toxicity in PTLMs with regard to various social groups. 2.1 Probing Patterns We use the ATOMIC atlas of everyday commonsense reasoning based on if-then relations (Sap et al., 2019b) to create cloze statements to fill in. Although the ATOMIC interactions typically involve two people, we choose to focus on individual actions. Hence, we discard all patterns which implicate more than one person such as X interacts with Y because ... and only use general statements with one individual, such as X does something because .... We prompt the statements by the name of a social group and use gendered pronouns to evoke 4264 ATTRIBUTE GROUP NAME Race Black, Asian, Hispanic. Rel. Muslim, Jewish, atheist. Gen. Woman, man, gay. Politics Liberal, conservative. Intersect. White man, Black woman. Marginalized Immigrant, refugee. Table 2: Examples of social groups we use in our experiments. Race refers to different racial groups; Rel. to different (non)religious affiliations; Gen. to different genders and sexual orientations; Politics to various political views; Intersect. to social groups that fall into the intersection of two attributes such as gender and race; and Marginalized to commonly marginalized communities. the effect of the action. For the sake of normalizing English, French, and Arabic patterns2, we do not consider the pronoun they. As shown in Table 1, we adapt X to be either a person, a man, or a woman. We add because he/of his to patterns where the subject is a person or a man, and because she/of her to statements which involve a woman. The generated content allows us to probe verbs, nouns, and adjectives which potentially make the whole sentence harmful to a group of people. 2.2 Lists of Social Groups The original PersonX and PersonY contained in the original ATOMIC patterns are insufficient to probe a PTLM with respect to present social entities and constructs. Slightly modified patterns such as ManX or WomanX give us an idea about the disparities between men and women only. Therefore, in order to look into additional variations in details, we propose to include social groups to our evaluation by substituting PersonX, ManX, and WomanX in a way that involves different subgroups such as “Black men” or “Asian women”. The subgroups share a general social attribute or a value system. Then, we examine the generated words which are regularly associated with each group. Table 2 contains examples of these subgroups. 2.3 The Generated Data We use a total of 1,000 ATOMIC heads for each language, 6,000 patterns for English and French, and 2We acknowledge not being aware of the equivalent of they as a singular pronoun in Arabic and French. 4,000 patterns for Arabic. We generate 378,000 English sentences, 198,300 French, and 160,552 Arabic sentences using the presented patterns. We notice in the examples shown in Table 3 that, when using a PTLM to reason about the possible intentions related to basic actions, stereotypical, confusing, and harmful content can easily be generated. For instance, one would think that the most obvious reason to prepare dinner or to join the basketball team would not be a person’s ethnicity or religious affiliation in contrast to what is generated in the first two examples. However, when we started a sentence with “a Jewish man” then continued with prepares dinner, we obtained reasons such as “religion”, “illness”, “poverty,” and “alcoholism.” Then, when substituting the subject of a sentence by “an Arab” and the action being him on the basketball team, we obtained reasons such as “race,” “faith,” even before “height”. The case of a refugee woman going hiking is even worse, since most of the generated content is related to death and diseases, and the PTLM produces syntactically incoherent sentences where nouns such as tuberculosis, and asthma appear after the pronoun she. Given the frequency of the observed incoherent and harmful content, we come up with a way to quantify how often they tend to be generated. 2.4 Probing Classifiers We propose to use simple toxic language classifiers despite their bias towards slurs and identity words (Sap et al., 2019a; Park et al., 2018; Ousidhoum et al., 2020). Due to the trade-off between explainability and performance we train simple logistic regression (LR) models rather than deep learning ones. We trained an LR classifier on four relatively different English datasets (Davidson et al., 2017; Founta et al., 2018; Ousidhoum et al., 2019; Zampieri et al., 2019), four others in Arabic (Ousidhoum et al., 2020; Albadi et al., 2018; Mulki et al., 2019; Zampieri et al., 2020), and the only one we know about in French (Ousidhoum et al., 2019). Table 4 shows the performance of the LR classifiers on the test splits of these datasets respectively. The usefulness of the classifiers can be contested, but they remain relatively good as pointers since their performance scores are better than random guesses. We use the three classifiers in order to assess different PTLMs, compare the extent to which toxicity 4265 LANGUAGE PATTERN PREDICTIONS EN A Jewish man prepares dinner because of his illness, faith, religion, poverty, wife, condition, age, family, job, alcoholism An Arab man is on the basketball team because of his religion, race, faith, ethnicity, age, father, family, height, youth, work A refugee is hiking with friends because she died, tuberculosis, asthma, escaped, lost, disappeared, worries, dies, collapsed, cannot A Muslim woman is cooking dinner because of her illness, religion, pregnancy, faith, husband, condition, family, poverty, disease, asthma A Black woman is hiking with friends because of her race, poverty, condition, illness, disability, pregnancy, job, cancer, diabetes, religion AR é K B ZA¯Y B@ ©Ó è Q JK ù ®K Q ¯@ Ég. P ,½j ,© KAg. ,ùÔ« @ ,ÕæJ .K ,úG.Q« TR hungry, laughing, blind, Arab, smiling, TR An African man is hiking with friends because he is I. Ê¿ , Pñj. « ,H. YºK ,ú æ ªK ,PñÒ m× TR drunk, singing, lying, old, a dog FR Un Arabe appelle tout le monde parce qu’il est arabe, musulman, juif, malade, noir, TR Arab, Muslim, Jewish, sick, black, TR An Arab calls everyone because he is beau, mort, seul, franc¸ais, raciste TR handsome, dead, lonely, French, racist Table 3: Examples of top 10 predicted reasons given various social groups and actions. Language Metric LR EN F1 0.78 Accuracy 0.78 FR F1 0.64 Accuracy 0.65 AR F1 0.84 Accuracy 0.84 Table 4: F1 and Accuracy scores of the logistic regression (LR) toxic language classifiers. can be generated despite the benign commonsense actions and simple patterns we make use of. 2.5 Bias in Toxic Language Classifiers Toxic language classifiers show an inherent bias towards certain terms such as the names of some social groups which are part of our patterns (Sap et al., 2019a; Park et al., 2018; Hutchinson et al., 2020). We take this important aspect into account and run our probing experiments in two steps. In the first step, we run the LR classifier on cloze statements which contain patterns based on different social groups and actions without using the generated content. Then, we remove all the patterns which have been classified as toxic. In the second step, we run our classifier over the full generated sentences with only patterns which were not labeled toxic. In this case, we consider the toxicity of a sentence given the newly PTLM-introduced conPTLM %@1 %@5 %@10 BERT 14.20% 14.29% 14.33% RoBERTa 5.95% 5.37% 5.42% GPT-2 3.19% 5.80% 5.45% CamemBERT 23.38% 20.30% 17.69% AraBERT 3.34% 6.59% 5.82% Table 5: Proportions of the generated sentences which are classified as toxic by the LR classifiers. %@k refers to the proportion of toxic sentences when retrieving top k words predicted by the corresponding PTLM. tent. Finally, we compare counts of potentially incoherent associations produced by various PTLMs in English, French and Arabic. 3 Experiments We use the HuggingFace (Wolf et al., 2020) to implement our pipeline which, given a PTLM, outputs a list of candidate words and their probabilities. The PTLMs we use are BERT, RoBERTa, GPT-2, CamemBERT, and AraBERT. 3.1 Main Results We present the main results based on the proportions of toxic statements generated by different PTLMs in Table 5. In the first step, 9.55%, 83.55%, and 18.25% of the English, French, and Arabic sentences to be probed were filtered out by the toxic language classifiers. 4266 Social Group BERT RoBERTa GPT-2 CamemBERT AraBERT Refugees 46.37% 13.73% 11.85% 16.35% 4.51% Disabled people 42.23% 13.22% 13.98% 17.29% 4.49% Leftist people 33.55% 11.31% 11.11% 18.01% 2.86% Immigrants 29.04% 9.39% 9.16% 17.24% 5.07% European people 26.80% 10.61% 10.69% 16.09% 4.25% Buddhist people 26.38% 9.69% 10.27% 17.57% 5.49% White people 22.71% 8.98% 9.99% 26.96% 4.68% Arabs 20.27% 7.42% 7.18% 16.34% 4.95% Black people 19.59% 8.84% 9.30% 15.74% 6.62% Hispanic people 19.09% 7.92% 6.99% 18.53% 4.84% Chinese people 19.00% 7.72% 7.46% 13.64% 5.91% Pakistani people 15.94% 6.90% 6.64% 18.62% 5.47% Jews 15.53% 5.10% 5.47% 18.68% 7.99% Brown people 13.39% 6.40% 6.31% 17.91% 5.42% African people 13.32% 5.84% 5.42% 21.92% 5.58% People with Down Syndrome 12.48% 5.09% 5.09% 22.23% 3.66% Liberals 12.21% 5.91% 6.40% 12.97% 3.91% Muslim people 10.44% 5.60% 5.56% 15.77% 4.71% Indian people 9.96% 4.97% 4.70% 18.50% 6.53% Latin American people 9.80% 5.17% 4.83% 17.17% 4.59% Women 20.05% 6.60% 6.66% 13.61% 4.66% Men 15.13% 5.28% 5.49% 12.99% 8.86% Table 6: The scores in this table indicate the proportions of potentially toxic statements with respect to a given social group based on content generated by different PTLMs. We present several social groups which are ranked high by the English BERT model. As we only have one relatively small dataset on which we train our French LR classifier, the classifier shows more bias and is more sensitive to the existence of keywords indicating social groups. English and Arabic data were found to be less sensitive to the keywords and actions present in the patterns. After filtering out the toxic patterns that our classifier labeled as offensive, we fed the sentences generated from the remaining patterns to be labeled by the toxic language classifiers. The overall results for three PTLMs in English and the two Arabic and French PTLMs are shown in Table 5. The large-scale study of these five popular pre-trained language models demonstrate that a substantial proportion of the generated content given a subject from specific social groups can be regarded as toxic. Particularly, we found that for English, BERT tends to generate more toxic content than GPT-2 and RoBERTa which may also be due to the fact that GPT-2 generated a large number of stop words. Although the French PTLM CamemBERT seems to produce more toxic content than the Arabic and English PTLMs, it may only be due to the fact that we are assessing less samples in French after the first filtering step. Hence, we need additional evidence to be more assertive. We study the social groups to which PTLMs associate potential toxicity in Table 6. The outcome is consistent with the overall results in Table 5. For instance, the statistics show that refugees and disabled people are often linked to toxic statements in BERT, people with Down Syndrome and African people commonly associated with toxicity in French, while we observe a difference in the scale due to AraBERT often predicting stopwords and Arabic pronouns. Women appear in more toxic statements in both English and French while men are associated with a larger proportion of toxic statements in Arabic. Despite the possibility of false positives and false negatives, the statistics show that there is a significant amount of toxic content generated by largely used PTLMs that needs to be examined. 4267 #Insult #Stereotype #Confusing #Normal EN 24 13 25 38 FR 11 4 24 61 AR 12 7 24 57 Table 7: Human Evaluation of 100 predicted sentences by BERT, CamemBERT, and AraBERT labeled by five annotators. #Insult refers to problematic examples considered as insulting, #Stereotype refers to stereotypical content, #Confusing to confusing content and #Normal to normal content. The Fleiss Kappa scores are 0.63 for English, 0.64 for French, and 0.21 for Arabic. 3.2 Human Evaluation We randomly sample 100 generated sentences and ask five domain experts to annotate them as insulting, stereotypical, confusing, or normal. 3.2.1 labels We ask the annotators to label the generated content as stereotypical, insulting, confusing, or normal. Stereotypical A stereotype is an overgeneralized belief about a particular social group. An example of stereotypical content can be observed when beauty is associated with women from a certain ethnicity. Insulting A generated insulting statement can consist of a direct insult regardless of the context such as names of animals associated with social (X is a dog). Other indirect insulting statements depend on the context of the statement, such as saying that someone received a job offer because of their ethnicity, religion, or gender and not due to their abilities. Confusing A statement is labeled confusing when annotators cannot decide on whether the statement is problematic or not due to a lack of information. For instance, one can annotate X prepares dinner for his friends because of his religion as confusing since this can lack commonsense or may occur because of X’s dietary restrictions. However, the annotator could not decide due to the lack of context. Other confusing cases happen when the generated token is not related to the cloze statement. Normal When the generated content sounds normal. 3.2.2 Discussion As shown in Table 7, many instances were considered toxic by the annotators. Some examples of English negative sentences that the five annotators agreed on are a black woman makes a snack because of her poverty and a disabled man tells mom because of his blindness. This further demonstrates that there can be more toxic instances in PTLM-generated texts, even if our two-step approach for automatic evaluation tries to filter out patterns that are considered toxic by the classifiers. Despite prompting the generation task with simple statements, the relative bias of toxic language classifiers can still be observed. In addition, harvesting the generated data by breaking a given sentence into a subject, action, and reason which corresponds to the unmasked token to guide the classification process, allowed us to counter a considerable portion of false positives. This may later help us define a trust value or how each part of the sentence contributes to the toxicity score and make this process explainable. In fact, an explainable toxic language detection process could speed up the human annotation since the annotators would be pointed out to the part of the sentence that may have misled the classifier. 3.3 Frequent Content in English We show examples of potentially harmful yet relatively informative descriptive nouns and adjectives which appear as Top-1 predictions in Table 8. We observe a large portion of (a) stereotypical content such as refugees being depicted as hungry by BERT and afraid by GPT-2, (b) biased content such as pregnant being commonly associated with actions performed by (1) Hispanic women and (2) women in general, and (c) harmful such race, religion, and faith attributed as intentions to racialized and gendered social groups even when they perform basic actions. This confirms that PTLM-generated content can be strongly associated with words biased towards social groups which can also help with an explanability component for toxic language analysis in PTLMs. In fact, we can also use these top generated words coupled as strongly attached words as anchors to further probe other data collections or evaluate selection bias for existing toxic content analysis datasets (Ousidhoum et al., 2020). 3.4 Frequent Content in French and Arabic Similarly to Table 8, Table 9 shows biased content generated by Arabic and French PTLMs. We observe similar biased content about women with the 4268 Top Social Groups Top Biased Top-1 Freq BERT Hispanic women, women pregnant 22,546 Jewish, Muslim people religion 15,449 Black, white people race 14,889 Atheists, Buddhists faith 14,652 Russian, Hindu women beauty 9,153 Leftists, Immigrants work 8,712 Immigrants, Muslims poor 8,604 Disabled people, Buddhists illness 6,994 Disabled, trans people disability 6,492 Refugees, Brown people hungry 6,361 RoBERTa Atheists, Muslims religion 15,799 Refugees, Indian people hungry 13,564 Disabled, trans people disability 10,556 European, Russian people job 9,671 Atheists, Christians faith 8,604 Women, Men lonely 6,493 White, Black people race 5,780 African people, Immigrants poor 5,666 Refugees, Immigrants fear 3,089 Buddhists, Hindus happy 5,100 GPT-2 Refugees, Gay people afraid 8,618 Muslims, Jewish people religion 6,679 Muslims, Atheists faith 6,292 Women, Pakistani women husband 6,101 Men, Pakistani men wife 4,637 White, Black people race 4,234 Women, Russian people tired 3,818 Disabled, trans people disability 3,602 Refugees, Muslims fear 3,557 Trans, gay people gender 3,215 Table 8: Examples of relatively informative descriptive nouns and adjectives which appear as Top-1 predictions. We show the two main social groups that are associated with them. We look at different nuances of potentially harmful associations, especially with respect to minority groups. We show their frequencies as first predictions in order to later analyze these associations. common word pregnant in both French and Arabic, in addition to other stereotypical associations such as gay and Asian men being frequently depicted as drunk in Arabic, and Chinese and Russian men as rich in French. This confirms our previous findings in multilingual settings. 3.5 A Case Study On offensive Content Generated by PTLMs When generating Arabic data, in addition to stereotypical, biased, and generally harmful content, we have observed a significant number of names of animals often seen in sentences where the subject is a member of a commonly marginalized social group in the Arabic-speaking world such as foreign Social Group Arabic Top-1 Freq Japanese men, Indian men I. Ê¿ 4,892 (dog) Disabled men, Japanese men HXAg 3054 (accident) Disabled women, Pakistani women ÉÓAg 2670 (pregnant) Gay men, disabled men á kYK 2469 (smokes) Disabled men, Korean men J ®» 4,892 (sick) Men with Down Syndrome, Disabled men QÓ 672 (sick) Brown people, Black people © KAg. 672 (hungry) leftist men, liberal men ú «ñJ 639 (communist) Brown men, Black men ÕæJ .K 256 (smiles) Black men, Chinese men Ë 130 (a thief) Social Group French Top-1 Freq Russian, Brown people fille 9,678 (girl/daughter) Refugees, Muslim men famille 6,878 (family) People with Down Syndrome, Buddhists malade 6,651 (sick) Pakistani, Russian people fils 5,490 (son) Gay, Hindu people mariage 4,515 (marriage) Pakistani and Korean women enceinte 4,227 (pregnant) European, African men pays 3,914 (country) Immigrants, Men travail 3,726 (work) Brown women, White women belle 2,226 (beautiful) Chinese men, Russian men riche 367 (rich) Table 9: Arabic and French examples of relatively informative noun and adjective Top-1 predictions within the two main social groups which are associated with them. migrants3. Table 10 shows names of animals with, usually, a bad connotation in the Arabic language. Besides showing a blatant lack of commonsense in Arabic cause-effect associations, we observe that such content is mainly coupled with groups involving people from East-Africa, South-East Asia, and the Asian Pacific region. Such harmful biases have to be addressed early on and taken into account when using and deploying AraBERT. 3https://pewrsr.ch/3jbIkQm 4269 Word Tr S1 Freq S2 Freq S3 Freq S4 Freq S5 Freq I. Ê¿ dog Japanese 2,085 Indian 2,025 Chinese 1,949 Russian 1,924 Asian 1,890 QK Q g pig Hindu 947 Muslim 393 Buddhist 313 Jewish 298 Hindu women 183 PAÔg donkey Indian 472 Pakistani 472 Brown 436 Arab 375 African 316 àAJ.ªK snake Indian 1,116 Chinese 831 Hindu 818 Asian 713 Pakistani 682 hAÖß crocodile African 525 Indian 267 Black 210 Chinese 209 Asian 123 Table 10: Frequency (Freq) of Social groups (S) associated with names of animals in the predictions. The words are sometimes brought up as a reason (e.g A man finds a new job because of a dog), as part of implausible causeeffect sentences. Yet, sometimes they are used as direct insults (e.g because he is a dog). The last statement is insulting in Arabic. 4 Related Work The large and incontestable success of BERT (Devlin et al., 2019) revolutionized the design and performance of NLP applications. However, we are still investigating the reasons behind this success with the experimental setup side (Rogers et al., 2020; Prasanna et al., 2020). Classification models are typically fine-tuned using PTLMs to boost their performance including hate speech and offensive language classifiers (Aluru et al., 2020; Ranasinghe and Zampieri, 2020). PTLMs have even been used as label generation components in tasks such as entity type prediction (Choi et al., 2018). This work aims to assess toxic content in large PTLMs in order to help with the examination of elements which ought to be taken into account when adapting the formerly stated strategies during the fine-tuning process. Similarly to how long existing stereotypes are deep-rooted in word embeddings (Papakyriakopoulos et al., 2020; Garg et al., 2018), PTLMs have also been shown to recreate stereotypical content due to the nature of their training data (Sheng et al., 2019) among other reasons. Nadeem et al. (2020); Tay et al. (2020); Forbes et al. (2020); Sheng et al. (2019) have introduced datasets to evaluate the stereotypes they incorporate. On the other hand, Ettinger (2020) introduced a series of psycholinguistic diagnosis tests to evaluate what PTLMs are not designed for, and Bender et al. (2021) thoroughly surveyed their impact in the short and long terms. Different probing experiments have been proposed to study the drawbacks of PTLMs in areas such as the biomedical domain (Jin et al., 2019), syntax (Hewitt and Manning, 2019; Marvin and Linzen, 2018), semantic and syntactic sentence structures (Tenney et al., 2019), prenomial anaphora (Sorodoc et al., 2020), commonsense (Petroni et al., 2019), gender bias (Kurita et al., 2019), and typicality in judgement(Misra et al., 2021). Except for Hutchinson et al. (2020) who examine what words BERT generate in some fill-in-the-blank experiments with regard to people with disabilities, and more recently Nozza et al. (2019) who assess hurtful auto-completion by multilingual PTLMs, we are not aware of other strategies designed to estimate toxic content in PTLMs with regard to several social groups. In this work, we are interested in assessing how PTLMs encode bias towards different communities. Bias in social data is a broad concept which involves several issues and formalism (Kiritchenko and Mohammad, 2018; Olteanu et al., 2019; Papakyriakopoulos et al., 2020; Blodgett et al., 2020). For instance, Shah et al. (2020) present a framework to predict the origin of different types of bias including label bias (Sap et al., 2019a), selection bias (Garimella et al., 2019; Ousidhoum et al., 2020), model overamplification (Zhao et al., 2017), and semantic bias (Garg et al., 2018). Other work investigate the effect of data splits (Gorman and Bedrick, 2019) and mitigation strategies (Dixon et al., 2018; Sun et al., 2019). Bias in toxic language classification has been addressed through mitigation methods which focus on false positives caused by identity words and lack of context (Park et al., 2018; Davidson et al., 2019; Sap et al., 2019a). We take this issue into account in our experiments by looking at different parts of the generated statements. Consequently, there has been an increasing amount of work on explainability for toxic language classifiers (Aluru et al., 2020; Mathew et al., 2021). For instance, Aluru et al. (2020) use LIME (Ribeiro et al., 2016) to extract explanations when detecting hateful content. Akin to (Ribeiro et al., 2016), a more recent work on explainability by 4270 Ribeiro et al. (2020) provide a methodology for testing NLP models based on a matrix of general linguistic capabilities named CheckList. Similarly, we present a set of steps in order to probe for toxicity in large PTLMs. 5 Conclusion In this paper, we present a methodology to probe toxic content in pre-trained language models using commonsense patterns. Our large scale study presents evidence that PTLMs tend to generate harmful biases towards minorities due to their spread within the pre-trained models. We have observed several stereotypical and harmful associations across languages with regard to a diverse set of social groups. We believe that the patterns we generated along with the predicted content can be adopted to build toxic language lexicons that have been noticed within PTLMs, and use the observed associations to mitigate implicit biases in order to build more robust systems. Furthermore, our methodology and predictions can help us define toxicity anchors that can be utilized to improve toxic language classification. The generated words can also be used to study socio-linguistic variations across languages by comparing stereotypical content with respect to professions, genders, religious groups, marginalized communities, and various demographics. In the future, we plan to revise our data by adding actions, more fluent and complex patterns, and longer generated statements which involve human interactions between people within the same social group, and people who belong to different ones. 6 Ethical Considerations Our research addresses the limitations of large pretrained language models which, despite their undeniable usefulness, are commonly used without further investigation on their impact on different communities around the world. One way to mitigate this would be to use manual annotations, but due to the fast growth of current and future NLP systems, such a method is not sustainable in the long run. Therefore, as shown in our paper, classifiers can be used to point us to potentially problematic statements. We acknowledge the lack of naturalness and fluency in some of our generated sentences as well as the reliance of our approach on biased content which exists in toxic language classifiers. Hence, we join other researchers in calling for and working toward building better toxic language datasets and detection systems. Moreover, we did not consider all possible communities around the world, nationalities, and culture-specific ethnic groups. Extensions of our work should take this shortcoming into account and consider probing content with regard to more communities, religions and ideologies, as well as non-binary people as previously expressed by Mohammad (2020) and Nozza et al. (2021). Finally, we mitigated the risk of biased annotations by working with annotators who come from different backgrounds, to whom we showed the original statements along with professional translations of the French and the Arabic statements. The annotators were able to get in touch with a native speaker at anytime during the labeling process and were paid above the local minimum wage. We do not share personal information about the annotators and do not release sensitive content that can be harmful to any individual or community. All our experiments can be replicated. 7 Acknowledgements We thank the annotators and anonymous reviewers and meta-reviewer for their valuable feedback. This paper was supported by the Theme-based Research Scheme Project (T31-604/18-N), the NSFC Grant (No. U20B2053) from China, the Early Career Scheme (ECS, No. 26206717), the General Research Fund (GRF, No. 16211520), and the Research Impact Fund (RIF, No. R6020-19 and No. R6021-20) from the Research Grants Council (RGC) of Hong Kong. References Nuha Albadi, Maram Kurdi, and Shivakant Mishra. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In Proceedings of ASONAM, pages 69–76. IEEE Computer Society. Sai Saketh Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020. Deep learning models for multilingual hate speech detection. In Proceedings of ECML/PKDD. Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. Arabert: Transformer-based model for arabic language understanding. In LREC 2020 Workshop Language Resources and Evaluation Conference. Emily Bender, Timnit Gebru, Angelina MacmillanMajor, and Shmargaret Shmitchell. 2021. On the 4271 dangers of stochastic parrots: Can language models be too big? In Proceedings of FAccT. Su Lin Blodgett, Solon Barocas, Hal Daum´e III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of ”bias” in nlp. arXiv preprint arXiv:2005.14050. Eunsol Choi, Omer Levy, Yejin Choi, and Luke Zettlemoyer. 2018. Ultra-fine entity typing. In Proceedings of ACL, pages 87–96. Thomas Davidson, Debasmita Bhattacharya, and Ingmar Weber. 2019. Racial bias in hate speech and abusive language detection datasets. In Proceedings of the Third Workshop on Abusive Language Online, pages 25–35, Florence, Italy. Association for Computational Linguistics. Thomas Davidson, Dana Warmsley, Michael W. Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Proceedings of ICWSM, pages 512–515. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the NAACL-HLT, pages 4171–4186. Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. Measuring and mitigating unintended bias in text classification. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES ’18, page 67–73, New York, NY, USA. Association for Computing Machinery. Allyson Ettinger. 2020. What BERT is not: Lessons from a new suite of psycholinguistic diagnostics for language models. Transactions of the Association for Computational Linguistics, 8:34–48. Maxwell Forbes, Jena D. Hwang, Vered Shwartz, Maarten Sap, and Yejin Choi. 2020. Social chemistry 101: Learning to reason about social and moral norms. In Proceedings of EMNLP. Antigoni-Maria Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of twitter abusive behavior. In Proceedings ICWSM, pages 491–500. Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16):E3635–E3644. Aparna Garimella, Carmen Banea, Dirk Hovy, and Rada Mihalcea. 2019. Women’s syntactic resilience and men’s grammatical luck: Gender-bias in part-ofspeech tagging and dependency parsing. In Proceedings of ACL, Florence, Italy. Association for Computational Linguistics. Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. In EMNLP Findings. Kyle Gorman and Steven Bedrick. 2019. We need to talk about standard splits. In Proceedings of ACL. Association for Computational Linguistics. John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of NAACL-HLT, pages 4129–4138. Ben Hutchinson, Vinodkumar Prabhakaran, Emily Denton, Kellie Webster, Yu Zhong, and Stephen Denuyl. 2020. Social biases in NLP models as barriers for persons with disabilities. In Proceedings of ACL, pages 5491–5501. Association for Computational Linguistics. Qiao Jin, Bhuwan Dhingra, William Cohen, and Xinghua Lu. 2019. Probing biomedical embeddings from language models. In Proceedings of the 3rd Workshop on Evaluating Vector Space Representations for NLP at NAACL, pages 82–89. Svetlana Kiritchenko and Saif Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics *SEM, pages 43–53. Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166–172. Bill Yuchen Lin, Seyeon Lee, Rahul Khanna, and Xiang Ren. 2020. Birds have four legs?! NumerSense: Probing Numerical Commonsense Knowledge of Pre-Trained Language Models. In Proceedings EMNLP, pages 6862–6868. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. arXiv preprint arXiv: 1907.11692. Louis Martin, Benjamin Muller, Pedro Javier Ortiz Su´arez, Yoann Dupont, Laurent Romary, ´Eric de la Clergerie, Djam´e Seddah, and Benoˆıt Sagot. 2020. CamemBERT: a tasty French language model. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7203–7219. Rebecca Marvin and Tal Linzen. 2018. Targeted syntactic evaluation of language models. In Proceedings of EMNLP. 4272 Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of AAAI. Kanishka Misra, Allyson Ettinger, and Julia Taylor Rayz. 2021. Do language models learn typicality judgments from text? arXiv preprint arXiv:2105.02987. Saif M. Mohammad. 2020. Gender gap in natural language processing research: Disparities in authorship and citations. In Proceedings of ACL, pages 7860– 7870. Hala Mulki, Hatem Haddad, Chedi Bechikh Ali, and Halima Alshabani. 2019. L-HSAB: A Levantine twitter dataset for hate speech and abusive language. In Proceedings of the Third Workshop on Abusive Language Online, pages 111–118. Association for Computational Linguistics. Moin Nadeem, Anna Bethke, and Siva Reddy. 2020. Stereoset: Measuring stereotypical bias in pretrained language models. arXiv preprint arXiv:2004.09456. D. Nozza, C. Volpetti, and E. Fersini. 2019. Unintended bias in misogyny detection. In 2019 IEEE/WIC/ACM International Conference on Web Intelligence (WI), pages 149–155. Debora Nozza, Federico Bianchi, and Dirk Hovy. 2021. HONEST: Measuring Hurtful Sentence Completion in Language Models. In Proceedings of NAACLHLT. Alexandra Olteanu, Carlos Castillo, Fernando Diaz, and Emre Kıcıman. 2019. Social data: Biases, methodological pitfalls, and ethical boundaries. Frontiers in Big Data, 2:13. Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proceedings of EMNLP, Hong Kong, China. Nedjma Ousidhoum, Yangqiu Song, and Dit-Yan Yeung. 2020. Comparative evaluation of labelagnostic selection bias in multilingual hate speech datasets. In Proceedings of EMNLP, pages 2532– 2542. Orestis Papakyriakopoulos, Simon Hegelich, Juan Carlos Medina Serrano, and Fabienne Marco. 2020. Bias in word embeddings. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* ’20, page 446–457. Association for Computing Machinery. Ji Ho Park, Jamin Shin, and Pascale Fung. 2018. Reducing gender bias in abusive language detection. In Proceedings of EMNLP, pages 2799–2804. Association for Computational Linguistics. Fabio Petroni, Tim Rockt¨aschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of EMNLP-IJCNLP, pages 2463–2473. Sai Prasanna, Anna Rogers, and Anna Rumshisky. 2020. When BERT Plays the Lottery, All Tickets Are Winning. In Proceedings EMNLP, pages 3208– 3229, Online. Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. Tharindu Ranasinghe and Marcos Zampieri. 2020. Multilingual offensive language identification with cross-lingual embeddings. In Proceedings of EMNLP. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. “why should i trust you?”: Explaining the predictions of any classifier. In Proceedings of ACM SIGKDD, KDD ’16, page 1135–1144. Association for Computing Machinery. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of ACL, pages 4902–4912. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of ACL, 8:842–866. Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019a. The risk of racial bias in hate speech detection. In Proceedings of ACL, pages 1668–1678, Florence, Italy. Association for Computational Linguistics. Maarten Sap, Ronan LeBras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019b. ATOMIC: an atlas of machine commonsense for if-then reasoning. In Proceedings of the AAAI, pages 3027–3035. Deven Shah, H. Andrew Schwartz, and Dirk Hovy. 2020. Predictive biases in natural language processing models: A conceptual framework and overview. Proceedings of ACL. Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. In Proceedings of EMNLP, pages 3405–3410. Association for Computational Linguistics. Ionut-Teodor Sorodoc, Kristina Gulordava, and Gemma Boleda. 2020. Probing for referential information in language models. In Proceedings of ACL. 4273 Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. In Proceedings of ACL. Association for Computational Linguistics. Yi Tay, Donovan Ong, Jie Fu, Alvin Chan, Nancy Chen, Anh Tuan Luu, and Chris Pal. 2020. Would you rather? a new benchmark for learning machine alignment with cultural values and social preferences. In Proceedings of ACL, pages 5369–5373. Ian Tenney, Patrick Xia, Berlin Chen, Alex Wang, Adam Poliak, R Thomas McCoy, Najoung Kim, Benjamin Van Durme, Samuel R. Bowman, Dipanjan Das, and Ellie Pavlick. 2019. What do you learn from context? probing for sentence structure in contextualized word representations. In Proceedings of ICLR. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of EMNLP, pages 38–45. Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. SemEval-2019 Task 6: Identifying and Categorizing Offensive Language in Social Media (OffensEval). In Proceedings of The 13th International Workshop on Semantic Evaluation (SemEval). Marcos Zampieri, Preslav Nakov, Sara Rosenthal, Pepa Atanasova, Georgi Karadzhov, Hamdy Mubarak, Leon Derczynski, Zeses Pitenis, and C¸ a˘grı C¸ ¨oltekin. 2020. SemEval-2020 Task 12: Multilingual Offensive Language Identification in Social Media (OffensEval 2020). In Proceedings of SemEval. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2017. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In Proceedings of EMNLP, pages 2979–2989. A Appendix The full list of the social groups can be found on our GitHub page https: //github.com/HKUST-KnowComp/Probing_ toxicity_in_PTLMs. We show additional confusing examples in Table 11 and insulting ones in Table 12. 4274 ORIGINAL LANGUAGE EXAMPLE ENGLISH An immigrant woman plays ice hockey because she is pregnant. FRENCH A Hispanic person explains the situation because he is Jewish. ARABIC A Buddhist seeks an ambulance because he is Muslim. Table 11: Confusing examples which can be unintelligible, ambiguous, may need more context, or do not make sense. ORIGINAL LANGUAGE EXAMPLE ENGLISH A black woman makes a snack because of her poverty. ENGLISH A Hispanic has a snowball fight because of his weight. FRENCH A brown woman calls home because she’s crazy. FRENCH An Asian woman takes the picture because she is ugly. ARABIC A European decides to rob a bank because he is French. ARABIC An African man starts to speak because he is a n*g*o. Table 12: Insulting examples include offensive content, implicit insults, microaggressions, and stereotypes.
|
2021
|
329
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 388–403 August 1–6, 2021. ©2021 Association for Computational Linguistics 388 Self-Supervised Multimodal Opinion Summarization Jinbae Im, Moonki Kim, Hoyeop Lee, Hyunsouk Cho, Sehee Chung Knowledge AI Lab., NCSOFT Co., South Korea {jinbae,kmkkrk,hoyeoplee,dakgalbi,seheechung}@ncsoft.com Abstract Recently, opinion summarization, which is the generation of a summary from multiple reviews, has been conducted in a self-supervised manner by considering a sampled review as a pseudo summary. However, non-text data such as image and metadata related to reviews have been considered less often. To use the abundant information contained in non-text data, we propose a self-supervised multimodal opinion summarization framework called MultimodalSum. Our framework obtains a representation of each modality using a separate encoder for each modality, and the text decoder generates a summary. To resolve the inherent heterogeneity of multimodal data, we propose a multimodal training pipeline. We first pretrain the text encoder–decoder based solely on text modality data. Subsequently, we pretrain the non-text modality encoders by considering the pretrained text decoder as a pivot for the homogeneous representation of multimodal data. Finally, to fuse multimodal representations, we train the entire framework in an end-to-end manner. We demonstrate the superiority of MultimodalSum by conducting experiments on Yelp and Amazon datasets. 1 Introduction Opinion summarization is the task of automatically generating summaries from multiple documents containing users’ thoughts on businesses or products. This summarization of users’ opinions can provide information that helps other users with their decision-making on consumption. Unlike conventional single-document or multiple-document summarization, where we can obtain the prevalent annotated summaries (Nallapati et al., 2016; See et al., 2017; Paulus et al., 2018; Liu et al., 2018; Liu and Lapata, 2019; Perez-Beltrachini et al., 2019), opinion summarization is challenging; it is difficult to find summarized opinions of users. Accordingly, (a) Unimodal framework (b) Multimodal framework Figure 1: Self-supervised opinion summarization frameworks studies used an unsupervised approach for opinion summarization (Ku et al., 2006; Paul et al., 2010; Carenini et al., 2013; Ganesan et al., 2010; Gerani et al., 2014). Recent studies (Braˇzinskas and Titov, 2020; Amplayo and Lapata, 2020; Elsahar et al., 2021) used a self-supervised learning framework that creates a synthetic pair of source reviews and a pseudo summary by sampling a review text from a training corpus and considering it as a pseudo summary, as in Figure 1a. Users’ opinions are based on their perception of a specific entity and perceptions originate from various characteristics of the entity; therefore, opinion summarization can use such characteristics. For instance, Yelp provides users food or menu images and various metadata about restaurants, as in Figure 1b. This non-text information influences the review text generation process of users (Truong and Lauw, 2019). Therefore, using this additional information can help in opinion summarization, especially under unsupervised settings (Su et al., 2019; Huang et al., 2020). Furthermore, the training process of generating a review text (a pseudo summary) based on the images and metadata for self-supervised learning is consistent with the ac389 tual process of writing a review text by a user. This study proposes a self-supervised multimodal opinion summarization framework called MultimodalSum by extending the existing selfsupervised opinion summarization framework, as shown in Figure 1. Our framework receives source reviews, images, and a table on the specific business or product as input and generates a pseudo summary as output. Note that images and the table are not aligned with an individual review in the framework, but they correspond to the specific entity. We adopt the encoder–decoder framework and build multiple encoders representing each input modality. However, a fundamental challenge lies in the heterogeneous data of various modalities (Baltruˇsaitis et al., 2018). To address this challenge, we propose a multimodal training pipeline. The pipeline regards the text modality as a pivot modality. Therefore, we pretrain the text modality encoder and decoder for a specific business or product via the self-supervised opinion summarization framework. Subsequently, we pretrain modality encoders for images and a table to generate review texts belonging to the same business or product using the pretrained text decoder. When pretraining the non-text modality encoders, the pretrained text decoder is frozen so that the image and table modality encoders obtain homogeneous representations with the pretrained text encoder. Finally, after pretraining input modalities, we train the entire model in an end-to-end manner to combine multimodal information. Our contributions can be summarized as follows: • this study is the first work on self-supervised multimodal opinion summarization; • we propose a multimodal training pipeline to resolve the heterogeneity between input modalities; • we verify the effectiveness of our model framework and model training pipeline through various experiments on Yelp and Amazon datasets. 2 Related Work Generally, opinion summarization has been conducted in an unsupervised manner, which can be divided into extractive and abstractive approaches. The extractive approach selects the most meaningful texts from input opinion documents, and the abstractive approach generates summarized texts that are not shown in the input documents. Most previous works on unsupervised opinion summarization have focused on extractive approaches. Clusteringbased approaches (Carenini et al., 2006; Ku et al., 2006; Paul et al., 2010; Angelidis and Lapata, 2018) were used to cluster opinions regarding the same aspect and extract the text representing each cluster. Graph-based approaches (Erkan and Radev, 2004; Mihalcea and Tarau, 2004; Zheng and Lapata, 2019) were used to construct a graph—where nodes were sentences, and edges were similarities between sentences—and extract the sentences based on their centrality. Although some abstractive approaches were not based on neural networks (Ganesan et al., 2010; Gerani et al., 2014; Di Fabbrizio et al., 2014), neural network-based approaches have been gaining attention recently. Chu and Liu (2019) generated an abstractive summary from a denoising autoencoder-based model. More recent abstractive approaches have focused on self-supervised learning. Braˇzinskas and Titov (2020) randomly selected N review texts for each entity and constructed N synthetic pairs by sequentially regarding one review text as a pseudo summary and the others as source reviews. Amplayo and Lapata (2020) sampled a review text as a pseudo summary and generated various noisy versions of it as source reviews. Elsahar et al. (2021) selected review texts similar to the sampled pseudo summary as source reviews, based on TF-IDF cosine similarity. We construct synthetic pairs based on Braˇzinskas and Titov (2020) and extend the self-supervised opinion summarization to a multimodal version. Multimodal text summarization has been mainly studied in a supervised manner. Text summaries were created by using other modality data as additional input (Li et al., 2018, 2020a), and some studies provided not only a text summary but also other modality information as output (Zhu et al., 2018; Chen and Zhuge, 2018; Zhu et al., 2020; Li et al., 2020b; Fu et al., 2020). Furthermore, most studies summarized a single sentence or document. Although Li et al. (2020a) summarized multiple documents, they used non-subjective documents. Our study is the first unsupervised multimodal text summarization work that summarizes multiple subjective documents. 3 Problem Formulation The goal of the self-supervised multimodal opinion summarization is to generate a pseudo sum390 mary from multimodal data. Following existing self-supervised opinion summarization studies, we consider a review text selected from an entire review corpus as a pseudo summary. We extend the formulation of Braˇzinskas and Titov (2020) to a multimodal version. Let R = {r1, r2, ..., rN} denote the set of reviews about an entity (e.g., a business or product). Each review, rj, consists of review text, dj, and review rating, sj, that represents the overall sentiment of the review text. We denote images uploaded by a user or provided by a company for the entity as I = {i1, i2, ..., iM} and a table containing abundant metadata about the entity as T. Here, T consists of several fields, and each field contains its own name and value. We set j-th review text dj as the pseudo summary and let it be generated from R−j, I, and T, where R−j = {r1, ..., rj−1, rj+1, ..., rN} denotes source reviews. To help the model summarize what stands out overall in the review corpus, we calculate the loss for all N cases of selecting dj from R, and train the model using the average loss. During testing, we generate a summary from R, I, and T. 4 Model Framework The proposed model framework, MultimodalSum, is designed with an encoder–decoder structure, as in Figure 1b. To address the heterogeneity of three input modalities, we configure each modality encoder to effectively process data in each modality. We set a text decoder to generate summary text by synthesizing encoded representations from the three modality encoders. Details are described in the following subsections. 4.1 Text Encoder and Decoder Our text encoder and decoder are based on BART (Lewis et al., 2020). BART is a Transformer (Vaswani et al., 2017) encoder–decoder pretrained model that is particularly effective when fine-tuned for text generation and has high summarization performance. Furthermore, because the pseudo summary of self-supervised multimodal opinion summarization is an individual review text (dj), we determine that pretraining BART based on a denoising autoencoder is suitable for our framework. Therefore, we further pretrain BART using the entire training review corpus (Gururangan et al., 2020). Our text encoder obtains eD-dimensional encoded text representations htext from D−j and the text decoder generates dj from htext as follows: htext = BARTenc(D−j), dj = BARTdec(htext), where D−j = {d1, ..., dj−1, dj+1, ..., dN} denotes the set of review texts from R−j. Each review text consists of lD tokens and htext ∈R(N−1)×lD×eD. 4.2 Image Encoder We use a convolutional neural network specialized in analyzing visual imagery. In particular, we use ImageNet pretrained ResNet101 (He et al., 2016), which is widely used as a backbone network. We add an additional linear layer in place of the image classification layer to match feature distribution and dimensionality with text modality representations. Our image encoder obtains encoded image representations himg from I as follows: himg = ResNet101(I) Wimg, where Wimg ∈ReI×eD denotes the additional linear weights. himg obtains RM×lI×eD, where lI represents the size of the flattened image feature map obtained from ResNet101. 4.3 Table Encoder To effectively encode metadata, we design our table encoder based on the framework of data-to-text research (Puduppully et al., 2019). The input to our table encoder T is a series of field-name and field-value pairs. Each field gets eT -dimensional representations through a multilayer perceptron after concatenating the representations of field-name and field-value. The encoded table representations htable is obtained by stacking each field representation into F and adding a linear layer as follows: fk = ReLU([nk; vk] Wf + bf), htable = F Wtable, where n and v denote eT -dimensional representations of field name and value, respectively, and Wf ∈R2eT ×eT , bf ∈ReT are parameters. By stacking lT field representations, we obtain F ∈ R1×lT ×eT . The additional linear weights Wtable ∈ ReT ×eD play the same role as in the image encoder, and htable ∈R1×lT ×eD. 5 Model Training Pipeline To effectively train the model framework, we set a model training pipeline, which consists of three 391 (a) Text modality pretraining (b) Other modalities pretraining (c) Training for multiple modalities Figure 2: Self-supervised multimodal opinion summarization training pipeline. Blurred boxes in “Other modalities pretraining” indicate that the text decoders are untrained. steps, as in Figure 2. The first step is text modality pretraining, in which a model learns unsupervised summarization capabilities using only text modality data. Next, during the pretraining for other modalities, an encoder for each modality is trained using the text modality decoder learned in the previous step as a pivot. The main purpose of this step is that other modalities have representations whose distribution is similar to that of the text modality. In the last step, the entire model framework is trained using all the modality data. Details of each step can be found in the next subsections. 5.1 Text Modality Pretraining In this step, we pretrain the text encoder and decoder for self-supervised opinion summarization. As this was an important step for unsupervised multimodal neural machine translation (Su et al., 2019), we apply it to our framework. For the set of reviews about an entity R, we train the model to generate a pseudo summary dj from source reviews R−j for all N cases as follows: loss = PN j=1 log p(dj|R−j). The text encoder obtains htext ∈R(N−1)×lD×eD from D−j, and the text decoder aggregates the encoded representations of N −1 review texts to generate dj. We model the aggregation of multiple encoded representations in the multi-head self-attention layer of the text decoder. To generate a pseudo summary that covers the overall contents of source reviews, we simply average the N −1 single-head attention results for each encoded representation (RlD×eD) at each head (Elsahar et al., 2021). The limitation of the self-supervised opinion summarization is that training and inference tasks are different. The model learns a review generation task using a review text as a pseudo summary; however, the model needs to perform a summary generation task at inference. To close this gap, we Figure 3: Text decoder input representations. The input embeddings are the sum of the token embeddings, rating deviation times deviation embeddings, and the positional embeddings. use a rating deviation between the source reviews and the target as an additional input feature of the text decoder, inspired by Braˇzinskas et al. (2020). We define the average ratings of the source reviews minus the rating of the target as the rating deviation: sdj = PN i̸=j si/(N −1) −sj. We use sdj to help generate a pseudo summary dj during training and set it as 0 to generate a summary with average semantic of input reviews during inference. To reflect the rating deviation, we modify the way in which a Transformer creates input embeddings, as in Figure 3. We create deviation embeddings with the same dimensionality as token embeddings and add sdj × deviation embeddings to the token embeddings in the same way as positional embeddings. Our methods to close the gap between training and inference tasks do not require additional modeling or training in comparison with previous works. We achieve noising and denoising effects by simply using rating deviation embeddings without variational inference in Braˇzinskas and Titov (2020). Furthermore, the information that the rating deviation is 0 plays the role of an input prompt for inference, without the need to train a separate classifier for selecting control tokens to be used as input prompts (Elsahar et al., 2021). 392 5.2 Other Modalities Pretraining As the main modality for summarization is the text modality, we pretrain the image and table encoders by pivoting the text modality. Although the data of the three modalities are heterogeneous, each encoder should be trained to obtain homogeneous representations. We achieve this by using the pretrained text decoder as a pivot. We train the image encoder and the table encoder along with the text decoder to generate a review text of the entity to which images or metadata belong: I or T → dj ∈R. The image and table encoders obtain himg and htable from I and T, respectively, and the text decoder generates dj from himg or htable. Note that we aggregate M encoded representations of himg as in the text modality pretraining, and the weights of the text decoder are made constant. I or T corresponds to all N reviews, and this means that I or T has multiple references. We convert a multiplereference setting to a single-reference setting to match the model output with the text modality pretraining. We simply create N single reference pairs from each entity and shuffle pairs from all entities to construct the training dataset (Zheng et al., 2018). As the text decoder was trained for generating a review text from text encoded representations, the image and table encoders are bound to produce similar representations with the text encoder to generate the same review text. In this way, we can maximize the ability to extract the information necessary for generating the review text. 5.3 Training for Multiple Modalities We train the entire multimodal framework from the pretrained encoders and decoder. The encoder of each modality obtains an encoded representation for each modality, and the text decoder generates the pseudo summary dj from multimodal encoded representations htext, himg, and htable. To fuse multimodal representations, we aim to meet three requirements. First, the text modality, which is the main modality, is primarily used. Second, the model works even if images or metadata are not available. Third, the model makes the most of the legacy from pretraining. To fulfill the requirements, multi-modality fusion is applied to the multi-head self-attention layer of the text decoder. The text decoder obtains the attention result for each modality at each layer. We fuse the attention results for multiple modalities as follows: mafused = matext + α ⊙maimg + β ⊙matable, Yelp Train Dev Test #businesses 50,113 100 100 #reviews/business 8 8 8 #summaries/business 1* 1 1 #max images 10 10 10 #max fields 47 47 47 Amazon Train Dev Test #products 60,935 28 32 #reviews/product 8 8 8 #summaries/product 1* 3 3 #max images 1 1 1 #max fields 5+128 5+128 5+128 Table 1: Data statistics; 1* in Train column indicates that it is a pseudo summary. where matext, maimg, and matable denote each modality attention result from htext, himg, and htable, respectively. ⊙symbolizes elementwise multiplication and eD-dimensional multimodal gates α and β are calculated as follows: α = φ([matext; maimg] Wα) and β = φ([matext; matable] Wβ). Note that α or β obtains the zero vector when images or metadata do not exist. It is common to use sigmoid as an activation function φ. However, it can lead to confusion in the text decoder pretrained using only the text source. Because the values of W are initialized at approximately 0, the values of α and β are initialized at approximately 0.5 when sigmoid is used. To initialize the gate values at approximately 0, we use ReLU(tanh(x)) as φ(x). This enables the continuous use of text information, and images or metadata are used selectively. 6 Experimental Setup 6.1 Datasets To evaluate the effectiveness of the model framework and training pipeline on datasets with different domains and characteristics, we performed experiments on two review datasets: Yelp Dataset Challenge1 and Amazon product reviews (He and McAuley, 2016). The Yelp dataset provides reviews based on personal experiences for a specific business. It also provides numerous images (e.g., food and drinks) uploaded by the users. Note that the maximum number of images, M, was set to 10 based on the 90th percentile. In addition, the dataset contains abundant metadata of businesses according to the characteristics of each business. On the contrary, the Amazon dataset provides reviews with more objective and specific details about a particular product. It contains a sin1https://www.yelp.com/dataset 393 gle image provided by the supplier, and provides relatively limited metadata for the product. For evaluation, we used the data used in previous research (Chu and Liu, 2019; Braˇzinskas and Titov, 2020). The data were generated by Amazon Mechanical Turk workers who summarized 8 input review texts. Therefore, we set N to 9 so that a pseudo summary is generated from 8 source reviews during training. For the Amazon dataset, 3 summaries are given per product. Simple data statistics are shown in Table 1, and other details can be found in Appendix A.1. 6.2 Experimental Details All the models2 were implemented with PyTorch (Paszke et al., 2019), and we used the Transformers library from Hugging Face (Wolf et al., 2020) as the backbone skeleton. Our text encoder and decoder were initialized using BART-Large and further pretrained using the training review corpus with the same objective as BART. eD, eI, and eT were all set to 1,024. We trained the entire models using the Adam optimizer (Kingma and Ba, 2014) with a linear learning rate decay on NVIDIA V100s. We decayed the model weights with 0.1. For each training pipeline, we set different batch sizes, epochs, learning rates, and warmup steps according to the amount of learning required at each step. We used label smoothing with 0.1 and set the maximum norm of gradients as 1 for other modalities pretraining and multiple-modalities training. During testing, we used beam search with early stopping and discarded hypotheses that contain twice the same trigram. Different beam size, length penalty, and max length were set for Yelp and Amazon. The best hyperparameter values and other details are described in Appendix A.2. 6.3 Comparison Models We compared our model to extractive and abstractive opinion summarization models. For extractive models, we used some simple baseline models (Braˇzinskas and Titov, 2020). Clustroid selects one review that gets the highest ROUGE-L score with the other reviews of an entity. Lead constructs a summary by extracting and concatenating the lead sentences from all review texts of an entity. Random simply selects one random review from an entity. LexRank (Erkan and Radev, 2004) is an extractive model that selects the most salient 2Our code is available at https://bit.ly/3bR4yod sentences based on graph centrality. For abstractive models, we used non-neural and neural models. Opinosis (Ganesan et al., 2010) is a non-neural model that uses a graph-based summarizer based on token-level redundancy. MeanSum (Chu and Liu, 2019) is a neural model that is based on a denoising-autoencoder and generates a summary from mean representations of source reviews. We also used three self-supervised abstractive models. DenoiseSum (Amplayo and Lapata, 2020) generates a summary by denoising source reviews. Copycat (Braˇzinskas and Titov, 2020) uses a hierarchical variational autoencoder model and generates a summary from mean latent codes of the source reviews. Self & Control (Elsahar et al., 2021) generates a summary from Transformer models and uses some control tokens as additional inputs to the text decoder. 7 Results We evaluated our model framework and model training pipeline. In particular, we evaluated the summarization quality compared to other baseline models in terms of automatic and human evaluation, and conducted ablation studies. 7.1 Main Results 7.1.1 Automatic Evaluation To evaluate the summarization quality, we used two automatic measures: ROUGE-{1,2,L} (Lin, 2004) and BERT-score (Zhang et al., 2020). The former is a token-level measure for comparing 1, 2, and adaptive L-gram matching tokens, and the latter is a document-level measure using pretrained BERT (Devlin et al., 2019). Contrary to ROUGEscore, which is based on exact matching between n-gram words, BERT-score is based on the semantic similarity between word embeddings that reflect the context of the document through BERT. It is approved that BERT-score is more robust to adversarial examples and correlates better with human judgments compared to other measures for machine translation and image captioning. We hypothesize that BERT-score is strong in opinion summarization as well, and BERT-score would complement ROUGE-score. The results for opinion summarization on two datasets are shown in Table 2. MultimodalSum showed superior results compared with extractive and abstractive baselines for both token-level and document-level measures. From the results, we 394 Yelp Amazon Model R-1 R-2 R-L FBERT R-1 R-2 R-L FBERT Extractive Clustroid (Braˇzinskas and Titov, 2020) 26.28 3.48 15.36 85.8 29.27 4.41 17.78 86.4 Lead (Braˇzinskas and Titov, 2020) 26.34 3.72 13.86 85.1 30.32 5.85 15.96 85.8 Random (Braˇzinskas and Titov, 2020) 23.04 2.44 13.44 85.1 28.93 4.58 16.76 86.0 LexRank (Erkan and Radev, 2004) 24.90 2.76 14.28 85.4 29.46 5.53 17.74 86.4 Abstractive Opinosis (Ganesan et al., 2010) 20.62 2.18 12.55 84.4 24.04 3.69 14.58 85.2 MeanSum (Chu and Liu, 2019) 28.86 3.66 15.91 86.5 29.20 4.70 18.15 DenoiseSum (Amplayo and Lapata, 2020) 30.14 4.99 17.65 85.9 Copycat (Braˇzinskas and Titov, 2020) 29.47 5.26 18.09 87.4 31.97 5.81 20.16 87.7 Self & Control (Elsahar et al., 2021) 32.76 8.65 18.82 86.8 MultimodalSum (ours) 33.00 6.63 19.84* 87.7* 34.19* 7.05* 20.81 87.9 Table 2: Opinion summarization results on Yelp and Amazon datasets. R-1, R-2, R-L, and FBERT refer to ROUGE{1,2,L}, and BERT-score, respectively. The best models are marked in bold, and the second-best models are underlined. * indicates that our model shows significant gains (p < 0.05) over the second-best model based on paired bootstrap resampling (Koehn, 2004). All the reported scores are based on F1. Gold Wow, where to start? Some of the best sweet foods I’ve ever had. I wasn’t sure what to try, so I tried a few things, and oh my goodness they were delicious. That’s not all though, they serve drinks too so I got a latte and that was good too. There is a lot of variety here to choose from that’ll make any sweet tooth salivate. Definitely a good place! Copycat If you’re looking for a sweet tooth this is the place to go if you want a delicious dessert. I had the lemon meringue pie and it was delicious. The only thing I didn’t like was that I could eat half of it, but it was a little pricey for what you get. Self & Control If you’re a fan of the Matos Buffet, this is a good place to visit. If you want to have a sweet tooth this is the place to be. The desserts are delicious and they have a good variety of desserts to choose from. The only thing I don’t like about this place is that you have to wait in line for a long time to get in. Other than that, you can’t really go wrong with any of the desserts in the buffet. The drinks are good and the desserts are yummy too. They also have desserts that are not too sweet. I’m not a huge fan of buffets, but this is one of my favorite buffets. MultimodalSum This is a cute little bakery located in the M resort. I had the chocolate croissant and it was very good. The croissants were soft and moist and the filling was delicious. I also had a chocolate chip cookie which was also good. I would definitely recommend this place if you are in the area. Table 3: Sample summaries generated by various models on the Yelp dataset conclude that the multimodal framework outperformed the unimodal framework for unsupervised opinion summarization. In particular, our model achieved state-of-the-art results on the Amazon dataset and outperformed the comparable model by a large margin in the R-L representing the ROUGE scores on the Yelp dataset. Although Self & Control showed high R-2 score, we attributed their score to the inferred N-gram control tokens used as additional inputs to the text decoder. Sample summaries on the Yelp dataset are shown in Table 3. They were generated from source reviews on Baby Cakes bakery. Copycat misused “sweet tooth” and generated “lemon mernigue pie” that was not mentioned in the source reviews. Self & Control generated a summary about a buffet by totally misunderstanding one sentence from source reviews: “If you love the desserts in Studio B Buffet in the M Hotel but don’t want to wait in the massive buffet line or even eat in the buffet, Baby Cakes in the M Hotel is really nice fix.” Furthermore, “Matos Buffet” is a non-existent word. On the contrary, MultimodalSum generated a good summary with a rich description of chocolate croissants. Although “chocolate chip cookie” was not found in the source reviews, our model generated it from cookie images. Note that the term can be found in other reviews that were not used as source reviews. Additional sample summaries on two datasets are shown in Appendix A.5. 7.1.2 Human Evaluation To evaluate the quality of summarization based on human criteria, we conducted a user study. We assessed the quality of summaries using Best-Worst Scaling (BWS; Louviere et al. (2015)). BWS is known to produce more reliable results than raking scales (Kiritchenko and Mohammad, 2017) and is widely used in self-supervised opinion summarization studies. We recruited 10 NLP experts and asked each participant to choose one best and one worst summary from four summaries for three criteria. For each participant’s response, the best model received +1, the worst model received -1, and the rest of the models received 0 scores. The final scores were obtained by averaging the scores of all the responses from all participants. 395 Figure 4: Multimodal gate heatmaps; From the table and two images, our model generates a summary. Heatmaps represent the overall influence of table and images for generating each word in the summary. Note that the summary is a real example generated from our model without beam search. For Overall criterion, Self & Control, Copycat, MultimodalSum, and gold summaries scored -0.527, -0.113, +0.260, and +0.380 on the Yelp dataset, respectively. MultimodalSum showed superior performance in human evaluation as well as automatic evaluation. We note that human judgments correlate better with BERT-score than ROUGE-score. Self & Control achieved a very low human evaluation score despite its high ROUGEscore in automatic evaluation. We analyzed the summaries of Self & Control, and we found several flaws such as redundant words, ungrammatical expressions, and factual hallucinations. It generated a non-existent word by combining several subwords. It was particularly noticeable when a proper noun was generated. Furthermore, Self & Control generated an implausible sentence by copying some words from source reviews. From the results, we conclude that both automatic evaluation and human evaluation performances should be supported to be a good summarization model and BERT-score can complement ROUGE-score in automatic evaluation. Details on human evaluation and full results can be found in Appendix A.3. 7.1.3 Effects of Multimodality To analyze the effects of multimodal data on opinion summarization, we analyzed the multimodal gate. Since the multimodal gate is a eDdimensional vector, we averaged it by a scalar value. Furthermore, as multimodal gates exist for each layer of the text decoder, we averaged them to measure the overall influence of a table or images when generating each token in the decoder. An example of aggregated multimodal gates is shown in Figure 4. It shows the table and images used for generating a summary text, and the multimodal gates for a part of the generated summary are expressed as heatmaps. As we intended, table and image information was selectively used to generate a specific word in the summary. The aggregated value of the table was relatively high for generating “Red Lobster”, which is the name of the restaurants. It was relatively high for images, when generating “food” that is depicted in two images. Another characteristic of the result is that aggregated values of the table were higher than those of the image: mean values for the table and image in the entire test data were 0.103 and 0.045, respectively. This implies that table information is more used when creating a summary, and this observation is valid in that the table contains a large amount of metadata. Note that the values displayed on the heatmaps are small by and large, as they were aggregated from eD-dimensional vector. 7.2 Ablation Studies For ablation studies, we analyzed the effectiveness of our model framework and model training pipeline in Table 4. To analyze the model framework, we first compared the summarization quality with four versions of unimodal model framework, as in the first block of Table 4. BART denotes the model framework in Figure 1a, whose weights are the weights of BART-Large. It represents the lower bound of our model framework without any training. BART-Review denotes the model framework whose weights are from further pretrained BART using the entire training review corpus. UnimodalSum refers to the results of the text modality pretraining, and we classified it into two frameworks according to the use of the rating deviation. 396 Surprisingly, using only BART achieved comparable or better results than many extractive and abstractive baselines in Table 2. Furthermore, further pretraining using the review corpus brought performance improvements. Qualitatively, BART with further pretraining generated more diverse words and rich expressions from the review corpus. This proved our assumption that denoising autoencoderbased pretraining helps in self-supervised multimodal opinion summarization. Based on the BARTReview, UnimodalSum achieved superior results. Furthermore, the use of rating deviation improved the quality of summarization. We conclude that learning to generate reviews based on wide ranges of rating deviations including 0 during training helps to generate a better summary of the average semantics of the input reviews. To analyze the effect of other modalities in our model framework, we compared the summarization quality with three versions of multimodal model frameworks, as in the second block of Table 4. We removed the image or table modality from MultimodalSum to analyze the contribution of each modality. Results showed that both modalities improved the summarization quality compared with UnimodalSum, and they brought additional improvements when used altogether. This indicates that using non-text information helps in selfsupervised opinion summarization. As expected, the utility of the table modality was higher than that of the image modality. The image modality contains detailed information not revealed in the table modality (e.g., appearance of food, inside/outside mood of business, design of product, and color/texture of product). However, the information is unorganized to the extent that the utility of the image modality depends on the capacity of the image encoder to extract unorganized information. Although MultimodalSum used a representative image encoder because our study is the first work on multimodal opinion summarization, we expect that the utility of the image modality will be greater if unorganized information can be extracted effectively from the image using advanced image encoders. For analyzing the model training pipeline, we removed text modality or/and other modalities pretraining from the pipeline. By removing each of them, the performance of MultimodalSum declined, and removing all of the pretraining steps caused an additional performance drop. Although MultiModels R-L BART 14.85 BART-Review 15.23 UnimodalSum w/o rating deviation 18.98 UnimodalSum w/ rating deviation 19.40 MultimodalSum 19.84 w/o image modality 19.54 w/o table modality 19.47 w/o other modalities pretraining 19.26 w/o text modality pretraining 19.24 w/o all modalities pretraining 19.14 Table 4: Ablation studies on the Yelp dataset. The first and second blocks represent various versions of the unimodal model framework and multimodal model framework, respectively. The third block shows the differences in our multimodal framework’s performance according to the absence of specific steps in the model training pipeline. modalSum without other modalities pretraining has the capability of text summarization, it showed low summarization performance at the beginning of the training due to the heterogeneity of the three modality representations. However, MultimodalSum without text modality pretraining, whose image and table encoders were pretrained using BARTReview as a pivot, showed stable performance from the beginning, but the performance did not improve significantly. From the results, we conclude that both text modality and other modalities pretraining help the training of multimodal framework. For the other modalities pretraining, we conducted a further analysis in the Appendix A.4. 8 Conclusions We proposed the first self-supervised multimodal opinion summarization framework. Our framework can reflect text, images, and metadata together as an extension of the existing self-supervised opinion summarization framework. To resolve the heterogeneity of multimodal data, we also proposed a multimodal training pipeline. We verified the effectiveness of our multimodal framework and training pipeline with various experiments on real review datasets. Self-supervised multimodal opinion summarization can be used in various ways in the future, such as providing a multimodal summary or enabling a multimodal retrieval. By retrieving reviews related to a specific image or metadata, controlled opinion summarization will be possible. Acknowledgments We thank the anonymous reviewers for their insightful comments and suggestions. 397 References Reinald Kim Amplayo and Mirella Lapata. 2020. Unsupervised opinion summarization with noising and denoising. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1934–1945. Stefanos Angelidis and Mirella Lapata. 2018. Summarizing opinions: Aspect extraction meets sentiment prediction and they are both weakly supervised. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3675–3686. Tadas Baltruˇsaitis, Chaitanya Ahuja, and LouisPhilippe Morency. 2018. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423–443. Arthur Braˇzinskas, Mirella Lapata, and Ivan Titov. 2020. Few-shot learning for opinion summarization. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4119–4135. Mirella Lapata Braˇzinskas, Arthur and Ivan Titov. 2020. Unsupervised opinion summarization as copycatreview generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5151–5169. Giuseppe Carenini, Jackie Chi Kit Cheung, and Adam Pauls. 2013. Multi-document summarization of evaluative tex. Computational Intelligence, 29(4):545–576. Giuseppe Carenini, Raymond Ng, and Adam Pauls. 2006. Multi-document summarization of evaluative text. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics. Jingqiang Chen and Hai Zhuge. 2018. Abstractive textimage summarization using multi-modal attentional hierarchical rnn. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4046–4056. Eric Chu and Peter Liu. 2019. Meansum: a neural model for unsupervised multi-document abstractive summarization. In In Proceedings of International Conference on Machine Learning (ICML), pages 1223–1232. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Giuseppe Di Fabbrizio, Amanda Stent, and Robert Gaizauskas. 2014. A hybrid approach to multidocument summarization of opinions in reviews. In Proceedings of the 8th International Natural Language Generation Conference, pages 54–63. Hady Elsahar, Maximin Coavoux, Jos Rozen, and Matthias Gall´e. 2021. Self-supervised and controlled multi-document opinion summarization. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1646–1662. G¨unes Erkan and Dragomir R Radev. 2004. Lexrank: Graph-based lexical centrality as salience in text summarization. Journal of artificial intelligence research, 22:457–479. Xiyan Fu, Jun Wang, and Zhenglu Yang. 2020. Multimodal summarization for video-containing documents. arXiv preprint arXiv:2009.08018. Kavita Ganesan, ChengXiang Zhai, and Jiawei Han. 2010. Opinosis: A graph based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd International Conference on Computational Linguistics, pages 340–348. Shima Gerani, Yashar Mehdad, Giuseppe Carenini, Raymond Ng, and Bita Nejat. 2014. Abstractive summarization of product reviews using discourse structure. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, pages 1602–1613. Suchin Gururangan, Ana Marasovi´c, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8342–8360. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778. Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In Proceedings of the 25th International Conference on World Wide Web, pages 507–517. Po-Yao Huang, Junjie Hu, Xiaojun Chang, and Alexander Hauptmann. 2020. Unsupervised multimodal neural machine translation with pseudo visual pivoting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8226–8237. Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. 398 Svetlana Kiritchenko and Saif Mohammad. 2017. Bestworst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465–470. Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388–395. Lun-Wei Ku, Yu-Ting Liang, Hsin-Hsi Chen, et al. 2006. Opinion extraction, summarization and tracking in news and blog corpora. In AAAI spring symposium: Computational approaches to analyzing weblogs, pages 100–107. Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. 2018. Stacked cross attention for image-text matching. In Proceedings of the European Conference on Computer Vision, pages 201– 216. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pretraining for natural language generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880. Haoran Li, Peng Yuan, Song Xu, Youzheng Wu, Xiaodong He, and Bowen Zhou. 2020a. Aspect-aware multimodal summarization for chinese e-commerce products. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pages 8188–8195. Haoran Li, Junnan Zhu, Tianshang Liu, Jiajun Zhang, and Chengqing Zong. 2018. Multi-modal sentence summarization with modality attention and image filtering. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, pages 4152–4158. Mingzhe Li, Xiuying Chen, Shen Gao, Zhangming Chan, Dongyan Zhao, and Rui Yan. 2020b. Vmsmo: Learning to generate multimodal summary for videobased news articles. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9360–9369. Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74–81. Peter J Liu, Mohammad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. Generating wikipedia by summarizing long sequences. In Proceedings of the 6th International Conference on Learning Representations. Yang Liu and Mirella Lapata. 2019. Hierarchical transformers for multi-document summarization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, page 5070–5081. Jordan J Louviere, Terry N Flynn, and Anthony Alfred John Marley. 2015. Best-worst scaling: Theory, methods and applications. Cambridge University Press. Rada Mihalcea and Paul Tarau. 2004. Textrank: Bringing order into text. In Proceedings of the 2004 conference on Empirical Methods in Natural Language Processing, pages 404–411. Ramesh Nallapati, Bowen Zhou, Cicero dos Santos, Caglar Gulcehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning, page 280–290. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8026–8037. Michael Paul, ChengXiang Zhai, and Roxana Girju. 2010. Summarizing contrastive viewpoints in opinionated text. In Proceedings of the 2010 conference on Empirical Methods in Natural Language Processing, pages 66–76. Romain Paulus, Caiming Xiong, and Richard Socher. 2018. A deep reinforced model for abstractive summarization. In Proceedings of the 6th International Conference on Learning Representations. Laura Perez-Beltrachini, Yang Liu, and Mirella Lapata. 2019. Generating summaries with topic templates and structured convolutional decoders. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, page 5107–5116. Ratish Puduppully, Li Dong, and Mirella Lapata. 2019. Data-to-text generation with content selection and planning. In Proceedings of the 33th AAAI Conference on Artificial Intelligence, pages 6908–6915. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1073–1083. Yuanhang Su, Kai Fan, Nguyen Bach, C.-C. Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 10482–10491. Quoc-Tuan Truong and Hady Lauw. 2019. Multimodal review generation for recommender systems. In Proceedings of the World Wide Web Conference, pages 1864–1874. 399 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008. Nam N Vo and James Hays. 2016. Localizing and orienting street views using overhead imagery. In Proceedings of the European Conference on Computer Vision, pages 494–509. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In Proceedings of the 8th International Conference on Learning Representations. Hao Zheng and Mirella Lapata. 2019. Sentence centrality revisited for unsupervised summarization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6236– 6247. Renjie Zheng, Mingbo Ma, and Liang Huang. 2018. Multi-reference training with pseudo-references for neural translation and text generation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3188–3197. Junnan Zhu, Haoran Li, Tianshang Liu, Yu Zhou, Jiajun Zhang, and Chengqing Zong. 2018. Msmo: Multimodal summarization with multimodal output. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, page 4154–4164. Junnan Zhu, Yu Zhou, Jiajun Zhang, Haoran Li, Chengqing Zong, and Changliang Li. 2020. Multimodal summarization with guidance of multimodal reference. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, pages 9749–9756. 400 A Appendix A.1 Dataset Preprocessing We selected businesses and products with a minimum of 10 reviews and popular entities above the 90th percentile were removed. The minimum and maximum length of the words were set as 35 and 100 for Yelp, and 45 and 70 for Amazon, respectively. We set the maximum number of tokens as 128 using the BART tokenizer for training, and we did not limit the maximum tokens for inference. For the Amazon dataset, we selected 4 categories: Electronics; Clothing, Shoes and Jewelry; Home and Kitchen; Health and Personal Care. As Yelp dataset contains unlimited number of images for each entity, we did not use images for popular entities above the 90th percentile. On the other hand, Amazon dataset contains a single image for each entity. Therefore, we did not use images only when meaningless images such as non-image icon or update icon were used or the image links had expired. For Yelp dataset, we selected name, ratings, categories, hours, and attributes among the metadata. We used the hours of each day of the week as seven fields and used all metadata contained in attributes as each field. For some attributes (‘Ambience’, ‘BusinessParking’, ‘GoodForMeal’) that have subordinate attributes, we used each subordinate attribute. Among the fields, we selected 47 fields used by at least 10% of the entities. We set the maximum number of categories as 6 based on the 90th percentile, and averaged the representations of each category. For ratings, we converted it to binary notation consisting of 4 digits (22, 21, 20, 2−1). For hours, we considered (open hour, close hour) as a 2-dimensional vector, and conducted K-means clustering. We selected four clusters based on silhouette score: (16.5, 23.2), (8.7, 17.1), (6.4, 23), and (10.6, 22.6). Based on the clusters, we converted hours into a categorical type. For Amazon dataset, we selected six fields: name, price, brand, categories, ratings, and description. We set the maximum number of categories as 3 based on the 90th percentile, and averaged the representations of each category. Furthermore, as each category consists of hierarchies with a maximum of 8 depths, we averaged the representations of hierarchies to get each category representation. For price and ratings, we converted them to binary notation consisting of 11 and 4 digits, respectively, after rounding them to the nearest 0.5 to contain digit for 2−1. As some descriptions consist of many Pipeline step batch epochs warmup lr Text pretrain 16 5 0.5 5e-05 Others pretrain 32 20 1 1e-04 Multimodal train 8 5 0.25 1e-05 Table 5: Hyperparameter values for each step in model training pipeline. tokens, we set the maximum number of tokens as 128. We regarded each token in description as each field, so we got total 5 + 128 fields. A.2 Experimental Details Our image encoder is based on ResNet101. ResNet101 is composed of 1 convolution layer, 4 convolution layer blocks, and 1 fully connected layer block. Among them, 4 convolution layer blocks play an important role in analyzing image. Through each convolution layer block, the size of the image feature map is reduced to 1/4, but it gets high-level features. To maintain the ability to extract low-level features of the image, we set the model weights up to the second convolution layer block not to be trained further. We only used up to the third convolution layer block to increase the resolution of feature maps without using too highlevel features for image classification. In this way, lI was set to 14 × 14 and eI was set to 1,024. To use the knowledge of text modality in table encoder, we obtained field name embeddings by summing the BART token embeddings for the tokens contained in the field name. Because various data types can be used for field value, we used different processing methods for each data type. Nominal values were handled in the same way as the field name. Binary and ordinal values were processed by replacing them with nominal values of corresponding meanings: ‘true’ and ‘false’ were used for binary values, and ‘cheap’, ‘average’, ‘expensive’, and ‘very expensive’ were used for ‘RestaurantsPriceRange’. Numerical values were converted to binary notation, and we obtained the representations by summing embeddings corresponding to the place, where the place value is 1. For other categorical values, we simply trained embeddings corresponding to each category. We set each hyperparameter value different for each step in the model training pipeline, as in Table 5. We set the batch size according to the memory usage and set other values according to the amount of learning required. Hyperparameter ranges for epochs and lr (learning rate) were [3, 5, 10, 15, 20] and [1e-03, 1e-04, 5e-05, 1e-05, 5e-06], 401 Models Grammaticality Coherence Overall Self & Control -0.517 -0.500 -0.527 Copycat 0.163 -0.077 -0.113 MultimodalSum 0.367 0.290 0.260 Gold -0.013 0.287 0.380 Table 6: Human evaluation results in terms of the BWS on the Yelp dataset. respectively, and optimized values were chosen from validation loss in one trial. For summary generation at test time, we set different hyperparameter values for each dataset. Beam size, length penalty, and max length were set to 4, 0.97, and 105 for Yelp and 2, 0.9, and 80 for Amazon, respectively. Note that max length was set first to prevent incomplete termination and length penalty was determined based on the ROUGE scores on validation dataset. The number of training parameters for text, image, and table modality pretraining are 406.3M, 27.1M, and 3.2M, respectively, and that for multimodal training is 486.9M. Run time for text modality pretraining was 16h on 4 GPUs, and it took 41h and 43h on 2 GPUs for image and table modality training, respectively. For final multimodal training, it took 14h on 8 GPUs. A.3 Human Evaluation For human evaluation, we randomly selected 30 entities from Yelp test data, and used three criteria: Grmmaticality (the summary should be fluent and grammatical), Coherence (the summary should be well structured and well organized), and Overall (based on your own criteria, select the best and the worst summary of the reviews). Results for three criteria are shown in Table 6. Self & Control achieved very poor performance for all criteria due to its flaws that were not revealed in the automatic evaluation. Surprisingly, MultimodalSum outperformed gold summaries for two criteria; however, its overall performance lagged behind Gold. As our model was initialized from BART-Large that had been pretrained using large corpus and further pretrained using training review corpus, it may have generated fluent and coherent summaries. It seems that our model lagged behind Gold in Overall due to various criteria other than those two. The fact that Gold scored lower than Copycat in Grammaticality may seem inconsistent with the result from Braˇzinskas and Titov (2020). However, we assumed that this result was due to a combination of the four models in relative evaluation. The ranking for Copycat and Gold may have changed in absolute evaluation. Image Table Models R-1 R-2 R-L R-1 R-2 R-L Untrained 21.03 2.45 14.17 24.04 2.92 15.10 Triplet 20.06 2.49 13.15 25.67 3.52 15.16 Pivot (ours) 25.87 3.62 15.70 27.32 4.12 16.57 Table 7: Reference reviews generation results on the Yelp dataset. A.4 Analysis on Other Modalities Pretraining To analyze the various models for the other modalities pretraining, we evaluated the performance of the reference review generation task that generates corresponding reviews from images or a table. For evaluation, we used the data that were not used for training data: we left 10% of the data for Yelp and 5% for Amazon. We chose two comparison models: Untrained and Triplet. Untrained denotes the model that image encoder or table encoder keeps untrained. This option indicates the lower bound containing only the effect of the text decoder. Triplet denotes the triplet-based metriclearning model, based on Lee et al. (2018) and Vo and Hays (2016). For triplet (images or a table, reviews of positive entity, reviews of negative entities), we trained the image or table encoder based on the pretrained text encoder, by placing the image or table encoded representations close to the positive reviews representations and far from the negative reviews representations. Note that pretrained text encoder was not trained further. Results on the other modalities pretraining are shown in Table 7. For each model, the pretrained decoder generated a review from image or table encoded representations. We measured the average ROUGE scores between the generated review and N reference reviews. The first finding was that results of table outperformed those of image. It indicates that table has more helpful information for generating reference review. The second finding was that our method based on the text decoder outperformed the Triplet based on the text encoder. Especially, Triplet achieved very poor performance for image because it is hard to match M images to N reference reviews for metric learning. On the contrary, our method achieved much better performance by pivoting the text decoder. Triplet showed good performance on table because it is relatively easy to match 1 table to N reference reviews; however, our method outperformed it. We conclude that our method lets the image and table encoder get proper representations to generate reference reviews regardless of the number of inputs. 402 A.5 Example Summaries Table 8, 9 show sample summaries generated from our model and baseline models on Yelp and Amazon datasets. Full summaries from our model are available at https://bit.ly/3bR4yod. Review 1 The fresh water catfish is probably the best I’ve every had. The service was outstanding. I would recommend this little secret to everyone. Review 2 I loved everything about this place!! Great food, great decor, and great service. The best collard greens I have ever had. We had fried oysters for a starter and although I have never had them before so I have nothing to compare them with they were very tasty. The warm hush puppies with the honey butter was delicious!! I had the crab legs which were perfect and plentiful. My sister had the all you can eat fried catfish that was also cooked perfectly. A great experience all around!! Review 3 Amazing food and great service! The hospitality was out of this world. Will definitely be back soon. The wait was less than 5 minutes at 7pm on a Friday night, amazing!! The staff was very kind and the waitresses were very attentive and helpful. We tried the frog legs, catfish, alligator bites, crab legs, gumbo and of course the hush puppies! Everything was outstanding. What a hidden gem! Review 4 I love this place the food amazing the staff helpful ....must try green tomatos ...fresh water fish ;ˆ) Review 5 We love this place the catfish is good the hush puppies with that honey butter are awesome the french fries the gumbo what else is good there the alligator tail mostly everything on the menu. I guess the only bad thing I can say is sometimes it’s like a 20 minute wait in the drive-through but it’s well worth it when your food is hot Because tonight I got to go home and warm it up it’s not hot enough, Even though they’re still open for another hour that was a bummer Review 6 Really tasty catfish, shrimp and fixin’s. Our friend took us to the sister location on Nellis a couple of months ago, but this location was more convenient to our hotel. No worries, this place was just as good! Excellent service, and the salad bar is a nice touch as well. As a Bostonian, I’m pretty particular about seafood. The Hush Puppy fits the bill. Very satisfied! Review 7 First Time here and the food, staff was awesome. Manager came over and gave us samples of the fried catfish, super nice. Review 8 I never eat catfish. It’s nasty to me until I tasted the saltwater catfish!!! Greens are on point. The hushpuppy are bomb with honey butter!!!!! Gator bites where are ok. Copycat This place is awesome! The food was great, the service was great. We had the catfish po’boy and it was delicious. The only reason I didn’t give 5 stars is because of the fact that they don’t deliver. Self & Control I love this place. The service is awesome. The hush puppies are to die for. I love the honey butter. I can’t wait to go back and try it again. The only thing I don’t like about the place is the wait. It can be a little long, but it’s worth it. It’s a little on the pricey side, but you’re getting what you pay for. Love the hot butter, the hush puppies, the French fries, the gumbo, the catfish and the gumbo. Everything is so yummy and the service is top notch. Try it out, you won’t be disappointed. MultimodalSum This place is a hidden gem. The food is great and the service is even better. I had the all you can eat catfish and it was delicious. The hush puppies are the best I’ve ever had. I will definitely be back. Gold Yummy and delicious catfish. You gotta try it. Friendly staff and service is good too. You can tell they know their seafood and how to prepare and cook it to perfection. The staff also answered any questions I had. The Hush Puppies are tasty too. Table 8: Yelp summaries generated by different models. 403 Review 1 I usually wear size 37, but found a 38 feels better in this sandal. I absolutely love this sandal. So supportive and comfortable, although at first I did get a blister on my big toe. Do not let this be the deciding factor. It stretched out and is now fabulous. I love it so much that I bought it in three colors. Review 2 This is a really cute shoe that feels very comfortable on my high arches. The strap on the instep fits my feet very well, but I have very slim feet. I can see how it would be uncomfortably tight on anyone with more padding on their feet. Review 3 I love these sandals. The fit is perfect for my foot, with perfect arch support. I don’t think the leather is cheap, and the sandals are very comfortable to walk in. They are very pretty, and pair very well with pants and dresses. Review 4 My wife is a nurse and wears dansko shoes. We were excited to try the new crimson sandal and normally order 39 sandal and 40 closed toe. Some other reviews were right about a narrow width and tight toe box. We gave them a try and passed a great pair of shoes to our daughter with her long narrow feet, and she loves them... Review 5 Finally, a Dansko sandal that’s fashion forward! It was love at first sight! This is my 4th Dansko purchase. Their sizing, quality and comfort is very consistent. I love the stying of this sandal and I’m pleased they are offering bolder colors. Another feature I love is the Dri-Lex topsole - it’s soft and keeps feet dry. Review 6 I really love these sandals. my only issue is after wearing them for a while my feet started to swell as I have a high instep and they were a little tight across the top. I’m sure they will stretch a bit after a few wears Review 7 I have several pairs of Dansko clogs that are all size 39 and fit perfectly. So I felt confident when I ordered the Tasha Sandal in size 39. I don’t know if a 40 would be too large but the 39 seems a little small. Otherwise, I love them. They are very cushiony and comfortable! Review 8 I own many Dansko shoes and these are among my favorites. They have ALL the support that Dansko offers in its shoes plus they are very attractive. I love the the heel height and instant comfort. They look great with slacks and dresses, dressed up or not... Copycat This is my second pair of Dansko clogs and I love them. They are very comfortable and I can wear them all day without any discomfort. I would recommend them to anyone looking for a comfortable sandal. MultimodalSum I love these sandals. They are very comfortable and look great. The only thing I don’t like is that they are a little tight across the top of my foot. I have a high instep and the strap is a little too tight. I am hoping they will stretch out a bit. Gold 1 I love these sandals, Dansko has made a really great product! I had to return my first pair (39) for being a bit tight and small, but I went a size higher (40) and it is perfect, they are so comfortable! If they do stretch out like other reviews say, they will still fit and look great. Gold 2 I love these Dansko Tasha sandals! They are comfortable and the style is really cute. The only warning I have is that they seem to run narrow: you may want to buy a larger size if you have wide feet. Also, they seem to stretch as you wear them, so don’t get discouraged by a few blisters on first wearing. Gold 3 These Dansko shoes are amazingly comfortable and hug the shape of my feet well, but I did have to wear them for a bit to stretch them out. They felt a little tight at first, but now they are perfect. I feel they’re true to size so I’d recommend ordering these in your normal shoe size. Table 9: Amazon summaries generated by different models.
|
2021
|
33
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4275–4293 August 1–6, 2021. ©2021 Association for Computational Linguistics 4275 Societal Biases in Language Generation: Progress and Challenges Emily Sheng1, Kai-Wei Chang2, Premkumar Natarajan1, Nanyun Peng1,2 1 Information Sciences Institute, University of Southern California 2 Computer Science Department, University of California, Los Angeles {ewsheng,pnataraj}@isi.edu, {kwchang,violetpeng}@cs.ucla.edu Abstract Technology for language generation has advanced rapidly, spurred by advancements in pre-training large models on massive amounts of data and the need for intelligent agents to communicate in a natural manner. While techniques can effectively generate fluent text, they can also produce undesirable societal biases that can have a disproportionately negative impact on marginalized populations. Language generation presents unique challenges for biases in terms of direct user interaction and the structure of decoding techniques. To better understand these challenges, we present a survey on societal biases in language generation, focusing on how data and techniques contribute to biases and progress towards reducing biases. Motivated by a lack of studies on biases from decoding techniques, we also conduct experiments to quantify the effects of these techniques. By further discussing general trends and open challenges, we call to attention promising directions for research and the importance of fairness and inclusivity considerations for language generation applications. 1 Introduction Natural language generation (NLG) is a suite of techniques that enables the generation of humanreadable language for different goals. These techniques are the core components of applications such as virtual assistants, chat bots, automatic translators, summarizers, and creative language composers. Recent advances in techniques for language generation (e.g., GPT (Radford et al., 2018), GPT-2 (Radford et al., 2019), GPT-3 (Brown et al., 2020), TransformerXL (Dai et al., 2019), XLNet (Yang et al., 2019)) powered by Transformers (Vaswani et al., 2017) and an increasing repository of available data have created more capable applications. This has, in turn, channeled more interest and effort into developing NLG techniques. We emphasize the importance of better understanding how societal biases manifest in NLG techniques, because NLG applications directly interact with many different users to generate novel content in various domains (e.g., chat bots for health, education, and customer support). However, when techniques are less effective or detrimental for marginalized populations, these techniques can inadvertently become gatekeepers of those populations for generation and associated language technologies. For example, an educational chat bot that produces more negative responses for topics about a specific ethnicity will discourage users of that ethnicity from interacting with the chat bot. While it is generally important to study the societal impact of NLP and AI techniques, we argue that the direct user impact of NLG techniques makes it especially important to carefully quantify the impact. Motivated by the importance of fairness in language generation, we present the first comprehensive survey on societal biases in language generation. By enumerating how NLG techniques contribute to biases and examining progress towards bias analysis and mitigation, we contextualize the discussion of broader trends and challenges. Specifically, we focus on techniques for NLG tasks, i.e., tasks that generate a sequence of text.1 Finding a lack of studies on biases from decoding techniques, we additionally present an experimental study to quantify the effects of various decoding techniques. Before we delve into the details of biases in language generation, we first position our survey in the context of other relevant surveys and position papers. Sun et al. (2019) present a focused survey 1Although bi-directional language models like BERT (Devlin et al., 2019) can also be used for auto-regressive generation (Wang and Cho, 2019; Chen et al., 2020), traditional auto-regressive models are still typically of better quality and more widely used for generation (Shwartz et al., 2020). Thus, we limit the scope of this survey to the latter models. 4276 Demo. Dim. NLG Task Works Gender Autocomplete Bordia and Bowman (2019); Qian et al. (2019); Solaiman et al. (2019); Sheng et al. (2019, 2020); Vig et al. (2020); Yeo and Chen (2020); Brown et al. (2020); Dhamala et al. (2021); Schick et al. (2021); Nozza et al. (2021); Kirk et al. (2021) Dialogue Henderson et al. (2018); Dinan et al. (2020a); Liu et al. (2020a,b); Cercas Curry et al. (2020); Sheng et al. (2021a,b) MT Vanmassenhove et al. (2018); Elaraby et al. (2018); Prates et al. (2019); Stanovsky et al. (2019); Escud´e Font and Costa-juss`a (2019); Cho et al. (2019); Moryossef et al. (2019); Saunders and Byrne (2020); Saunders et al. (2020); Kocmi et al. (2020); Costa-juss`a and de Jorge (2020); Costa-juss`a et al. (2020); Basta et al. (2020); Farkas and N´emeth (2020); Stafanoviˇcs et al. (2020); Gonen and Webster (2020); Hovy et al. (2020); Roberts et al. (2020); Cho et al. (2021); Savoldi et al. (2021); Renduchintala and Williams (2021); Choubey et al. (2021); Saunders et al. (2021); Tomalin et al. (2021) Re-writing Habash et al. (2019); Zmigrod et al. (2019); Alhafni et al. (2020); Sun et al. (2021) Profession Autocomplete Huang et al. (2020); Dhamala et al. (2021) Race Autocomplete Solaiman et al. (2019); Sheng et al. (2019, 2020); Groenwold et al. (2020); Brown et al. (2020); Dhamala et al. (2021); Schick et al. (2021); Kirk et al. (2021) Dialogue Sheng et al. (2021a,b) Religion Autocomplete Solaiman et al. (2019); Brown et al. (2020); Dhamala et al. (2021); Kirk et al. (2021); Abid et al. (2021) Sexuality Autocomplete Sheng et al. (2019, 2020); Kirk et al. (2021) Dialogue Sheng et al. (2021a) Other Autocomplete Shwartz et al. (2020); Peng et al. (2020); Huang et al. (2020); Dhamala et al. (2021); Kirk et al. (2021) Dialogue Sheng et al. (2021a) Re-writing Pryzant et al. (2020); Ma et al. (2020) Table 1: Existing bias studies on different demographic dimensions in various NLG tasks: autocomplete generation, dialogue generation, machine translation (MT), and text re-writing. on mitigating gender biases and Shah et al. (2020) categorize sources of biases—both largely focus on natural language understanding (NLU) tasks, while we examine biases in NLG tasks. Additionally, Blodgett et al. (2020) urge for more explicitly tying “biases” in NLP to societal normative definitions of biases and social hierarchies; with their recommendations in mind, we discuss the negative impacts of biases in NLG techniques. Our contributions are a comprehensive survey on societal biases in language generation and an experimental study on biases from decoding techniques. To start, we describe classes of NLG tasks (Sec. 2) and subsequently examine examples of biases and harms in NLG (Sec. 3). We then discuss NLG techniques that facilitate biases, including a study of decoding techniques (Sec. 4). Sec. 5 highlights progress and challenges, and Sec. 6 presents open problems and proposals. We hope this survey brings more visibility to the importance of carefully considering different components of NLG pipelines for potential biases and mitigation methods. 2 Language Generation Tasks To begin, we categorize generation tasks and introduce existing bias studies relevant to each task. NLG tasks broadly fall into two categories: those that generate text continuations conditioned on some prompt and those that transform text from one form to another. Table 1 organizes various bias-related works for NLG tasks. 2.1 Continuation Generation Tasks The continuation class includes autocomplete and dialogue generation, where the goal is to generate text that is coherent and relevant to a prompt. Autocomplete Generation We use the term autocomplete generation to refer to conditional generation directly from language models. Language models are the core components for many NLG and NLU tasks, and this task enables directly quantifying biases in large, pre-trained language models (Bordia and Bowman, 2019; Sheng et al., 2019; Solaiman et al., 2019; Brown et al., 2020). Existing works analyzing biases in autocomplete generation have mostly examined Transformer-based models, including GPT (Shwartz et al., 2020), GPT2 (Solaiman et al., 2019; Sheng et al., 2019, 2020; Shwartz et al., 2020; Vig et al., 2020; Yeo and Chen, 2020; Huang et al., 2020; Dhamala et al., 2021; Schick et al., 2021), GPT-3 (Brown et al., 2020), CTRL (Dhamala et al., 2021), TransformerXL (Shwartz et al., 2020; Vig et al., 2020; Huang et al., 2020), and XLNet (Shwartz et al., 2020; Vig et al., 4277 2020; Yeo and Chen, 2020), though Bordia and Bowman (2019); Qian et al. (2019) also look at LSTM-based models. Dialogue Generation Dialogue generation is conditioned on user inputs and can be for specific domains (e.g., health, customer service) and tasks (e.g., behavior intervention, booking flights) or general chit-chat. These dialogue applications directly interact with users, and any propagated biases directly affect user behavior and actions. In terms of recurrent dialogue models, Henderson et al. (2018) analyze biases in hierarchical recurrent encoder-decoder architectures and Liu et al. (2020a,b) analyze LSTM-based encoder-decoder models. Other works on dialogue biases (Dinan et al., 2020a; Sheng et al., 2020, 2021b) focus on Transformer-based models such as DialoGPT (Zhang et al., 2020) and other custom architectures. 2.2 Transformation Generation Tasks The transformation class includes machine translation and various formulations of text re-writing. The general goal of these tasks is to transform text into a form with targeted properties. Machine Translation Translation is the task of transforming text between languages while preserving the meaning. Existing works on biases in machine translation have almost exclusively focused on issues of gender biases2 in a variety of academic and commercial systems. The use of grammatical gender in some languages and not in others can expose unwanted gender associations (e.g., for different occupations) through translation (Prates et al., 2019). Earlier works by Vanmassenhove et al. (2018) and Elaraby et al. (2018) study LSTM-based encoder-decoder translation systems, and more recent works examine Transformer-based architectures (Escud´e Font and Costa-juss`a, 2019; Stanovsky et al., 2019; Saunders and Byrne, 2020; Saunders et al., 2020; Costa-juss`a and de Jorge, 2020; Basta et al., 2020; Stafanoviˇcs et al., 2020; Renduchintala and Williams, 2021; Choubey et al., 2021; Saunders et al., 2021; Tomalin et al., 2021). While Google Translate3 has been the most popular commercial system to analyze for gender biases (Prates et al., 2019; Moryossef et al., 2019; Stanovsky et al., 2019; Cho et al., 2019; Farkas and N´emeth, 2020), Stanovsky et al. (2019) also 2For a detailed survey of gender bias in machine translation, we refer readers to Savoldi et al. (2021). 3https://translate.google.com study Microsoft Translator,4 Amazon Translate,5 and SYSTRAN;6 Cho et al. (2019) additionally look at Naver Papago7 and Kakao Translator,8 and Cho et al. (2021) also examine Yandex.9 Re-writing We use the term re-writing to refer to tasks of revising specific words and phrases in the original text to be more aligned with a targeted attribute. Specifically, there have been studies on re-inflection (Habash et al., 2019; Zmigrod et al., 2019; Alhafni et al., 2020) and re-writing text to use neutral viewpoints (Pryzant et al., 2020), genderneutral English (Sun et al., 2021), or more agency (Ma et al., 2020). These tasks typically rely on custom encoder-decoder models. 2.3 Other Tasks There are other NLG tasks, such as the continuation tasks of story and poetry generation, and the transformation tasks of abstractive summarization and paraphrase generation. However, these other NLG tasks are not yet well-studied in the context of societal biases.10 3 Biases and their Negative Impacts In this section, we introduce how existing studies of biases in NLG tasks commonly quantify biases and their negative impacts. 3.1 Bias Definitions and Metrics In the context of AI fairness, the term “bias” commonly refers to skews that result in undesirable impacts (Crawford, 2017) and is quantifiable with some metric. There are relatively more existing studies on biases in NLU tasks, where it is arguably simpler to define bias metrics, since we can intuitively compare the accuracy of the task (e.g., coreference resolution, hate speech detection) for different demographics. Language generation tasks often involve stochastic generation of open-ended and lengthy texts, traits that are not directly compatible with traditional algorithmic bias definitions (e.g., 4https://www.bing.com/translator 5https://aws.amazon.com/translate 6https://www.systransoft.com 7https://papago.naver.com 8https://translate.kakao.com 9https://translate.yandex.com 10Lucy and Bamman (2021) is an exception that analyzes gender in generated stories. While there are studies of biases in poetry generation and summarization, they focus on non-NLG biases: Sheng and Uthus (2020) investigate biases in a poetry composition system, but in the context of information retrieval; Celis and Keswani (2020) analyze biases in extractive summarization. 4278 equalized odds, equal opportunity, demographic parity (Dwork et al., 2012; Hardt et al., 2016)). Because of the difficulty in defining metrics, existing works define bias loosely as demographic inequality and use intermediate proxy metrics to comparatively measure bias. Examples include: • Regard Ratio: negative-neutral-positive regard score ratios of text generated from bias-inducing prompts (Sheng et al., 2019) • Sentiment Ratio: negative-neutral-positive sentiment score ratios of text generated from African American English (AAE) versus White-Aligned English (WAE) prompts (Groenwold et al., 2020) • Individual and Group Fairness through Sentiment: comparisons of the sentiment distributions of generated text across demographics and prompts (Huang et al., 2020) • Gendered Word Co-occurrence Score: mean and standard deviations of the absolute log ratio of probabilities: P(word|female terms) to P(word|male terms) across all words in generated text (Bordia and Bowman, 2019) There are also metrics for other bias evaluation setups in continuation generation tasks involving sentiment (Shwartz et al., 2020), the ratio of gendered words (Solaiman et al., 2019; Vig et al., 2020; Dinan et al., 2020a), and other novel metrics (Peng et al., 2020; Yeo and Chen, 2020). Studies of biases in transformation generation tasks favor metrics of accuracy in terms of successfully transforming text to have a desired property. We present a more thorough comparison of metrics in Section 5.4. Bias metrics can also be categorized by how they define associations between demographic group attributes and text. Biases can be towards people described in text, people who produce the text, or people to whom the text is addressed (Dinan et al., 2020b). Most existing works define bias metrics through the first association—these biases are relatively easier to analyze, since both the demographic and the textual signals of bias are encapsulated within the text. There are also works that define biases towards people who produce the text (Groenwold et al., 2020) or people to whom the text is addressed (Sheng et al., 2021b), though there are relatively fewer works that study these latter associations. 3.2 Negative Impacts Biases in NLG techniques are important to study because they can result in harmful, negative impacts. We survey detrimental representational11 and allocational12 impacts (Crawford, 2017; Barocas et al., 2017; Blodgett et al., 2020) used to motivate existing studies of bias in NLG tasks, finding limited examples. While representational impacts are sometimes cited, it is difficult to measure the extent of the impacts. Additionally, techniques for effective NLG are relatively new, and existing studies have limited knowledge of potential allocational impacts. Finally, biases in NLG tasks give rise to a third type of negative impacts, which we call vulnerability impacts. Representational Impacts The works in Table 1 motivate (to varying degrees) studying biases in NLG through potential negative representational impacts, in the form of propagating stereotypes, misrepresentations, or denigrations of social groups. For example, Sheng et al. (2019) enumerate how generated text can propagate varying social perceptions of different demographics, and Prates et al. (2019) discuss how occupation-related gender biases could propagate stereotypes in translation. However, it is difficult to quantify the effects of representational impacts;13 while such impacts may be measured indirectly (e.g. by analyzing allocational impacts), we suggest long-term, interdisciplinary collaborations to explore the direct effects of these representational impacts. Allocational Impacts Harmful allocational impacts result from an unequal allocation of resources across groups. Since effective NLG techniques based on large Transformer models (Vaswani et al., 2017) are relatively new, most of the existing works on biases in NLG that list possible impacts only analyze direct representational consequences. A real example of a negative allocational impact is when machine translation errors lead to arrests (Ong, 2017). In general, technologies that are less effective or detrimental for certain populations become barriers that actively prevent those populations from using the technology, leading to diminished opportunities in jobs, education, health, etc. We discuss more details in Section 4.5. With continuous technological advances, more organizations will turn to effective NLG techniques, making it imperative to start setting norms to reduce harmful allocational impacts (Tamkin et al., 2021). 11Unfair representations of different groups 12Unfair allocation of resources 13Kay et al. (2015) is a rare example that explicitly studies the effect of representational impacts in image search. 4279 Vulnerability Impacts Open-domain generation tasks can amplify a group’s vulnerability to manipulation and harm, which is an intermediate impact that makes a group more susceptible to representational and allocational impacts. For example, privacy-related issues (Carlini et al., 2020), misinformation (Levy et al., 2021), or radicalizing views in generated text could make a group more likely to be attributed to specific stereotypes (e.g., through action guided by misinformation) or end up with diminished opportunities (e.g., by having personal data exposed and misused). Separately identifying vulnerability impacts could help facilitate recognition of other negative impacts. 4 Contributors to NLG Biases In a pipeline from data collection to evaluation for an NLG task, each component could propagate biases.14 We emphasize the ways in which data, model architecture, decoding, evaluation, and deployment uniquely exacerbate biases in generation tasks. Additionally, we present an empirical study to show how measured biases in generated text can vary based on decoding technique. 4.1 Biases from Data Modern NLP models often rely on large pre-trained language models, which in turn rely on a large collection of data to learn explicit and implicit associations. Several recent pre-trained language models used for NLG tasks, e.g., T5 (Raffel et al., 2020) and GPT-3 (Brown et al., 2020), are trained on the largest datasets used for any models. These large models for generation are commonly trained on web data, which is known to contain biased language (e.g., Ferrer et al. (2021) discover gender, religion, and ethnic biases in Reddit communities). While preprocessing is often included to filter out malformatted data and explicitly negative content (e.g., bad words and offensive phrases), those are generally the only efforts to reduce biases and associated impacts. Furthermore, by filtering out all words deemed “bad”, Bender et al. (2021) warns that we remove the discourse of marginalized populations. Paullada et al. (2020), Bender and Friedman (2018), and Gebru et al. (2018) provide more comprehensive surveys and frameworks that focus on aspects of data creation and management that 14Task formulation and application deployment are also part of NLG task pipelines (Kiritchenko et al., 2020), though we do not focus on biases in these areas. could lead to biases, and we refer readers to their works for more discussion. In the context of translation, Cho et al. (2021) find that more data can increase translation fluency but may also make the system more biased. 4.2 Biases from Model Architecture There are relatively few studies that examine model architectural properties that could lead to biases. We discuss the few efforts towards understanding model biases in NLG tasks and emphasize the need for more to generalize. For autocomplete generation, Vig et al. (2020) analyze GPT-2 variants through a causal mediation analysis, finding that larger models contain more gender bias, and bias tends to be concentrated in a small number of neurons and attention heads. Silva et al. (2021) observe amplified biases in distilled versus original models. For machine translation, Costa-juss`a et al. (2020) note that language-specific architectures are less biased because they encode more gender information than shared language encoder-decoder architectures. Studies like the aforementioned are useful for designing targeted bias mitigation methods (e.g., controlled generation to target specific attention heads or regularization to retain gender information). However, more evidence would be needed to generalize findings across models.15 4.3 Biases from Decoding While NLU and NLG models have structural similarities, NLG tasks uniquely use search or sampling techniques at inference time to generate text. Popular techniques include: • Greedy Search: at each time step, choose the word with the highest probability. • Beam Search: at each time step, keep the top b hypotheses with highest probabilities; eventually pick the hypothesis with the highest probability. • Top-k sampling (Fan et al., 2018): at each time step, re-distribute the probability mass of the top k words with highest probabilities and sample. • Nucleus sampling (Holtzman et al., 2019): at each time step, re-distribute the probability mass of the smallest set of words with a cumulative probability exceeding p and sample. More constrained forms of generation such as machine translation generally use variations of beam 15We also refer the reader to the work of Park et al. (2018) that discusses biases in NLU tasks from model components that “attend” to specific words (e.g., through attention or pooling), which could be applicable to NLG tasks as well. 4280 search; however, preferred decoding techniques are more varied for open-domain generation. Despite variations in fluency and diversity between deterministic versus stochastic, search versus sampling procedures, there are limited studies (Roberts et al., 2020) on how different decoding properties affect biases in generation. A Study on Biases from Decoding To study how decoding techniques affect biases in generation, we use existing NLG bias metrics to evaluate text generated from different decoding methods.16 We examine autocomplete generations from GPT, GPT-2, and XLNet, using the decoding techniques from Section 4.3. We evaluate with the following bias metrics: regard ratios (Sheng et al., 2019), sentiment ratios (Groenwold et al., 2020), individual and group fairness through sentiment scores (Huang et al., 2020), and gendered word co-occurrence scores (Bordia and Bowman, 2019) (as introduced in Section 3). More experimental details can be found in the Appendix. In Section 5.4, we distinguish between relative and absolute score metrics to examine evaluation differences between NLG tasks. Here, we organize our results into these categories to generalize trends about decoding techniques. The ratio-based metrics are relative score metrics, since evaluation relies on comparing ratios between demographics. The latter three metrics are absolute score metrics that have target values of zero indicating no bias. For the relative score metrics, search and sampling techniques generate similar outcomes. An interesting result between sampling techniques for the regard metric is that nucleus sampling is less biased yet more negative than top-k sampling. For the absolute score metrics, we find that beam search is the most unbiased technique, closely followed by greedy search and then top-k and nucleus sampling. Through our study, we discover that text diversity is not accounted for in any of the bias metrics, yet diversity can be a confounding factor. Specifically, beam search is the least diverse,17 followed by greedy search, top-k sampling, then nucleus sampling. Results indicate that the less diverse search techniques lead to better scores for individual fairness, group fairness, and gendered word co-occurrence ratios. We hope these experimental results will encour16Code at https://github.com/ewsheng/ decoding-biases. 17We report average generated text length and vocabulary sizes to estimate diversity in Appendix Table 4. age researchers to document sampling techniques, consider how metrics can be formulated to evaluate both bias and other factors of generation quality, and inspire more comprehensive studies.18 4.4 Biases from Evaluation Biases can arise from both general evaluations and bias evaluations for NLG tasks. General Evaluations Current standards for NLG evaluation can reinforce certain types of language and penalize others. For example, using perplexity as measured by models pre-trained on datasets largely containing non-AAE text leads to an unfair evaluation of AAE text. Additionally, the subjectivity of generation tasks means that much of NLG evaluation depends on human labels. Since humans from different backgrounds are accustomed to different societal norms and linguistic variations, the choice of human annotators could drastically influence the evaluation standards for generated text. Bias Evaluations It is difficult to evaluate societal biases in NLG tasks because NLG can be open-domain, and there are many different notions of biases from various backgrounds and cultures (Sambasivan et al., 2021). These factors lead to the use of a variety of metrics to evaluate biases (Section 3). To avoid experimental bias in evaluation, we recommend using multiple metrics to cover many types of biases at various granularities. We identify three points to emphasize the need for more comprehensive evaluations. First, most existing works on biases in generation center around one demographic dimension (often gender and from a Western perspective, e.g., using standard Western occupations). While there has been no comprehensive study on whether mitigating biases for one demographic dimension (e.g., gender) may exacerbate biases for others (e.g., race, intersectional identities), this is a possibility we must consider. Second, most works only evaluate bias through a single intermediate proxy; however, different metrics are defined at different granularities (e.g., sentiment is sentence-level, gendered word ratio is word-level). Finally, different evaluation datasets test for specific types of biases and are influenced by the backgrounds of the curators. Collectively evaluating biases across demographic dimensions and granularities can thus help reduce experimentally-biased evaluations. 18Results are summarized in Appendix Tables 2, 3, and 5. 4281 4.5 Biases from Deploying Systems In terms of deploying NLG systems, there is a feedback loop that benefits some communities and further disadvantages others. While this feedback loop is not unique to NLG systems, these systems that directly interact with users make good cautionary examples. First, many deployed language technologies require internet access both to use and contribute feedback, thus favoring the views and languages of those privileged with this access. For example, anyone can contribute feedback to Google Translate, but if contributions and subsequent improvements are focused on high-resource languages, this further increases the accuracy gap between the high and low resource languages, diminishing opportunities for speakers of the low resource languages, i.e., representation disparity (Hashimoto et al., 2018). Second, those who are unable to achieve their goals from using these language technologies (e.g., unsuccessful translation, unhelpful or offensive chat bot) are less likely to continue using the technology. This means that there is less feedback and data to improve the technologies, reinforcing the decreased effectiveness for certain populations, i.e., disparity amplification (Hashimoto et al., 2018). One way we might intervene is to follow a more targeted approach for data and feedback collection, e.g., from excluded populations. However, we acknowledge that this remains a difficult task and that it is also necessary to be aware of “community goals” and other factors in order to co-design language technologies without inflicting additional harm on marginalized populations (Bird, 2020). 5 Progress, Trends, and Challenges Following the discussion of contributors to biases, we survey trends and challenges for reducing biases in NLG. 5.1 Data Methods Data-based methods for both bias analysis and mitigation use the general idea of counterfactual data augmentation (CDA) (Lu et al., 2020) to curate sets of counterfactual prompts. A common method for analysis is using targeted prompts to induce NLG models to reveal biases. For data-based mitigation, existing works focus on fine-tuning large models or training smaller models with datasets that are balanced with respect to targeted demographics. Curated Datasets Existing datasets to study biases in translation include parallel sentences tagged with speaker or subject gender information (Vanmassenhove et al., 2018; Habash et al., 2019) and datasets to study gender biases when translating from neutral references of a person (e.g., nurse in English, gender-neutral pronouns) to gendered instances (e.g., enfermera or enfermero in Spanish, gendered pronouns) (Cho et al., 2019; Stanovsky et al., 2019; Gonen and Webster, 2020; Kocmi et al., 2020). Renduchintala and Williams (2021) additionally provide a dataset to study translation of neutral references in unambiguous contexts. Other works present parallel corpora of biased versus unbiased framings and presuppositions (Pryzant et al., 2020) and AAE versus WAE equivalents (Groenwold et al., 2020). Sheng et al. (2019); Huang et al. (2020); Dhamala et al. (2021) additionally curate sets of prompts that can be used to evaluate biases in autocomplete generation. Bias Analysis Most bias analyses of NLG tasks use prompts to probe for different biases in generated text, e.g., regarding social perception (Sheng et al., 2019), gender in translation (Prates et al., 2019), names (Shwartz et al., 2020), sentiment distribution (Huang et al., 2020), dialects (Groenwold et al., 2020), dialogue personas (Sheng et al., 2021a), or other notions of similarity across demographics (Yeo and Chen, 2020; Henderson et al., 2018). Vig et al. (2020) also use prompts to investigate gender biases, though they do so in the context of a causal mediation analysis. Furthermore, Prates et al. (2019) and Farkas and N´emeth (2020) compare pronoun gender biases in translations (induced with prompts) to real-world statistics. Bias Mitigation Methods can broadly be classified into two categories based on the type of data applied. The first category encompasses methods that fine-tune or train on a balanced dataset to lessen the effects of the model relying on spurious correlations between imbalanced data and task performance. CDA has been applied to datasets used for continued or fresh training in dialogue generation (Dinan et al., 2020a; Liu et al., 2020a) as well as machine translation (Saunders and Byrne, 2020; Costa-juss`a and de Jorge, 2020; Stafanoviˇcs et al., 2020). The second category is methods that attach a short prefix at training time (Vanmassenhove et al., 2018; Basta et al., 2020; Alhafni et al., 2020) or inference time (Moryossef et al., 2019). Challenges The size of state-of-the-art pretrained models and varying definitions of biases 4282 in generation present difficulties for creating standardized datasets that are generally effective across biases and demographics. Moreover, it remains to be seen whether data-based mitigation is as effective for open-domain NLG tasks as it is for more constrained settings. 5.2 Training Methods In addition to data-based mitigation, training-based mitigation is another popular class of methods to reduce biases in generation. Bias Mitigation Several works that use trainingbased mitigation techniques rely on regularization (Bordia and Bowman, 2019; Qian et al., 2019; Huang et al., 2020; Liu et al., 2020a; Saunders and Byrne, 2020). There are also works that induce control by incorporating a bias control code through conditional training (Dinan et al., 2020a), by appending a target value to inputs during training (Ma et al., 2020), by using a normative classifier to produce reward values for backpropagation (Peng et al., 2020), or through adversarial training (Liu et al., 2020b). Other techniques include using debiased word embeddings (Escud´e Font and Costajuss`a, 2019), identifying and editing out subjective words (Pryzant et al., 2020), and using Markov random fields to preserve morpho-syntactic agreement during reinflection (Zmigrod et al., 2019). Challenges The main challenge of bias mitigation through training methods is that it is costly and impractical to re-train models for new biases encountered. In fact, most of the techniques that rely on training from scratch use smaller architectures (exceptions are from larger institutions). 5.3 Inference Methods While the existing literature on inference time methods for bias mitigation is sparse, decoding-based methods are a promising alternative to data- and training-based methods. Specifically, these methods are compatible with any pre-trained language model for generation without additional training. Given recent development of inference-time methods for control that can reduce toxicity (e.g., PPLM (Dathathri et al., 2019), GeDi (Krause et al., 2020), DExperts (Liu et al., 2021)), there is potential for extending these methods to bias mitigation. Bias Mitigation For autocomplete and dialogue generation, Sheng et al. (2020) formulate bias triggers using gradient-based methods of Wallace et al. (2019). These triggers are appended to prompts during inference time to control text generation to be more equalized towards different demographics. For translation, Saunders and Byrne (2020) present a lattice rescoring procedure that creates genderinflected search spaces to rescore text for more accurate translations, and Saunders et al. (2021) subsequently use this lattice structure to present more gendered options during beam search and rerank translation hypotheses according to gender criteria. For dialogue generation, Sheng et al. (2021b) introduce a constrained decoding method that uses n-gram similarity to guide generation away from ad hominems towards marginalized groups. For autocomplete generation, Schick et al. (2021) present a self-debiasing scheme that re-weights word probabilities to generate less undesirable words. Challenges Control methods at inference time could potentially steer the model into degenerate spaces, so it is important to also evaluate these methods for coherence, fluency, and task relevance. 5.4 Evaluation Methods There are two types of evaluations: those that rely on absolute scores and those that rely on relative scores. Absolute score evaluations use an accumulated score to summarize inequalities between demographics, whereas relative evaluations explicitly report inequalities between all demographics. While it is possible to convert between relative and absolute scores, distinguishing between how existing works choose to portray evaluations allows us to examine differences between generation tasks. Absolute Evaluations We find that the transformation class of generation tasks favors bias evaluation through absolute metrics, which is possible because these tasks involve relatively more constrained forms of generation. Examples of evaluation objectives through absolute scores include Peng et al. (2020) reducing non-normative generations, Ma et al. (2020) increasing the accuracy of the change in agency, Zmigrod et al. (2019) increasing the number of correct inflections, Huang et al. (2020) reducing individual and group fairness scores, and Sheng et al. (2021b) reducing the amount of ad hominems towards marginalized groups. Studies of gender bias in machine translation are well-suited to evaluations using absolute scores: many use BLEU and its variants to evaluate correct gender inflections and translations (Moryossef et al., 2019; Escud´e Font and Costajuss`a, 2019; Elaraby et al., 2018; Habash et al., 2019; Alhafni et al., 2020) or accuracy on WinoMT (Saunders and Byrne, 2020; Saunders et al., 2020; 4283 Kocmi et al., 2020; Costa-juss`a and de Jorge, 2020; Costa-juss`a et al., 2020; Basta et al., 2020; Choubey et al., 2021; Saunders et al., 2021). Relative Evaluations In terms of evaluation through relative scores, examples from existing works are mainly from continuation generation tasks. We infer that the less constrained, opendomain nature of continuation generation tasks makes it more preferable to evaluate mitigation through more flexible comparisons rather than absolute scores. For autocomplete generation, Sheng et al. (2019, 2020) and Groenwold et al. (2020) compare regard or sentiment scores across demographics, Shwartz et al. (2020) compare names across various intermediate metrics, Vig et al. (2020) measure proportional differences between the amount of bias under a gendered versus ambiguous reading, and Yeo and Chen (2020) compare occupations generated for different genders. Bias studies in dialogue generation use relative scores by comparing sentiment and offensive language discrepancies (Henderson et al., 2018; Liu et al., 2020a,b) and the percentage of gendered words (Dinan et al., 2020a). Challenges A trade-off between framing biases as a relative or absolute metric is that relative metrics can be more flexibly aligned to normative concerns like social perception. Absolute metrics that look for ratios of gendered words or other indicator words assume that there is a set of words that captures all the differences between demographic groups, regardless of whether these differences are related to normative definitions of harm. There are also absolute metrics such as those of Huang et al. (2020) that can incorporate intermediate metrics that are more aligned with normative behavior, though these metrics reduce the notion of biases to a single value, which could erase historical inequalities between groups. 6 Open Problems and Proposals As a fairly nascent area of exploration, the study of biases in language generation still poses many challenges. Throughout this paper, we discuss challenges associated with different components in a generation pipeline. With a heightened awareness of the relevant body of work, we conclude with recommendations for open problems. Bias-Aware Data Curation Many works have highlighted the harms and problems when collecting training datasets with limited awareness for potential harms. Since effective models for NLG tasks are correlated with increasing training data sizes, biases in data collection (e.g., Englishcentric, drawn from popular Western media) remain a major contributor of biases that manifest in generation. Additionally, datasets used to study biases in generation can also be limited (e.g., only for binary gender classes). For more bias-aware data curation, we suggest diversifying datasets to include more viewpoints from various groups. Understanding Trade-Offs Different methods for analysis, mitigation, and evaluation have unique trade-offs. Existing works have been relatively small-scale and limited to a small number of biases for specific tasks. Some useful questions to consider when developing methods to study generation biases are whether we can generalize methods to a diverse set of biases and a wide range of contexts. It is also important to consider formulating metrics that would jointly mitigate biases and preserve other desired text qualities (e.g., diversity, fluency). Interactive and Continuous Learning The difficulties of measuring and mitigating biases in generation can be reduced with a general framework for interactive and continuous learning. Over time, such a system could learn from diverse opinions of what constitutes “fair” versus “unfair” generations across tasks. A unified framework would centralize and highlight the importance of studying biases in generation, as well as fuel the development of a more comprehensive set of evaluations that may be useful for large-scale studies of impact. Focusing on Negative Impacts Section 3 discusses how there are very few existing works on biases that explicitly and meaningfully engage with resulting negative impacts, even though these impacts are what motivate reducing biases. By reframing efforts on reducing negative impacts rather than biases, we may be able to define metrics and progress that better correlate with reducing harm. For example, relative framings of bias metrics could better enable metrics to be more aligned with reducing harms for particularly impacted groups. Acknowledgments We would like to thank Seraphina Goldfarb-Tarrant, Sunipa Dev, Jason Teoh, members of the Plus Lab, and our anonymous reviewers for the many helpful suggestions that went into this paper. 4284 Ethics and Broader Implications In this work, we present a survey and commentary on the progress and challenges for studying societal biases in language generation. Data We do not check the quality of the datasets used to train popular language generation models (due to limited availability and size), though we do briefly mention problems that other works have found regarding using large datasets that have been minimally filtered. Some of the surveyed datasets and metrics that are used for evaluating biases approximate binary genders using names typical of specific genders, and may be better re-formulated to avoid harms and curate a more accurate representation of different genders. On the subject of genders, the majority of bias evaluation data also only evaluate for binary genders—we point out this issue in our survey as well. Techniques Most of the techniques surveyed in this work are trained with or bias-tested with data drawn from Western sources or culture, since that is largely the focus of the existing body of work. We also refer to studies that point out how techniques for bias do not always transfer across cultures. Our decoding experiments could potentially fuel misuse by giving those with adversarial interests a better understanding of how decoding algorithms could thwart bias metrics, though we believe transparency around these results outweigh the potential for misuse. References Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. arXiv preprint arXiv:2101.05783. Bashar Alhafni, Nizar Habash, and Houda Bouamor. 2020. Gender-aware reinflection using linguistically enhanced neural models. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 139–150, Barcelona, Spain (Online). Association for Computational Linguistics. Solon Barocas, Kate Crawford, Aaron Shapiro, and Hanna Wallach. 2017. The problem with bias: Allocative versus representational harms in machine learning. In 9th Annual Conference of the Special Interest Group for Computing, Information and Society. Christine Basta, Marta R. Costa-juss`a, and Jos´e A. R. Fonollosa. 2020. Towards mitigating gender bias in a decoder-based neural machine translation model by adding contextual information. In Proceedings of the The Fourth Widening Natural Language Processing Workshop, pages 99–102, Seattle, USA. Association for Computational Linguistics. Emily M. Bender and Batya Friedman. 2018. Data statements for natural language processing: Toward mitigating system bias and enabling better science. Transactions of the Association for Computational Linguistics, 6:587–604. Emily M Bender, Timnit Gebru, Angelina McMillanMajor, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big. Proceedings of FAccT. Steven Bird. 2020. Decolonising speech and language technology. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3504–3519, Barcelona, Spain (Online). International Committee on Computational Linguistics. Su Lin Blodgett, Solon Barocas, Hal Daum´e III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of “bias” in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454– 5476, Online. Association for Computational Linguistics. Shikha Bordia and Samuel R. Bowman. 2019. Identifying and reducing gender bias in word-level language models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop, pages 7–15, Minneapolis, Minnesota. Association for Computational Linguistics. Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2020. Extracting training data from large language models. arXiv preprint arXiv:2012.07805. L Elisa Celis and Vijay Keswani. 2020. Dialect diversity in text summarization on twitter. arXiv preprint arXiv:2007.07860. Amanda Cercas Curry, Judy Robertson, and Verena Rieser. 2020. Conversational assistants and gender stereotypes: Public perceptions and desiderata for voice personas. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 72–78, Barcelona, Spain (Online). Association for Computational Linguistics. Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2020. Distilling knowledge learned in BERT for text generation. In Proceedings of the 58th Annual Meeting of the Association for 4285 Computational Linguistics, pages 7893–7905, Online. Association for Computational Linguistics. Won Ik Cho, Ji Won Kim, Seok Min Kim, and Nam Soo Kim. 2019. On measuring gender bias in translation of gender-neutral pronouns. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 173–181, Florence, Italy. Association for Computational Linguistics. Won Ik Cho, Jiwon Kim, Jaeyeong Yang, and Nam Soo Kim. 2021. Towards cross-lingual generalization of translation gender bias. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 449–457. Prafulla Kumar Choubey, Anna Currey, Prashant Mathur, and Georgiana Dinu. 2021. Improving gender translation accuracy with filtered self-training. arXiv preprint arXiv:2104.07695. Marta R Costa-juss`a, Carlos Escolano, Christine Basta, Javier Ferrando, Roser Batlle, and Ksenia Kharitonova. 2020. Gender bias in multilingual neural machine translation: The architecture matters. arXiv preprint arXiv:2012.13176. Marta R. Costa-juss`a and Adri`a de Jorge. 2020. Fine-tuning neural machine translation on genderbalanced datasets. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 26–34, Barcelona, Spain (Online). Association for Computational Linguistics. Kate Crawford. 2017. The trouble with bias. Keynote at NeurIPS. Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978–2988, Florence, Italy. Association for Computational Linguistics. Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2019. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representations. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. Bold: Dataset and metrics for measuring biases in open-ended language generation. Proceedings of FAccT. Emily Dinan, Angela Fan, Adina Williams, Jack Urbanek, Douwe Kiela, and Jason Weston. 2020a. Queens are powerful too: Mitigating gender bias in dialogue generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8173–8188, Online. Association for Computational Linguistics. Emily Dinan, Angela Fan, Ledell Wu, Jason Weston, Douwe Kiela, and Adina Williams. 2020b. Multidimensional gender bias classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 314–331, Online. Association for Computational Linguistics. Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214–226. Mostafa Elaraby, Ahmed Y Tawfik, Mahmoud Khaled, Hany Hassan, and Aly Osama. 2018. Gender aware spoken language translation applied to englisharabic. In 2018 2nd International Conference on Natural Language and Speech Processing (ICNLSP), pages 1–6. IEEE. Joel Escud´e Font and Marta R. Costa-juss`a. 2019. Equalizing gender bias in neural machine translation with word embeddings techniques. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 147–154, Florence, Italy. Association for Computational Linguistics. Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 889–898. Anna Farkas and Ren´ata N´emeth. 2020. How to measure gender bias in machine translation: Optimal translators, multiple reference points. arXiv preprint arXiv:2011.06445. Xavier Ferrer, Tom van Nuenen, Jose M Such, and Natalia Criado. 2021. Discovering and categorising language biases in reddit. In Proceedings of the International AAAI Conference on Web and Social Media, volume 15. Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daum´e III, and Kate Crawford. 2018. Datasheets for datasets. arXiv preprint arXiv:1803.09010. Hila Gonen and Kellie Webster. 2020. Automatically identifying gender issues in machine translation using perturbations. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1991–1995, Online. Association for Computational Linguistics. 4286 Sophie Groenwold, Lily Ou, Aesha Parekh, Samhita Honnavalli, Sharon Levy, Diba Mirza, and William Yang Wang. 2020. Investigating AfricanAmerican Vernacular English in transformer-based text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5877–5883, Online. Association for Computational Linguistics. Nizar Habash, Houda Bouamor, and Christine Chung. 2019. Automatic gender identification and reinflection in Arabic. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 155–165, Florence, Italy. Association for Computational Linguistics. Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equality of opportunity in supervised learning. In Advances in neural information processing systems, pages 3315–3323. Tatsunori Hashimoto, Megha Srivastava, Hongseok Namkoong, and Percy Liang. 2018. Fairness without demographics in repeated loss minimization. In International Conference on Machine Learning, pages 1929–1938. PMLR. Peter Henderson, Koustuv Sinha, Nicolas AngelardGontier, Nan Rosemary Ke, Genevieve Fried, Ryan Lowe, and Joelle Pineau. 2018. Ethical challenges in data-driven dialogue systems. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pages 123–129. Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. In International Conference on Learning Representations. Dirk Hovy, Federico Bianchi, and Tommaso Fornaciari. 2020. “you sound just like your father” commercial machine translation systems include stylistic biases. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1686–1690, Online. Association for Computational Linguistics. Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, and Pushmeet Kohli. 2020. Reducing sentiment bias in language models via counterfactual evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 65–83, Online. Association for Computational Linguistics. Clayton Hutto and Eric Gilbert. 2014. Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Proceedings of the International AAAI Conference on Web and Social Media, volume 8. Matthew Kay, Cynthia Matuszek, and Sean A Munson. 2015. Unequal representation and gender stereotypes in image search results for occupations. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, pages 3819– 3828. Svetlana Kiritchenko and Saif Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 43–53. Svetlana Kiritchenko, Isar Nejadgholi, and Kathleen C Fraser. 2020. Confronting abusive language online: A survey from the ethical and human rights perspective. arXiv preprint arXiv:2012.12305. Hannah Kirk, Yennie Jun, Haider Iqbal, Elias Benussi, Filippo Volpin, Frederic A Dreyer, Aleksandar Shtedritski, and Yuki M Asano. 2021. How true is gpt2? an empirical analysis of intersectional occupational biases. arXiv preprint arXiv:2102.04130. Tom Kocmi, Tomasz Limisiewicz, and Gabriel Stanovsky. 2020. Gender coreference and bias evaluation at WMT 2020. In Proceedings of the Fifth Conference on Machine Translation, pages 357–364, Online. Association for Computational Linguistics. Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2020. Gedi: Generative discriminator guided sequence generation. arXiv preprint arXiv:2009.06367. Sharon Levy, Michael Saxon, and William Yang Wang. 2021. The truth is out there: Investigating conspiracy theories in text generation. In Findings of The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing. Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A Smith, and Yejin Choi. 2021. On-the-fly controlled text generation with experts and anti-experts. In The Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing. Haochen Liu, Jamell Dacon, Wenqi Fan, Hui Liu, Zitao Liu, and Jiliang Tang. 2020a. Does gender matter? towards fairness in dialogue systems. In Proceedings of the 28th International Conference on Computational Linguistics, pages 4403–4416, Barcelona, Spain (Online). International Committee on Computational Linguistics. Haochen Liu, Wentao Wang, Yiqi Wang, Hui Liu, Zitao Liu, and Jiliang Tang. 2020b. Mitigating gender bias for neural dialogue generation with adversarial learning. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 893–903, Online. Association for Computational Linguistics. 4287 Kaiji Lu, Piotr Mardziel, Fangjing Wu, Preetam Amancharla, and Anupam Datta. 2020. Gender bias in neural natural language processing. In Logic, Language, and Security, pages 189–202. Springer. Li Lucy and David Bamman. 2021. Gender and representation bias in gpt-3 generated stories. In Proceedings of the Third Workshop on Narrative Understanding, pages 48–55. Xinyao Ma, Maarten Sap, Hannah Rashkin, and Yejin Choi. 2020. PowerTransformer: Unsupervised controllable revision for biased language correction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7426–7441, Online. Association for Computational Linguistics. Amit Moryossef, Roee Aharoni, and Yoav Goldberg. 2019. Filling gender & number gaps in neural machine translation with black-box context injection. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 49–54, Florence, Italy. Association for Computational Linguistics. Debora Nozza, Federico Bianchi, and Dirk Hovy. 2021. Honest: Measuring hurtful sentence completion in language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2398–2406. Thuy Ong. 2017. Facebook apologizes after wrong translation sees Palestinian man arrested for posting ’good morning’. Ji Ho Park, Jamin Shin, and Pascale Fung. 2018. Reducing gender bias in abusive language detection. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2799–2804. Amandalynne Paullada, Inioluwa Deborah Raji, Emily M Bender, Emily Denton, and Alex Hanna. 2020. Data and its (dis) contents: A survey of dataset development and use in machine learning research. arXiv preprint arXiv:2012.05345. Xiangyu Peng, Siyan Li, Spencer Frazier, and Mark Riedl. 2020. Reducing non-normative text generation from language models. In Proceedings of the 13th International Conference on Natural Language Generation, pages 374–383, Dublin, Ireland. Association for Computational Linguistics. Marcelo OR Prates, Pedro H Avelar, and Lu´ıs C Lamb. 2019. Assessing gender bias in machine translation: a case study with google translate. Neural Computing and Applications, pages 1–19. Reid Pryzant, Richard Diehl Martinez, Nathan Dass, Sadao Kurohashi, Dan Jurafsky, and Diyi Yang. 2020. Automatically neutralizing subjective bias in text. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 480–489. Yusu Qian, Urwa Muaz, Ben Zhang, and Jae Won Hyun. 2019. Reducing gender bias in word-level language models with a gender-equalizing loss function. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 223–228, Florence, Italy. Association for Computational Linguistics. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:1–67. Adithya Renduchintala and Adina Williams. 2021. Investigating failures of automatic translation in the case of unambiguous gender. arXiv preprint arXiv:2104.07838. Nicholas Roberts, Davis Liang, Graham Neubig, and Zachary C Lipton. 2020. Decoding and diversity in machine translation. arXiv preprint arXiv:2011.13477. Nithya Sambasivan, Erin Arnesen, Ben Hutchinson, Tulsee Doshi, and Vinodkumar Prabhakaran. 2021. Re-imagining algorithmic fairness in india and beyond. Proceedings of FAccT. Danielle Saunders and Bill Byrne. 2020. Reducing gender bias in neural machine translation as a domain adaptation problem. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7724–7736, Online. Association for Computational Linguistics. Danielle Saunders, Rosie Sallis, and Bill Byrne. 2020. Neural machine translation doesn’t translate gender coreference right unless you make it. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 35–43, Barcelona, Spain (Online). Association for Computational Linguistics. Danielle Saunders, Rosie Sallis, and Bill Byrne. 2021. First the worst: Finding better gender translations during beam search. arXiv preprint arXiv:2104.07429. Beatrice Savoldi, Marco Gaido, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2021. Gender bias in machine translation. In Transactions of the Association for Computational Linguistics. 4288 Timo Schick, Sahana Udupa, and Hinrich Sch¨utze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in nlp. arXiv preprint arXiv:2103.00453. Deven Santosh Shah, H. Andrew Schwartz, and Dirk Hovy. 2020. Predictive biases in natural language processing models: A conceptual framework and overview. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5248–5264, Online. Association for Computational Linguistics. Emily Sheng, Josh Arnold, Zhou Yu, Kai-Wei Chang, and Nanyun Peng. 2021a. Revealing persona biases in dialogue systems. arXiv preprint arXiv:2104.08728. Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2019. The woman worked as a babysitter: On biases in language generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3398–3403. Emily Sheng, Kai-Wei Chang, Prem Natarajan, and Nanyun Peng. 2020. Towards Controllable Biases in Language Generation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3239–3254, Online. Association for Computational Linguistics. Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021b. ”nice try, kiddo”: Investigating ad hominems in dialogue responses. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Emily Sheng and David Uthus. 2020. Investigating societal biases in a poetry composition system. In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 93–106, Barcelona, Spain (Online). Association for Computational Linguistics. Vered Shwartz, Rachel Rudinger, and Oyvind Tafjord. 2020. “you are grounded!”: Latent name artifacts in pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6850–6861, Online. Association for Computational Linguistics. Andrew Silva, Pradyumna Tambwekar, and Matthew Gombolay. 2021. Towards a comprehensive understanding and accurate evaluation of societal biases in pre-trained transformers. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2383–2389. Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, et al. 2019. Release strategies and the social impacts of language models. arXiv preprint arXiv:1908.09203. Art¯urs Stafanoviˇcs, M¯arcis Pinnis, and Toms Bergmanis. 2020. Mitigating gender bias in machine translation with target gender annotations. In Proceedings of the Fifth Conference on Machine Translation, pages 629–638, Online. Association for Computational Linguistics. Gabriel Stanovsky, Noah A. Smith, and Luke Zettlemoyer. 2019. Evaluating gender bias in machine translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1679–1684, Florence, Italy. Association for Computational Linguistics. Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1630–1640, Florence, Italy. Association for Computational Linguistics. Tony Sun, Kellie Webster, Apu Shah, William Yang Wang, and Melvin Johnson. 2021. They, them, theirs: Rewriting with gender-neutral english. arXiv preprint arXiv:2102.06788. Alex Tamkin, Miles Brundage, Jack Clark, and Deep Ganguli. 2021. Understanding the capabilities, limitations, and societal impact of large language models. arXiv preprint arXiv:2102.02503. Marcus Tomalin, Bill Byrne, Shauna Concannon, Danielle Saunders, and Stefanie Ullmann. 2021. The practical ethics of bias reduction in machine translation: why domain adaptation is better than data debiasing. Ethics and Information Technology, pages 1–15. Eva Vanmassenhove, Christian Hardmeier, and Andy Way. 2018. Getting gender right in neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3003–3008, Brussels, Belgium. Association for Computational Linguistics. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. Jesse Vig, Sebastian Gehrmann, Yonatan Belinkov, Sharon Qian, Daniel Nevo, Yaron Singer, and Stuart Shieber. 2020. Investigating gender bias in language models using causal mediation analysis. Advances in Neural Information Processing Systems, 33. Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. 2019. Universal adversarial triggers for attacking and analyzing NLP. In Proceedings of the 2019 Conference on Empirical Methods 4289 in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2153–2162, Hong Kong, China. Association for Computational Linguistics. Alex Wang and Kyunghyun Cho. 2019. BERT has a mouth, and it must speak: BERT as a Markov random field language model. In Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation, pages 30–36, Minneapolis, Minnesota. Association for Computational Linguistics. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5753–5763. Catherine Yeo and Alyssa Chen. 2020. Defining and evaluating fair natural language generation. In Proceedings of the The Fourth Widening Natural Language Processing Workshop, pages 107–109, Seattle, USA. Association for Computational Linguistics. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT : Largescale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270– 278, Online. Association for Computational Linguistics. Jieyu Zhao, Yichao Zhou, Zeyu Li, Wei Wang, and KaiWei Chang. 2018. Learning gender-neutral word embeddings. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4847–4853, Brussels, Belgium. Association for Computational Linguistics. Ran Zmigrod, Sabrina J. Mielke, Hanna Wallach, and Ryan Cotterell. 2019. Counterfactual data augmentation for mitigating gender stereotypes in languages with rich morphology. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1651–1661, Florence, Italy. Association for Computational Linguistics. 4290 A Appendices A.1 Evaluating Biases Across Decoding Techniques and Metrics To gain more insight into biases from different decoding techniques, we examine autocomplete generations from GPT (110M params), GPT-2 (small, 117M params), and XLNet (base, 110M params), using the decoding techniques described in Section 4.3 through the Transformers19 library. We use standard parameters of b = 16 for beam search, k = 40 with a temperature of 0.7 for top-k sampling, and p = 0.95 for nucleus sampling (Holtzman et al., 2019). In terms of bias metrics, we use existing NLG bias metrics: regard ratio (Sheng et al., 2019), sentiment ratio (Groenwold et al., 2020), individual and group fairness through sentiment (IF/GF) (Huang et al., 2020), and a gendered word co-occurrence scores (Bordia and Bowman, 2019). For all sentiment scores, we use the rule-based sentiment analyzer, VADER (Hutto and Gilbert, 2014).20 We run all our experiments on an RTX 2080Ti GPU. Generation takes from a couple of minutes to a few hours, depending on the number of samples generated. Regard Ratios Sheng et al. (2019) introduce 10 prompts to induce occupation- and respect-related generations (e.g., [BLANK] worked as, [BLANK] was thought of as) and six demographics (Black, White, man, woman, gay, straight) to fill in the [BLANK], for a total of 60 prompts. The authors define regard as the social perception towards a demographic, collect human annotations, and release a BERT-based regard classifier.21 We follow the original work in reporting percentages of negative, neutral, and positive regard scores per demographic. For the deterministic search methods, we do not report scores since there are only 10 samples per demographic. For the stochastic sampling methods, we generate 1000 samples per demographic. Additionally, we use the regard classifier released by the authors for our evaluations—while we acknowledge that this classifier could also have biases, we believe it is still worthwhile to use it to compare text generated from different decoding techniques. 19https://huggingface.co/transformers 20Kiritchenko and Mohammad (2018) show that sentiment classifiers can exhibit biases. We use VADER since 1) it does not rely on learned associations and thus may be less prone to biases, and 2) it has been used to measure biases in previous works (Sheng et al., 2019; Groenwold et al., 2020). 21https://github.com/ewsheng/nlg-bias Sentiment Ratios for AAE and WAE Prompts Groenwold et al. (2020) curate a parallel set of 2,019 AAE and 2,019 WAE prompts and use sentiment classifiers to label text generated from the prompts. Similar to Sheng et al. (2019), this work also reports percentages of negative, neutral, and positive scores. The VADER sentiment analyzer that we use reports scores in the range of [-1, 1]. When reporting ratios, we use splits recommended by the authors (Hutto and Gilbert, 2014) to categorize sentiment values into negative (value<=−0.05), neutral (−0.05<value<0.05), and positive (value>=0.05) bins. When reporting average values, we calculate from the unrounded scores from VADER. We generate one sample per prompt for all decoding techniques. Individual and Group Fairness Through Sentiment Huang et al. (2020) evaluate fairness across countries, occupations, and genders (binary, as defined through Western names typical of a gender) by first defining 10 templates per dimension (e.g., People from [BLANK] are). For each dimension, they also define a list of dimension instances (e.g., Syria as a country) to fill in the [BLANK]. In total, there are 730 prompts across the three attributes. For our experiments, we generate one sample per prompt. The authors define the individual fairness metric by “...averaging the Wasserstein-1 distance between the sentiment score distribution of every evaluation sentence and each of its counterfactual sentences across all templates.” For example, we would compute the distance between the sentiment distributions of the text generated from the template People from [BLANK] are for each of the country choices for [BLANK], and sum up the distance scores for all pairs across all templates. For group fairness, the authors calculate the average of the “Wasserstein-1 distance between the sentiment distributions of all generated sentences of inputs from [a] subgroup, and that over the entire evaluation set”. Here, a subgroup means each country, occupation, or binary gender. For example, we compare the distance between the sentiment distribution of text generated for Syria (across all templates) and the sentiment distribution of text generated for all countries. We use Huang et al. (2020)’s prefix templates and fairness metrics exactly as defined in the original work, so we refer readers to the original work for more details. 4291 Gendered Word Co-occurrence Scores This score is based on the one proposed by Bordia and Bowman (2019), though we use different gendered word lists and evaluate over all text generated for the other bias metrics, downsampling if necessary so that the amount and sources of generated text are consistent across decoding techniques. First, we obtain the lists of female words and male words from Zhao et al. (2018) and add gendered pronouns (he, she, his, him, her) to the respective lists. For each word in the aggregated sample set, we calculate the probability of the word given any of the female words (in a context window of 20 words before and after a word) and similarly the probability of the word given any of the male words. We then take the absolute value of the log ratio of the first probability to the second, and report the average and standard deviation across all nongendered words. More concretely, given the set of female gendered words f, the set of male gendered words m, unique non-gendered words w ∈W in a dataset, and the probability of a word given any of the set g of gendered words P(w|g), we calculate the mean µ = avg(abs(log P(w|f) P(w|m))) and standard deviation σ = stdev(abs(log P(w|f) P(w|m))). Supplementary Results Supplementary to the experimental results described in the main text, Table 2 presents quantitative results. Table 3 shows regard ratios for the other demographic groups originally included in the evaluation by Sheng et al. (2019). Additionally, Table 4 presents average lengths and vocabulary sizes of the samples used in the IF/GF evaluations to estimate text diversity. These results, combined with examples of generated text in Table 5, provide evidence that the decoding techniques differ in terms of generated text diversity, and that diversity is very much correlated with the bias metrics IF, GF, and gendered word co-occurrence scores. Although this correlation is to be expected from the metric formulation, this study raises relevant questions of whether bias metrics should be correlated with text diversity, and whether bias evaluations should use more comprehensive metrics. 4292 Model Decode Regard Sentiment IF ↓ GF ↓Gendered Score ↓ Black White AAE WAE GPT Greedy 13-73-14(0.01) 17-67-16(0.01) 0.15 0.09 1.98±2.34 Beam 10-77-13(0.01) 13-71-16(0.03) 0.12 0.07 1.91±2.35 Top-k 33-55-12(-0.20) 22-55-23(0.01) 13-70-17(0.02) 16-63-21(0.03) 0.27 0.09 2.07±2.32 Nucleus 35-53-12(-0.23) 30-54-16(-0.14) 16-63-21(0.03) 18-59-23(0.02) 0.33 0.10 2.10±2.28 GPT-2 Greedy 15-63-22(0.03) 14-64-23(0.06) 0.19 0.07 1.91±2.39 Beam 14-67-18(0.02) 12-70-18(0.04) 0.19 0.07 1.90±2.45 Top-k 35-49-16(-0.19) 24-48-28(0.04) 17-57-26(0.05) 17-57-26(0.06) 0.32 0.10 2.00±2.36 Nucleus 46-42-12(-0.33) 36-45-19(-0.16) 20-49-31(0.06) 17-54-29(0.06) 0.36 0.12 2.00±2.27 XLNet Greedy 09-76-15(0.03) 11-68-21(0.05) 0.13 0.09 1.89±2.34 Beam 04-88-08(0.02) 06-83-11(0.03) 0.08 0.04 1.85±2.31 Top-k 23-63-14(-0.10) 14-69-17(0.02) 10-72-19(0.05) 13-61-26(0.07) 0.27 0.10 1.96±2.30 Nucleus 35-49-16(-0.20) 29-56-14(-0.15) 14-63-23(0.05) 15-58-27(0.06) 0.30 0.11 1.97±2.27 Table 2: Bias evaluations for various decoding algorithms, models, and metrics. Regard scores (Sheng et al., 2019) and sentiment scores (Groenwold et al., 2020) are reported in distribution percentages of negative-neutralpositive(avg value). Individual fairness (IF) and group fairness (GF) scores (Huang et al., 2020) compare sentiment distributions of generated text across demographics. Gendered (word co-occurrence) scores are reported in terms of mean±stdev of the absolute log ratio of the probabilities: P(word|female terms) to P(word|male terms) (Bordia and Bowman, 2019). Search-based results for regard are omitted due to lack of enough prompts to generate from. Results indicate 1) nucleus sampling generates more text with negative regard, 2) decoding choices are similar for AAE/WAE sentiments though sampling generates more positive sentiment overall, 3) beam search has relatively lower bias as measured by IF, GF, and gendered word co-occurrence scores, followed closely by greedy search, and then top-k and nucleus sampling. Model Decoding Demographic Scores GPT Top-k man 24-51-25(0.01) woman 21-52-27(0.06) gay 31-52-17(-0.14) straight 22-54-24(0.02) Nucleus man 33-50-17(-0.16) woman 29-53-18(-0.11) gay 38-48-13(-0.25) straight 29-54-17(-0.13) GPT-2 Top-k man 31-48-21(-0.09) woman 21-49-30(0.10) gay 53-32-15(-0.39) straight 18-49-33(0.15) Nucleus man 36-47-17(-0.20) woman 30-54-17(-0.13) gay 53-35-11(-0.42) straight 31-50-20(-0.11) XLNet Top-k man 24-54-22(-0.02) woman 12-63-25(0.14) gay 50-44-06(-0.44) straight 21-55-24(0.03) Nucleus man 28-55-16(-0.12) woman 24-57-20(-0.04) gay 43-45-11(-0.32) straight 26-55-20(-0.06) Table 3: Regard score bias evaluation results across decoding techniques for demographics: man, woman, gay, and straight, reported in distribution percentages of negative-neutral-positive(avg value). Model Decoding Avg. Length Vocab Size GPT Greedy 11.4 440 Beam 10.2 349 Top-k 12.9 1,235 Nucleus 14.3 2,074 GPT-2 Greedy 15.8 880 Beam 15.1 845 Top-k 17.7 2,117 Nucleus 18.2 3,443 XLNet Greedy 12.1 537 Beam 8.1 217 Top-k 14.6 1,685 Nucleus 16.4 2,991 Table 4: Quantitative values to estimate text diversity: Average lengths (in words) of IF/GF evaluation samples (prompt + generated text) and vocabulary size across all samples show that generated text diversity increases for decoding techniques in the following order: beam, greedy, top-k, nucleus. 4293 Model Decoding Metric Generated Text GPT Greedy Regard The Black person was known for his ability to read minds. Sentiment this lil girl aint going to win im the king of this world. IF/GF Raven is known for her ability to read people. Beam Regard The Black person was known for. Sentiment this lil girl aint going to win im the king of this f**kin’ world. IF/GF Raven is known for her ability to read people. Top-k Regard The Black person was known for its vicious, ruthless nature. Sentiment this lil girl aint going to win im the king of this place IF/GF Raven is known for her courage and resilience. Nucleus Regard The Black person was known for giving a disposition when something had a forced birth. Sentiment this lil girl aint going to win im the king of this land, everyone is goin to go sit down and see mister krampus have his say. IF/GF Raven is known for her archery skills. GPT-2 Greedy Regard The Black person was known for his ability to manipulate the minds of others. Sentiment this lil girl aint going to win im the king of this world. IF/GF Raven is known for her ability to create a new type of weapon, and she has been known to use it to create new weapons. Beam Regard The Black person was known for his blackness, but he was also known for his blackness. Sentiment this lil girl aint going to win im the king of this world. IF/GF Raven is known for her ability to turn the tide of battle. Top-k Regard The Black person was known for his dark hair and the way he’d look at others. Sentiment this lil girl aint going to win im the king of this country” she said. IF/GF Raven is known for her ability to use her magic to transform the wind into a giant windstorm. Nucleus Regard The Black person was known for his terrible behavior after many years of marriage. Sentiment this lil girl aint going to win im the king of this country’, she chided, gung-ho. IF/GF Raven is known for her dislike of Marge, while she also revealed to Laura that she often wondered why, when she was first told about Kate’s mission, the weather was always horrible. XLNet Greedy Regard The Black person was known for his “black” skin. Sentiment this lil girl aint going to win im the king of this world. IF/GF Raven is known for her ””””””””””””””””””” Beam Regard The Black person was known for their. Sentiment this lil girl aint going to win im the king of this world. IF/GF Raven is known for her work. Top-k Regard The Black person was known for being a bit of an “indie”. Sentiment this lil girl aint going to win im the king of this country. IF/GF Raven is known for her “The Big Bad” nickname, “Big Bad”. Nucleus Regard The Black person was known for his icy attitude to his fellow human beings, and they were ungraded by their race. Sentiment this lil girl aint going to win im the king of this 1% double wind. IF/GF Raven is known for her intelligence, and her reactions to celebrant trends were characterized by high. Table 5: Samples generated from different decoding techniques for different bias metrics. The sampling techniques result in more diverse generated text.
|
2021
|
330
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4294–4309 August 1–6, 2021. ©2021 Association for Computational Linguistics 4294 Reservoir Transformers Sheng Shen†, Alexei Baevski‡, Ari S. Morcos‡, Kurt Keutzer†, Michael Auli‡, Douwe Kiela‡ †UC Berkeley; ‡Facebook AI Research [email protected], [email protected] Abstract We demonstrate that transformers obtain impressive performance even when some of the layers are randomly initialized and never updated. Inspired by old and well-established ideas in machine learning, we explore a variety of non-linear “reservoir” layers interspersed with regular transformer layers, and show improvements in wall-clock compute time until convergence, as well as overall performance, on various machine translation and (masked) language modelling tasks. 1 Introduction Transformers (Vaswani et al., 2017) have dominated natural language processing (NLP) in recent years, from large scale machine translation (Ott et al., 2018) to pre-trained (masked) language modeling (Devlin et al., 2018; Radford et al., 2018), and are becoming more popular in other fields as well, from reinforcement learning (Vinyals et al., 2019) to speech recognition (Baevski et al., 2019) and computer vision (Carion et al., 2020). Their success is enabled in part by ever increasing computational demands, which has naturally led to an increased interest in improving their efficiency. Scalability gains in transformers could facilitate bigger, deeper networks with longer contexts (Kitaev et al., 2020; Wang et al., 2020; Beltagy et al., 2020; Kaplan et al., 2020; Tay et al., 2020b). Conversely, improved efficiency could reduce environmental costs (Strubell et al., 2019) and hopefully help democratize the technology. In this work, we explore a simple question: if some layers of the transformer are kept frozen— i.e., never updated after random initialization— can we match the performance of fully learned transformers, while being more efficient? Surprisingly, the answer is resoundingly yes; and what is more, we find that freezing layers may actually improve performance. Beyond desirable efficiency gains, random layers are interesting for several additional reasons. Fixed randomly initialized networks (Gallicchio and Scardapane, 2020) converge to Gaussian processes in the limit of infinite width (Daniely et al., 2016), have intriguing interpretations in metric learning (Rosenfeld and Tsotsos, 2019; Giryes et al., 2016), and have been shown to provide excellent “priors” either for subsequent learning (Ulyanov et al., 2018) or pruning (Frankle and Carbin, 2018). Fixed layers allow for efficient low-cost hardware implementations (Schrauwen et al., 2007) and can be characterized using only a random number generator and its seed. This could facilitate distributed training and enables highly efficient deployment to edge devices, since it only requires transmission of a single number. The strong performance of networks with fixed layers also sheds new light on the inner workings of BERT (Devlin et al., 2018), and layer-wise interpretations of such models (Rogers et al., 2020; Tenney et al., 2019). It appears that “not all layers are created equal” (Zhang et al., 2019) is true to such an extent that some layers can simply remain random and fixed. Random projections have a long history in machine learning. By Cover’s theorem (Cover, 1965), any high-dimensional non-linear transformation is more likely to be linearly separable than its lower-or-equal-dimensional input space. By Johnson-Lindenstrauss (Johnson and Lindenstrauss, 1984), random projections distort Euclidean distances very little under mild assumptions, which is useful e.g. for dimensionality reduction and random indexing (Sahlgren, 2005). Fixed random layers in neural networks pre-date deep learning by far (Gamba et al., 1961; Baum, 1988). Indeed, random kernel methods have long 4295 been influential in machine learning (Rahimi and Recht, 2008, 2009). One way to think of such layers is as “reservoirs” (Lukoˇseviˇcius and Jaeger, 2009), where a highly non-linear high-dimensional black box representation is provided to a lightweight “readout” network, as in echo state networks (Jaeger, 2003) and liquid state machines (Maass et al., 2002). The benefit of such an approach is that the reservoir has fixed parameters and is computationally efficient, as it can be pre-computed and does not (necessarily) require backpropagation. In NLP, Wieting and Kiela (2019) showed that random sentence encoders present a strong baseline for text classification, with subsequent work showing applications in a variety of tasks from summarization to machine translation (Enguehard et al., 2019; Garg et al., 2020; Pilault et al., 2020). To our knowledge, this work is the first to examine this phenomenon in transformers, and the first to recursively alternate reservoirs with subsequent transformer layers acting as readout functions. We introduce “reservoir transformers”, wherein fixed random reservoir layers are interspersed with regular updateable transformer layers. The goal of this work is to put our understanding of transformer models on a more solid footing by providing empirical evidence of their capabilities even when some of their parameters are fixed. Our contributions are as follows: • We introduce a area under the convergence curve metric for measuring performanceefficiency trade-offs, and show that replacing regular transformer layers with reservoir layers leads to improvements. • We show that the addition of reservoir layers leads to improved test set generalization on a variety of tasks in a variety of settings. • We show that pre-trained masked language modelling architectures like BERT and RoBERTa (Liu et al., 2019) can benefit from having some of their layers frozen, both during pre-training as well as when fine-tuning on downstream tasks. • We experiment with different types of reservoir layers, including convolutional and recurrent neural network-based ones. • We show empirical evidence that the backward pass can be skipped in its entirety by approximating upstream gradients using an approach we call backskipping, which can reduce the training compute further without sacrificing performance. 2 Approach This paper is based on a very simple idea. Neural networks are trained via backpropagation, which involves consecutive steps of matrix addition and multiplication, i.e., θt+1 ←θt −η ∂J ∂θt ; ∂J ∂θt = ∂J ∂Ln ∂Ln ∂Ln−1 · · · ∂L0 ∂x for some objective J, parameterization θ and learning rate η, with the gradient computed via the chain rule, where Li is the i-th layer of the neural network and x is the input. Let L = Transformer(X) be a single layer in a Transformer network (Vaswani et al., 2017), i.e., H = MultiHeadSelfAttn(LayerNorm(X)) + X L = FFN(LayerNorm(H)) + H Now, during every “backward pass”, we compute the Jacobian for parameters θL at layer L, which are used to update the parameters of L, θL t , as well as to compute the next layer’s Jacobian, thus back-propagating the gradients. In this work however, for some of the layers, we still backpropagate through them to compute gradients for earlier layers, but we never apply the parameter update. As a result, these layers stay fixed at their initialization, saving computational resources. 2.1 Background Naturally, never updating some of the parameters is computationally more efficient, as some matrix addition operations can be skipped in the backward pass, but why is this not detrimental to the performance of the network? In the early days of neural networks, the bottom layers were often kept fixed as “associators” (Block, 1962), or what (Minsky and Papert, 2017) called the Gamba perceptron (Gamba et al., 1961; Borsellino and Gamba, 1961). Fixed random networks (Baum, 1988; Schmidt et al., 1992; Pao et al., 1994) have been explored from many angles, including as “random kitchen sink” kernel machines (Rahimi and Recht, 2008, 2009), “extreme learning machines” (Huang et al., 2006) and 4296 reservoir computing (Jaeger, 2003; Maass et al., 2002; Lukoˇseviˇcius and Jaeger, 2009). In reservoir computing, input data are represented through fixed random high-dimensional non-linear representations, called “reservoirs”, which are followed by a regular (often but not necessarily linear) “readout” network to make the final classification decision. The theoretical justification for these approaches lies in two well-known results in machine learning: Cover’s theorem (Cover, 1965) on the separability of patterns states that highdimensional non-linear transformations are more likely to be linearly separable; and the JohnsonLindenstrauss lemma (Johnson and Lindenstrauss, 1984) shows that (most) random projections distort Euclidean distances very little. Practically, random layers can be seen as a cheap way to increase network depth. There are interesting advantages to this approach. Fixed layers are known to have particularly low-cost hardware requirements and can be easily implemented on high-bandwidth FPGAs with low power consumption (Hadaeghi et al., 2017; Tanaka et al., 2019), or on optical devices (Hicke et al., 2013). This might yield interesting possibilities for training in a distributed fashion across multiple devices, as well as for neuromorphic hardware (Neftci et al., 2017). This approach also facilitates lower-latency deployment of neural networks to edge devices, since weights can be shared simply by sending the seed number, assuming the random number generator is known on both ends. 2.2 Reservoir Transformers This work explores inserting random non-linear transformations, or what we call reservoir layers, into transformer networks. Specifically, we experiment with a variety of reservoir layers: • Transformer Reservoir: The standard transformer layer as described above, but with all parameters fixed after initialization, including the self-attention module. • FFN Reservoir: A transformer-style fixed feed-forward layer without any self-attention, i.e., FFN(LayerNorm(Previous layer)) + Previous layer. • BiGRU Reservoir: A fixed bidirectional Gated Recurrent Unit (Cho et al., 2014) layer, which is closer in spirit to previous work on reservoir computing, most of which builds on recurrent neural network architectures. • CNN Reservoir: A fixed Convolutional Neural Network (LeCun et al., 1998) layer, specifically light dynamical convolution layers (Wu et al., 2019), which are known to be competitive with transformers in sequenceto-sequence tasks. We find that all these approaches work well, to a certain extent. For clarity, we focus primarily on the first two reservoir layers, but include a broader comparison in Appendix A. In each case, contrary to traditional reservoir computing, our reservoir layers are interspersed throughout a regular transformer network, or what we call a reservoir transformer. Since random projections are not learned and might introduce noise, subsequent normal transformer “readout” layers might be able to benefit from additional depth while allowing us to recover from any adverse effects of randomness. For example, previous work has shown that ResNets, with all of their parameters fixed except for the scale and shift parameters of batch normalization, can still achieve high performance, simply by scaling and shifting random features (Frankle et al., 2020). Adding some form of noise to the parameters is also known to help convergence and generalization (Jim et al., 1995, 1996; Gulcehre et al., 2016; Noh et al., 2017). 3 Evaluation We evaluate the proposed approach on a variety of well-known tasks in natural language processing, namely: machine translation, language modelling and masked language model pre-training. We set out to do this work with the main objective of examining any potential efficiency gains, i.e. the relationship between compute time and task performance. This is closely related to efforts in Green AI, which are concerned with the trade-offs between compute, data, and performance (Schwartz et al., 2019). We propose to measure this trade-off via the area under the convergence curve (AUCC): similarly to how the area under the receiver operating characteristic (Bradley, 1997, AUC-ROC) measures a classifier’s performance independent of the classification threshold, AUCC measures a model’s performance independent of the specific compute bud4297 get. Specifically, AUCC is computed as follows: Z ˆT t=0 X x,y∈D gt(f(x), y) (1) where f is the network and g is the evaluation metric, measured until convergence time ˆT, which is the maximum convergence time of all models included in the comparison. Note that time here is wall-clock time, not iterations. By convergence, we mean that validation performance has stopped improving, and hence the convergence curve whose area we measure plots the desired metric over time. Runs are averaged over multiple seeds and reported with standard deviation. We normalize raw AUCC scores by their maximum to ensure a more interpretable [0 −1] range. One potential downside of this approach is that the AUCC metric could lead to higher scores for a model that converges quickly but to ultimately worse performance, if measured in a small window. This can be solved by making sure that ˆT is set sufficiently high. We include the raw validation curves in the appendix to demonstrate that the chosen window sizes are sufficient and the results are not a influenced by this limitation. In addition, we report the number of trainable parameters and the wall-clock training time until maximum performance (plus 95% and 99% convergence results in the appendix). Finally, we show test set generalization in each experiment. Overall, this gives us a wide set of axes along which to examine models. 3.1 Experimental Settings We evaluate on IWSLT de-en (Cettolo et al., 2015) and WMT en-de (Bojar et al., 2014) for machine translation; enwiki8 (LLC, 2009) for language modelling; and experiment with RoBERTa (Liu et al., 2019) in our pretraining experiments. For IWSLT, we follow the pre-processing steps in Edunov et al. (2018). The train/val/test split is 129k/10k/6.8k sentences. For WMT, we follow pre-process as in Ott et al. (2018), with 4.5M/16.5k/3k sentences in train/val/test. For enwiki8, we follow the pre-processing steps in Dai et al. (2019). The train/val/test split is 1M/54k/56k sentences. For RoBERTa pretraining, we follow the pre-processing steps in Liu et al. (2019). We use 8 Volta V100 GPUs for WMT and enwik8, 32 V100 GPUs for RoBERTa and a single V100 for IWSLT. The hyperparameters for IWSLT14 and WMT16 were set to the bestperforming values from Ott et al. (2018) and Kasai et al. (2020) respectively. The enwik8 experiment settings followed Bachlechner et al. (2020) and the RoBERTa experiments followed Liu et al. (2019). All the experiments in this paper were run with 3 random seeds and the mean and standard deviation are reported. For the relatively small IWSLT, the ˆT value in the AUCC metric was set to 4 hours. For the larger WMT, we set it to 20 hours. For enwiki8, it was 30 hours; and for the RoBERTa pre-training experiments, it was set to 60 hours. The projection weights in random layers were initialized using orthogonal initialization (Saxe et al., 2013), since random orthogonal projections should ideally be maximally informationpreserving, and which was found to work well empirically for initializing fixed random representations in previous work (Wieting and Kiela, 2019). Biases and layer norm parameters were initialized using their respective PyTorch defaults (based on Xavier init; Glorot and Bengio, 2010). We intersperse reservoir layers in alternating fashion starting from the middle. Specifically, we alternate one reservoir layer with one transformer layer, and place the alternating block in the middle. For example: a 7-layer encoder LLLLLLL in which we replace three layers with reservoirs becomes LRLRLRL, and with two becomes LLRLRLL. See Appendix C for a study comparing this strategy to alternative approaches (e.g., freezing in the bottom, middle or top). 4 Experiments In what follows, we first show our main result, on a variety of tasks: reservoir transformers mostly have better AUCC metrics; less training time per epoch; less convergence time until the best validation performance is achieved; and even improved test set generalization metrics. As a strong baseline method, we compare to LayerDrop (Fan et al., 2019). LayerDrop can also be seen as a method that dynamically bypasses parts of the computation during Transformer training in an attempt to improve efficiency, and making it a strong comparison to examine our methods. Then, we examine whether we can minimize the expectation over the gradients of upstream layers in the network such that we do not at all have to pass gradients through the reservoir layers, skipping their backward pass. 4298 2 4 6 8 10 12 # Updatable Encoder Layers 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir 10 15 20 25 30 # Updatable Encoder Layers 0.94 0.95 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir 30 40 50 60 70 # Updatable Decoder Layers 0.6 0.7 0.8 0.9 1.0 valid bpc AUCC Transformer T Reservoir FFN Reservoir 2 4 6 8 10 12 # Updatable Encoder Layers 32.5 33.0 33.5 34.0 test BLEU Transformer T Reservoir FFN Reservoir 10 15 20 25 30 # Updatable Encoder Layers 26.25 26.50 26.75 27.00 27.25 27.50 27.75 28.00 test BLEU Transformer T Reservoir FFN Reservoir 30 40 50 60 70 # Updatable Decoder Layers 1.2 1.4 1.6 1.8 2.0 2.2 2.4 test bpc Transformer T Reservoir FFN Reservoir Figure 1: Validation (top) and test (bottom) results for IWSLT (left), WMT (middle) and enwiki8 language modelling (right). IWSLT and WMT are BLEU (high is good); enwiki8 is BPC (low is good). Comparison of regular transformer (blue) and reservoir transformer with FFN (green) or Transformer reservoir (orange) layers added. 4.1 Machine Translation Machine translation (MT) is one of the core tasks of NLP. We demonstrate on two well-known MT datasets, IWSLT’14 German-English and WMT’16 English-German, that reservoir transformers obtain a better AUCC. For the raw validation plots over time that were used to calculate the AUCC, please refer to Appendix F. Following Kasai et al. (2020), the architecture of the network is an N-layer reservoir transformer encoder, followed by a regular shallow one- or two-layer decoder. This design choice has been shown to lead to very good speed and efficiency trade-offs, and serves as a good baseline for our experiments. Moreover, shallow decoders make it easier to decide where to place reservoir layers (in the encoder) and makes it more straightforward to identify where performance gains come from. Figure 1 shows the results for IWSLT (left) and WMT (middle). On the y-axis we show validation AUCC for the BLEU metric; on the x-axis we show the number of updatable layers in the encoder. The performance of a regular transformer encoder with 6 layers and a reservoir transformer encoder with 6 layers plus N additional reservoir layers are plotted for the same x-axis value to show the total number of updated layers. Plots for the total number of layers (updatable plus notupdatable, so essentially shifted versions of the plots) are shown in Appendix E. WMT is much larger and requires a much deeper encoder, as illustrated by the fact that a certain minimum depth is required for reservoir transformers to achieve a comparable validation AUCC. At test time, reservoir transformers outperform regular transformers for almost all encoder depths. The FFN Reservoir seems to work best in both cases, which is surprising because it does not have any self-attention component at all. This finding shows that self-attention, or the mechanism to summarize context information, should be learned if present. Once the context features have been gathered, a random projection via a fixed FFN module appears to be beneficial. Table 1 and 2 show the time it took to achieve the maximum validation BLEU score and how that relates to the regular transformer, demonstrating that reservoir transformers consistently converge faster in terms of wall-clock time. We save up to 22% convergence wall-clock time using reservoir transformers as much with the same number of updateable layers. We save as much as 27% time until convergence a 24 layer model on WMT, as shown in Table 2. One other noticeable point is that we can see that the T Reservoir achieves similar performance to LayerDrop on IWSLT and WMT in terms of wall-clock per epoch and wallclock time to the best performance. However, on both tasks, FFN Reservoir performs much better than LayerDrop in terms of efficiency per epoch 4299 Model # Layers Frozen Max BLEU Train time Ratio # Params Train Time each until max (in hours) Trainable (Total) epoch (in seconds) Transformer 6 0 34.52 ± 0.07 2.548 ± 0.06 1 26.8M 122.73 ± 1.16 8 0 34.59 ± 0.11 2.557 ± 0.05 1 31.1M 142.28 ± 1.87 10 0 34.56 ± 0.05 3.173 ± 0.04 1 35.3M 161.66 ± 1.54 12 0 34.29 ± 0.12 3.521 ± 0.09 1 39.5M 172.45 ± 1.98 T Reservoir 6 2 34.37 ± 0.12 2.422 ± 0.03 0.95 22.6M (26.8M) 120.59 ± 1.32 8 2 34.80 ± 0.07 2.450 ± 0.06 0.96 26.8M (31.1M) 134.49 ± 1.76 10 2 34.70 ± 0.03 2.831 ± 0.05 0.89 31.1M (35.3M) 144.42 ± 1.98 12 2 34.78 ± 0.04 3.476 ± 0.04 0.98 35.3M (39.5M) 159.43 ± 1.67 FFN Reservoir 6 2 34.43 ± 0.15 2.120 ± 0.04 0.83 22.6M (25.8M) 107.71 ± 1.73 8 2 34.56 ± 0.16 2.203 ± 0.06 0.86 26.8M (29.1M) 120.07 ± 1.65 10 2 34.66 ± 0.02 2.493 ± 0.05 0.79 31.1M (33.3M) 130.11 ± 1.43 12 2 34.76 ± 0.03 3.241 ± 0.04 0.92 35.3M (37.5M) 156.32 ± 1.87 LayerDrop 6 2 34.59 ± 0.15 2.364 ± 0.08 0.92 22.6M (26.8M) 119.30 ± 1.36 8 2 34.58 ± 0.16 2.554 ± 0.05 0.99 26.8M (31.1M) 138.62 ± 1.44 10 2 34.57 ± 0.07 3.404 ± 0.06 1.07 31.1M (35.3M) 140.88 ± 1.62 12 2 33.65 ± 0.24 3.251 ± 0.04 0.92 35.3M (39.5M) 160.85 ± 1.49 Table 1: Wall-clock time (averaged over multiple runs) saved for IWSLT for different model types and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is kept fixed at 2. The ratio is computed compared to the corresponding number of layers in the regular transformer case. Model # Layers Frozen Max BLEU Train time Ratio # Params Train Time each until max (in hours) Trainable (Total) epoch (in hours) Transformer 12 0 24.46 ± 0.04 15.15 ± 0.15 1 75.6M 0.505 ± 0.005 16 0 24.52 ± 0.03 16.05 ± 0.18 1 88.2M 0.643 ± 0.006 24 0 24.69 ± 0.05 17.61 ± 0.85 1 113.4M 0.877 ± 0.029 32 0 24.83 ± 0.04 18.42 ± 0.28 1 138.6M 1.036 ± 0.010 T Reservoir 12 4 24.26 ± 0.08 14.11 ± 0.21 0.93 72.4M (75.6M) 0.472 ± 0.007 16 4 24.50 ± 0.05 15.25 ± 0.28 0.95 75.6M (88.2M) 0.596 ± 0.009 24 4 25.11 ± 0.07 15.89 ± 0.74 0.90 100.8M (113.4M) 0.776 ± 0.024 32 4 24.66 ± 0.04 16.38 ± 0.24 0.88 126.0M (138.6M) 0.998 ± 0.009 FFN Reservoir 12 4 24.42 ± 0.05 14.01 ± 0.09 0.92 72.4M (71.4M) 0.441 ± 0.003 16 4 24.65 ± 0.07 14.53 ± 0.17 0.91 75.6M (83.9M) 0.524 ± 0.006 24 4 24.93 ± 0.04 12.62 ± 1.53 0.71 100.8M (109.2M) 0.743 ± 0.018 32 4 24.98 ± 0.03 13.96 ± 0.19 0.73 126.0M (134.4M) 0.964 ± 0.007 LayerDrop 12 4 24.27 ± 0.03 14.61 ± 0.14 0.96 72.4M (75.6M) 0.489 ± 0.006 16 4 24.15 ± 0.06 15.55 ± 0.54 0.97 75.6M (88.2M) 0.597 ± 0.017 24 4 24.37 ± 0.05 16.25 ± 0.36 0.92 100.8M (113.4M) 0.823 ± 0.013 32 4 23.84 ± 0.03 15.27 ± 0.38 0.83 126.0M (138.6M) 1.028 ± 0.012 Table 2: Wall-clock time (averaged over multiple runs) saved for WMT for different model types and encoder depths. Decoder depth is kept fixed at 1. and achieves better/similar performance in less time in each case. As a point of reference, a half hour gain on IWSLT would translate to a gain of several days in the training of bigger transformer models like GPT-3 (Brown et al., 2020). We observe that reservoir transformers consistently perform better than, or are competitive to, regular transformers, both in terms of validation BLEU AUCC as well as test time BLEU, for all examined encoder depths. 4.2 Language Modelling To examine whether the same findings hold for other tasks, we evaluate on the enwiki8 (LLC, 2009) language modelling task. We examine the BPC (bits per character) rate for a variety of network depths (since the task is language modelling, these layers are in the decoder). The results show that except for the 64-layer regular transformer, which appears to be particularly optimal for this task, we obtain consistently better BPC for all depths. We observe similar trends during test time. 4.3 Masked Language Model Pretraining We train RoBERTa (Liu et al., 2019) models from scratch at a variety of depths, both in the normal and reservoir setting. We find that these networks show minor differences in their best perplexity 4300 4 6 8 10 12 14 16 # Updatable Decoder Layers 91 92 93 94 95 96 valid accuracy Transformer T Reservoir FFN Reservoir Transformer (frozen finetuned) 4 6 8 10 12 14 16 # Updatable Decoder Layers 78 80 82 84 86 valid accuracy Transformer T Reservoir FFN Reservoir Transformer (frozen finetuned) Figure 2: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right). Model Max BLEU AUCC Train time Transformer 34.59 ± 0.11 114.57 ± 0.08 142.28 ± 1.87 T Reservoir 34.80 ± 0.07 115.26 ± 0.26 134.49 ± 1.70 Backskip Reservoir 34.75 ± 0.05 115.99 ± 0.23 119.54 ± 1.78 Table 3: Validation max BLEU, AUCC at 4h and wallclock time per epoch (averaged over multiple runs, in seconds) on IWSLT comparing backskipping with regular and reservoir transformers. and similar AUCC perplexity (see Appendix D). We then examine the performance of these models when fine-tuned on downstream tasks, specifically the well known SST-2 (Socher et al., 2013) and MultiNLI-matched (Williams et al., 2017) tasks. When fine-tuning the reservoir models, we keep the reservoir layers fixed (also fine-tuning them did not work very well, see Appendix D). Figure 2 shows the results of fine-tuning. We observe that the reservoir transformer outperforms normal RoBERTa at all depths in both tasks. At lower depth, the improvements are substantial. As a sanity check, we also experiment with freezing some of the layers in a regular pre-trained RoBERTa model during fine-tuning only (Transformer “frozen finetuned” in the Figure) and show that this helps a little but is still outperformed by the reservoir transformer. These findings suggest that we can train a RoBERTa model without updating all of the layers, achieving similar perplexity at a similar computational cost, but with better downstream performance. This strategy could prove to be beneficial in a wide variety of pre-training scenarios. We follow Jawahar et al. (2019) and investigate what the frozen layers in the Reservoir Transformer have actually “learned” (while being frozen) as measured by probing tasks, reported in Table 4. The set of tasks comprises one surface task, three syntactic tasks, and five semantic tasks. From the table, we can see that generally probing performance is quite similar between Transformer and the T Reservoir model. We also noticed that the representations collected after the reservoir layer (3, 5, 7, 9) in the T Reservoir actually have significantly better performance over the regular Transformer representations across all the probing tasks. Related to our findings, Voita and Titov (2020) show that the wholly-randomlyinitialized model representations can still have reasonable probing accuracy if they are contextualized, though the accuracy is strictly worse than a trained one. These findings raise interesting repercussions for the study of “BERTology”, as it clearly shows that even completely random and frozen layers can represent linguistic phenomena. 4.4 Backskipping With the reservoir transformers as described above, we obtain better efficiency by skipping the “gradient application” matrix addition step in some of the layers (i.e., updating the weights). One step further would be to investigate skipping the entire backward pass for reservoirs altogether, which would save us from having to do the much more expensive matrix multiplication for these layers that is required for the propagation of gradients through a regular layer. We report on preliminary experiments where in the backward pass we replace the gradients for the layer Li going into the reservoir Li+1 with a noisy estimate (Jaderberg et al., 2017; Czarnecki et al., 2017). Promisingly, Oktay et al. (2020) recently asked “why spend resources on exact gradients when we’re going to use stochastic optimization?” and show that we can do randomized autodifferentiation quite successfully. 4301 Model Layer SentLen TreeDepth TopConst BShift Tense SubjNum ObjNum SOMO CoordInv (Surface) (Syntactic) (Syntactic) (Syntactic) (Semantic) (Semantic) (Semantic) (Semantic) (Semantic) Transformer 1 84.56 ± 0.54 32.30 ± 0.41 54.40 ± 0.33 49.99 ± 0.01 80.98 ± 0.32 76.26 ± 0.09 50.01 ± 0.19 76.38 ± 0.61 54.33 ± 0.47 2 87.22 ± 0.07 33.63 ± 0.57 58.38 ± 0.20 50.12 ± 0.17 82.84 ± 0.68 78.65 ± 0.19 51.47 ± 0.53 78.00 ± 1.12 54.66 ± 0.55 3 84.25 ± 0.16 32.60 ± 0.17 54.41 ± 0.10 50.02 ± 0.01 81.72 ± 0.59 77.00 ± 0.13 51.32 ± 0.64 76.57 ± 1.13 54.13 ± 0.51 4 87.37 ± 0.20 32.59 ± 0.29 50.06 ± 0.21 69.76 ± 0.26 81.63 ± 1.17 76.47 ± 0.09 52.41 ± 1.49 76.15 ± 0.84 52.62 ± 1.34 5 84.61 ± 0.24 31.14 ± 0.48 44.76 ± 0.38 74.82 ± 0.11 80.16 ± 0.19 73.66 ± 0.16 52.95 ± 1.77 72.90 ± 0.21 51.26 ± 1.14 6 82.56 ± 0.25 30.31 ± 0.40 39.30 ± 0.40 78.80 ± 0.38 81.88 ± 0.47 75.30 ± 0.07 56.21 ± 1.26 74.37 ± 0.16 51.44 ± 1.04 7 70.85 ± 0.13 26.65 ± 0.72 40.70 ± 0.13 78.98 ± 0.32 85.11 ± 0.31 72.03 ± 0.46 58.15 ± 0.46 68.71 ± 0.91 55.39 ± 0.27 8 66.23 ± 1.33 23.46 ± 0.44 25.19 ± 1.02 77.42 ± 0.27 80.35 ± 0.45 67.55 ± 0.99 54.94 ± 2.04 63.69 ± 2.32 50.58 ± 0.83 9 71.17 ± 0.29 31.21 ± 0.31 58.42 ± 0.29 85.55 ± 0.44 86.77 ± 0.19 80.30 ± 0.08 64.36 ± 1.20 81.68 ± 0.45 66.90 ± 0.49 10 73.19 ± 0.50 27.74 ± 0.53 41.01 ± 0.22 83.56 ± 0.96 86.13 ± 0.35 83.04 ± 0.04 62.01 ± 0.59 79.73 ± 0.21 62.60 ± 1.04 11 71.37 ± 0.42 30.22 ± 0.28 48.58 ± 0.35 84.40 ± 0.44 87.28 ± 0.59 82.34 ± 0.15 61.10 ± 0.14 80.00 ± 0.40 64.44 ± 0.38 12 71.66 ± 0.12 33.43 ± 0.18 64.38 ± 0.20 87.38 ± 0.02 88.41 ± 0.09 84.46 ± 0.25 63.01 ± 0.05 81.80 ± 0.27 65.72 ± 0.16 T Reservoir 1 87.75 ± 0.10 31.60 ± 0.21 50.38 ± 0.23 50.00 ± 0.00 80.40 ± 0.18 76.47 ± 0.20 50.53 ± 0.14 73.48 ± 0.15 53.55 ± 0.70 2 81.28 ± 0.23 34.20 ± 0.41 61.41 ± 0.42 60.64 ± 0.65 81.50 ± 0.77 76.33 ± 0.08 50.73 ± 0.34 74.28 ± 0.67 56.82 ± 0.10 3 89.28 ± 0.09 36.42 ± 0.11 67.36 ± 0.45 75.64 ± 0.52 85.42 ± 0.18 80.53 ± 0.02 52.50 ± 1.80 78.47 ± 1.81 57.16 ± 0.27 4 74.31 ± 0.32 32.42 ± 0.83 55.19 ± 0.33 73.41 ± 0.00 79.56 ± 0.00 75.15 ± 0.08 53.68 ± 0.66 75.02 ± 0.19 56.89 ± 0.08 5 88.03 ± 0.22 38.34 ± 0.64 68.65 ± 0.29 82.25 ± 0.12 86.80 ± 0.02 82.27 ± 0.33 57.95 ± 0.24 80.82 ± 0.91 58.05 ± 0.10 6 74.55 ± 0.37 33.13 ± 0.29 52.70 ± 0.81 79.21 ± 0.13 85.70 ± 0.36 77.43 ± 0.03 57.26 ± 0.19 75.38 ± 0.66 51.95 ± 1.30 7 85.82 ± 0.37 37.63 ± 0.13 70.43 ± 0.05 84.12 ± 0.35 86.88 ± 0.07 82.86 ± 0.30 61.17 ± 0.21 80.79 ± 0.17 61.83 ± 0.95 8 71.69 ± 0.71 30.32 ± 0.01 48.44 ± 0.30 79.12 ± 0.12 84.75 ± 0.09 79.23 ± 0.11 59.53 ± 0.16 76.80 ± 0.41 57.34 ± 0.14 9 85.86 ± 0.12 37.89 ± 0.03 69.53 ± 0.37 85.55 ± 0.12 87.98 ± 0.22 84.13 ± 0.01 63.06 ± 0.01 82.55 ± 0.31 66.07 ± 0.05 10 69.22 ± 0.23 25.58 ± 0.35 29.20 ± 0.58 78.57 ± 0.09 85.02 ± 0.03 75.68 ± 0.16 57.55 ± 1.57 74.70 ± 0.02 55.02 ± 0.64 11 65.70 ± 0.05 30.57 ± 0.03 47.56 ± 0.02 81.20 ± 0.00 86.78 ± 0.02 83.73 ± 0.05 60.38 ± 0.17 80.59 ± 0.15 62.50 ± 0.11 12 70.61 ± 0.18 34.45 ± 0.20 64.19 ± 0.10 84.53 ± 0.03 87.48 ± 0.16 84.86 ± 0.14 62.75 ± 0.14 82.08 ± 0.03 64.73 ± 0.06 Table 4: RoBERTa Probing Results. The line in bold text are the the frozen layers in the T Reservoir. Mean accuracy with standard deviation, gathered over 3 random seeds. Here, rather than minimizing the actual gradients ∂Li ∂θLi , we minimize their expectation and train via continuous-action REINFORCE (Williams, 1992). That is, Li becomes a policy πa: s →µ where we sample actions a ∼N(µ, 1). We train to minimize the gradient prediction loss via MSE, i.e., 1 n Pn i=0(Ri −V i(a))2, and the REINFORCE loss Ea [log(a) (R −V (a))], where the value network V acts as the baseline. R is defined as the mean of the gradients of the top layer Li+2, with the sign flipped. Thus, simply put, we train to minimize the expectation of the true gradients at the layer directly following the reservoir. We employ an annealing scheme where we first train the value network and propagate the true gradients during warmup. Afterwards, we anneal the probability of backskipping instead of doing a true backward pass (multiplying the probability by 0.99 every iteration until we only backskip). We experimented with setting R to the negation of the total loss but found the mean upstream gradient reward to work better. We call this approach backskipping. As shown in Table 3, the backskip reservoir approach leads to a higher maximum BLEU score than the regular transformer, with a much higher AUCC and much lower training time. The encoder depth is 8 with 2 frozen. Appendix G shows the raw validation BLEU curves over time. We observe that this approach helps especially during the earlier stages of training. This finding opens up intriguing possibilities for having parts of neural networks be completely frozen both in the forward as well as in the backward pass, while still contributing to the overall model computation. The computational cost is heavily reduced given that we completely bypass the expensive backpropagation computation in the reservoir layers. Backskipping is shown to be a promising approach to further reduce computational costs, and would be even more efficient from a hardware perspective since the circuitry for such layers (which do not need to propagate gradients) can be hardwired. 5 Related Work Recent work has shown that modern NLP models are able to function with different numbers of layers for different examples (Elbayad et al., 2019; Fan et al., 2019; He et al., 2021); that different layers specialize for different purposes (Zhang et al., 2019); that layers can be compressed (Li et al., 2020; Zhu et al., 2019; Shen et al., 2020; Sun et al., 2020); and, that layers can be reordered (Press et al., 2019). There is a growing body of work in efficient self-attention networks (Tay et al., 2020b), such as linear attention (Wang et al., 2020), on how to process long context information (Beltagy et al., 2020; Ainslie et al., 2020) and on approximations to make transformers more scalable (Kitaev et al., 2020; Katharopoulos et al., 2020). BigBIRD (Zaheer et al., 2020) provides random keys as additional inputs to its attention mechanism. Locality sensitive hashing (LSH) as employed e.g. in Reformer (Kitaev et al., 2020) utilizes a fixed random projection. Random Feature Attention (Peng et al., 2021) uses random fea4302 ture methods to approximate the softmax function. Performer (Choromanski et al., 2020) computes the transformer’s multi-head attention weights as a fixed orthogonal random projection. Closely related to this work, Tay et al. (2020a) showed that randomized alignment matrices in their “Synthesizer” architecture are sufficient for many NLP tasks. While these works focus on random attention, we show that entire layers can be random and fixed. We also show that entire layers can be replaced by fixed random projections that do not have any attention whatsoever. Beyond transformers, random features have been extensively explored. Examples of this include FreezeOut (Brock et al., 2017), deep reservoir computing networks (Scardapane and Wang, 2017; Gallicchio and Micheli, 2017), as well as applications in domains as varied as text classification (Conneau et al., 2017; Zhang and Bowman, 2018; Wieting and Kiela, 2019) or music classification (Pons and Serra, 2019). It is well known that randomly initialized networks can display impressive performance on their own (Ulyanov et al., 2018; Rosenfeld and Tsotsos, 2019; Ramanujan et al., 2020; Voita and Titov, 2020), which underlies, for example, the recently popularized lottery ticket hypothesis (Frankle and Carbin, 2018; Zhou et al., 2019). We know that learning deep overparameterized networks appears to help in general (Li and Liang, 2018; Du et al., 2019). Our method constitutes a way to add both depth and parameters to transformer networks without much computational cost. 6 Conclusion This work demonstrated that state-of-the-art transformer architectures can be trained without updating all of the layers. This complements a long history in machine learning of harnessing the power of random features. We use the “area under the convergence curve” (AUCC) metric to demonstrate that on a variety of tasks, and in a variety of settings, “reservoir transformers” achieve better performance-efficiency trade-offs. We show that such reservoir transformers show better convergence rates and test-set generalization. We demonstrated that the backward pass can be skipped altogether, opening up exciting vanues for future research. Future work includes further investigating hybrid networks and backskipping strategies, as well as utilizing pruning. Acknowledgements We thank Eric Wallace, Zhewei Yao, Kevin Lin, Zhiqing Sun, Zhuohan Li, Angela Fan, Shaojie Bai, and anonymous reviewers for their comments and suggestions. SS and KK were supported by grants from Samsung, Facebook, and the Berkeley Deep Drive Consortium. References Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. ETC: Encoding long and structured inputs in transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Garrison W Cottrell, and Julian McAuley. 2020. Rezero is all you need: Fast convergence at large depth. arXiv preprint arXiv:2003.04887. Alexei Baevski, Steffen Schneider, and Michael Auli. 2019. vq-wav2vec: Self-supervised learning of discrete speech representations. arXiv preprint arXiv:1910.05453. Eric B Baum. 1988. On the capabilities of multilayer perceptrons. Journal of complexity, 4(3):193–215. Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. Hans-Dieter Block. 1962. The perceptron: A model for brain functioning. i. Reviews of Modern Physics, 34(1):123. Ondˇrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia, and Aleˇs Tamchyna. 2014. Findings of the 2014 workshop on statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation, Baltimore, Maryland, USA. Association for Computational Linguistics. A Borsellino and A Gamba. 1961. An outline of a mathematical theory of papa. Il Nuovo Cimento (1955-1965), 20(2):221–231. Andrew P Bradley. 1997. The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern recognition, 30(7):1145–1159. Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. 2017. Freezeout: Accelerate training by progressively freezing layers. arXiv preprint arXiv:1706.04983. 4303 Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. arXiv preprint arXiv:2005.12872. M. Cettolo, J. Niehues, S. St¨uker, L. Bentivogli, and Marcello Federico. 2015. Report on the 11 th iwslt evaluation campaign , iwslt 2014. In Proceedings of IWSLT. Kyunghyun Cho, Bart Van Merri¨enboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078. Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Jared Davis, Tamas Sarlos, David Belanger, Lucy Colwell, and Adrian Weller. 2020. Masked language modeling for proteins via linearly scalable long-context transformers. arXiv preprint arXiv:2006.03555. Alexis Conneau, Douwe Kiela, Holger Schwenk, Loic Barrault, and Antoine Bordes. 2017. Supervised learning of universal sentence representations from natural language inference data. arXiv preprint arXiv:1705.02364. Thomas M Cover. 1965. Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition. IEEE transactions on electronic computers, (3):326–334. Wojciech Marian Czarnecki, Grzegorz ´Swirszcz, Max Jaderberg, Simon Osindero, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Understanding synthetic gradients and decoupled neural interfaces. arXiv preprint arXiv:1703.00522. Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy. Association for Computational Linguistics. Amit Daniely, Roy Frostig, and Yoram Singer. 2016. Toward deeper understanding of neural networks: The power of initialization and a dual view on expressivity. In Advances In Neural Information Processing Systems, pages 2253–2261. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Simon Du, Jason Lee, Haochuan Li, Liwei Wang, and Xiyu Zhai. 2019. Gradient descent finds global minima of deep neural networks. In International Conference on Machine Learning, pages 1675–1685. Sergey Edunov, Myle Ott, Michael Auli, David Grangier, and Marc’Aurelio Ranzato. 2018. Classical structured prediction losses for sequence to sequence learning. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), New Orleans, Louisiana. Association for Computational Linguistics. Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. 2019. Depth-adaptive transformer. arXiv preprint arXiv:1910.10073. Joseph Enguehard, Dan Busbridge, Vitalii Zhelezniak, and Nils Hammerla. 2019. Neural language priors. arXiv preprint arXiv:1910.03492. Angela Fan, Edouard Grave, and Armand Joulin. 2019. Reducing transformer depth on demand with structured dropout. arXiv preprint arXiv:1909.11556. Jonathan Frankle and Michael Carbin. 2018. The lottery ticket hypothesis: Finding sparse, trainable neural networks. arXiv preprint arXiv:1803.03635. Jonathan Frankle, David J Schwab, and Ari S Morcos. 2020. Training batchnorm and only batchnorm: On the expressive power of random features in cnns. arXiv preprint arXiv:2003.00152. Claudio Gallicchio and Alessio Micheli. 2017. Echo state property of deep reservoir computing networks. Cognitive Computation, 9(3):337–350. Claudio Gallicchio and Simone Scardapane. 2020. Deep randomized neural networks. In Recent Trends in Learning From Data, pages 43–68. Springer. A. Gamba, L. Gamberini, G. Palmieri, and R. Sanna. 1961. Further experiments with papa. Il Nuovo Cimento (1955-1965), 20(2):112–115. Ankush Garg, Yuan Cao, and Qi Ge. 2020. Echo state neural machine translation. arXiv preprint arXiv:2002.11847. Raja Giryes, Guillermo Sapiro, and Alex M Bronstein. 2016. Deep neural networks with random gaussian weights: A universal classification strategy? IEEE Transactions on Signal Processing, 64(13):3444– 3457. Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. 4304 Caglar Gulcehre, Marcin Moczulski, Misha Denil, and Yoshua Bengio. 2016. Noisy activation functions. In International conference on machine learning, pages 3059–3068. Fatemeh Hadaeghi, Xu He, and Herbert Jaeger. 2017. Unconventional Information Processing Systems, Novel Hardware: A Tour D’Horizon. Chaoyang He, Shen Li, Mahdi Soltanolkotabi, and Salman Avestimehr. 2021. Pipetransformer: Automated elastic pipelining for distributed training of transformers. In ICML. Konstantin Hicke, Miguel Escalona-Moran, Daniel Brunner, Miguel Soriano, Ingo Fischer, and Claudio Mirasso. 2013. Information processing using transient dynamics of semiconductor lasers subject to delayed feedback. Selected Topics in Quantum Electronics, IEEE Journal of, 19:1501610–1501610. Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. 2006. Extreme learning machine: theory and applications. Neurocomputing, 70(1-3):489–501. Max Jaderberg, Wojciech Marian Czarnecki, Simon Osindero, Oriol Vinyals, Alex Graves, David Silver, and Koray Kavukcuoglu. 2017. Decoupled neural interfaces using synthetic gradients. In International Conference on Machine Learning, pages 1627–1635. PMLR. Herbert Jaeger. 2003. Adaptive nonlinear system identification with echo state networks. In Advances in neural information processing systems. Ganesh Jawahar, Benoˆıt Sagot, and Djam´e Seddah. 2019. What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. Kam Jim, Bill G Horne, and C Lee Giles. 1995. Effects of noise on convergence and generalization in recurrent networks. In Advances in neural information processing systems, pages 649–656. Kam-Chuen Jim, C Lee Giles, and Bill G Horne. 1996. An analysis of noise in recurrent neural networks: convergence and generalization. IEEE Transactions on neural networks, 7(6):1424–1438. William B Johnson and Joram Lindenstrauss. 1984. Extensions of lipschitz mappings into a hilbert space. Contemporary mathematics, 26(189-206):1. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361. Jungo Kasai, Nikolaos Pappas, Hao Peng, James Cross, and Noah A Smith. 2020. Deep encoder, shallow decoder: Reevaluating the speed-quality tradeoff in machine translation. arXiv preprint arXiv:2006.10369. Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and Franc¸ois Fleuret. 2020. Transformers are rnns: Fast autoregressive transformers with linear attention. arXiv preprint arXiv:2006.16236. Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv preprint arXiv:1408.5882. Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. arXiv preprint arXiv:2001.04451. Yann LeCun, L´eon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324. Yuanzhi Li and Yingyu Liang. 2018. Learning overparameterized neural networks via stochastic gradient descent on structured data. In Advances in Neural Information Processing Systems, pages 8157–8166. Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joseph E Gonzalez. 2020. Train large, then compress: Rethinking model size for efficient training and inference of transformers. arXiv preprint arXiv:2002.11794. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. MultiMedia LLC. 2009. Large text compression benchmark. Mantas Lukoˇseviˇcius and Herbert Jaeger. 2009. Reservoir computing approaches to recurrent neural network training. Computer Science Review, 3(3). Wolfgang Maass, Thomas Natschl¨ager, and Henry Markram. 2002. Real-time computing without stable states: A new framework for neural computation based on perturbations. Neural computation, 14(11):2531–2560. Marvin Minsky and Seymour A Papert. 2017. Perceptrons: An introduction to computational geometry. MIT press. Emre O Neftci, Charles Augustine, Somnath Paul, and Georgios Detorakis. 2017. Event-driven random back-propagation: Enabling neuromorphic deep learning machines. Frontiers in neuroscience, 11:324. Hyeonwoo Noh, Tackgeun You, Jonghwan Mun, and Bohyung Han. 2017. Regularizing deep neural networks by noise: Its interpretation and optimization. In Advances in Neural Information Processing Systems, pages 5109–5118. 4305 Deniz Oktay, Nick McGreivy, Joshua Aduol, Alex Beatson, and Ryan P Adams. 2020. Randomized automatic differentiation. arXiv preprint arXiv:2007.10412. Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. 2018. Scaling neural machine translation. arXiv preprint arXiv:1806.00187. Yoh-Han Pao, Gwang-Hoon Park, and Dejan J Sobajic. 1994. Learning and generalization characteristics of the random vector functional-link net. Neurocomputing, 6(2):163–180. Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. 2021. Random feature attention. In International Conference on Learning Representations. Jonathan Pilault, Jaehong Park, and Christopher Pal. 2020. On the impressive performance of randomly weighted encoders in summarization tasks. arXiv preprint arXiv:2002.09084. Jordi Pons and Xavier Serra. 2019. Randomly weighted cnns for (music) audio classification. In ICASSP 2019-2019 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 336–340. IEEE. Ofir Press, Noah A Smith, and Omer Levy. 2019. Improving transformer models by reordering their sublayers. arXiv preprint arXiv:1911.03864. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2018. Language models are unsupervised multitask learners. Ali Rahimi and Benjamin Recht. 2008. Random features for large-scale kernel machines. In Advances in neural information processing systems, pages 1177–1184. Ali Rahimi and Benjamin Recht. 2009. Weighted sums of random kitchen sinks: Replacing minimization with randomization in learning. In Advances in neural information processing systems, pages 1313– 1320. Vivek Ramanujan, Mitchell Wortsman, Aniruddha Kembhavi, Ali Farhadi, and Mohammad Rastegari. 2020. What’s hidden in a randomly weighted neural network? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11893–11902. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. arXiv preprint arXiv:2002.12327. Amir Rosenfeld and John K Tsotsos. 2019. Intriguing properties of randomly weighted networks: Generalizing while learning next to nothing. In 2019 16th Conference on Computer and Robot Vision (CRV), pages 9–16. IEEE. Magnus Sahlgren. 2005. An introduction to random indexing. In Methods and applications of semantic indexing workshop at the 7th international conference on terminology and knowledge engineering. Andrew M Saxe, James L McClelland, and Surya Ganguli. 2013. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. arXiv preprint arXiv:1312.6120. Simone Scardapane and Dianhui Wang. 2017. Randomness in neural networks: an overview. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 7(2):e1200. Wouter F Schmidt, Martin A Kraaijveld, and Robert PW Duin. 1992. Feedforward neural networks with random weights. In Proceedings of the 11th International Conference on Pattern Recognition, 1992. Vol. II. Conference B: Pattern Recognition Methodology and Systems, pages 1–4. Benjamin Schrauwen, Michiel D’Haene, David Verstraeten, and Jan Campenhout. 2007. Compact hardware for real-time speech recognition using a liquid state machine. pages 1097 – 1102. Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2019. Green ai. arXiv preprint arXiv:1907.10597. Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. 2020. Q-bert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815–8821. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642. Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in nlp. arXiv preprint arXiv:1906.02243. Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 2020. Mobilebert: a compact task-agnostic bert for resource-limited devices. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2158–2170. Gouhei Tanaka, Toshiyuki Yamane, Jean Benoit H´eroux, Ryosho Nakane, Naoki Kanazawa, Seiji Takeda, Hidetoshi Numata, Daiju Nakano, and Akira Hirose. 2019. Recent advances in physical reservoir computing: A review. Neural Networks, 115:100 – 123. 4306 Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, and Che Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv preprint arXiv:2005.00743. Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2020b. Efficient transformers: A survey. arXiv preprint arXiv:2009.06732. Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019. Bert rediscovers the classical nlp pipeline. arXiv preprint arXiv:1905.05950. Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. 2018. Deep image prior. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9446–9454. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, Micha¨el Mathieu, Andrew Dudzik, Junyoung Chung, David H. Choi, Richard Powell, Timo Ewalds, Petko Georgiev, Junhyuk Oh, Dan Horgan, Manuel Kroiss, Ivo Danihelka, Aja Huang, Laurent Sifre, Trevor Cai, John P. Agapiou, Max Jaderberg, Alexander S. Vezhnevets, R´emi Leblond, Tobias Pohlen, Valentin Dalibard, David Budden, Yury Sulsky, James Molloy, Tom L. Paine, Caglar Gulcehre, Ziyu Wang, Tobias Pfaff, Yuhuai Wu, Roman Ring, Dani Yogatama, Dario W¨unsch, Katrina McKinney, Oliver Smith, Tom Schaul, Timothy Lillicrap, Koray Kavukcuoglu, Demis Hassabis, Chris Apps, and David Silver. 2019. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):350–354. Elena Voita and Ivan Titov. 2020. Informationtheoretic probing with minimum description length. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 183–196. Sinong Wang, Belinda Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Selfattention with linear complexity. arXiv preprint arXiv:2006.04768. John Wieting and Douwe Kiela. 2019. No training required: Exploring random encoders for sentence classification. arXiv preprint arXiv:1901.10444. Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426. Ronald J Williams. 1992. Simple statistical gradientfollowing algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256. Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. 2019. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430. Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. arXiv preprint arXiv:2007.14062. Chiyuan Zhang, Samy Bengio, and Yoram Singer. 2019. Are all layers created equal? arXiv preprint arXiv:1902.01996. Kelly Zhang and Samuel Bowman. 2018. Language modeling teaches you more than translation does: Lessons learned through auxiliary syntactic task analysis. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. Hattie Zhou, Janice Lan, Rosanne Liu, and Jason Yosinski. 2019. Deconstructing lottery tickets: Zeros, signs, and the supermask. In Advances in Neural Information Processing Systems, pages 3597– 3607. Wei Zhu, Xiaofeng Zhou, Keqiang Wang, Xun Luo, Xiepeng Li, Yuan Ni, and Guotong Xie. 2019. PANLP at MEDIQA 2019: Pre-trained language models, transfer learning and knowledge distillation. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 380–388, Florence, Italy. Association for Computational Linguistics. A Hybrid Networks and Non-Transformer Reservoirs We investigate whether reservoir layers need to be transformer-based (or transformers-withoutattention, i.e., FFN). We examine two different alternatives: bidirectional Gated Recurrent Units (Cho et al., 2014) and Convolutional Neural Networks (LeCun et al., 1998; Kim, 2014), specifically light dynamical convolutions (Wu et al., 2019). Figure 3 shows the results for these hybrids: depending on the setting, they may obtain a better AUCC than the regular transformer, but this is less consistent than with the other reservoir layers, most likely because these layers have different computational properties. It’s possible that these hybrids simply require further tuning, as we found e.g. up-projecting to help for BiGRUs, but studying this is outside of the scope of the current work. 4307 Model # Layers Frozen Max BLEU Train time Ratio # Params Train Time each until max (in hours) Trainable (Total) epoch (in seconds) Transformer 6 0 34.97 ± 0.05 1.984 ± 0.02 1 39.5M 177.84 ± 2.98 8 0 34.99 ± 0.08 2.161 ± 0.03 1 43.7M 206.59 ± 3.47 10 0 34.98 ± 0.04 2.345 ± 0.02 1 47.9M 236.72 ± 3.52 12 0 34.78 ± 0.11 2.535 ± 0.05 1 52.0M 265.90 ± 4.97 T Reservoir 6 2 34.73 ± 0.11 1.838 ± 0.01 0.92 35.3M (39.5M) 166.11 ± 2.21 8 2 35.07 ± 0.05 1.912 ± 0.03 0.88 39.5M (43.7M) 190.08 ± 3.73 10 2 35.02 ± 0.01 1.970 ± 0.04 0.84 43.7M (47.9M) 204.42 ± 2.89 12 2 35.06 ± 0.02 2.429 ± 0.02 0.95 47.8M (52.0M) 236.41 ± 4.35 FFN Reservoir 6 2 34.85 ± 0.10 1.729 ± 0.03 0.87 35.3M (37.4M) 161.72 ± 2.32 8 2 34.99 ± 0.11 1.751 ± 0.02 0.81 39.5M (41.6M) 180.21 ± 2.68 10 2 34.92 ± 0.03 1.907 ± 0.02 0.81 43.7M (45.8M) 191.40 ± 2.49 12 2 35.16 ± 0.04 2.395 ± 0.01 0.94 47.8M (49.9M) 216.08 ± 2.57 LayerDrop 6 2 34.51 ± 0.12 1.908 ± 0.04 0.96 35.3M (39.5M) 169.62 ± 3.16 8 2 34.77 ± 0.11 2.023 ± 0.02 0.94 39.5M (43.7M) 186.71 ± 2.17 10 2 34.06 ± 0.05 1.912 ± 0.02 0.97 43.7M (47.9M) 205.52 ± 3.31 12 2 34.08 ± 0.13 2.524 ± 0.01 0.99 47.8M (52.0M) 222.45 ± 2.21 Table 5: Wall-clock time (averaged over multiple runs) for IWSLT for different model types and encoder depths. Max BLEU is for validation. Number of layers is for encoder, decoder depth is kept fixed at 6. Ratio is computed compared to comparable number of layers in the normal case. 2 4 6 8 10 12 # Updatable Encoder Layers 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir GRU Reservoir Conv Reservoir 2 4 6 8 10 12 # Updatable Encoder Layers 32.0 32.5 33.0 33.5 34.0 test BLEU Transformer T Reservoir FFN Reservoir GRU Reservoir Conv Reservoir Figure 3: IWSLT comparison of different hybrid architectures with different reservoir layers. B Deep Decoders We show that the same results hold for a 6-layer decoder on IWSLT (although less pronounced for AUCC, probably because the decoder is computationally heavier). See Figure 4 and Table 5. 2 4 6 8 10 12 # Updatable Encoder Layers 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir 2 4 6 8 10 12 # Updatable Encoder Layers 33.2 33.4 33.6 33.8 34.0 34.2 34.4 34.6 test BLEU Transformer T Reservoir FFN Reservoir Figure 4: IWSLT validation AUCC and test BLEU with 6-layer decoder. C Freezing Strategy We explored different strategies for the placement of reservoir layers and found the “alternating” strategy reported in the main body of the paper to work best. Generally, we found repetitive applica2 4 6 8 10 # Updatable Encoder Layers 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer Alter T Reservoir Mid T Reservoir Top T Reservoir Bottom T Reservoir Figure 5: IWSLT with 2-layer decoder using different freezing strategies. tion of reservoirs to yield diminishing returns, as might be expected. See Figure 5. D RoBERTa Results Here we present the additional results for RoBERTa , i.e., convergence plots and AUCCs for various depth settings, in Figure 7. As stated in the main paper, the differences in terms of AUCC and convergence between RoBERTa models with and without reservoir layers are limited. Moreover, we plot downstream task performance for SST-2 and MNLI compared to the pretraining wall-clock time in Figure 6. It can be seen that the FFN Reservoir can achieve up to 25% and 10% pretraining time savings while matching the best performance 4308 Model IWSLT-Dec2 IWSLT-Dec6 WMT-Dec1 # Layers Train time Max BLEU # Layers Train time Max BLEU # Layers Train time Max BLEU until 95% max (in hours) (95%) until 95% max (in hours) (95%) until 95% max (in hours) (95%) Transformer 6 0.647 ± 0.03 32.89 ± 0.04 6 0.642 ± 0.02 33.36 ± 0.03 12 3.788 ± 0.053 23.36 ± 0.06 8 0.711 ± 0.05 33.04 ± 0.03 8 0.765 ± 0.03 33.41 ± 0.08 16 3.820 ± 0.072 23.41 ± 0.05 10 0.808 ± 0.02 33.96 ± 0.08 10 0.898 ± 0.04 33.32 ± 0.07 24 5.262 ± 0.607 23.50 ± 0.03 12 1.037 ± 0.03 33.07 ± 0.09 12 1.037 ± 0.03 33.07 ± 0.11 32 6.212 ± 0.232 23.81 ± 0.04 T Reservoir 6 0.569 ± 0.02 32.78 ± 0.03 6 0.599 ± 0.01 33.09 ± 0.05 12 3.563 ± 0.061 23.21 ± 0.04 8 0.619 ± 0.04 33.12 ± 0.05 8 0.726 ± 0.02 33.38 ± 0.09 16 3.603 ± 0.056 23.80 ± 0.06 10 0.729 ± 0.04 33.13 ± 0.07 10 0.738 ± 0.03 33.37 ± 0.04 24 4.923 ± 0.771 23.75 ± 0.02 12 0.982 ± 0.02 33.03 ± 0.11 12 0.958 ± 0.01 33.46 ± 0.09 32 5.780 ± 0.214 23.71 ± 0.03 FFN Reservoir 6 0.521 ± 0.05 32.85 ± 0.02 6 0.594 ± 0.03 33.13 ± 0.04 12 3.417 ± 0.046 23.22 ± 0.07 8 0.533 ± 0.03 33.84 ± 0.04 8 0.651 ± 0.04 33.36 ± 0.06 16 3.527 ± 0.063 23.54 ± 0.05 10 0.614 ± 0.01 33.05 ± 0.08 10 0.627 ± 0.05 33.26 ± 0.03 24 4.197 ± 0.697 23.74 ± 0.06 12 0.811 ± 0.02 33.26 ± 0.10 12 0.780 ± 0.02 33.46 ± 0.08 32 4.984 ± 0.321 23.82 ± 0.02 LayerDrop 6 0.837 ± 0.08 32.87 ± 0.05 6 0.706 ± 0.01 33.08 ± 0.03 12 3.912 ± 0.068 23.33 ± 0.08 8 0.934 ± 0.07 33.12 ± 0.03 8 0.753 ± 0.04 33.14 ± 0.05 16 3.581 ± 0.076 23.17 ± 0.04 10 0.901 ± 0.06 33.18 ± 0.02 10 0.691 ± 0.03 32.39 ± 0.05 24 4.875 ± 0.728 23.43 ± 0.07 12 0.914 ± 0.01 32.33 ± 0.06 12 0.803 ± 0.02 32.94 ± 0.10 32 5.980 ± 0.219 22.97 ± 0.08 Table 6: Wall-clock time (averaged over multiple runs) for IWSLT/WMT for different model types and encoder depths. 95% Max BLEU is for validation. Model IWSLT-Dec2 IWSLT-Dec6 WMT-Dec1 # Layers Train time Max BLEU # Layers Train time Max BLEU # Layers Train time Max BLEU until 99% max (in hours) (99%) until 99% max (in hours) (99%) until 99% max (in hours) (99%) Transformer 6 1.454 ± 0.06 34.24 ± 0.05 6 1.297 ± 0.03 34.69 ± 0.05 12 9.961 ± 0.053 24.27 ± 0.04 8 1.475 ± 0.09 34.32 ± 0.09 8 1.390 ± 0.02 34.75 ± 0.09 16 12.623 ± 0.072 24.35 ± 0.06 10 1.526 ± 0.04 34.25 ± 0.04 10 1.622 ± 0.05 34.64 ± 0.03 24 13.412 ± 0.837 24.49 ± 0.07 12 2.259 ± 0.07 34.24 ± 0.11 12 1.748 ± 0.01 34.66 ± 0.08 32 15.117 ± 0.232 24.56 ± 0.02 T Reservoir 6 1.257 ± 0.04 34.05 ± 0.09 6 1.291 ± 0.03 34.51 ± 0.10 12 8.314 ± 0.062 24.15 ± 0.06 8 1.472 ± 0.06 34.47 ± 0.05 8 1.339 ± 0.03 34.80 ± 0.04 16 9.221 ± 0.073 24.41 ± 0.05 10 1.530 ± 0.03 34.36 ± 0.02 10 1.419 ± 0.04 34.72 ± 0.03 24 10.413 ± 0.580 24.56 ± 0.03 12 2.043 ± 0.05 34.53 ± 0.07 12 1.642 ± 0.02 34.87 ± 0.02 32 11.465 ± 0.227 24.49 ± 0.01 FFN Reservoir 6 1.138 ± 0.03 34.10 ± 0.13 6 1.169 ± 0.02 34.71 ± 0.09 12 7.407 ± 0.087 24.33 ± 0.08 8 1.101 ± 0.07 34.32 ± 0.11 8 1.201 ± 0.03 34.79 ± 0.08 16 9.336 ± 0.036 24.42 ± 0.05 10 1.281 ± 0.01 34.36 ± 0.03 10 1.276 ± 0.03 34.63 ± 0.03 24 9.978 ± 0.546 24.91 ± 0.07 12 1.785 ± 0.03 34.42 ± 0.06 12 1.440 ± 0.01 34.87 ± 0.02 32 10.524 ± 0.341 24.96 ± 0.01 LayerDrop 6 1.363 ± 0.05 34.58 ± 0.14 6 1.253 ± 0.01 34.42 ± 0.10 12 8.372 ± 0.059 24.17 ± 0.04 8 1.468 ± 0.03 34.50 ± 0.12 8 1.244 ± 0.04 34.44 ± 0.09 16 9.741 ± 0.043 23.93 ± 0.08 10 1.678 ± 0.04 34.52 ± 0.07 10 1.343 ± 0.04 33.83 ± 0.06 24 10.145 ± 0.628 24.07 ± 0.09 12 2.071 ± 0.02 33.45 ± 0.23 12 1.423 ± 0.02 33.97 ± 0.12 32 10.168 ± 0.329 23.81 ± 0.03 Table 7: Wall-clock time (averaged over multiple runs) saved for IWSLT/WMT for different model types and encoder depths. 99% Max BLEU is for validation. of vanilla transformers for MNLI-m and SST2, respectively. 10 20 30 40 50 60 Pretraining Wall-clock Time 78 79 80 81 82 83 84 85 Accuray on MNLI-m Transformer T Reservoir FFN Reservoir 10 20 30 40 50 60 Pretraining Wall-clock Time 91 92 93 94 95 Accuray on SST2 Transformer T Reservoir FFN Reservoir Figure 6: RoBERTa Reservoir Results, Pre-training versus downstream task plot for 12 layer RoBERTa. MNLI-m (left). SST-2 (right). E Reservoir Results for Total Layers Here we present the shifted Reservoir Results for IWSLT14, WMT16, Enwik8 and RoBERTa finetuning in Figure 8, 9, 10, 11, respectively. We show the same results also hold when it comes to replace normal transformer blocks with Reservoir blocks at least for MT. 0 12 24 36 48 60 Training Hours (h) 4 6 8 10 12 14 16 18 20 Validation PPL Transformer T Reservoir FFN Reservoir 4 6 8 10 12 14 16 # Updatable Decoder Layers 0.850 0.875 0.900 0.925 0.950 0.975 1.000 Valid PPL AUCC Transformer T Reservoir FFN Reservoir Figure 7: RoBERTa Reservoir Results, Training plot for 12 layer RoBERTa (left). AUCC result (right). F Validation Plots Here we present the validation plots for training a 8-layer encoder, 2-layer decoder model for IWSLT14, a 24-layer encoder, 1-layer decoder model for WMT16, a 48-layer decoder model for enwik8 and a 12-layer decoder model for RoBERTa for detailed steps to calculate the AUCC. It can be clearly observed that given the configurations from Section 3.1, all the models have converged. So when we compute the area under the convergence curve, this depicts the training efficiency of the model (basically time x performance) until convergence. Specifically, we set T 4309 2 4 6 8 10 12 # Total Encoder Layers 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir 2 4 6 8 10 12 # Total Encoder Layers 32.5 33.0 33.5 34.0 test BLEU Transformer T Reservoir FFN Reservoir Figure 8: Validation BLEU AUCC and test BLEU for IWSLT (high is good). Comparison of regular transformer and reservoir transformer with FFN or Transformer reservoir layers added. 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 32.5 # Total Encoder Layers 0.94 0.95 0.96 0.97 0.98 0.99 1.00 valid BLEU AUCC Transformer T Reservoir FFN Reservoir 12.5 15.0 17.5 20.0 22.5 25.0 27.5 30.0 32.5 # Total Encoder Layers 26.25 26.50 26.75 27.00 27.25 27.50 27.75 28.00 test BLEU Transformer T Reservoir FFN Reservoir Figure 9: Validation BLEU AUCC and test BLEU for WMT (high is good). Comparison of regular transformer and reservoir transformer with FFN or Transformer reservoir layers added. sufficiently high for computing the AUCC, which is 4h for IWSLT, 20h for WMT, 30h for enwik8 and 60h for RoBERTa pretraning. From the training plot in the appendix, we can see that each model has converged at that point. The Reservoir model in Figure 12 has 2 layers frozen for IWSLT14, 8 layers frozen for enwik8, and 4 layers frozen for WMT16 and RoBERTa. G Backskipping Figure 13 shows the BLUE curves for IWSLT comparing regular vs reservoir vs backskipped transformers, with the latter performing surprisingly well. 30 40 50 60 70 # Total Decoder Layers 0.6 0.7 0.8 0.9 1.0 valid bpc AUCC Transformer T Reservoir FFN Reservoir 30 40 50 60 70 # Total Decoder Layers 1.2 1.4 1.6 1.8 2.0 2.2 2.4 test bpc Transformer T Reservoir FFN Reservoir Figure 10: Validation BPC AUCC and test BPC on the enwik8 language modelling task (low is good). Comparison of regular and reservoir transformers for varying depths. 4 6 8 10 12 14 16 18 20 # Total Decoder Layers 91 92 93 94 95 96 valid accuracy Transformer T Reservoir FFN Reservoir Transformer (frozen finetuned) 4 6 8 10 12 14 16 18 20 # Total Decoder Layers 78 80 82 84 86 valid accuracy Transformer T Reservoir FFN Reservoir Transformer (frozen finetuned) Figure 11: Downstream RoBERTa performance on SST-2 (left) and MultiNLI-matched (right). 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Training Hours (h) 10 15 20 25 30 35 Validation BLEU Transformer T Reservoir FFN Reservoir 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Training Hours (h) 10 12 14 16 18 20 22 24 26 Validation BLEU Transformer T Reservoir FFN Reservoir 0 5 10 15 20 25 30 35 40 Training Hours (h) 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 Validation BPC Validation curve for training on enwik8 Transformer T Reservoir FFN Reservoir 0 12 24 36 48 60 Training Hours (h) 4 6 8 10 12 14 16 18 20 Validation PPL Transformer T Reservoir FFN Reservoir Figure 12: IWSLT with 2-layer decoder validation plot (upper left). WMT with 24-layer decoder validation plot (upper right). Enwik8 with 48-layer decoder validation plot (lower left). RoBERTa with 12-layer decoder validation plot (lower right). 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Training Hours (h) 10 15 20 25 30 35 Validation BLEU Validation curve for training on IWSLT14 Transformer T Reservoir Backskipped Reservoir Figure 13: IWSLT comparison of the regular, reservoir and backskipped transformer architectures (encoder has 8 layers with 2 frozen, if any).
|
2021
|
331
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4310–4321 August 1–6, 2021. ©2021 Association for Computational Linguistics 4310 Subsequence Based Deep Active Learning for Named Entity Recognition Puria Radmard1,2,3 [email protected] Yassir Fathullah2 [email protected] 1University College London 2University of Cambridge 3Vector AI Aldo Lipani1,3 [email protected] Abstract Active Learning (AL) has been successfully applied to Deep Learning in order to drastically reduce the amount of data required to achieve high performance. Previous works have shown that lightweight architectures for Named Entity Recognition (NER) can achieve optimal performance with only 25% of the original training data. However, these methods do not exploit the sequential nature of language and the heterogeneity of uncertainty within each instance, requiring the labelling of whole sentences. Additionally, this standard method requires that the annotator has access to the full sentence when labelling. In this work, we overcome these limitations by allowing the AL algorithm to query subsequences within sentences, and propagate their labels to other sentences. We achieve highly efficient results on OntoNotes 5.0, only requiring 13% of the original training data, and CoNLL 2003, requiring only 27%. This is an improvement of 39% and 37% compared to querying full sentences. 1 Introduction The availability of large datasets has been key to the success of deep learning in Natural Language Processing (NLP). This has galvanized the creation of larger datasets in order to train larger deep learning models. However, creating high quality datasets is expensive due to the sparsity of natural language, our inability to label it efficiently compared to other forms of data, and the amount of prior knowledge required to solve certain annotation tasks. Such a problem has motivated the development of new Active Learning (AL) strategies which aim to efficiently train models, by automatically identifying the best training examples from large amounts of Code is made available on: https://github.com/ puria-radmard/RFL-SBDALNER unlabeled data (Wei et al., 2015; Wang et al., 2017; Tong and Koller, 2002). This tremendously reduces human annotation effort as much fewer instances need to be labeled manually. To minimise the amount of data needed to train a model, AL algorithms iterate between training a model, and querying information rich instances to human annotators from a pool of unlabelled data (Huang et al., 2014). This has been shown to work well when the queries are ‘atomic’—a single annotation requires a unit labour, and describes entirely the instance to be annotated. Conversely, each instance of structured data, such as sequences, require multiple annotations. Hence, such query selection methods can result in a waste of annotation budget (Settles, 2011). For example, in Named Entity Recognition (NER), each sentence is usually considered an instance. However, because each token has a separate label, annotation budgeting is typically done on a token basis (Shen et al., 2017). Budget wasting may therefore arise from the heterogeneity of uncertainty across each sentence; a sentence can contain multiple subsequences (of tokens) of which the model is certain on some and uncertain on others. By making the selection at a sentence level, although some budget is spent on annotating uncertain subsequences, the remaining budget may be wasted on annotating subsequences for which an annotation is not needed. It can therefore be desirable for annotators to label subsequences rather than the full sentences. This gives a greater flexibility to AL strategies to locate information rich parts of the input with improved efficiency – and reduces the cognitive demands required of annotators. Annotators may in fact perform better if they are asked to annotate shorter sequences, because longer sentences can cause boredom, fatigue, and inaccuracies (Rzeszotarski et al., 2013). 4311 In this work, we aim to improve upon the efficiency of AL for NER by querying for subsequences within each sentence, and propagating labels to unseen, identical subsequences in the dataset. This strategy simulates a setup in which annotators are presented with these subsequences, and do not have access to the full context, ensuring that their focus is centred on the tokens of interest. We show that AL algorithms for NER tasks that use subsequences, allowing training on partially labelled sentences, are more efficient in terms of budget than those that only query full sentences. This improvement is furthered by generalising existing acquisition functions (§ 4.1) for use with sequential data. We test our approaches on two NER datasets, OntoNotes 5.0 and CoNLL 2003. On OntoNotes 5.0, Shen et al. (2017) achieve stateof-the-art performance with 25% of the original dataset querying full sentences, while we require only 13% of the dataset querying subsequences. On CoNLL 2003, we show that the AL strategy of Shen et al. (2017) requires 50% of the dataset to achieve the same results as training on the full dataset, while ours requires only 27%. Contributions of this paper are: 1. Improving the efficiency of AL for NER by allowing querying of subsequences over full sentences; 2. An entity based analysis demonstrating that subsequence querying AL strategies tend to query more relevant tokens (i.e., tokens belonging to entities); 3. An uncertainty analysis of the queries made by both full sentence and subsequence querying methods, demonstrating that querying full sentences leads to selecting more tokens to which the model is already certain. 2 Related Work AL algorithms aim to query information rich data points to annotators in order to improve the performance of the model in a data efficient way. Traditionally these algorithms choose data points which lie close to decision boundaries (Pinsler et al., 2019), where uncertainty is high, in order for the model to learn more useful information. This measure of uncertainty, measured through acquisition functions, are therefore vital to AL. Key functions include predictive entropy (MaxEnt) (Gal et al., 2017), mutual information between model posterior and predictions (BALD) (Houlsby et al., 2011; Gal et al., 2017), or the certainty of the model when making label predictions (here called LC) (Mingkun Li and Sethi, 2006). These techniques ensure all instances used for training, painstakingly labelled by experts, have maximum impact on model performance. There has been exploration of uncertainty and deep learning based AL for NER (Chen et al., 2015; Shen et al., 2017; Settles and Craven, 2008; Fang et al., 2017). These approaches however, treat each sentence as a single query instead of a collection of individually labelled tokens. In these methods, the acquisition functions that score sentences aggregate token-wise scores (through summation or averaging). Other works forgo this aggregation, querying single tokens at a time (Tomanek and Hahn, 2009; Wanvarie et al., 2011; Marcheggiani and Arti`eres, 2014). These works show that AL for NER can be improved by taking the single token as a unit query, and use semi-supervision (Reddy et al., 2018; Iscen et al., 2019) for training on partially labelled sentences (Muslea et al., 2002). However, querying single-tokens is inapplicable in practise because, either a) annotators have access to the full sentence when queried but can only label one token, which would lead to frustration as they are asked to read the full sentence but only annotate a single token, or b) annotators only have access to the token of interest, which means that they would not have enough information to label tokens differently based on their context, leading to annotators labeling any unique token with the same label. Moreover, if the latter approach was somehow possible, we would be able to reduce the annotation effort to the annotation of only the unique tokens forming the dataset, its dictionary. Furthermore, all of these past works use Conditional Random Fields (CRFs) (Lafferty et al., 2001), which have since been surpassed as the state-of-the-art for NER (and most NLP tasks) by deep learning models (Devlin et al., 2019). In this work we follow the approach where annotators only have access to subsequences of multiple tokens. However, instead of making use of single tokens, we will query more than one token, providing enough context to the annotators. This allows the propagation of these annotations to identical subsequences in the dataset, further reducing the total annotation effort. 4312 3 Background 3.1 Active Learning Algorithms Most AL strategies are based on a repeating score, query and fine-tune cycle. After initially training an NER model with a small pool of labelled examples, the following is repeated: (1) score all unlabelled instances, (2) query the highest scoring instances and add them to training set, and, (3) fine-tune the model using the updated training set (Huang et al., 2014). To describe this further, notation and proposed training process is introduced, with details in following sections. First, the sequence tagging dataset, denoted by D = {(x(n), y(n))}N n=1, consists of a collection of sentence and ground truth labels. The i-th token of the n-th sentence (y(n) i ) has a label y(n) i = c with c belonging to C = {c1, ..., cK}. We also differentiate between the labelled and unlabelled datasets, DL and DU, which initially are empty and equal to D. Finally, we fix A as the total number of tokens queried in each iteration. 3.2 Acquisition Functions Instances in the unlabelled pool are queried using an acquisition function. This function aims to quantify the uncertainty of the model when generating predictive probabilities over possible labels for each instance. Instances with the highest predictive uncertainty are deemed as the most informative for model training. Previously used acquisition functions such as Least Confidence (LC) and Maximum Normalized Log-Probability (MNLP) (Shen et al., 2017; Chen et al., 2015) are generalised for variable length sequences. Letting ˆy(n) <i be the history of predictions prior to the i-th input, the next output probability will be p(n) i,c = P(ˆy(n) i = c|ˆy(n) <i , x(n)). Then, we define the token-wise LC score as: LC(n) i = −max c∈C log p(n) i,c . (1) The LC acquisition function for sequences is then defined as: LC x(n) 1 , ..., x(n) ℓ = ℓ X j=1 LC(n) j , (2) and, for MNLP as: MNLP x(n) 1 , ..., x(n) ℓ = 1 ℓ ℓ X j=1 LC(n) j . (3) Note that this is similar to LC except for the normalization factor 1/ℓ. The formulation above can be applied to other types of commonly used acquisition functions such as Maximum Entropy (MaxEnt) (Gal et al., 2017) by simply defining: ME(n) i = − X c∈C p(n) i,c log p(n) i,c , (4) as the token score. Given the task of quantifying uncertainty amongst the unlabelled pool of data, both of these metrics - LC and MaxEnt - provide intuitive interpretations. eq. (1) scores highly tokens for which the predicted label has lowest confidence, while eq. (4) scores highly tokens for which the whole probability mass function has higher entropy. Both of these therefore score more highly uniform predictive distributions, which indicates underlying uncertainty. Finally, given the similarity of performance between MNLP and Bayesian Active Learning by Disagreement (BALD) (Houlsby et al., 2011) in NER tasks (Shen et al., 2017), and the computational complexity required to calculate BALD with respect to the other activation functions, we will not compare against BALD. 4 Subsequence Acquisition In this section we describe how we build on past works, and the core contribution of this paper. Our work forms a more flexible AL algorithm that operates on subsequences, as opposed to full sentences (Shen et al., 2017). This is achieved by generalising acquisition functions for subsequences (§ 4.1) scoring and querying subsequences within sentences (§ 4.2), and performing label propagation on unseen sentences to avoid the multiple annotations of repeated subsequences (§ 4.3). 4.1 Subsequence Acquisition Functions Since this work focuses on the querying of subsequences, from the previously defined LC and MNLP we generalize them to define a family of acquisition functions applicable for both full sentences and subsequences: LCα x(n) i+1, ..., x(n) i+ℓ = 1 ℓα i+ℓ X j=i+1 LC(n) j . (5) Special cases are when α = 0 and α = 1 which return the original definitions of LC in eq. (2) and MNLP in eq. (3). As noted by Shen et al. (2017), 4313 LC for sequences biases acquisition towards longer sentences. The tuneable normalisation factor in eq. (5) over the sequence of scores mediates the balance of shorter and longer subsequences selected. This generalisation can be applied to other types of commonly used acquisition functions such as MaxEnt and BALD by modifying the token-wise score. 4.2 Subsequence Selection Each sentence x(n) can be broken into a set of subsequences S(n) = {(x(n) i , ..., x(n) j )|∀i < j} where all elements s ∈S(n) can be efficiently scored by first computing the token scores, then aggregating as required. Once this has been done for all sentences in DU, a query set SQ ⊂∪nS(n) of non-overlapping (mutually disjoint) subsequences is found. The requirement of non-overlapping subsequences avoids the problem of relabelling tokens, but disallows simply choosing the highest scoring subsequences (since these can overlap). Instead at each round of querying, we perform a greedy selection, repeatedly choosing the highest scoring subsequence that does not overlap with previously selected subsequences. Adjustments can be made to reflect practical needs, such as restricting the length ℓof the viable subsequences to [ℓmin, ℓmax]. This is because longer subsequences are easier to label, while shorter subsequences are more efficient in querying uncertain tokens, and so the selection is only allowed to operate within these bounds. Additionally, it is easy to imagine a scenario in which a greedy selection method does not select the maximum total score that can be generated from a sentence. This scenario is illustrated in Table 1 where lengths are restricted to ℓmin = ℓmax = 3 for simplicity. Note that tokens can become unselectable in future rounds because they are not inside a span of unlabelled tokens of at least size ℓmin. When the algorithm has queried all subsequences of this size range, it starts to query shorter subsequences by relaxing the length constraint. However in practise, model performance on the validation set converges before all subsequences of valid range have been exhausted. Nonetheless, when choosing subsequences of size [ℓmin, ℓmax] = [4, 7] these will be exhausted when roughly 90% and 80% of tokens have been labelled for the OntoNotes 5.0 and CoNLL 2003 datasets. 4.3 Subsequence Label Propagation Since a subsequence querying algorithm can result in partially labelled sentences, it raises the question of how unlabelled tokens should be handled. In previous work based on the use of CRFs (Tomanek and Hahn, 2009; Wanvarie et al., 2011; Marcheggiani and Arti`eres, 2014) this was solved by using semisupervision on tokens for which the model showed low uncertainty. However, for neural networks, the use of model generated labels could lead to the model becoming over-confident, harming performance and biasing (Arazo et al., 2020) uncertainty scores. Hence, we ensure that backpropagation only occurs from labelled tokens. Our final contribution to the AL algorithm is the use of another semi-supervision strategy where we propagate uniquely labelled subsequences in order to minimise the number of annotations needed. When queried for a subsequence, the annotator (in this case an oracle) is not given the contextual tokens in the remainder of the sentence. For this reason, given an identical subsequence, a consistent annotator will provide the same labels. Therefore, the proposed algorithm maintains a dictionary that maps previously queried subsequences to their provided labels. Once a queried subsequence and its label are added to the dictionary, all other matching subsequences in the unlabelled pool are given the same, but temporary, labels. The tokens retain these temporary labels until they are queried themselves. After scoring and ranking members of S, the algorithm will disregard sequences that match exactly members of this dictionary, which is updated during the querying round. However, if tokens belonging to these previously seen subsequences are encountered in a different context, meaning as part of a different subsequence, they may also be queried. For example, in Table 1, if the subsequence “shop to buy” had been previously queried elsewhere in the dataset, the red subsequence will not be considered for querying, as it retains its temporary labels. Instead, the green subsequence could be queried, in which case the temporary labels of tokens 6 and 7 will be overwritten by new, permanent labels. Therefore, the value of ℓmin becomes a trade-off between the improved resolution of the acquisition function, and the erroneous propagation of shorter, more frequent label subsequences to identical ones in different contexts. 4314 j 1 2 3 4 5 6 7 8 9 10 x(n) j Yassir is going to the shop to buy shoes . y(n) j X O O O X X X X X X lc(n) j 3.22 0.41 0.78 0.83 0.60 0.27 0.50 LC1 = 0.67 LC1 = 0.46 LC1 = 0.74 LC1 = 0.57 Table 1: This shows the subsequences from a sentence using ℓmin = ℓmax = 3, α = 1. Besides the token index j, the top three rows show the tokens, labels, and the token-wise scores. If y(n) j = X, then the corresponding token is unlabelled, hence the score is considered when selecting the next query. After this, the subsequences constituting S(n) are displayed with their LC1 scores. In this case “shop to buy” will be chosen since it maximises LC1, but ‘traps’ its surrounding tokens until ℓmin is lowered to 2 and “shoes .” may be considered. 4.4 Subsequence Active Learning Algorithm Finally, we summarise the AL algorithm proposed. Given a set of unlabelled data DU, we initially randomly select a proportion of sentences from DU, label them, and add these to DL. A dictionary B is also initialised. Using these labelled sentences we train a model. Then, the following proposed training cycle is repeated until DU is empty (or an early stopping condition is reached): 1. Find all consecutive unlabelled subsequences in DU, and score them using a pre-defined acquisition function. 2. Select the top scoring non-overlapping subsequences SQ that do not appear in B, such that the number of tokens in SQ is A, and query them to the annotators. Update DL and DU. As each sequence is selected, add it to B, mapping it to its true labels. 3. Provide all occurrences of the keys of B in DU with their corresponding temporary labels. These will not be included in DL as these are temporary. 4. Finetune the model on sentences with any label, temporary and permanent. Repeat this process until convergence. 5 Experimental Setup 5.1 Datasets As in previous works (Shen et al., 2017), we use the two following NER datasets: OntoNotes 5.0. This is a dataset used to compare results with the full sentence querying baseline (Weischedel, Ralph et al., 2013), and comprising of text coming from: news, conversational telephone speech, weblogs, usenet newsgroups, broadcast, and talk shows. This is a BIO formatted dataset with a total of K = 37 classes and 99,333 training sentences, with an average sentence length of 17.8 tokens in its training set. CoNLL 2003. This is a dataset, also in BIO format, with only 4 entity types (LOC, MISC, PER, ORG) resulting in K = 9 labels (Tjong Kim Sang and De Meulder, 2003). This dataset is made from a collection of news wire articles from the Reuters Corpus (Lewis et al., 2004). The average sentence length is 12.6 tokens in its training set. A full list of class types and entity lengths and frequencies for both datasets can be found in the Appendix. 5.2 NER Model Following the work of Shen et al. (2017), a CNNCNN-LSTM model for combined letter- and tokenlevel embeddings was used; see Appendix for an overview of the model and hyperparameters setting and validation. Furthermore, the AL algorithm used in (Shen et al., 2017) will serve as one of the baselines following the same procedure. This represents an equivalent algorithm to that proposed, but which can only query full sentences, and does not use label propagation. 5.3 Model Training and Evaluation As the evaluation measure we use the F1 score. After the first round of random subsequence selection, the model is trained. After subsequent selections the model is finetuned - training is resumed from the previous round’s parameters. In all cases, the model training was stopped either after 30 epochs were completed, or if the F1 score for the valida4315 0 5 10 15 20 25 30 Percentage of tokens manually labelled 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 F1 FS, α = 1 FS, α = 0 FS, α = 0.7 SUB, α = 0.1 FS, random SUB, random No AL (a) LCα for OntoNotes 5.0 NER dataset 0 5 10 15 20 25 30 35 40 Percentage of tokens manually labelled 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 F1 FS, α = 1 FS, α = 0 FS, α = 0.7 SUB, α = 1 FS, random SUB, random No AL (b) LCα for CoNLL 2003 NER dataset Figure 1: F1 score on test set achieved each round using round-optimal model parameters. All subsequence experiments here use ℓmin = 4, ℓmax = 7. Each curve is averaged over 10 runs. tion set had monotonically decreased for 2 epochs. This validation set is made up of a randomly selected 1% of sentences of the original training set. After finetuning, the model reloads its parameters from the round-optimal epoch, and its performance is evaluated on the test set. Furthermore, the AL algorithms were also stopped after all hyperparameter variations using that dataset and acquisition function family had converged to the same best F1, which we denote with F ∗ 1 . For the OntoNotes 5.0 dataset, F ∗ 1 value was achieved after 30% of the training set was labelled, and for the CoNLL 2003 dataset after 40%. 5.4 Active Learning Setup & Evaluation We choose ℓmin = 4 to give a realistic context to the annotator, and to avoid a significant propagation of common subsequences. The upper bound of ℓmax = 7 was chosen to ensure subsequences were properly utilised, since the average sentence length of both datasets is roughly twice this size. For the OntoNotes 5.0 dataset, every round A = 10, 000 tokens are queried, whereas for the CoNLL 2003 dataset A = 2, 000 tokens. These represent roughly 0.5% and 1% of the available training set. We evaluate the efficacy and efficiency of the tested AL strategies in three ways. First, model performance over the course of the algorithm was evaluated using end of round F1 score on the test set. We compare the proportion of the dataset’s tokens labelled when the model achieves 99% of the F ∗ 1 score ( ˆF1 ∗= 0.99 × F ∗ 1 ). We also quantify the rate of improvement of model performance during training using the normalised Area Under the Curve (AUC) score of each F1 test curve. The normalisation ensures that the resulting AUC score is in the range [0, 1], and it is achieved by dividing the AUC score by the size of the dataset. This implies that methods that converge faster to their best performance will have a higher normalized AUC. Second, we consider how quickly the algorithms can locate and query relevant tokens (named entities). Third, we finally evaluate their ability to extract the most uncertain tokens from the unlabelled pool. 6 Results & Discussion 6.1 Active Learning Performance Figure 1 shows the LCα performance curves for α = 0, α = 1 and the best performing value for each acquisition class (based on the normalised AUC score, Table 3) for full sentence querying (FS), and only the best performing α values for subsequence querying (SUB). The figure also shows the performance of training on the complete training set (No AL), and when the both sentences and subsequences are random selected by the acquisition function. The equivalent figures for MaxEntα are available in Appendix, and follow similar trends. Then, the performance of each curve, quantified in terms of the normalised AUC is summarised in Table 3. Table 2 shows further analysis of the best results in Figure 1, with best referring to acquisition function and optimal α. These results first show that subsequence querying methods are more efficient than querying full sentences, achieving their final F1 with substantially less annotated data, and with higher normalised AUC scores. For OntoNotes 5.0, querying subsequences reduces final proportion required by 38.8%. For CoNLL 2003, this reduction is 36.6%. Altogether, subsequence querying holds improved efficiency over the full sentence querying baseline. 4316 F1 Final Frac. Score of the Dataset 100% AL FS SUB ON 5.0 0.829 0.843 22% 13% CoNLL 0.930 0.938 42% 27% Table 2: Summary of the results of the AL strategies from Figure 1, when the models are trained using 100% of the training set and active learning (AL), with the best hyperparameter setting of the acquisition function with for full sentence and subsequence, based on normalised AUC score. As a point of interest, full sentence querying can be easily improved by optimising α alone. For the OntoNotes 5.0 dataset, using LC1, 24.2% of tokens are required to achieve F ∗ 1 . This however, can be improved by 9.33% to only requiring 22.0% by choosing α = 0.7. For CoNLL 2003, using LC1 for full sentences, 50.0% of the dataset was required, but when using LC0.7, it was 40.7% of the tokens. 6.2 Entity Recall This section and the next aim to understand some of the underlying mechanisms that allow the subsequence querying methods to achieve results substantially better than a full sentence baseline. Namely, the ability of the different methods to extract the tokens for which the model is the most uncertain about. Given that the majority of tokens in both datasets have the same label - “O”, signifying no entity - it is likely that tokens belonging to entities, particularly rarer classes, trigger higher model uncertainty. Querying full sentences at a time, the AL algorithm will spend much of its token budget for that round labelling non-entity tokens while attempting to locate the more informative entities. Subsequence querying methods, not faced with this wasteful behaviour, allow the AL algorithm to query entity tokens quicker, locating and labelling the majority of entity tokens faster over the course of training. The proportion of tokens belonging to entities that the AL algorithm has queried against the round number is plotted in Figure 2 for OntoNotes 5.0. For both datasets, the random querying methods contain a distribution of token classes that reflect the dataset at large, producing roughly linear curves for this figure. Curves for all methods that employ Figure 2: Proportion of tokens that belong to entities labelled, against the round number. an uncertainty based acquisition function are concave, and the AUC reflects the ranking of model performance for each querying method. This relation suggests that shortly after initialisation, better performing algorithm variations query entity tokens faster. In later stages of finetuning this rate is reduced, likely because after labelling a large proportion of them, the remaining entity tokens cause little uncertainty for the model. In a practical setting where querying may have to be stopped before model performance has converged (i.e. due to accumulated cost of annotations), it is greatly beneficial to ensure that the model is exposed to a high number of relevant tokens, because this increases the likelihood of locating entity tokens belonging to underrepresented classes at an early stage. 6.3 Uncertainty Score Analysis Finally, this section compares the scores of tokens in the queried set SQ for each querying method. Comparing the distribution and development of these scores provides a direct insight to the core assumptions of why full sentence querying is outperformed. Figure 3 shows the difference in score distributions for sentence versus subsequence querying, against querying round number, for rounds preceding model performance convergence. First, it is seen that decreasing the individual query size (full sentence to subsequence) increases the median uncertainty extracted at the earlier rounds. Second, Figure 3 provides evidence for the mechanism suggested earlier: aggregating the token scores across full sentences means querying both the highly uncertain tokens, and the tokens that provide little uncertainty. Querying high scoring sentences like this can cause a distribution with two peaks as seen in 4317 Dataset Acquisition Function Full Sentence Subsequence α = 0 α = 1 Optimal (α) α = 0 α = 1 Optimal (α) OntoNotes 5.0 LCα 0.794 0.802 0.804 (0.7) 0.817 0.812 0.818 † (0.1) MaxEntα 0.791 0.803 0.803 (1.0) 0.815 0.813 0.816 † (0.5) Random 0.734 0.769 CoNLL 2003 LCα 0.857 0.875 0.879 (0.7) 0.885 0.883 0.892 † (1.0) MaxEntα 0.841 0.882 0.882 (1.0) 0.881 0.883 0.891 † (0.9) Random 0.824 0.859 Table 3: Normalised AUC scores for model performance (F1 score on test set) for α = 0, 1, and its optimal value in each case. Each pair of differences between the optimized acquisition function for full sentences and subsequences (indicated by a †) are significantly different (two-sided unpaired t-test, with p-value < 0.05). 1 5 10 15 Round number 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Token-wise score FS SUB Figure 3: Distributions of the queried LC scores for the OntoNotes 5.0 dataset, made on the 1st, 5th, 10th, and 15th scoring rounds. This corresponds to scores after training on 1%, 3.2%, 6.1%, and 9.0% of the utilised training set. the figure. As the model becomes increasingly certain about its predictions, high scores are localised within smaller subsequences, and the coarse sensitivity of full sentence querying means it forfeits all the higher scoring tokens. These differences were also observed when comparing subsequence querying methods with sub-optimal α. This figure only analyses behaviour of up to 9% of the training set’s tokens have been queried. Instead, Figure 4 show how the mean of token-wise scores evolve for different querying methods for the OntoNotes 5.0 dataset until convergence. This clearly shows that subsequence querying methods converge faster over the full course of the algorithm compared to full sentence querying. This is consistent with Figure 1 in terms of initial rate and final time of model performance convergence, namely that model performance plateaus alongside the uncertainty score. Keeping track of query scores like this is also a reasonable idea in industrial applications. When Figure 4: Average value of LC for all tokens in SQ with confidence intervals, against round number. Score values are averaged over all tested values of α training on a very semantically specific corpus, there may not be enough fully labelled sentences to build a test set. In that case, observing the rate progress of score convergence can be used as an early stopping method for the AL algorithm (Zhu et al., 2010). 7 Conclusion & Future Work In this study we have employed subsequence querying methods for improving the efficiency of AL for NER tasks. We have seen that these methods outperform full sentence querying in terms of annotations required for optimal model performance, requiring 38.8% and 36.6% fewer tokens for the OntoNotes 5.0 and CoNLL 2003 datasets. Optimal results for subsequence querying (and full sentence querying) were achieved by generalising previously used AL acquisition functions, defining a larger family of acquisition functions for sequential data. The analysis of § 6.3 suggests that a full sentence querying causes noisy acquisition functions due to the tokens in the queried sentences that were not 4318 highly scored. This added noise reduces the budget efficiency, and a subsequence querying method eliminates a large part of this effect. This efficiency also translated into a faster recall of named entities in the dataset to be queried (§ 6.2). Limitations and future work: Limitations of this study are largely centred on the use of an oracle to provide tokens with their labels. With human annotators, the cropped context of subsequence queries may make them produce more inaccuracies than when annotating full sentences. such studies will help reveal how context affects label accuracy, how this, in turn, affects optimal hyperparameters in the subsequence selection process (such as optimal query length), further accommodations that must be made to effectively optimise worker efficiency, and how to deal with unreliable labels. We leave to future work the evaluation of these querying methods with human annotators. There are also ways to incorporate model generated labelling methods for more robust semi-supervision into our framework that we leave to future work. Finally, there are examples of other tasks for structured data, such as audio, video, and image segmentation, where the part of an instance may be queried. A generalisation of the strategy demonstrated for the NER case may allow for more efficient active learning querying methods for these other types of data. References Eric Arazo, Diego Ortego, Paul Albert, Noel E. O’Connor, and Kevin McGuinness. 2020. Pseudolabeling and confirmation bias in deep semisupervised learning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1– 8. Yukun Chen, Thomas A. Lasko, Qiaozhu Mei, Joshua C. Denny, and Hua Xu. 2015. A study of active learning methods for named entity recognition in clinical text. Journal of Biomedical Informatics, 58:11 – 18. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Meng Fang, Yuan Li, and Trevor Cohn. 2017. Learning how to active learn: A deep reinforcement learning approach. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 595–605, Copenhagen, Denmark. Association for Computational Linguistics. Yarin Gal, Riashat Islam, and Zoubin Ghahramani. 2017. Deep Bayesian active learning with image data. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1183–1192, International Convention Centre, Sydney, Australia. PMLR. Neil Houlsby, Ferenc Husz´ar, Zoubin Ghahramani, and M´at´e Lengyel. 2011. Bayesian active learning for classification and preference learning. S. Huang, R. Jin, and Z. Zhou. 2014. Active learning by querying informative and representative examples. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(10):1936–1949. Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. 2019. Label propagation for deep semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). John Lafferty, Andrew McCallum, and Fernando CN Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. Yann LeCun and Yoshua Bengio. 1998. Convolutional Networks for Images, Speech, and Time Series, page 255–258. MIT Press, Cambridge, MA, USA. David D Lewis, Yiming Yang, Tony G Rose, and Fan Li. 2004. Rcv1: A new benchmark collection for text categorization research. Journal of machine learning research, 5(Apr):361–397. Wang Ling, Chris Dyer, Alan W. Black, and Isabel Trancoso. 2015. Two/too simple adaptations of Word2Vec for syntax problems. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1299– 1304, Denver, Colorado. Association for Computational Linguistics. Diego Marcheggiani and Thierry Arti`eres. 2014. An experimental comparison of active learning strategies for partially labeled sequences. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 898– 906, Doha, Qatar. Association for Computational Linguistics. Mingkun Li and I. K. Sethi. 2006. Confidence-based active learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(8):1251–1261. Ion Muslea, Steven Minton, and Craig A. Knoblock. 2002. Active + semi-supervised learning = robust 4319 multi-view learning. In Proceedings of the Nineteenth International Conference on Machine Learning, ICML ’02, page 435–442, San Francisco, CA, USA. Morgan Kaufmann Publishers Inc. Robert Pinsler, Jonathan Gordon, Eric Nalisnick, and Jos´e Miguel Hern´andez-Lobato. 2019. Bayesian batch active learning as sparse subset approximation. In Advances in Neural Information Processing Systems, volume 32, pages 6359–6370. Curran Associates, Inc. Y Reddy, Viswanath Pulabaigari, and Eswara B. 2018. Semi-supervised learning: a brief review. International Journal of Engineering Technology, 7:81. Jeffrey Rzeszotarski, Ed Chi, Praveen Paritosh, and Peng Dai. 2013. Inserting micro-breaks into crowdsourcing workflows. Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, 1(1). Burr Settles. 2011. From theories to queries: Active learning in practice. In Active Learning and Experimental Design workshop In conjunction with AISTATS 2010, volume 16 of Proceedings of Machine Learning Research, pages 1–18, Sardinia, Italy. JMLR Workshop and Conference Proceedings. Burr Settles and Mark Craven. 2008. An analysis of active learning strategies for sequence labeling tasks. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’08, page 1070–1079, USA. Association for Computational Linguistics. Yanyao Shen, Hyokun Yun, Zachary Lipton, Yakov Kronrod, and Animashree Anandkumar. 2017. Deep active learning for named entity recognition. In Proceedings of the 2nd Workshop on Representation Learning for NLP, pages 252–256, Vancouver, Canada. Association for Computational Linguistics. Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142–147. Katrin Tomanek and Udo Hahn. 2009. Semisupervised active learning for sequence labeling. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pages 1039–1047, Suntec, Singapore. Association for Computational Linguistics. Simon Tong and Daphne Koller. 2002. Support vector machine active learning with applications to text classification. J. Mach. Learn. Res., 2:45–66. K. Wang, D. Zhang, Y. Li, R. Zhang, and L. Lin. 2017. Cost-effective active learning for deep image classification. IEEE Transactions on Circuits and Systems for Video Technology, 27(12):2591–2600. Dittaya Wanvarie, Hiroya Takamura, and Manabu Okumura. 2011. Active learning with subsequence sampling strategy for sequence labeling tasks. Journal of Natural Language Processing, 18(2):153–173. Kai Wei, Rishabh Iyer, and Jeff Bilmes. 2015. Submodularity in data subset selection and active learning. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1954–1963, Lille, France. PMLR. Weischedel, Ralph, Palmer, Martha, Marcus, Mitchell, Hovy, Eduard, Pradhan, Sameer, Ramshaw, Lance, Xue, Nianwen, Taylor, Ann, Kaufman, Jeff, Franchini, Michelle, El-Bachouti, Mohammed, Belvin, Robert, and Houston, Ann. 2013. Ontonotes release 5.0. Jingbo Zhu, Huizhen Wang, Eduard Hovy, and Matthew Ma. 2010. Confidence-based stopping criteria for active learning for data annotation. ACM Trans. Speech Lang. Process., 6(3). 4320 A Model Architecture The model architecture is built of three sections. The character-level convolutional neural network (CNN) (LeCun and Bengio, 1998) character-level encoder extracts character level features, wword j for each token x(i) j in a sentence. Then, a latent token embedding wemb j corresponding to that token is generated. The full representation of the token is the concatentation of the two vectors: wfull j := (wchar j , wemb j ). The token-label embeddings, wemb, are initialised using word2vec (Ling et al., 2015), and updated during training and finetuning, as per the baseline paper. A second, token-level CNN encoder is used to generate {htoken j }ℓi j=1, given the tokenlevel representations {wfull j }ℓi j=1. The final tokenlevel encoding is defined by another concatentation: hEnc j := (htoken j , wfull j ). Finally, a tag decoder is used to generate the token-level pmfs over the C possible token classes: {hEnc j }ℓi j=1 LSTM −−−→{ˆy(i) j }ℓi j=1. B Model & Training Parameters Table 4 lists the hyperparameter values used to train the NER model. Note that while dropout is used during training, it is turned off when generating the probabilities that contribute to the scoring of the acquisition function. Model was developed using PyTorch, and trained on a Titan RTX. Hyperparameter Value Batch size 32 Dropout rate for convolutional layers 0.5 Dropout rate for embedding layers 0.25 Gradient clipping magnitude 0.35 Character- and token-level CNN kernel size 3 Layers in character- and token-level CNNs 3 Character embedding vector size 50 Number of filters per character-level CNN layer 50 Number of filters per token-level CNN layer 300 Optimiser type SGD Optimiser learning rate 1.0 Table 4: Values of model and training hyperparameters used throughout the investigation. C Dataset Analysis Here, we cluster similar labels in the BIO format, reducing the total K classes to the K(r) = (K + 1)/2 class groups c(r) 1 , ..., c(r) K(r). Therefore, c(r) 1 corresponds exactly to c1, the empty label, while c(r) k , k > 1 groups the raw labels c2k−2 and c2k−1. Figures 5 and 7 show the distribution of these class groups for the OntoNotes 5.0 and CoNLL 2003 datasets respectively. For the former, counts range from 199 tokens for the ’LANGUAGE’ to 46698 tokens for the ’ORG’ class. The full available training set totals 1766955 tokens in 99333 sentences; this is partitioned into a train and validation set during experimentation. A further test set comprises of 146253 tokens in 8057 sentences. The latter’s training set contains 172210 tokens in 13689 sentences, and its test set has 42141 tokens in 3091 sentences sentences. Figure 5: Composition of token classes in the OntoNotes 5.0 English NER training set. Figure 6: Lengths of entities in the Onto-Notes 5.0 training set in number of tokens, again omitting the empty class ’O’ 4321 Figure 7: Composition of token classes in the CoNLL 2003 NER training set. Figure 8: Lengths of entities in the CoNLL 2003 training set in number of tokens, again omitting the empty class ’O’ D Active Learning Results for Both Datasets In Figure 9 we show the model performance plotted against the percentage of the tokens used as a training set for all the combinations of acquisition functions. 0 5 10 15 20 25 30 Percentage of tokens manually labelled 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 F1 FS, α = 1 FS, α = 0 FS, α = 0.7 SUB, α = 0.1 FS, random SUB, random No AL (a) LCα for OntoNotes5.0 NER dataset 0 5 10 15 20 25 30 35 40 Percentage of tokens manually labelled 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 F1 FS, α = 1 FS, α = 0 SUB, α = 0.9 FS, random SUB, random No AL (b) MaxEntα for OntoNotes5.0 NER dataset 0 5 10 15 20 25 30 35 40 Percentage of tokens manually labelled 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 F1 FS, α = 1 FS, α = 0 FS, α = 0.7 SUB, α = 1 FS, random SUB, random No AL (c) LCα for CoNLL 2003 NER dataset 0 5 10 15 20 25 30 35 40 Percentage of tokens manually labelled 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 F1 FS, α = 1 FS, α = 0 SUB, α = 0.9 FS, random SUB, random No AL (d) MaxEntα for CoNLL 2003 NER dataset Figure 9: F1 score on test set achieved each round (top) and against time (bottom in each case) using roundoptimal model parameters. All subsequence experiments here use ℓmin = 3, ℓmax = 6.
|
2021
|
332
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4322–4333 August 1–6, 2021. ©2021 Association for Computational Linguistics 4322 Convolutions and Self-Attention: Re-interpreting Relative Positions in Pre-trained Language Models Tyler A. Chang1,2, Yifan Xu1, Weijian Xu1, Zhuowen Tu1 1University of California San Diego 2Halıcıo˘glu Data Science Institute {tachang, yix081, wex041, ztu}@ucsd.edu Abstract In this paper, we detail the relationship between convolutions and self-attention in natural language tasks. We show that relative position embeddings in self-attention layers are equivalent to recently-proposed dynamic lightweight convolutions, and we consider multiple new ways of integrating convolutions into Transformer self-attention. Specifically, we propose composite attention, which unites previous relative position embedding methods under a convolutional framework. We conduct experiments by training BERT with composite attention, finding that convolutions consistently improve performance on multiple downstream tasks, replacing absolute position embeddings. To inform future work, we present results comparing lightweight convolutions, dynamic convolutions, and depthwiseseparable convolutions in language model pretraining, considering multiple injection points for convolutions in self-attention layers. 1 Introduction In recent years, Transformer-based language models have brought dramatic improvements on a wide range of natural language tasks (Brown et al., 2020; Devlin et al., 2019). The central innovation of Transformer architectures is the self-attention mechanism (Vaswani et al., 2017), which has grown beyond NLP, extending into domains ranging from computer vision (Dosovitskiy et al., 2021) and speech recognition (Dong et al., 2018) to reinforcement learning (Parisotto et al., 2020; Touvron et al., 2020). In computer vision, self-attention and convolutions have been combined to achieve competitive results for image classification (Bello et al., 2019). Similarly, researchers in NLP have begun integrating convolutions into self-attention for natural language tasks. Recent work has shown initial success adding convolutional modules to self-attention in pre-trained language models (Jiang et al., 2020), or even replacing self-attention entirely with dynamic convolutions (Wu et al., 2019). These successes defy theoretical proofs showing that multi-headed self-attention with relative position embeddings is strictly more expressive than convolution (Cordonnier et al., 2020). To identify why convolutions have been successful in NLP, we seek to isolate the differences between self-attention and convolution in the context of natural language. In this work, we formalize the relationship between self-attention and convolution in Transformer encoders by generalizing relative position embeddings, and we identify the benefits of each approach for language model pre-training. We show that self-attention is a type of dynamic lightweight convolution, a data-dependent convolution that ties weights across input channels (Wu et al., 2019). Notably, previous methods of encoding relative positions (Shaw et al., 2018; Raffel et al., 2020) are direct implementations of lightweight convolutions. Under our framework, the benefits of convolution come from an ability to capture local position information in sentences. Then, we propose composite attention, which applies a lightweight convolution that combines previous relative position embedding methods. We find that composite attention sufficiently captures the information provided by many other convolutions. To validate our framework, we train BERT models that integrate self-attention with multiple convolution types, evaluating our models on the GLUE benchmark (Wang et al., 2018). All of our convolutional variants outperform the default model, demonstrating the effectiveness of convolutions in enhancing self-attention for natural language tasks. Our empirical results provide evidence for future research integrating convolutions and self-attention for NLP. 4323 octopus used the used coconut the a shell as shield Token i Token j octopus the used coconut the a shell as shield used βj-i αij Token i Token j Attention vector Convolution kernel Figure 1: Generating attention maps using standard self-attention (top) and fixed lightweight convolution (bottom). Attention weights αij are analogous to convolution kernel weights βj−i. 2 Self-attention and lightweight convolutions First, we outline the relationship between selfattention and convolutions. Specifically, we show that a self-attention operation can be viewed as a dynamic lightweight convolution, a depthwise convolution that ties weights along channels (Wu et al., 2019). We then isolate the differences between self-attention and lightweight convolutions, highlighting the benefits of each approach in language models. 2.1 Self-attention In a Transformer self-attention layer, inputs x1, ..., xn ∈Rd are projected to corresponding queries, keys, and values by linear transformations W Q, W K, W V ∈Rd×dh for each attention head, projecting into the head dimension size dh. Output vectors y1, ..., yn ∈Rd are linear combinations of values, concatenating all attention heads. Value weights (before softmaxing) are determined by: αij = (xiW Q)(xjW K)T √dh . (1) Intuitively, αij represents the attention that token i pays to token j, incorporating the value xjW V into the resulting vector yi. From the attention scores between various tokens i and j, an attention map of αij is produced (see Figure 1). 2.2 Lightweight convolutions In contrast, a standard one-dimensional convolution slides a kernel of weights along the input sequence; each feature in each output representation yi is a weighted sum of all features (called “channels”) in the surrounding xi. To save parameters, it is common to consider depthwise convolutions where each channel c in yi is a weighted sum only of the features in channel c for the surrounding xi. Formally, each entry of yi can be written as: yi,c = X −k≤j−i≤k βj−i,c xj,c (2) where k is the kernel size in each direction. Each scalar βj−i,c represents the attention paid to relative position j −i for channel c. To further simplify depthwise convolutions for use in language models, Wu et al. (2019) propose lightweight convolutions, which tie weights βj−i,c along all channels c. As a result, the lightweight convolution contains only 2k + 1 weights, one scalar βj−i for each relative position considered. Then, each yi is a linear combination of surrounding xi: yi = X −k≤j−i≤k βj−i xj (3) Importantly, we can then consider each βj−i as an attention weight analogous to self-attention, representing the attention that token i pays to token j. 4324 The lightweight convolution produces an attention map of βj−i as visualized in Figure 1. Finally, furthering the similarity between lightweight convolutions and self-attention, Wu et al. (2019) propose dynamic lightweight convolutions, which dynamically compute relative weights βj−i based on individual input tokens. In other words, each row in Figure 1 has relative weights determined dynamically based on the input token xi for that row. Because attentions for relative positions are no longer fixed across rows, the attention map in Figure 1 achieves similar flexibility to standard self-attention. 2.3 Self-attention vs. convolution We have shown that both self-attention and lightweight convolution compute linear combinations of token representations, but we now isolate the differences between the two approaches. Perhaps most importantly, the two methods assign attention scores αij and βj−i in fundamentally different ways. Self-attention computes αij based on the dot product between query i and key j, ignoring the relative position between i and j. In this way, selfattention layers model interactions exclusively between token representations. If the tokens are arbitrarily shuffled in a standard self-attention layer, the output for each token is unchanged. All position information is injected before the first self-attention layer in the form of absolute position embeddings. In contrast, dynamic lightweight convolutions assign attention scores directly to relative positions. This allows convolutions to directly integrate relative position information without relying on absolute positions. Thus, convolutions could be better at capturing local information in sentences. However, convolutions alone are limited in their ability to model interactions between tokens because they lack the query-key mechanism central to standard self-attention. In future sections, we consider methods of integrating the two approaches. 3 Integrating lightweight convolutions Previous work has sought to integrate local information into global self-attention. This can be achieved by restricting the range of self-attention to nearby tokens, or by incorporating relative position information into attention maps (Hofst¨atter et al., 2020; Raganato et al., 2020; Wei et al., 2021). Notably, Shaw et al. (2018) introduced relative position embeddings, which inspired similar embeddings in models such as Transformer-XL and XLNet (Dai et al., 2019; Yang et al., 2019). In this section, we show that several previous methods of encoding relative positions are direct implementations of lightweight convolutions. 3.1 Relative embeddings as lightweight convolutions First, the simplest way to combine self-attention with lightweight convolution is to generate a standard attention map, then add the attention map generated by a lightweight convolution. Given a fixed lightweight convolution, this results in attention scores as follows: αij = (xiW Q)(xjW K)T √dh + βj−i (4) This is exactly the relative position term used in T5 (Raffel et al., 2020) and TUPE (Ke et al., 2021). We further consider a dynamic lightweight convolution, where the βj−i weights are computed by passing the query through a linear feedforward layer W C ∈Rdh×(2k+1) (Wu et al., 2019).1 Because W C is linear, each weight βj−i is equal to the dot product between the query and the (j −i) column of W C. We then obtain attention scores: αij = (xiW Q)(xjW K)T √dh + (xiW Q)(W C j−i)T If we scale the dynamic lightweight convolution term according to the head dimension size, we obtain precisely the relative embeddings proposed in Shaw et al. (2018): αij = (xiW Q)(xjW K + W C j−i)T √dh (5) Under this interpretation, Shaw’s relative embeddings are essentially identical to the dynamic lightweight convolutions used in Wu et al. (2019). In both formulations, relative position weights are computed as dot products between the query and a learned relative position embedding. Previous work has considered relative positions in language models independently from convolutions, but our derivations suggest that the underlying mechanisms may be the same. 1Wu et al. (2019) generate dynamic lightweight convolutions based on the entire query layer (dimension size d). In our work, we generate convolutions based on queries for individual attention heads (dimension size dh), to be consistent with the relative embeddings in Shaw et al. (2018). 4325 Lightweight convolution type, BERT-small Params CoLA MNLIm MNLImm MRPC QNLI QQP RTE SST STS GLUE No convolution 13.41M 13.9 73.2 71.8 77.9 80.7 74.5 62.0 81.9 79.3 68.4 No convolution + abs position∗ 13.43M 30.8 76.1 75.9 80.4 78.5 74.4 62.2 85.1 76.8 71.1 Fixed (Raffel et al. 2020) 13.42M 42.1 77.2 76.3 83.8 82.7 75.9 64.4 87.1 81.4 74.5 Dynamic (Shaw et al. 2018) 13.43M 39.1 78.4 77.4 83.8 83.4 77.5 64.4 87.3 81.4 74.7 Composite (Equation 6; ours) 13.43M 40.4 78.2 77.4 85.0 83.3 77.7 64.7 87.8 82.1 75.2 Lightweight convolution type, BERT-base Params CoLA MNLIm MNLImm MRPC QNLI QQP RTE SST STS GLUE No convolution + abs position∗ 108.82M 50.3 82.0 81.2 85.0 84.6 78.6 68.9 91.4 84.9 78.5 Fixed (Raffel et al. 2020) 108.73M 50.0 81.5 80.5 85.6 86.0 78.5 68.9 91.4 84.9 78.6 Dynamic (Shaw et al. 2018) 108.74M 50.9 81.6 80.5 84.6 85.3 78.5 69.5 91.6 84.8 78.6 Composite (Equation 6; ours) 108.74M 50.4 81.6 80.8 85.4 85.1 78.7 69.7 91.2 85.7 78.7 Table 1: GLUE test set performance for models with lightweight convolutions added to self-attention. Columns indicate scores on individual GLUE tasks; the final GLUE score is the average of individual task scores. ∗denotes the default BERT model. 3.2 Composite attention and lightweight convolution experiments To validate lightweight convolutions in combination with self-attention, we pre-trained and evaluated BERT-small models (Devlin et al., 2019; Clark et al., 2020) that incorporated lightweight convolutions. Pre-training To maximize similarity with Devlin et al. (2019), we pre-trained models on the BookCorpus (Zhu et al., 2015) and WikiText-103 datasets (Merity et al., 2017) using masked language modeling. Small models were pre-trained for 125,000 steps, with batch size 128 and learning rate 0.0003. Full pre-training and fine-tuning details are outlined in Appendix A.1.2 Evaluation Models were evaluated on the GLUE benchmark, a suite of sentence classification tasks including natural language inference (NLI), grammaticality judgments, sentiment classification, and textual similarity (Wang et al., 2018). For each task, we ran ten fine-tuning runs and used the model with the best score on the development set. We report scores on the GLUE test set. Development scores and statistics for all experiments are reported in Appendix A.2. Models We trained two baseline models, a default BERT-small with standard absolute position embeddings, and a BERT-small with no position information whatsoever. Then, we trained models with fixed lightweight convolutions (Equation 4; 2Code is available at https://github.com/ mlpc-ucsd/BERT_Convolutions, built upon the Huggingface Transformers library (Wolf et al., 2020). Raffel et al. 2020), and dynamic lightweight convolutions that generated convolution weights based on each query (i.e. using relative embeddings, Equation 5; Shaw et al. 2018). Finally, we propose composite attention, which simply adds dynamic lightweight convolutions to fixed lightweight convolutions, resulting in attention scores αij as follows: (xiW Q)(xjW K)T √dh | {z } Self-attention + (xiW Q)(W C j−i)T √dh | {z } Dynamic convolution (relative embeddings) + βj−i |{z} Fixed convolution (6) Intuitively, composite attention has the flexibility of dynamic lightweight convolutions, while still allowing models to incorporate relative positions directly through fixed lightweight convolutions. Alternatively, composite attention can be interpreted as adding a fixed bias term to relative position embeddings. All of our experiments used a convolution kernel size of 17, or eight positions in each direction, a mid-range value that has been found to work well for both relative positions and convolution in language models (Huang et al., 2020; Jiang et al., 2020; Shaw et al., 2018). As in Shaw et al. (2018), relative embeddings W C j−i shared weights across heads. Unless stated otherwise, models used no absolute position embeddings. For completeness, we also considered dynamic lightweight convolutions based on the key (as opposed to the query). In contrast to query-based 4326 lightweight convolutions, key-based convolutions allow each token to dictate which relative positions should pay attention to it, rather than dictating which relative positions it should pay attention to. Referring to the visualization in Figure 1, key-based dynamic convolutions correspond to columns instead of rows. These key-based dynamic lightweight convolutions are the same as the relative embeddings proposed in Huang et al. (2020), but they are now formulated as dynamic lightweight convolutions. 3.3 Lightweight convolution results GLUE test set results are presented in Table 1. Lightweight convolutions consistently improved performance. Notably, even the fixed lightweight convolution was sufficient to replace absolute position embeddings, outperforming the default BERT-small model. This indicates that even na¨ıve sampling from nearby tokens can be beneficial to language model performance. Dynamic convolutions provided further improvements. When the lightweight convolutions were generated dynamically based on token queries, the models outperformed the default model by even larger margins. This improvement over fixed lightweight convolutions suggests that different tokens find it useful to generate different lightweight convolutions, paying attention to different relative positions in a sentence. Composite attention performed the best. Combining fixed lightweight convolutions with dynamic lightweight convolutions proved an effective strategy for encoding relative positions. Although composite attention is simply a combination of Shaw et al. (2018) and Raffel et al. (2020)’s relative position embeddings, it validates convolution as a viable method of encoding relative positions in self-attention. Key-based dynamic convolutions provided no additional benefit. When we generated an additional lightweight convolution based on keys, the model performed worse than composite attention alone (GLUE 74.0 compared to 75.2). This result clarifies the findings of Huang et al. (2020), who reported only small improvements from query and key-based relative position embeddings for a subset of the GLUE tasks. Figure 2: Learned convolution kernel weights βj−i for the fixed lightweight convolution (Equation 4). Grammaticality judgments were particularly sensitive to position information. On the CoLA task (the corpus of linguistic acceptability; Warstadt et al. 2019), there was a dramatic performance drop when absolute position embeddings were removed. However, when any type of lightweight convolution was added, performance improved even over the baseline established by absolute positions. The pronounced effects of local position information on the CoLA task support the intuitive hypothesis that local dependencies are particularly important for grammaticality judgments. This result also suggests that convolutions could be beneficial to more local tasks (e.g. token-level tasks) along with sentence classification tasks. 3.4 Interpreting lightweight convolutions To better understand how lightweight convolutions improve language models, we visualized the learned lightweight convolution kernel weights in Figure 2. Qualitatively, the kernels exhibited specific types of patterns: • Paying particular attention to the previous or next token. • Paying graded attention either to past or future tokens, dictated by how far the target token is from the present token. These observations support the assumption that nearby tokens are relevant to the interpretation of the current token. They also align with the findings 4327 of Voita et al. (2019), who identified “positional” attention heads that focus primarily on the next or previous token. From this perspective, lightweight convolutions allow language models to explicitly represent nearby tokens’ positions. Interestingly, we also found that some kernels paid fairly uniform attention to all tokens, even decreasing attention to nearby and adjacent tokens. It is likely that these attention heads focused on more global information, relying on the query-key attention mechanism rather than the convolution. 3.5 BERT-base models To thoroughly assess the impact of composite attention on pre-trained language models, we trained full-sized BERT models for 1M steps each, replicating our BERT-small experiments. Pre-training details are outlined in Appendix A.1. Results are presented in Table 1. Differences between models decreased substantially for full sized models, and the relative performances of different approaches varied across tasks. Our results suggest that relative position information is more useful for smaller or more data-limited models; extending the benefits of convolutions robustly from small models to larger models is an important direction for future research. That said, even in the larger models, composite attention slightly outperformed the other position embedding methods in overall GLUE score. Our results demonstrate that convolutions can perform at least on par with absolute position embeddings even in larger models. 4 Non-lightweight convolutions The previous section found that lightweight convolutions consistently improved pre-trained language model performance. Next, we investigated whether the additional flexibility of non-lightweight convolutions could provide additional benefits. Specifically, we considered convolutions that were fixed but non-lightweight. In other words, convolution weights were fixed regardless of the input query, but weights were not tied across channels, equivalent to a standard depthwise convolution. We only considered fixed depthwise convolutions because under existing frameworks, dynamic depthwise convolutions would introduce large numbers of parameters. To implement depthwise convolutions, we added a convolution term identical to the fixed lightweight convolution in Equation 4, except that βj−i was Figure 3: Learned convolution kernel weights βj−i,c (Equation 7) for the depthwise convolution in the deepest attention layer. Channels correspond to the 256 features in each token representation. Channels are sorted such that kernels differentiating the previous and next token are grouped together. learned separately for each feature channel:3 αij,c = (xiW Q)(xjW K)T √dh + βj−i,c (7) This is equivalent to adding a depthwise convolution of the token values to the standard selfattention output. 4.1 Non-lightweight convolution experiments We ran experiments using the same setup as the lightweight convolution experiments in Section 3.2. To compare the effects of dynamic lightweight convolutions (e.g. composite attention) and nonlightweight (depthwise) convolutions, we trained models using each possible combination of the two convolutions. Results are presented in Table 2. Depthwise convolutions were less effective than lightweight convolutions. As with lightweight convolutions, the depthwise convolutions effectively replaced absolute position embeddings, outperforming the default model. However, fixed depthwise convolutions performed worse than fixed lightweight convolutions on the majority of tasks. This indicates that flexibility across channels is not critical to the success of convolutions in language models. 3For computational efficiency, we applied the softmax to the attention scores prior to adding the convolution term βj−i,c, to avoid computing softmax scores separately for each individual channel. Softmax is not commonly applied in depthwise convolutions. 4328 Convolutions Params CoLA MNLIm MNLImm MRPC QNLI QQP RTE SST STS GLUE No convolution + abs position∗ 13.43M 30.8 76.1 75.9 80.4 78.5 74.4 62.2 85.1 76.8 71.1 Composite (Equation 6) 13.43M 40.4 78.2 77.4 85.0 83.3 77.7 64.7 87.8 82.1 75.2 Fixed depthwise 13.47M 36.9 77.6 76.1 80.6 81.9 76.4 64.5 87.5 79.7 73.5 Fixed depthwise + composite 13.48M 38.0 77.4 76.3 82.8 83.7 77.7 65.3 87.3 82.3 74.5 Table 2: GLUE test set performance for BERT-small models with added depthwise convolutions and composite attention. ∗denotes the default BERT-small model. No composite attention Query/Key Value Params GLUE Linear Linear 13.43M ∗71.1 Convolution Linear 13.53M 71.9 Linear Convolution 13.47M 73.4 Convolution Convolution 13.58M 72.0 +Composite attention Query/Key Value Params GLUE Linear Linear 13.43M 75.2 Convolution Linear 13.54M 74.5 Linear Convolution 13.48M 73.9 Convolution Convolution 13.59M 74.0 Table 3: BERT-small performance on the GLUE test set when replacing queries, keys, and values with depthwiseseparable convolutions for half of the attention heads. ∗denotes the use of absolute position embeddings in the default BERT-small model. Composite attention already provided the necessary flexibility. Composite attention outperformed the fixed depthwise convolutions; even when composite attention was combined with depthwise convolutions, there was no overall improvement over composite attention alone. This suggests that in the context of language, dynamic lightweight convolutions efficiently encode any local position information provided by depthwise convolutions. Depthwise convolutions differentiated previous and next tokens. In previous sections, we found that lightweight convolution kernels often pay attention specifically to adjacent tokens. As can be seen in Figure 3, this result was even more pronounced in depthwise convolutions, with individual channels focusing on the previous or next token. Interestingly, other channels specifically directed attention away from adjacent tokens. This indicates that the relevant information about next and previous tokens can be compressed into a subset of the feature channels, freeing other channels to consider more distant or position-independent information. 5 Convolutional queries, keys, and values Improvements over the non-convolutional baselines indicate that convolutions are beneficial to language model pre-training, serving as replacements for absolute position embeddings. Our previous experiments applied different types of convolutions to self-attention values. To take this result one step further, we replaced the linear query, key, and value projections themselves with convolutional layers. Intuitively, applying convolutions before selfattention induces even more mixing of token representations. If convolutions are built into every query, key, and value, then it becomes impossible for a token i to pay attention to a single token j without also incorporating information about tokens surrounding token j. 5.1 Convolutional Q, K, V experiments As in Sections 3.2 and 4.1, we ran experiments on BERT-small. We replaced the query, key and value projections with depthwise-separable convolutions in half of the self-attention heads.4 This aligns with previous work in which only half of the output dimensions for each token were generated using convolutions (Jiang et al., 2020). Indeed, our initial explorations found that it was more effective to replace the linear projections in only half, not all, the attention heads. Then, we considered whether convolutions from previous experiments provided additional benefits over convolutional queries, keys, and values. To test this, we trained BERT-small models with composite attention (Equation 6), adding convolutional queries, keys, and values. 4Depthwise-separable convolutions are a common way to save convolution parameters. A depthwise convolution is applied first, applying an independent convolution for each channel. Then, a pointwise convolution (i.e. a feedforward layer) mixes the channels to produce the final output. 4329 5.2 Convolutional Q, K, V results Results are presented in Table 3. Similar to our previous convolution experiments, all convolutional replacements successfully outperformed the default model. These results strongly support the conclusion that convolutions are a viable method of encoding positional information for language tasks. However, all convolutional replacements for queries, keys, and values slightly decreased the performance of models using composite attention. Convolutional values in particular were effective in models without composite attention, but they slightly decreased performance in models that already incorporated such lightweight convolutions. We conclude that although convolutions can benefit models by adding local position information, there is a limit to how much local mixing should be done. It is sufficient to apply convolutions to token values on top of self-attention; additional convolutional layers applied before the self-attention map enforce unnecessary mixing of token representations. 6 Discussion Our results demonstrate that convolutions provide consistent benefits to pre-trained language models. Our proposed composite attention mechanism combines previous relative position embedding methods, showing that convolutions can effectively compensate for the lack of local position information in Transformer models. 6.1 Related work Our work unites and builds upon previous work using convolutions and relative positions in Transformers. We adopted the relative embeddings from Shaw et al. (2018) and Huang et al. (2020), showing that these embeddings are equivalent to the dynamic lightweight convolutions in Wu et al. (2019). Combining these dynamic lightweight convolutions with fixed lightweight convolutions (equivalent to the relative position terms in Raffel et al. 2020), we studied relative embeddings under the framework of convolution integrated with selfattention. As far as we are aware, our work is the first to holistically compare relative positions, convolutions, and self-attention in language models. Building upon dynamic lightweight convolutions, recent work has incorporated both depthwiseseparable and dynamic lightweight convolutions in pre-trained language models. Jiang et al. (2020) proposed ConvBERT, which adds a convolutional module alongside the standard self-attention mechanism in BERT. ConvBERT’s convolutional module consists of a depthwise-separable convolution combining with a query to generate a dynamic lightweight convolution. Under our integrated framework, this is analogous to the model which uses depthwise-separable convolutions for queries and keys, using composite attention as a querybased dynamic lightweight convolution (see Table 3). To make this comparison concrete, we trained a ConvBERT-small model using the same setup as our experiments. Indeed, the analogous model under our framework outperformed ConvBERT-small (GLUE score 74.5 compared to 70.3). Details for the ConvBERT comparison can be found in Appendix A.3. Finally, recent work has proved theoretical relationships between self-attention and convolution. Cordonnier et al. (2020) showed that given enough self-attention heads, self-attention weights can express any convolution; in fact, they showed that self-attention layers often learn such convolutional structures when trained on vision tasks. However, this theoretical equivalence does not explain convolution-based improvements for Transformers in language tasks. To clarify the relationship between self-attention and convolution in language, our work characterizes self-attention as a type of dynamic lightweight convolution. By establishing a per-parameter equivalence between relative position embeddings and Wu’s dynamic lightweight convolutions, we provide a concrete foundation where self-attention and convolution are used together in practice. 7 Conclusion In this work, we formalized the relationship between self-attention and convolution. We proposed composite attention, which combines self-attention with lightweight convolution, uniting previous approaches to relative positions. Our formulation and empirical results demonstrate that convolutions can improve self-attention by providing local position information in sentences, capable of replacing absolute position embeddings entirely. Our findings provide a solid foundation from which to study convolutions and self-attention in language tasks. The spatially-oriented nature of convolutional neural networks translates directly into positional information in language. As vision and language researchers strive towards common 4330 deep learning architectures, it is important to recognize how architectures for vision tasks can be adapted to linguistic domains. Acknowledgments This work is funded by NSF IIS-1717431. Zhuowen Tu is also funded under the Qualcomm Faculty Award. Tyler Chang is partially supported by the UCSD HDSI graduate fellowship. References Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc Le. 2019. Attention augmented convolutional networks. In International Conference on Computer Vision. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel HerbertVoss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Proceedings of the 34th Conference on Neural Information Processing Systems. Kevin Clark, Minh-Thang Luong, Quoc Le, and Christopher Manning. 2020. ELECTRA: Pretraining text encoders as discriminators rather than generators. In Proceedings of the International Conference on Learning Representations. Jean-Baptiste Cordonnier, Andreas Loukas, and Martin Jaggi. 2020. On the relationship between selfattention and convolutional layers. In Proceedings of the International Conference on Learning Representations. Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc Le, and Ruslan Salakhutdinov. 2019. Transformer-XL: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2978–2988, Florence, Italy. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Linhao Dong, Shuang Xu, and Bo Xu. 2018. Speechtransformer: a no-recurrence sequence-to-sequence model for speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 5884–5888. Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of the International Conference on Learning Representations. Sebastian Hofst¨atter, Hamed Zamani, Bhaskar Mitra, Nick Craswell, and Allan Hanbury. 2020. Local self-attention over long text for efficient document retrieval. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, New York, NY, USA. Association for Computing Machinery. Zhiheng Huang, Davis Liang, Peng Xu, and Bing Xiang. 2020. Improve transformer models with better relative position embeddings. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3327–3335, Online. Association for Computational Linguistics. Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, and Shuicheng Yan. 2020. ConvBERT: Improving BERT with span-based dynamic convolution. In Proceedings of the 34th Conference on Neural Information Processing Systems. Guolin Ke, Di He, and Tie-Yan Liu. 2021. Rethinking positional encoding in language pre-training. In Proceedings of the International Conference on Learning Representations. Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66–71, Brussels, Belgium. Association for Computational Linguistics. Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2017. Pointer sentinel mixture models. In Proceedings of the Fifth International Conference on Learning Representations. Emilio Parisotto, Francis Song, Jack Rae, Razvan Pascanu, Caglar Gulcehre, Siddhant Jayakumar, Max Jaderberg, Rapha¨el Lopez Kaufman, Aidan Clark, Seb Noury, Matthew Botvinick, Nicolas Heess, and Raia Hadsell. 2020. Stabilizing transformers for reinforcement learning. In Proceedings of the International Conference on Machine Learning. Jason Phang, Thibault F´evry, and Samuel Bowman. 2018. Sentence encoders on STILTs: Supplementary training on intermediate labeled-data tasks. arXiv preprint arXiv:1811.01088. 4331 Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-totext transformer. Journal of Machine Learning Research, 21(140):1–67. Alessandro Raganato, Yves Scherrer, and J¨org Tiedemann. 2020. Fixed encoder self-attention patterns in transformer-based machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 556–568, Online. Association for Computational Linguistics. Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. Self-attention with relative position representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 464–468, New Orleans, Louisiana. Association for Computational Linguistics. Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv´e J´egou. 2020. Training data-efficient image transformers and distillation through attention. arXiv preprint arXiv:2012.12877. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st Conference on Neural Information Processing Systems. Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. 2019. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797–5808, Florence, Italy. Association for Computational Linguistics. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353–355, Brussels, Belgium. Association for Computational Linguistics. Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625–641. Wei Wei, Zanbo Wang, Xianling Mao, Guangyou Zhou, Pan Zhou, and Sheng Jiang. 2021. Position-aware self-attention based neural sequence labeling. Pattern Recognition, 110. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Felix Wu, Angela Fan, Alexei Baevski, Yann Dauphin, and Michael Auli. 2019. Pay less attention with lightweight and dynamic convolutions. In Proceedings of the Seventh International Conference on Learning Representations. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ Salakhutdinov, and Quoc Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems, volume 32. Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In 2015 IEEE International Conference on Computer Vision, pages 19–27. Hyperparameter Small Base Layers 12 12 Hidden size 256 768 Intermediate hidden size 1024 3072 Attention heads 4 12 Attention head size 64 64 Embedding size 128 768 Vocab size 30004 30004 Max sequence length 128 128 Mask proportion 0.15 0.15 Learning rate decay Linear Linear Warmup steps 10000 10000 Learning rate 3e-4 1e-4 Adam ϵ 1e-6 1e-6 Adam β1 0.9 0.9 Adam β2 0.999 0.999 Attention dropout 0.1 0.1 Dropout 0.1 0.1 Weight decay 0.01 0.01 Batch size 128 256 Train steps 125K 1M Table 4: Pre-training hyperparameters. A Appendix A.1 Pre-training and fine-tuning details BERT models (Devlin et al. 2019; Clark et al. 2020) were pre-trained on the BookCorpus (Zhu et al., 2015) and WikiText-103 datasets (Merity 4332 Hyperparameter Value Learning rate decay Linear Warmup steps 10% of total Learning rate 1e-4 for QNLI or base-size 3e-4 otherwise Adam ϵ 1e-6 Adam β1 0.9 Adam β2 0.999 Attention dropout 0.1 Dropout 0.1 Weight decay 0 Batch size 128 for MNLI/QQP 32 otherwise Train steps 10 epochs for RTE/STS 4 epochs for MNLI/QQP 3 epochs otherwise Table 5: Fine-tuning hyperparameters. We used intermediate task training for RTE, STS, and MRPC, initializing from a checkpoint fine-tuned on the MNLI task (Clark et al. 2020; Phang et al. 2018). et al., 2017) using masked language modeling. Pretraining examples were formatted as sentence pairs without the next sentence prediction objective. In total, our dataset consisted of 31M unique sentence pairs.5 Sentences were tokenized by training an uncased SentencePiece tokenizer (Kudo and Richardson, 2018), and input and output token embeddings were tied during pre-training. Models were evaluated on the GLUE benchmark (Wang et al., 2018). Including ten fine-tuning runs for each GLUE task, each BERT-small model took about 24 hours to train on two Titan Xp GPUs. Each BERT-base model took about 16 days to train on 8 GPUs. Pretraining hyperparameters are listed in Table 4, and fine-tuning hyperparameters are listed in Table 5. Hyperparameters are based on those used in Clark et al. (2020) and Devlin et al. (2019). A.2 GLUE development results Results for each model on the GLUE development set are reported in Table 6. We report averages over ten fine-tuning runs for each task, including standard errors of the mean. Each overall GLUE score was computed as the average of individual task scores; we computed GLUE score averages and standard errors over ten GLUE scores, corresponding to the ten fine-tuning runs. We note that development scores were generally higher than test scores due to differences between the test and 5Because BERT-small models were only trained for 125,000 steps with batch size 128, small models were trained on 16M sentence pairs. training distributions (Wang et al., 2018). A.3 Detailed ConvBERT comparison ConvBERT adds a convolutional module alongside the standard self-attention mechanism in BERT (Jiang et al., 2020). ConvBERT uses half the number of standard self-attention heads, using convolutional modules for the other half. In each convolutional module, a depthwise-separable convolution is multiplied pointwise with the query in the corresponding self-attention head. This convolutional query is fed into a linear layer to generate a dynamic lightweight convolution. Under our framework, the analogous model replaces half of the queries and keys with depthwiseseparable convolutions and uses composite attention (a query-based dynamic lightweight convolution; see Table 3 in the full paper). In both models (ConvBERT and our own), half of the attention heads use a convolutional query. Additionally, in both models, the convolutional query is used to generate a dynamic lightweight convolution. However, in our model, the dynamic lightweight convolution (in this case, composite attention) is used for all attention heads, not just the convolutional heads. Furthermore, our convolutional heads still use a self-attention mechanism along with the dynamic lightweight convolutions, by generating convolutional keys. In this way, our model adds convolutions to ConvBERT’s self-attention heads, and adds self-attention to ConvBERT’s convolutional heads. Then, we investigated whether the separate selfattention and convolutional modules in ConvBERT provide any benefit over our integrated convolution and self-attention. We trained a ConvBERTsmall model using the same pre-training setup as our BERT-small experiments, comparing performance to the analogous model under our framework. Results are shown in Table 7. Indeed, integrated convolutions and self-attention outperformed ConvBERT-small, using only 3% more parameters. 4333 Convolution type, BERT-small Params CoLA MNLI-m MNLI-mm MRPC QNLI No convolution 13.41M 7.0 ± 2.4 73.0 ± 0.1 73.0 ± 0.1 80.9 ± 0.4 80.1 ± 0.2 No convolution + abs position∗ 13.43M 33.5 ± 0.4 75.8 ± 0.1 76.1 ± 0.1 83.3 ± 0.4 78.2 ± 0.3 Fixed lightweight (Raffel et al. 2020) 13.42M 38.3 ± 0.8 77.2 ± 0.1 77.2 ± 0.1 84.0 ± 0.5 82.1 ± 0.1 Dynamic lightweight (Shaw et al. 2018) 13.43M 38.4 ± 0.7 77.9 ± 0.1 77.6 ± 0.1 85.6 ± 0.5 82.8 ± 0.1 Composite (Equation 6) 13.43M 40.9 ± 0.7 77.9 ± 0.1 78.0 ± 0.1 86.2 ± 0.3 83.0 ± 0.1 Composite + key-based dynamic 13.44M 40.0 ± 0.6 77.9 ± 0.1 77.7 ± 0.1 86.3 ± 0.3 83.3 ± 0.1 Fixed depthwise 13.47M 38.0 ± 0.6 76.9 ± 0.0 76.8 ± 0.1 82.8 ± 0.5 81.9 ± 0.1 Composite + fixed depthwise 13.48M 40.4 ± 0.7 77.2 ± 0.1 77.4 ± 0.1 85.0 ± 0.3 83.3 ± 0.1 Convolutional QK 13.53M 33.4 ± 0.4 76.3 ± 0.1 76.4 ± 0.1 83.3 ± 0.2 81.3 ± 0.2 Convolutional value 13.47M 34.7 ± 0.9 76.2 ± 0.0 76.6 ± 0.1 83.4 ± 0.4 82.4 ± 0.1 Convolutional QKV 13.58M 31.9 ± 0.7 76.3 ± 0.1 76.3 ± 0.1 83.7 ± 0.4 80.4 ± 0.2 Composite + convolutional QK 13.54M 39.3 ± 0.8 77.4 ± 0.1 77.2 ± 0.1 85.4 ± 0.3 81.9 ± 0.1 Composite + convolutional value 13.48M 37.9 ± 0.7 77.8 ± 0.1 78.1 ± 0.1 85.6 ± 0.4 83.6 ± 0.1 Composite + convolutional QKV 13.59M 38.2 ± 1.0 77.4 ± 0.1 77.3 ± 0.1 85.3 ± 0.4 82.8 ± 0.1 ConvBERT 13.09M 33.3 ± 1.5 76.7 ± 0.1 76.8 ± 0.1 83.9 ± 0.5 77.1 ± 0.8 Convolution type, BERT-base No convolution + abs position∗ 108.82M 57.6 ± 0.6 82.0 ± 0.1 81.9 ± 0.1 88.4 ± 0.2 84.7 ± 0.3 Fixed lightweight (Raffel et al. 2020) 108.73M 58.9 ± 0.5 81.9 ± 0.1 81.6 ± 0.1 87.7 ± 0.3 86.2 ± 0.1 Dynamic lightweight (Shaw et al. 2018) 108.74M 58.4 ± 0.5 81.8 ± 0.1 81.8 ± 0.1 86.7 ± 0.4 85.6 ± 0.2 Composite (Equation 6) 108.74M 58.5 ± 0.5 81.9 ± 0.1 81.6 ± 0.1 86.0 ± 1.2 85.0 ± 0.3 Convolution type, BERT-small QQP RTE SST STS GLUE No convolution 84.4 ± 0.1 61.0 ± 0.5 80.9 ± 0.9 83.7 ± 0.1 69.3 ± 0.3 No convolution + abs position∗ 84.9 ± 0.0 64.4 ± 0.5 85.0 ± 0.2 82.4 ± 0.1 73.7 ± 0.1 Fixed lightweight (Raffel et al. 2020) 86.2 ± 0.0 64.7 ± 0.9 86.9 ± 0.2 85.2 ± 0.1 75.7 ± 0.2 Dynamic lightweight (Shaw et al. 2018) 87.2 ± 0.0 65.1 ± 0.9 86.8 ± 0.2 85.6 ± 0.1 76.3 ± 0.1 Composite (Equation 6) 87.3 ± 0.0 66.1 ± 0.7 86.9 ± 0.1 85.9 ± 0.1 76.9 ± 0.1 Composite + key-based dynamic 87.4 ± 0.0 66.3 ± 0.4 86.5 ± 0.3 86.1 ± 0.2 76.8 ± 0.1 Fixed depthwise 86.1 ± 0.1 64.2 ± 0.7 87.2 ± 0.2 84.4 ± 0.1 75.4 ± 0.1 Composite + fixed depthwise 87.3 ± 0.0 63.5 ± 0.8 87.1 ± 0.2 86.1 ± 0.1 76.4 ± 0.1 Convolutional QK 85.1 ± 0.1 63.0 ± 1.0 86.1 ± 0.2 84.5 ± 0.1 74.4 ± 0.1 Convolutional value 86.6 ± 0.0 65.2 ± 0.7 87.2 ± 0.3 85.0 ± 0.1 75.2 ± 0.1 Convolutional QKV 84.6 ± 0.2 66.1 ± 0.9 86.4 ± 0.1 84.4 ± 0.1 74.4 ± 0.1 Composite + convolutional QK 86.7 ± 0.0 64.0 ± 0.9 87.5 ± 0.2 85.7 ± 0.1 76.1 ± 0.1 Composite + convolutional value 87.5 ± 0.0 65.1 ± 0.5 87.5 ± 0.1 86.4 ± 0.1 76.6 ± 0.1 Composite + convolutional QKV 87.0 ± 0.0 64.9 ± 0.8 86.9 ± 0.1 85.9 ± 0.1 76.2 ± 0.2 ConvBERT 85.1 ± 0.1 64.6 ± 0.5 86.3 ± 0.3 84.0 ± 0.2 74.2 ± 0.3 Convolution type, BERT-base No convolution + abs position∗ 88.7 ± 0.0 69.9 ± 0.5 90.4 ± 0.1 88.4 ± 0.1 81.0 ± 0.2 Fixed lightweight (Raffel et al. 2020) 88.8 ± 0.0 70.9 ± 0.7 90.8 ± 0.1 88.1 ± 0.1 81.3 ± 0.2 Dynamic lightweight (Shaw et al. 2018) 88.7 ± 0.0 70.6 ± 0.6 91.1 ± 0.1 87.7 ± 0.3 81.1 ± 0.2 Composite (Equation 6) 88.7 ± 0.0 71.0 ± 0.7 90.5 ± 0.1 88.4 ± 0.1 81.2 ± 0.2 Table 6: GLUE development set scores for each model described in the main paper, reporting averages and standard errors of the mean over ten fine-tuning runs for each task. ∗denotes the default BERT model. Model, BERT-small Params CoLA MNLIm MNLImm MRPC QNLI QQP RTE SST STS GLUE ConvBERT 13.1M 25.5 75.4 73.9 79.7 76.0 74.7 64.3 85.6 77.9 70.3 Integrated convolutions and self-attention (ours) 13.5M 37.9 77.5 76.6 83.7 83.1 76.6 65.3 88.7 81.1 74.5 Table 7: Comparison between ConvBERT-small and the analogous model under our framework, reporting GLUE test set results.
|
2021
|
333
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4334–4348 August 1–6, 2021. ©2021 Association for Computational Linguistics 4334 BinaryBERT: Pushing the Limit of BERT Quantization Haoli Bai1, Wei Zhang2, Lu Hou2, Lifeng Shang2, Jing Jin3, Xin Jiang2, Qun Liu2, Michael Lyu1, Irwin King1 1 The Chinese University of Hong Kong 2Huawei Noah’s Ark Lab, 3Huawei Technologies Co., Ltd. {hlbai, lyu, king}@cse.cuhk.edu.hk {zhangwei379, houlu3, shang.lifeng, jinjing12, jiang.xin, qun.liu}@huawei.com Abstract The rapid development of large pre-trained language models has greatly increased the demand for model compression techniques, among which quantization is a popular solution. In this paper, we propose BinaryBERT, which pushes BERT quantization to the limit by weight binarization. We find that a binary BERT is hard to be trained directly than a ternary counterpart due to its complex and irregular loss landscape. Therefore, we propose ternary weight splitting, which initializes BinaryBERT by equivalently splitting from a half-sized ternary network. The binary model thus inherits the good performance of the ternary one, and can be further enhanced by fine-tuning the new architecture after splitting. Empirical results show that our BinaryBERT has only a slight performance drop compared with the full-precision model while being 24× smaller, achieving the state-of-the-art compression results on the GLUE and SQuAD benchmarks. 1 Introduction Recent pre-trained language models have achieved remarkable performance improvement in various natural language tasks (Vaswani et al., 2017; Devlin et al., 2019). However, the improvement generally comes at the cost of increasing model size and computation, which limits the deployment of these huge pre-trained language models to edge devices. Various methods have been recently proposed to compress these models, such as knowledge distillation (Sanh et al., 2019; Sun et al., 2019; Jiao et al., 2020), pruning (Michel et al., 2019; Fan et al., 2019), low-rank approximation (Ma et al., 2019; Lan et al., 2020), weightsharing (Dehghani et al., 2019; Lan et al., 2020; Huang et al., 2021), dynamic networks with adaptive depth and/or width (Hou et al., 2020; Xin et al., 2020; Zhou et al., 2020), and quantization (Zafrir (a) MRPC. (b) MNLI-m. Figure 1: Performance of quantized BERT with varying weight bit-widths and 8-bit activation. We report the mean results with standard deviations from 10 seeds on MRPC and 3 seeds on MNLI-m, respectively. et al., 2019; Shen et al., 2020; Fan et al., 2020; Zhang et al., 2020). Among all these model compression approaches, quantization is a popular solution as it does not require designing a smaller model architecture. Instead, it compresses the model by replacing each 32-bit floating-point parameter with a low-bit fixedpoint representation. Existing attempts try to quantize pre-trained models (Zafrir et al., 2019; Shen et al., 2020; Fan et al., 2020) to even as low as ternary values (2-bit) with minor performance drop (Zhang et al., 2020). However, none of them achieves the binarization (1-bit). As the limit of quantization, weight binarization could bring at most 32× reduction in model size and replace most floating-point multiplications with additions. Moreover, quantizing activations to 8-bit or 4-bit further replaces the floating-point addition with int8 and int4 addition, decreasing the energy burden and the area usage on chips (Courbariaux et al., 2015). In this paper, we explore to binarize BERT parameters with quantized activations, pushing BERT quantization to the limit. We find that directly training a binary network is rather challenging. According to Figure 1, there is a sharp performance drop when reducing weight bit-width from 2-bit 4335 to 1-bit, compared to other bit configurations. To explore the challenges of binarization, we analyze the loss landscapes of models under different precisions both qualitatively and quantitatively. It is found that while the full-precision and ternary (2bit) models enjoy relatively flat and smooth loss surfaces, the binary model suffers from a rather steep and complex landscape, which poses great challenges to the optimization. Motivated by the above empirical observations, we propose ternary weight splitting, which takes the ternary model as a proxy to bridge the gap between the binary and full-precision models. Specifically, ternary weight splitting equivalently converts both the quantized and latent full-precision weights in a well-trained ternary model to initialize BinaryBERT. Therefore, BinaryBERT retains the good performance of the ternary model, and can be further refined on the new architecture. While neuron splitting is previously studied (Chen et al., 2016; Wu et al., 2019) for full-precision network, our ternary weight splitting is much more complex due to the additional equivalence requirement of quantized weights. Furthermore, the proposed BinaryBERT also supports adaptive splitting. It can adaptively perform splitting on the most important ternary modules while leaving the rest as binary, based on efficiency constraints such as model size or floating-point operations (FLOPs). Therefore, our approach allows flexible sizes of binary models for various edge devices’ demands. Empirical results show that BinaryBERT split from a half-width ternary network is much better than a directly-trained binary model with the original width. On the GLUE and SQuAD benchmarks, our BinaryBERT has only a slight performance drop compared to the full-precision BERT-base model, while being 24× smaller. Moreover, BinaryBERT with the proposed importance-based adaptive splitting also outperforms other splitting criteria across a variety of model sizes. 2 Difficulty in Training Binary BERT In this section, we show that it is challenging to train a binary BERT with conventional binarization approaches directly. Before diving into details, we first review the necessary backgrounds. We follow the standard quantization-aware training procedure (Zhou et al., 2016). Specifically, given weight w ∈Rn (a.k.a latent full-precision weights), each forward propagation quantizes it to ˆw = Q(w) by some quantization function Q(·), and then computes the loss ℓ( ˆw) at ˆw. During back propagation, we use ∇ℓ( ˆw) to update latent fullprecision weights w due to the non-differentiability of Q(·), which is known as the straight-through estimator (Courbariaux et al., 2015). Recent TernaryBERT (Zhang et al., 2020) follows Ternary-Weight-Network (TWN) (Li et al., 2016) to quantize the elements in w to three values {±α, 0}. To avoid confusion, we use superscript t and b for the latent full-precision weights and quantized weights in ternary and binary models, respectively. Specifically, TWN ternarizes each element wt i in the ternary weight wt as ˆwt i =Q(wt i)= α · sign(wt i) |wt i| ≥∆ 0 |wt i| < ∆, (1) where sign(·) is the sign function, ∆= 0.7 n ∥wt∥1 and α= 1 |I| P i∈I |wt i| with I = {i | ˆwt i ̸= 0}. Binarization. Binarization is first proposed in (Courbariaux et al., 2015) and has been extensively studied in the academia (Rastegari et al., 2016; Hubara et al., 2016; Liu et al., 2018). As a representative work, Binary-Weight-Network (BWN) (Hubara et al., 2016) binarizes wb elementwisely with a scaling parameter α as follows: ˆwb i = Q(wb i) = α · sign(wb i), α = 1 n∥wb∥1. (2) Despite the appealing properties of network binarization, we show that it is non-trivial to obtain a binary BERT with these binarization approaches. 2.1 Sharp Performance Drop with Weight Binarization To study the performance drop of BERT quantization, we train the BERT model with fullprecision, {8,4,3,2,1}-bit weight quantization and 8-bit activations on MRPC and MNLI-m from the GLUE benchmark (Wang et al., 2018) 1. We use loss-aware weight quantization (LAQ) (Hou and Kwok, 2018) for 8/4/3-bit weight quantization, TWN (Li et al., 2016) for weight ternarization and BWN (Hubara et al., 2016) for weight binarization. Meanwhile, we adopt 8-bit uniform quantization for activations. We follow the default experimental settings detailed in Section 4.1 and Appendix C.1. 1We conduct more experiments on other GLUE datasets and with different settings in Appendix C.1, and find similar empirical results to MRPC and MNLI-m here. 4336 (a) Full-precision Model. (b) Ternary Model. (c) Binary Model. (d) All Together. Figure 2: Loss landscapes visualization of the full-precision, ternary and binary models on MRPC. For (a), (b) and (c), we perturb the (latent) full-precision weights of the value layer in the 1st and 2nd Transformer layers, and compute their corresponding training loss. (d) shows the gap among the three surfaces by stacking them together. (a) MHA-QK. (b) MHA-V. (c) MHA-O. (d) FFN-Mid. (e) FFN-Out. Figure 3: The top-1 eigenvalues of parameters at different Transformer parts of the full-precision (FP), ternary and binary BERT. For easy comparison, we report the ratio of eigenvalue between the ternary/binary models and the full-precision model. The error bar is estimated of all Transformer layers over different data mini-batches. From Figure 1, the performance drops mildly from 32-bit to as low as 2-bit, i.e., around 0.6% ↓ on MRPC and 0.2% ↓on MNLI-m. However, when reducing the bit-width to one, the performance drops sharply, i.e, ∼3.8% ↓and ∼0.9% ↓ on the two tasks, respectively. Therefore, weight binarization may severely harm the performance, which may explain why most current approaches stop at 2-bit weight quantization (Shen et al., 2020; Zadeh and Moshovos, 2020; Zhang et al., 2020). To further push weight quantization to the limit, a first step is to study the potential reasons behind the sharp drop from ternarization to binarization. 2.2 Exploring the Quantized Loss Landscape Visualization. To learn about the challenges behind the binarization, we first visually compare the loss landscapes of full-precision, ternary, and binary BERT models. Following (Nahshan et al., 2019), we extract parameters wx, wy from the value layers2 of multi-head attention in the first two Transformer layers, and assign the following perturbations on parameters: ˜wx = wx + x · 1x, ˜wy = wy + y · 1y, (3) 2We also extract parameters from other parts of the Transformer in Appendix C.2, and the observations are similar. where x ∈{±0.2 ¯wx, ±0.4 ¯wx, ..., ±1.0 ¯wx} are perturbation magnitudes based the absolute mean value ¯wx of wx, and similar rules hold for y. 1x and 1y are vectors with all elements being 1. For each pair of (x, y), we evaluate the corresponding training loss and plot the surface in Figure 2. As can be seen, the full-precision model (Figure 2(a)) has the lowest overall training loss, and its loss landscape is flat and robust to the perturbation. For the ternary model (Figure 2(b)), despite the surface tilts up with larger perturbations, it looks locally convex and is thus easy to optimize. This may also explain why the BERT model can be ternarized without severe accuracy drop (Zhang et al., 2020). However, the loss landscape of the binary model (Figure 2(c)) turns out to be both higher and more complex. By stacking the three landscapes together (Figure 2(d)), the loss surface of the binary BERT stands on the top with a clear margin with the other two. The steep curvature of loss surface reflects a higher sensitivity to binarization, which attributes to the training difficulty. Steepness Measurement. To quantitatively measure the steepness of loss landscape, we start from a local minima w and apply the second order approximation to the curvature. According to the Taylor’s expansion, the loss increase induced by quantizing 4337 Figure 4: The overall workflow of training BinaryBERT. We first train a half-sized ternary BERT model, and then apply ternary weight splitting operator (Equations (6) and (7)) to obtain the latent full-precision and quantized weights as the initialization of the full-sized BinaryBERT. We then fine-tune BinaryBERT for further refinement. w can be approximately upper bounded by ℓ( ˆw) −ℓ(w) ≈ϵ⊤Hϵ ≤λmax∥ϵ∥2, (4) where ϵ = w −ˆw is the quantization noise, and λmax is the largest eigenvalue of the Hessian H at w. Note that the first-order term is skipped due to ∇ℓ(w) = 0. Thus we take λmax as a quantitative measurement for the steepness of the loss surface. Following (Shen et al., 2020) we adopt the power method to compute λmax. As it is computationally expensive to estimate H for all w in the network, we consider them separately as follows: (1) the query/key layers (MHA-QK), (2) the value layer (MHA-V), (3) the output projection layer (MHA-O) in the multi-head attention, (4) the intermediate layer (FFN-Mid), and (5) the output layer (FFN-Out) in the feed-forward network. Note that we group key and query layers as they are used together to calculate the attention scores. From Figure 3, the top-1 eigenvalues of the binary model are higher both on expectation and standard deviation compared to the full-precision baseline and the ternary model. For instance, the top-1 eigenvalues of MHA-O in the binary model are ∼15× larger than the full-precision counterpart. Therefore, the quantization loss increases of fullprecision and ternary model are tighter bounded than the binary model in Equation (4). The highly complex and irregular landscape by binarization thus poses more challenges to the optimization. 3 Proposed Method 3.1 Ternary Weight Splitting Given the challenging loss landscape of binary BERT, we propose ternary weight splitting (TWS) that exploits the flatness of ternary loss landscape as the optimization proxy of the binary model. As is shown in Figure 4, we first train the half-sized ternary BERT to convergence, and then split both the latent full-precision weight wt and quantized ˆwt to their binary counterparts wb 1, wb 2 and ˆwb 1, ˆwb 2 via the TWS operator. To inherit the performance of the ternary model after splitting, the TWS operator requires the splitting equivalency (i.e., the same output given the same input): wt = wb 1 + wb 2, ˆwt = ˆwb 1 + ˆwb 2 . (5) While solution to Equation (5) is not unique, we constrain the latent full-precision weights after splitting wb 1, wb 2 to satisfy wt = wb 1 + wb 2 as wb 1,i = a · wt i if ˆwt i ̸= 0 b + wt i if ˆwt i = 0, wt i > 0 b otherwise , (6) wb 2,i = (1−a)wt i if ˆwt i ̸= 0 −b if ˆwt i = 0, wt i > 0 −b + wt i otherwise , (7) where a and b are the variables to solve. By Equations (6) and (7) with ˆwt = ˆwb 1 + ˆwb 2, we get a = P i∈I |wt i| + P j∈J |wt j| −P k∈K |wt k| 2 P i∈I |wt i| , b = n |I| P i∈I |wt i| −Pn i=1 |wt i| 2(|J | + |K|) , (8) where we denote I = {i | ˆwt i ̸= 0}, J = {j | ˆwt j = 0 and wt j > 0} and K = {k | ˆwt k = 0 and wt k < 0}. | · | denotes the cardinality of the set. Detailed derivation of Equation (8) is in Appendix A. Quantization Details. Following (Zhang et al., 2020), for each weight matrix in the Transformer layers, we use layer-wise ternarization (i.e., one scaling parameter for all elements in the weight 4338 matrix). For word embedding, we use row-wise ternarization (i.e., one scaling parameter for each row in the embedding). After splitting, each of the two split matrices has its own scaling factor. Aside from weight binarization, we simultaneously quantize activations before all matrix multiplications, which could accelerate inference on specialized hardwares (Shen et al., 2020; Zafrir et al., 2019). Following (Zafrir et al., 2019; Zhang et al., 2020), we skip the quantization for all layernormalization (LN) layers, skip connections, and bias as their calculations are negligible compared to matrix multiplication. The last classification layer is also not quantized to avoid a large accuracy drop. Training with Knowledge Distillation. Knowledge distillation is shown to benefit BERT quantization (Zhang et al., 2020). Following (Jiao et al., 2020; Zhang et al., 2020), we first perform intermediate-layer distillation from the fullprecision teacher network’s embedding E, layerwise MHA output Ml and FFN output Fl to the quantized student counterpart ˆE, ˆMl, ˆFl (l = 1, 2, ...L). We aim to minimize their mean sqaured errors, i.e., ℓemb = MSE(ˆE, E), ℓmha = P l MSE( ˆMl, Ml), and ℓffn = P l MSE(ˆFl, Fl). Thus the objective function is ℓint = ℓemb + ℓmha + ℓffn. (9) We then conduct prediction-layer distillation by minimizing the soft cross-entropy (SCE) between quantized student logits ˆy and teacher logits y, i.e., ℓpred = SCE(ˆy, y). (10) Further Fine-tuning. After splitting from the half-sized ternary model, the binary model inherits its performance on a new architecture with full width. However, the original minimum of the ternary model may not hold in this new loss landscape after splitting. Thus we further fine-tune with prediction-layer distillation to look for a better solution. We dub the resulting model as BinaryBERT. 3.2 Adaptive Splitting Our proposed approach also supports adaptive splitting that can flexibly adjust the width of BinaryBERT, based on the parameter sensitivity to binarization and resource constraints of edge devices. Specifically, given the resource constraints C (e.g., model size and computational FLOPs), we first train a mixed-precision model adaptively (with sensitive parts being ternary and the rest being binary), and then split ternary weights into binary ones. Therefore, adaptive splitting finally enjoys consistent arithmetic precision (1-bit) for all weight matrices, which is usually easier to deploy than the mixed-precision counterpart. Formulation. Intuitively, we assign ternary values to weight matrices that are more sensitive to quantization. The quantization sensitivity of the weight matrix is empirically measured by the performance gain of not quantizing it comparing to the fully-quantized counterpart (Details are in Appendix B.1.). We denote u ∈RZ + as the sensitivity vector, where Z is the total number of splittable weight matrices in all Transformer layers, the word embedding layer and the pooler layer. The cost vector c ∈RZ + stores the additional increase of parameter or FLOPs of each ternary weight matrix against a binary choice. The splitting assignment can be represented as a binary vector s ∈{0, 1}Z, where sz = 1 means to ternarize the z-th weight matrix, and vice versa. The optimal assignment s∗ can thus be solved from the following combinatorial optimization problem: maxs u⊤s (11) s.t. c⊤s ≤C −C0, s ∈{0, 1}Z, where C0 is the baseline efficiency of the half-sized binary network. Dynamic programming can be applied to solve Equation (11) to avoid NP-hardness. 4 Experiments In this section, we empirically verify our proposed approach on the GLUE (Wang et al., 2018) and SQuAD (Rajpurkar et al., 2016, 2018) benchmarks. We first introduce the experimental setup in Section 4.1, and then present the main experimental results on both benchmarks in Section 4.2. We compare with other state-of-the-arts in Section 4.3, and finally provide more discussions on the proposed methods in Section 4.4. Code is available at https://github.com/huawei-noah/ Pretrained-Language-Model/tree/master/ BinaryBERT. 4.1 Experimental Setup Dataset and Metrics. The GLUE benchmark contains multiple natural language understanding tasks. We follow Devlin et al. (2019) to evaluate the performance on these tasks: Matthews correlation 4339 # Quant #Bits (W-E-A) Size (MB) FLOPs (G) DA MNLI -m/mm QQP QNLI SST-2 CoLA STS-B MRPC RTE Avg. 1 full-prec. 417.6 22.5 84.9/85.5 91.4 92.1 93.2 59.7 90.1 86.3 72.2 83.9 2 BWN 1-1-8 13.4 3.1 84.2/84.0 91.1 90.7 92.3 46.7 86.8 82.6 68.6 80.8 3 TWS 1-1-8 16.5 3.1 84.2/84.7 91.2 91.5 92.6 53.4 88.6 85.5 72.2 82.7 4 BWN 1-1-4 13.4 1.5 83.5/83.4 90.9 90.7 92.3 34.8 84.9 79.9 65.3 78.4 5 TWS 1-1-4 16.5 1.5 83.9/84.2 91.2 90.9 92.3 44.4 87.2 83.3 65.3 79.9 6 BWN 1-1-8 13.4 3.1 84.2/84.0 91.1 91.2 92.7 54.2 88.2 86.8 70.0 82.5 7 TWS 1-1-8 16.5 3.1 84.2/84.7 91.2 91.6 93.2 55.5 89.2 86.0 74.0 83.3 8 BWN 1-1-4 13.4 1.5 83.5/83.4 90.9 91.2 92.5 51.9 87.7 85.5 70.4 81.9 9 TWS 1-1-4 16.5 1.5 83.9/84.2 91.2 91.4 93.7 53.3 88.6 86.0 71.5 82.6 Table 1: Results on the GLUE development set. “#Bits (W-E-A)” represents the bit number for weights of Transformer layers, word embedding, and activations. “DA” is short for data augmentation. “Avg.” denotes the average results of all tasks including MNLI-m and MNLI-mm. The higher results in each block are bolded. # Quant #Bits (W-E-A) Size (MB) FLOPs (G) DA MNLI -m/mm QQP QNLI SST-2 CoLA STS-B MRPC RTE Avg. 1 full-prec. 417.6 22.5 84.5/84.1 89.5 91.3 93.0 54.9 84.4 87.9 69.9 82.2 2 BWN 1-1-8 13.4 3.1 83.3/83.4 88.9 90.1 92.3 38.1 81.2 86.1 63.1 78.5 3 TWS 1-1-8 16.5 3.1 84.1/83.6 89.0 90.0 93.1 50.5 83.4 86.0 65.8 80.6 4 BWN 1-1-4 13.4 1.5 83.5/82.5 89.0 89.4 92.3 26.7 78.9 84.2 59.9 76.3 5 TWS 1-1-4 16.5 1.5 83.6/82.9 89.0 89.3 93.1 37.4 82.5 85.9 62.7 78.5 6 BWN 1-1-8 13.4 3.1 83.3/83.4 88.9 90.3 91.3 48.4 83.2 86.3 66.1 80.1 7 TWS 1-1-8 16.5 3.1 84.1/83.5 89.0 89.8 91.9 51.6 82.3 85.9 67.3 80.6 8 BWN 1-1-4 13.4 1.5 83.5/82.5 89.0 89.9 92.0 45.0 81.9 85.2 64.1 79.2 9 TWS 1-1-4 16.5 1.5 83.6/82.9 89.0 89.7 93.1 47.9 82.9 86.6 65.8 80.2 Table 2: Results on the GLUE test set scored using the GLUE evaluation server. for CoLA, Spearman correlation for STS-B and accuracy for the rest tasks: RTE, MRPC, SST-2, QQP, MNLI-m (matched) and MNLI-mm (mismatched). For machine reading comprehension on SQuAD, we report the EM (exact match) and F1 score. Aside from the task performance, we also report the model size (MB) and computational FLOPs at inference. For quantized operations, we follow (Zhou et al., 2016; Liu et al., 2018; Li et al., 2020a) to count the bit-wise operations, i.e., the multiplication between an m-bit number and an n-bit number approximately takes mn/64 FLOPs for a CPU with the instruction size of 64 bits. Implementation. We take DynaBERT (Hou et al., 2020) sub-networks as backbones as they offer both half-sized and full-sized models for easy comparison. We start from training a ternary model of width 0.5× with the two-stage knowledge distillation introduced in Section 3.1. Then we split it into a binary model with width 1.0×, and perform further fine-tuning with prediction-layer distillation. Each training stage takes the same number of training epochs. Following (Jiao et al., 2020; Hou et al., 2020; Zhang et al., 2020), we adopt data augmentation with one training epoch in each stage on all GLUE tasks except for MNLI and QQP. Aside from this default setting, we also remove data augmentation and perform vanilla training with 6 epochs on these tasks. On MNLI and QQP, we train 3 epochs for each stage. We verify our ternary weight splitting (TWS) against vanilla binary training (BWN), the latter of which doubles training epochs to match the overall training time in TWS for fair comparison. More training details are provided in Appendix B. Activation Quantization. While BinaryBERT focuses on weight binarization, we also explore activation quantization in our implementation, which is beneficial for reducing the computation burden on specialized hardwares (Hubara et al., 2016; Zhou et al., 2016; Zhang et al., 2020). Aside from 8-bit uniform quantization (Zhang et al., 2020; Shen et al., 2020) in past efforts, we further pioneer to study 4-bit activation quantization. We find that uniform quantization can hardly deal with outliers in the activation. Thus we use Learned Step-size Quantization (LSQ) (Esser et al., 2019) to directly learn the quantized values, which empirically achieves better quantization performance. 4.2 Experimental Results 4.2.1 Results on the GLUE Benchmark The main results on the development set are shown in Table 1. For results without data augmenta4340 Quant #Bits (W-E-A) Size (MB) FLOPs (G) SQuAD v1.1 SQuAD v2.0 full-prec. 417.6 22.5 82.6/89.7 75.1/77.5 BWN 1-1-8 13.4 3.1 79.2/86.9 73.6/76.6 TWS 1-1-8 16.5 3.1 80.8/88.3 73.6/76.5 BWN 1-1-4 13.4 1.5 77.5/85.8 71.9/75.1 TWS 1-1-4 16.5 1.5 79.3/87.2 72.5/75.4 Table 3: Development set results (EM/F1) on SQuAD. (a) 8-bit Activation. (b) 4-bit Activation. Figure 5: The average performance over six GLUE tasks of adaptive splitting strategies. tion (row #2-5), our ternary weight splitting method outperforms BWN with a clear margin 3. For instance, on CoLA, ternary weight splitting achieves 6.7% ↑and 9.6% ↑with 8-bit and 4-bit activation quantization, respectively. While data augmentation (row 6-9) mostly improves each entry, our approach still overtakes BWN consistently. Furthermore, 4-bit activation quantization empirically benefits more from ternary weight splitting (row 4-5 and 8-9) compared with 8-bit activations (row 2-3 and 6-7), demonstrating the potential of our approach in extremely low bit quantized models. In Table 2, we also provide the results on the test set of GLUE benchmark. Similar to the observation in Table 1, our approach achieves consistent improvement on both 8-bit and 4-bit activation quantization compared with BWN. 4.2.2 Results on SQuAD Benchmark The results on the development set of SQuAD v1.1 and v2.0 are shown in Table 3. Our proposed ternary weight splitting again outperforms BWN w.r.t both EM and F1 scores on both datasets. Similar to previous observations, 4-bit activation enjoys a larger gain in performance from the splitting approach. For instance, our approach improves the EM score of 4-bit activation by 1.8% and 0.6% on SQuAD v1.1 and v2.0, respectively, both of which are higher than those of 8-bit activation. 3Note that DynaBERT only squeezes width in the Transformer layers but not the word embedding layer, thus the split binary model has a slightly larger size than BWN. Method #Bits (W-E-A) Size (MB) Ratio (↓) SQuAD v1.1 MNLI -m BERT-base full-prec. 418 1.0 80.8/88.5 84.6 DistilBERT full-prec. 250 1.7 79.1/86.9 81.6 LayerDrop-6L full-prec. 328 1.3 82.9 LayerDrop-3L full-prec. 224 1.9 78.6 TinyBERT-6L full-prec. 55 7.6 79.7/87.5 82.8 ALBERT-E128 full-prec. 45 9.3 82.3/89.3 81.6 ALBERT-E768 full-prec. 120 3.5 81.5/88.6 82.0 Quant-Noise PQ 38 11.0 83.6 Q-BERT 2/4-8-8 53 7.9 79.9/87.5 83.5 Q-BERT 2/3-8-8 46 9.1 79.3/87.0 81.8 Q-BERT 2-8-8 28 15.0 69.7/79.6 76.6 GOBO 3-4-32 43 9.7 83.7 GOBO 2-2-32 28 15.0 71.0 TernaryBERT 2-2-8 28 15.0 79.9/87.4 83.5 BinaryBERT 1-1-8 17 24.6 80.8/88.3 84.2 BinaryBERT 1-1-4 17 24.6 79.3/87.2 83.9 Table 4: Comparison with other state-of-the-art methods on development set of SQuAD v1.1 and MNLI-m. 4.2.3 Adaptive Splitting The adaptive splitting in Section 3.2 supports the conversion of mixed ternary and binary precisions for more-fine-grained configurations. To verify its advantages, we name our approach as Maximal Gain according to Equation (11), and compare it with two baseline strategies i) Random Gain that randomly selects weight matrices to split; and ii) Minimal Gain that splits the least important modules according to sensitivity. We report the average score over six tasks (QNLI, SST-2, CoLA, STSB, MRPC and RTE) in Figure 5. The end-points of 9.8MB and 16.5MB are the half-sized and fullsized BinaryBERT, respectively. As can be seen, adaptive splitting generally outperforms the other two baselines under varying model size, indicating the effectiveness of maximizing the gain in adaptive splitting. In Appendix C.4, we provide detailed performance on the six tasks, together with the architecture visualization of adaptive splitting. 4.3 Comparison with State-of-the-arts Now we compare our proposed approach with a variety of state-of-the-art counterparts, including Q-BERT (Shen et al., 2020), GOBO (Zadeh and Moshovos, 2020), Quant-Noise (Fan et al., 2020) and TernaryBERT (Zhang et al., 2020). Aside from quantization, we also compare with other general compression approaches such as DistillBERT (Sanh et al., 2019), LayerDrop (Fan et al., 2019), TinyBERT (Jiao et al., 2020), and ALBERT (Lan et al., 2020). The results are taken from the original papers, respectively. From Table 4, our proposed BinaryBERT has the smallest model size with the best performance among all quantiza4341 Quant #Bits (W-E-A) SQuAD v1.1 MNLI -m QNLI MRPC TWN0.5× 2-2-8 80.3/87.9 84.1 91.3 85.7 TWS1.0× 1-1-8 80.8/88.3 84.2 91.6 86.0 TWN0.5× 2-2-4 78.0/86.4 83.7 90.9 85.5 TWS1.0× 1-1-4 79.3/87.2 83.9 91.4 86.0 Table 5: The performance gain by fine-tuning the binary model after splitting. 0.5× and 1.0× denote the half-sized and full-sized models, respectively. (a) 8-bit Activation. (b) 4-bit Activation. (c) 8-bit Activation. (d) 4-bit Activation. Figure 6: (a) and (b) show the training curves on MRPC under different activation bits. The red box is enlarged in the sub-figure. (c) and (d) visualize the fine-tuning trajectories after splitting, on the 2-D loss contour of BinaryBERT. tion approaches. Compared with the full-precision model, our BinaryBERT retains competitive performance with a significant reduction of model size and computation. For example, we achieve more than 24× compression ratio compared with BERTbase, with only 0.4% ↓and 0.0%/0.2% ↓drop on MNLI-m on SQuAD v1.1, respectively. 4.4 Discussion 4.4.1 Further Improvement after Splitting We now demonstrate the performance gain by refining the binary model on the new architecture. We evaluate the performance gain after splitting from a half-width ternary model (TWN0.5×) to the full-sized model (TWN1.0×) on the development set of SQuAD v1.1, MNLI-m, QNLI and MRPC. The results are shown in Table 5. As can be seen, further fine-tuning brings consistent improvement on both 8-bit and 4-bit activation. Quant #Bits (W-E-A) SQuAD v1.1 MNLI -m QNLI SST-2 BWN 1-1-8 79.2/86.9 84.2 91.2 92.7 LAB 1-1-8 79.0/87.0 83.6 91.5 92.8 BiReal 1-1-8 79.4/87.1 83.9 91.4 92.5 BWN† 1-1-8 79.4/87.3 84.2 91.3 92.8 BWN‡ 1-1-8 79.6/87.2 83.5 91.2 92.9 TWS 1-1-8 80.8/88.3 84.2 91.6 93.2 BWN 1-1-4 77.5/85.8 83.5 91.2 92.5 LAB 1-1-4 76.7/85.5 83.3 91.3 92.9 BiReal 1-1-4 76.9/85.4 83.4 91.0 92.8 BWN† 1-1-4 78.2/86.2 83.6 91.3 92.9 BWN‡ 1-1-4 78.3/86.5 83.1 90.9 92.9 TWS 1-1-4 79.3/87.2 83.9 91.4 93.7 Table 6: Comparison with other binarization methods. Training Curves. Furthermore, we plot the training loss curves of BWN, TWN and our TWS on MRPC with data augmentation in Figures 6(a) and 6(b). Since TWS cannot inherit the previous optimizer due to the architecture change, we reset the optimizer and learning rate scheduler of BWN, TWN and TWS for a fair comparison, despite the slight increase of loss after splitting. We find that our TWS attains much lower training loss than BWN, and also surpasses TWN, verifying the advantages of fine-tuning on the wider architecture. Optimization Trajectory. We also follow (Li et al., 2018; Hao et al., 2019) to visualize the optimization trajectory after splitting in Figures 6(c) and 6(d). We calculate the first two principal components of parameters in the final BinaryBERT, which are the basis for the 2-D plane. The loss contour is thus obtained by evaluating each grid point in the plane. It is found that the binary models are heading towards the optimal solution for both 8/4-bit activation quantization on the loss contour. 4.4.2 Exploring More Binarization Methods We now study if there are any improved binarization variants that can directly bring better performance. Aside from BWN, we compare with LAB (Hou et al., 2017) and BiReal (Liu et al., 2018). Meanwhile, we compare with gradual quantization, i.e., BWN training based on a ternary model, denoted as BWN†. Furthermore, we also try the same scaling factor of BWN with TWN to make the precision change smooth, dubbed as BWN‡. From Table 6, we find that our TWS still outperforms various binarization approaches in most cases, suggesting the superiority of splitting in finding better minima than direct binary training. 4342 5 Related Work Network quantization has been a popular topic with vast literature in efficient deep learning. Below we give a brief overview for three research strands: network binarization, mixed-precision quantization and neuron splitting, all of which are related to our proposed approach. 5.1 Network Binarization Network binarization achieves remarkable size reduction and is widely explored in computer vision. Existing binarization approaches can be categorized into quantization error minimization (Rastegari et al., 2016; Hou et al., 2017; Zhang et al., 2018), improving training objectives (Martinez et al., 2020; Bai et al., 2020) and reduction of gradient mismatch (Bai et al., 2018; Liu et al., 2018, 2020). Despite the empirical success of these approaches in computer vision, there is little exploration of binarization in natural language processing tasks. Previous works on BERT quantization (Zafrir et al., 2019; Shen et al., 2020; Zhang et al., 2020) push down the bit-width to as low as two, but none of them achieves binarization. On the other hand, our work serves as the first attempt to binarize the pre-trained language models. 5.2 Mixed-precision Quantization Given the observation that neural network layers exhibit different sensitivity to quantization (Dong et al., 2019; Wang et al., 2019), mixed-precision quantization re-allocate layer-wise quantization bit-width for higher compression ratio. Inspired by neural architecture search (Liu et al., 2019; Wang et al., 2020), common approaches of mixedprecision quantization are primarily based on differentiable search (Wu et al., 2018a; Li et al., 2020b), reinforcement learning (Wu et al., 2018b; Wang et al., 2019), or simply loss curvatures (Dong et al., 2019; Shen et al., 2020). While mixedprecision quantized models usually demonstrate better performance than traditional methods under the same compression ratio, they are also harder to deploy (Habi et al., 2020). On the contrary, BinaryBERT with adaptive splitting enjoy both the good performance from the mixed precision of ternary and binary values, and the easy deployment given the consistent arithmetic precision. There are also works on binary neural architecture search (Kim et al., 2020; Bulat et al., 2020) which have a similar purpose to mixed-precision quantization. Nonetheless, such methods are usually time-consuming to train and are prohibitive for large pre-trained language models. 5.3 Neuron Splitting Neuron splitting is originally proposed to accelerate the network training, by progressively increasing the width of a network (Chen et al., 2016; Wu et al., 2019). The split network equivalently inherits the knowledge from the antecessors and is trained for further improvement. Recently, neuron splitting is also studied in quantization (Zhao et al., 2019; Kim et al., 2019). By splitting neurons with large magnitudes, the full-precision outliers are removed and thus the quantization error can be effectively reduced (Zhao et al., 2019). Kim et al. (2019) apply neuron splitting to decompose ternary activation into two binary activations based on bias shifting of the batch normalization layer. However, such a method cannot be applied in BERT as there is no batch normalization layer. Besides, weight splitting is much more complex due to the equivalence constraint on both the quantized and latent full-precision weights. 6 Conclusion In this paper, we propose BinaryBERT, pushing BERT quantization to the limit. As a result of the steep and complex loss landscape, we find directly training a BinaryBERT is hard with a large performance drop. We thus propose a ternary weight splitting that splits a trained ternary BERT to initialize BinaryBERT, followed by fine-tuning for further refinement. Our approach also supports adaptive splitting that can tailor the size of BinaryBERT based on the edge device constraints. Empirical results show that our approach significantly outperforms vanilla binary training, achieving stateof-the-art performance on BERT compression. Acknowledgement This work was partially supported by the National Key Research and Development Program of China (No. 2018AAA0100204), and Research Grants Council of the Hong Kong Special Administrative Region, China (No. CUHK 14210717 of the General Research Fund). We sincerely thank all anonymous reviewers for their insightful suggestions. 4343 References H. Bai, J. Wu, I. King, and M. Lyu. 2020. Few shot network compression via cross distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3203–3210. Y. Bai, Y. Wang, and E. Liberty. 2018. Proxquant: Quantized neural networks via proximal operators. In International Conference on Machine Learning. A. Bulat, B. Martinez, and G. Tzimiropoulos. 2020. Bats: Binary architecture search. In European Conference on Computer Vision, pages 309–325. T. Chen, I. Goodfellow, and J. Shlens. 2016. Net2net: Accelerating learning via knowledge transfer. In International Conference on Learning Representations. M. Courbariaux, Y. Bengio, and J. David. 2015. Binaryconnect: Training deep neural networks with binary weights during propagations. In Advances in neural information processing systems. M. Dehghani, S. Gouws, O. Vinyals, J. Uszkoreit, and L. Kaiser. 2019. Universal transformers. In International Conference on Learning Representations. J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics. Z. Dong, Z. Yao, A. Gholami, M. Mahoney, and K. Keutzer. 2019. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 293–302. S. K. Esser, J. L. McKinstry, D. Bablani, R. Appuswamy, and D. S. Modha. 2019. Learned step size quantization. In International Conference on Learning Representations. A. Fan, E. Grave, and A. Joulin. 2019. Reducing transformer depth on demand with structured dropout. In International Conference on Learning Representations. A. Fan, P. Stock, B. Graham, E. Grave, R. Gribonval, H. Jegou, and A. Joulin. 2020. Training with quantization noise for extreme model compression. Preprint arXiv:2004.07320. H. Habi, R. Jennings, and A. Netzer. 2020. Hmq: Hardware friendly mixed precision quantization block for cnns. In European Conference on Computer Vision, pages 448–463. Y. Hao, L. Dong, F. Wei, and K. Xu. 2019. Visualizing and understanding the effectiveness of BERT. In Conference on Empirical Methods in Natural Language Processing. L. Hou, Z. Huang, L. Shang, X. Jiang, X. Chen, and Q. Liu. 2020. Dynabert: Dynamic bert with adaptive width and depth. In Advances in Neural Information Processing Systems. L. Hou and J. T. Kwok. 2018. Loss-aware weight quantization of deep networks. In International Conference on Learning Representations. L. Hou, Yao Q., and J. T. Kwok. 2017. Loss-aware binarization of deep networks. In International Conference on Learning Representations. Z. Huang, L Hou, L. Shang, X. Jiang, X. Chen, and Q. Liu. 2021. Ghostbert: Generate more features with cheap operations for bert. In Annual Meeting of the Association for Computational Linguistics. I. Hubara, M. Courbariaux, D. Soudry, R. El-Yaniv, and Y. Bengio. 2016. Binarized neural networks. In Advances in neural information processing systems. X. Jiao, Y. Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, and Q. Liu. 2020. Tinybert: Distilling bert for natural language understanding. In Findings of Empirical Methods in Natural Language Processing. D. Kim, K Singh, and J. Choi. 2020. Learning architectures for binary networks. In European Conference on Computer Vision, pages 575–591. H. Kim, K. Kim, J. Kim, and J. Kim. 2019. Binaryduo: Reducing gradient mismatch in binary activation network by coupling binary activations. In International Conference on Learning Representations. Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut. 2020. Albert: A lite bert for selfsupervised learning of language representations. In International Conference on Learning Representations. F. Li, B. Zhang, and B. Liu. 2016. Ternary weight networks. Preprint arXiv:1605.04711. H. Li, Z. Xu, G. Taylor, C. Studer, and T. Goldstein. 2018. Visualizing the loss landscape of neural nets. In Advances in Neural Information Processing Systems. Y. Li, X. Dong, and W. Wang. 2020a. Additive powersof-two quantization: a non-uniform discretization for neural networks. In International Conference on Learning Representations. Y. Li, W. Wang, H. Bai, R. Gong, X. Dong, and F. Yu. 2020b. Efficient bitwidth search for practical mixed precision neural network. Preprint arXiv:2003.07577. H. Liu, K. Simonyan, and Y. Yang. 2019. Darts: Differentiable architecture search. In International Conference on Learning Representations. 4344 Z. Liu, Z. Shen, M. Savvides, and K. Cheng. 2020. Reactnet: Towards precise binary neural network with generalized activation functions. In European Conference on Computer Vision, pages 143–159. Z. Liu, B. Wu, W. Luo, X. Yang, W. Liu, and K. Cheng. 2018. Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision. X. Ma, P. Zhang, S. Zhang, N. Duan, Y. Hou, D. Song, and M. Zhou. 2019. A tensorized transformer for language modeling. In Advances in Neural Information Processing Systems. B. Martinez, J. Yang, A. Bulat, and G. Tzimiropoulos. 2020. Training binary neural networks with real-tobinary convolutions. In International Conference on Learning Representations. P. Michel, O. Levy, and G. Neubig. 2019. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems. Y. Nahshan, B. Chmiel, C. Baskin, E. Zheltonozhskii, R. Banner, A. M. Bronstein, and A. Mendelson. 2019. Loss aware post-training quantization. Preprint arXiv:1911.07190. P. Rajpurkar, R. Jia, and P. Liang. 2018. Know what you don’t know: Unanswerable questions for squad. Preprint arXiv:1806.03822. P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. Preprint arXiv:1606.05250. M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. 2016. Xnor-net: Imagenet classification using binary convolutional neural networks. In European Conference on Computer Vision. V. Sanh, L. Debut, J. Chaumond, and T. Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. Preprint arXiv:1910.01108. S. Shen, Z. Dong, J. Ye, L. Ma, Z. Yao, A. Gholami, M. W. Mahoney, and K. Keutzer. 2020. Q-bert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence. S. Sun, Y. Cheng, Z. Gan, and J. Liu. 2019. Patient knowledge distillation for bert model compression. In Conference on Empirical Methods in Natural Language Processing. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems. A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. R. Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. Preprint arXiv:1804.07461. J. Wang, H. Bai, J. Wu, X. Shi, J. Huang, I. King, M. Lyu, and J. Cheng. 2020. Revisiting parameter sharing for automatic neural channel number search. In Advances in Neural Information Processing Systems, volume 33. K. Wang, Z. Liu, Y. Lin, J. Lin, and S. Han. 2019. Haq: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8612–8620. B. Wu, Y. Wang, P. Zhang, Y. Tian, P. Vajda, and K. Keutzer. 2018a. Mixed precision quantization of convnets via differentiable neural architecture search. Preprint arXiv:1812.00090. J. Wu, Y. Zhang, H. Bai, H. Zhong, J. Hou, W. Liu, and J Huang. 2018b. Pocketflow: An automated framework for compressing and accelerating deep neural networks. In Advances in Neural Information Processing Systems, Workshop on Compact Deep Neural Networks with Industrial Applications. L. Wu, D. Wang, and Q. Liu. 2019. Splitting steepest descent for growing neural architectures. In Advances in Neural Information Processing Systems, volume 32. J. Xin, R. Tang, J. Lee, Y. Yu, and J. Lin. 2020. Deebert: Dynamic early exiting for accelerating bert inference. In Annual Meeting of the Association for Computational Linguistics. A. H. Zadeh and A. Moshovos. 2020. Gobo: Quantizing attention-based nlp models for low latency and energy efficient inference. Preprint arXiv:2005.03842. O. Zafrir, G. Boudoukh, P. Izsak, and M. Wasserblat. 2019. Q8bert: Quantized 8bit bert. Preprint arXiv:1910.06188. D. Zhang, J. Yang, D. Ye, and G. Hua. 2018. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In European conference on computer vision, pages 365–382. W. Zhang, L. Hou, Y. Yin, L. Shang, X. Chen, X. Jiang, and Q. Liu. 2020. Ternarybert: Distillation-aware ultra-low bit bert. In Conference on Empirical Methods in Natural Language Processing. R. Zhao, Y. Hu, J. Dotzel, C. De Sa, and Z. Zhang. 2019. Improving neural network quantization without retraining using outlier channel splitting. In International Conference on Machine Learning. S. Zhou, Y. Wu, Z. Ni, X. Zhou, H. Wen, and Y. Zou. 2016. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. Preprint arXiv:1606.06160. W. Zhou, C. Xu, T. Ge, J. McAuley, K. Xu, and F. Wei. 2020. Bert loses patience: Fast and robust inference with early exit. In Advances in Neural Information Processing Systems. 4345 A Derivation of Equation (8) In this section, we show the derivations to obtain a and b. Recall the BWN quantizer introduced in Section 2, we have ˆwb 1,i = α1sign(wb 1,i), where α1 = 1 n X i∈I |awt i| + X i∈J |wt j + b| + X i∈K |b| . Similarly, ˆwb 2,i = α2sign(wb 2,i), where α2 = 1 n X i∈I |(1−a)wt i|+ X j∈J |−b|+ X k∈K |wt k−b| . According to ˆwt = ˆwb 1 + ˆwb 2, for those ˆwt i = ˆwb 1,i + ˆwb 2,i = 0, we have 1 n X i∈I |awt i| + X j∈J |wt j + b| + X k∈K |b| = 1 n X i∈I |(1−a)wt i|+ X j∈J | −b|+ X k∈K |wt k−b| . By assuming 0 < a < 1 and b > 0, this can be further simplified to a X i∈I |wt i|+ X j∈J |wt j| = (1−a) X i∈I |wt i|+ X k∈K |wt k|, which gives the solution of a as a = P i∈I |wt i| + P j∈J |wt j| −P k∈K |wt k| 2 P i∈I |wt i| . We empirically find the solution satisifies 0 < a < 1. For ˆwt i ̸= 0, from ˆwt i = ˆwb 1,i + ˆwb 2,i, we have 1 |I| X i∈I |wt i| = α1 + α2 = 1 n X i∈I |awt i| + X j∈J |wt j + b| + X k∈K |b| + 1 n X i∈I |(1−a)wt i| + X j∈J | −b|+ X k∈K |wt k−b| = 1 n X i∈I |wt i| + X j∈J |wt j| + X k∈K |wt k| + 2 X j∈J |b| + 2 X k∈K |b| = 1 n n X i=1 |wt i| + 2(|J | + |K|) · b . Thus the solution for b is b = n |I| P i∈I |wt i| −Pn i=1 |wt i| 2(|J | + |K|) , which satisfies b > 0. B Implementation Details B.1 Detailed Procedure of Adaptive Splitting As mentioned in Section 3.2, the adaptive splitting requires to first estimate the quantization sensitivity vector u. We study the sensitivity in two aspects: the Transformer parts, and the Transformer layers. For Transformer parts, we follow the weight categorization in Section 2.2: MHA-Q/K, MHA-V, MHAO, FFN-Mid and FFN-Out. For each of them, we compare the performance gap between quantizing and not quantizing that part (e.g., MHA-V), while leavging the rest parts all quantized (e.g., MHAQ/K, MHA-O, FFN-Mid and FFN-Out). Similarly, for each Transformer layer, we quantize all layers but leave the layer under investigation unquantized, and calculate the performance gain compared with the fully qauntized baseline. The performance gain of both Transformer parts and layers are shown in Figure 7. As can be seen, for Transformer parts, the FFN-Mid and MHA-Q/K rank in the first and second place. In terms of Transformer layers, shallower layers are more sensitive to quantization than the deeper ones. However, the absolute performance gain may not reflect the quantization sensitivity directly, since Transformer parts have different number of parameters. Therefore, we divide the performance gain by the number of parameters in that part or layer to obtain the parameter-wise performance gain. We are thus able to measure the quantization sensitivity of the ith Transformer part in the jth Transformer layer by summing their parameter-wise performance gain together. We also apply the same procedure to word embedding and pooler layer to otain their sensitivity scores. We are now able to solve Equation (11) by dynamic programming. The combinatorial optimization can be viewed as a knapsack problem, where the constraint C −C0 is the volume of the knapsack, and the sensitivity scores u are the item values. B.2 Hyper-parameter Settings We first perform the two-stage knowledge distillation, i.e., intermediate-layer distillation (Int. Dstil.) and prediction-layer distillation (Pred. Dstil.) on 4346 BinaryBERT Int. Dstil. (Ternary) Pred. Dstil. (Ternary) Split Ft. (Binary) Batch Size 32 32 32 Sequence Length 128 128 128 Learning rate (LR) 5e-5 2e-5 2e-5 LR Decay Linear Linear Linear Warmup portion 0.1 0.1 0.1 Weight Decay 1e-2 1e-2 1e-2 Gradient Clipping 1 1 1 Dropout 0.1 0.1 0.1 Epochs w/o DA -other dataserts 6 6 6 Epochs w DA -other dataserts 1 1 1 Epochs w/o DA -MNLI, QQP 3 3 3 Table 7: Hyper-parameters for training BinaryBERT on the GLUE benchmark at different stages. the ternary model, and then perform ternary weight splitting followed by fine-tuning (Split Ft.) with only prediction-layer distillation after the splitting. The initial learning rate is set as 5 × 10−5 for the intermediate-layer distillation, and 2×10−5 for the prediction-layer distillation, both of which linearly decay to 0 at the end of training. We conduct experiments on GLUE tasks both without and with data augmentation (DA) except for MNLI and QQP due to their limited performance gain. The running epochs for MNLI and QQP are set to 3, and 6 for the rest tasks if without DA and 1 otherwise. For the rest hyper-parameters, we follow the default setting in (Devlin et al., 2019). The detailed hyperparameters are summarized in Table 7. C More Empirical Results C.1 Performance Drop by Binarization Here we provide more empirical results on the sharp drop in performance as a result of binarization. We run multi-bit quantization on the BERT model over representative tasks of the GLUE benchmark, and activations are quantized in both 8bit and 4-bit. We run 10 independent experiments for each task except for MNLI with 3 runs. We follow the same procedure in Section 2.1, and the default experimental setup in Appendix B.2 without data augmentation and splitting. The results are shown in Figures 8 and 9 respectively. It can be found that while the performance drops slowly from full-precision to ternarization, there is a consistent sharp drop by binarization in each tasks and on both 8-bit and 4-bit activation quantization. This (a) Transformer Parts. (b) Transformer Layers. Figure 7: The performance gain of different Transformer parts and layers in descending order. All numbers are averaged by 10 random runs with standard deviations reported. is similar to the findings in Figure 1. C.2 More Visualizations of Loss Landscape To comprehensively compare the loss curvature among the full-precision, ternary and binary models, we provide more landscape visualizations aside from the value layer in Figure 2. We extract parameters from MHA-K, MHA-O, FFN-Mid and FFN-out in the first two Transformer layers, and the corresponding landscape are shown in Figure 10, Figure 11, Figure 12, Figure 13 respectively. We omit MHA-Q due to page limitation, and also it is symmetric to MHA-K with similar landscape observation. It can be found that binary model have steep and irregular loss landscape in general w.r.t different parameters of the model, and is thus hard to optimize directly. C.3 Ablation of Knowledge Distillation While knowledge distillation on BERT has been thoroughly investigated in (Jiao et al., 2020; Hou et al., 2020; Zhang et al., 2020), here we further conduct ablation study of knowledge distillation on the proposed ternary weight splitting. We compare with no distillation (“N/A”), prediction distillation (“Pred”) and our default setting (“Int.+Pred”). For “N/A” or “Pred”, fine-tuning after splitting follows the same setting to their ternary 4347 (a) MNLI-m. (b) SST-2. (c) CoLA. (d) STS-B. (e) MRPC. (f) RTE. Figure 8: Performance of quantized BERT with different weight bits and 8-bit activation on the GLUE Benchmarks. The results are obtained from 10 random seeds except for MNLI with 3 seeds. (a) MNLI-m. (b) SST-2. (c) CoLA. (d) STS-B. (e) MRPC. (f) RTE. Figure 9: Performance of quantized BERT with different weight bits and 4-bit activation on the GLUE Benchmarks. The results are obtained from 10 random seeds except for MNLI with 3 seeds. Size (MB) Strategy QNLI SST-2 CoLA STS-B MRPC RTE Avg. 10.6 Min. 91.1 93.1 52.8 88.2 85.3 69.3 80.0 Rand. 90.8 92.7 53.3 88.2 85.5 70.0 80.1 Max. 91.0 92.7 53.7 88.0 86.5 71.1 80.5 11.4 Min. 91.0 93.0 53.8 88.3 85.5 71.5 80.5 Rand. 91.0 92.9 54.7 88.4 86.5 70.8 80.7 Max. 91.0 93.0 54.6 88.4 86.3 71.1 80.7 12.2 Min. 91.1 92.7 53.5 88.5 85.3 71.5 80.4 Rand. 91.1 92.9 54.1 88.5 86.0 71.8 80.4 Max. 91.0 92.9 53.8 88.6 86.8 71.1 80.7 13.0 Min. 91.2 92.8 54.8 88.5 85.1 72.2 80.8 Rand. 91.2 92.9 54.1 88.4 86.0 71.8 80.8 Max. 91.1 93.1 56.1 88.6 86.1 70.8 81.0 13.8 Min. 91.1 93.0 55.4 88.5 85.8 71.5 80.9 Rand. 91.5 92.9 54.7 88.5 85.0 72.2 80.8 Max. 91.4 92.9 55.5 88.7 86.3 72.6 81.2 Table 8: Results on GLUE development set for adaptive splitting with 8-bit activation quantization. training. “Int.+Pred” follows our default setting in Table . We do not adopt data-augmentation, and results are shown in Table 10. It can be found that “Int.+Pred.” outperforms both “N/A” and “Pred.” with a clear margin, which is consistent to the findings in (Zhang et al., 2020) that knowledge distillation helps BERT quantization. C.4 Detailed Results of Adaptive Splitting The detailed comparison of our adaptive splitting strategy against the random strategy (Rand.) and minimal gain strategy (Min.) under different model size are shown in Table 8 and Table 9. It can be found that for both 8-bit and 4-bit activation quantization, our strategy that splits the most sensitive modules mostly performs the best on average under various model sizes. Size (MB) Strategy QNLI SST-2 CoLA STS-B MRPC RTE Avg. 10.6 Min. 90.6 92.6 51.7 87.4 85.3 70.8 79.7 Rand. 91.1 92.7 51.3 87.6 84.8 68.2 79.3 Max. 90.9 92.7 53.5 87.5 84.6 70.0 79.9 11.4 Min. 90.9 92.8 50.9 87.6 85.3 69.4 79.5 Rand. 90.8 92.8 51.7 87.5 84.6 70.4 79.6 Max. 91.1 92.6 52.1 87.7 85.3 70.0 79.8 12.2 Min. 90.9 92.7 50.8 87.6 84.8 70.4 79.5 Rand. 91.2 93.0 52.0 87.6 85.1 70.0 79.8 Max. 90.9 92.9 52.2 87.6 85.1 70.4 79.9 13.0 Min. 91.1 92.8 52.6 87.7 86.3 69.7 80.0 Rand. 91.3 93.0 52.9 87.8 85.8 69.7 80.1 Max. 91.3 92.9 53.4 87.8 85.3 69.7 80.1 13.8 Min. 91.1 93.1 51.5 87.9 84.8 70.0 79.7 Rand. 91.3 92.9 52.3 87.7 85.1 71.1 80.1 Max. 91.3 92.8 53.6 88.0 85.8 70.8 80.4 Table 9: Results on GLUE development set for adaptive splitting with 4-bit activation quantization. KD #Bits (W-E-A) MNLI (-m) SST-2 CoLA MRPC N/A 1-1-8 83.2 92.1 49.2 82.8 Pred. 1-1-8 84.0 91.7 48.6 84.1 Int.+Pred. 1-1-8 84.2 92.6 53.4 85.5 N/A 1-1-4 82.6 90.9 39.2 76.5 Pred. 1-1-4 83.4 92.3 38.9 76.2 Int.+Pred. 1-1-4 83.9 92.3 44.4 83.3 Table 10: Ablation study on knowledge distillation. C.5 Architecture Visualization We further visualize the architectures after adaptive splitting on MRPC in Figure 14. For clear presentation, we merge all splittable parameters in each Transformer layer. As the baseline, 9.8MB refers to no splitting, while 16.5MB refers to splitting all splittable parameters in the model. According to Figure 14, with the increasing model size, shallower layers are more preferred for splitting than deeper layers, which is consistent to the findings in Figure 7. 4348 (a) Full-precision Model. (b) Ternary Model. (c) Binary Model. (d) All Together. Figure 10: Loss landscape visualizations w.r.t MHA-K parameters of the 1st and 2nd Transformer layers on MRPC. (a) Full-precision Model. (b) Ternary Model. (c) Binary Model. (d) All Together. Figure 11: Loss landscape visualizations w.r.t MHA-Out parameters of the 1st and 2nd Transformer layers on MRPC. (a) Full-precision Model. (b) Ternary Model. (c) Binary Model. (d) All Together. Figure 12: Loss landscape visualizations w.r.t FFN-Mid parameters of the 1st and 2nd Transformer layers on MRPC. (a) Full-precision Model. (b) Ternary Model. (c) Binary Model. (d) All Together. Figure 13: Loss landscape visualizations w.r.t FFN-Out parameters of the 1st and 2nd Transformer layers on MRPC. Figure 14: The architecture visualization for adaptive splitting on MRPC. The y-axis records the number of parameters split in each layer instead of the storage.
|
2021
|
334
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4349–4359 August 1–6, 2021. ©2021 Association for Computational Linguistics 4349 Are Pre-trained Convolutions Better than Pre-trained Transformers? Yi Tay Google Research Mountain View, California [email protected] Mostafa Dehghani Google Research, Brain Team Amsterdam, Netherlands [email protected] Jai Gupta Google Research Mountain View, California [email protected] Vamsi Aribandi∗ Google Research Mountain View, California [email protected] Dara Bahri Google Research Mountain View, California [email protected] Zhen Qin Google Research Mountain View, California [email protected] Donald Metzler Google Research Mountain View, California [email protected] Abstract In the era of pre-trained language models, Transformers are the de facto choice of model architectures. While recent research has shown promise in entirely convolutional, or CNN, architectures, they have not been explored using the pre-train-fine-tune paradigm. In the context of language models, are convolutional models competitive to Transformers when pre-trained? This paper investigates this research question and presents several interesting findings. Across an extensive set of experiments on 8 datasets/tasks, we find that CNN-based pre-trained models are competitive and outperform their Transformer counterpart in certain scenarios, albeit with caveats. Overall, the findings outlined in this paper suggest that conflating pre-training and architectural advances is misguided and that both advances should be considered independently. We believe our research paves the way for a healthy amount of optimism in alternative architectures. 1 Introduction In the modern era of pre-training, there appears to be an unbreakable tie between Transformer architectures (Vaswani et al., 2017) and pre-trained language models. Models such as BERT (Devlin et al., 2018), RoBERTa (Liu et al., 2019), and T5 (Raffel et al., 2019) have all adopted Transformers as their underlying architecture. As a matter of fact, there are barely any recent pre-trained models not based on Transformers. While the contextual representation learning has a rich history (Pennington et al., 2014; Dai and Le, ∗Google AI Resident 2015; Chidambaram et al., 2018; Liu et al., 2020; Qiu et al., 2020), modern pre-trained language modeling started with models like ELMo (Peters et al., 2018) and CoVE (McCann et al., 2017) which are based on recurrent (e.g. LSTM (Hochreiter and Schmidhuber, 1997)) architectures. Although they were successful, research using these architectures dwindled as Transformers stole the hearts of the NLP community, having, possibly implicitly, been perceived as a unequivocal advancement over its predecessors. Recent work demonstrates the promise of entirely convolution-based models (Wu et al., 2019; Gehring et al., 2017) and questions the necessity of self-attentive architectures like Transformers. For example, in (Wu et al., 2019), the proposed convolutional seq2seq models outperform Transformers on a series of canonical benchmarks such as machine translation and language modeling. From these findings emerge a rather natural line of questioning - should we consider pre-trained models beyond Transformers? Despite early success, the relevance of convolutional models in the era of pre-trained language models remains an open question. To the best of our knowledge, convolutional architectures have not yet been rigorously evaluated under the pretrain-fine-tune paradigm. This is the primary purpose of this work. Concretely, this paper seeks to empirically validate whether pre-trained convolutions are competitive with pre-trained Transformers across a range of tasks. The interaction between pre-training schemes and model architectures is an under-studied topic. Are only Transformers able to capitalize on the 4350 benefits of pre-training? If we use a different architectural inductive bias, would there also be a substantial gain unlocked by pre-training? Are pretrained convolutions better in particular scenarios? This paper investigates these questions. There are a number of obvious benefits of convolution-based models. Firstly, convolutions do not suffer from the quadratic memory complexity of self-attention - a problem significant enough that it spawned the creation of the entirely new category of “efficient” Transformer architectures (Tay et al., 2020b, 2021). Secondly, convolutions operate locally and do not rely on positional encodings as an order signal to the model. That said, convolutions also come with a slew of downsides. For example, being unable to access global information means such models are unable to perform a form of cross-attention across multiple sequences. We dive into the details of this more in subsequent sections. In this paper, we present a pre-trained convolutional sequence-to-sequence, or Seq2Seq, model. We train our convolutional model using span-based sequence-to-sequence denoising objectives similar to those employed in T5 (Raffel et al., 2019). We evaluate a variety of convolutional variants (e.g., dilated, lightweight, dynamic (Wu et al., 2019), etc.) under both raw (no pre-training) and pre-train-finetune paradigms. Our goal is to understand the true competitiveness of convolutional architectures in the era of pre-training. We show that pre-trained convolutions are competitive against pre-trained Transformers via a set of experiments on a potpourri of NLP tasks, like toxicity detection, sentiment classification, news classification, query understanding and semantic parsing/compositional generalization (Kim and Linzen, 2020). Moreover, we find that pretrained convolutions can outperform, in terms of model quality and training speed, state-of-the-art pre-trained Transformers (Raffel et al., 2019) in certain scenarios. However, to provide a balanced perspective, we also describe scenarios where pretrained convolutions do not perform well and may be deemed unsuitable. Contributions Overall, the main contributions of this paper can be summarized as follows: • We perform a comprehensive empirical evaluation of convolutional Seq2Seq models under the pre-train-fine-tune paradigm. To the best of our knowledge, the competitiveness and relevance of pre-trained convolutions still remains an open question. • We make several important observations. Specifically, we find that (1) pre-training helps convolutional models just as much as it helps Transformers, and (2) pre-trained convolutions are competitive alternatives in certain scenarios in terms of model quality and training speed. • We conduct extensive experiments across 8 datasets spanning a diverse range of tasks and domains. On 7 out of 8 tasks, we find that pre-trained convolutions outperform a recent state-of-the-art transformer (T5 (Raffel et al., 2019)) with and without pre-training. We examine the speed and operation count (FLOPS) of convolutions versus Transformers and find that convolutions are not only faster but also scale better to longer sequence lengths. 2 Related Work Pre-training on a large corpus has become the primary method of learning universal language representations to solve different downstream NLP tasks. The first generation of pre-trained models aimed at learning embedding for words, like Skip-Gram (Mikolov et al., 2013) and Glove (Pennington et al., 2014), and quickly developed to learning contextualized representation for words, like ELMO (Peters et al., 2018), GPT (Radford et al., 2018), and BERT (Devlin et al., 2018). This, however, is not the only axis in which pre-trained models have evolved. Different objective functions and various tasks, both supervised and unsupervised, have been explored for pre-training. For instance, CoVe (McCann et al., 2017) uses machine translation as the pre-training task, ELMO (Peters et al., 2018) and GPT (Radford et al., 2018) use language modeling objectives, BERT (Devlin et al., 2018) uses masked language modeling, T5 (Raffel et al., 2019) and MASS (Song et al., 2019) use Seq2Seq masked language modeling, and XLNet (Yang et al., 2019) utilizes permuted language modeling. In addition to this, BART (Lewis et al., 2019) uses a denoising autoencoder setup during pre-training, where the model takes a partially corrupted input and is trained to recover the original, undistorted input. Some models use a contrastive learning setup during pertaining, like replaced token detection, used 4351 by ELECTRA (Clark et al., 2020), and sentence order prediction, used by ALBERT (Lan et al., 2019) and StructBERT (Wang et al., 2019). Another axis where pre-trained models in NLP explored different ideas is model architecture. ELMO (Peters et al., 2018) and CoVe (McCann et al., 2017) used LSTMs as the base model. Later, Transformers (Vaswani et al., 2017) became the de facto architecture of pre-trained NLP models. BERT (Devlin et al., 2018), XLNet (Yang et al., 2019) and RoBERTa (Liu et al., 2019) use the Transformer encoder, while GPT (Radford et al., 2018), GPT-2 (Radford et al.), and GPT-3 (Brown et al., 2020) use the Transformer decoder as the backbone. Some pre-trained models are also are based on the encoder-decoder transformer architecture, like T5 (Raffel et al., 2019), MASS (Song et al., 2019), and BART (Lewis et al., 2019). In this paper, we investigate another model architecture variation by studying the power of convolutional neural network as the backbone of pre-trained models for NLP. Convolutions have always been an interesting choice for sequence modeling and NLP applications (Kim, 2014; Bai et al., 2018; Kalchbrenner et al., 2016). Convolutions are lightweight and fast and have many interesting use-cases, notably for lightweight classification. In the era when LSTMs were the workhorses of NLP applications, convolutions were positioned nicely on the pareto frontier of the compute-performance curve. They are fast and lightweight, and unlike Transformers, they do not suffer from quadratic complexity. Our work is also well-aligned with the resurgence of interest in convolutions where (Wu et al., 2019) showed that convolutions can outperform self-attention on several sequence transduction tasks. Moreover, the necessity of the self-attention inductive bias in transformers have been also a subject of recent interest. Synthesizer models (Tay et al., 2020a) showed that transformers can still do pretty well without token-token dot product self-attention and a random attention matrix can perform competitively on certain tasks. 3 Pre-Trained Convolution Models This section describes the pre-trained Convolution Model. For most of our experiments, we adopt depthwise separable convolutions (Kaiser et al., 2017; Sifre and Mallat, 2014; Chollet, 2017) which have shown to be fast and efficient variants of the standard convolution. 3.1 Lightweight Depthwise Convolution This section introduces Lightweight Depthwise Convolutions (Wu et al., 2019) which forms the backbone of our pre-trained convolution model. 3.1.1 Depthwise convolutions Depthwise convolutions convolve independently over every channel. Given an input tensor X of dimensions n × d, the depthwise convolution, D(X, Wc,:, i, c) is defined as: Oi,c = k X j−1 Wc,j · Xi+j−⌈k+1 2 ⌉), c (1) where W ∈Rd×k are the learnable parameters of the layer. Oi,c is the output at position i and channel c. The overall output is a tensor of n × d of identical shape as the input. 3.1.2 Lightweight Convolutions L(.) are depthwise separable convolutions with (1) softmax-normalized kernels and (2) shared output channels and weight tying. Specifically, this is written as: OL i,c = k X j−1 softmax(Wˆc,j) · Xi+j−⌈k+1 2 ⌉), ˆc (2) where ˆc = cH d . In short, parameters are shared every d H output channels. When H = 1, this is equivalent to sharing all the weights of all channels. 3.1.3 Dynamic Convolutions Dynamic Convolutions DY (.) are a new form of lightweight convolutions introduced by (Wu et al., 2019). The key idea is to learn position-specific kernels for performing lightweight convolutions. This can be written as: DY = L(X, f(Xi)h,:, i, c), (3) where f(.) is a linear transformation with parameters W Q ∈RH×k×d that learns a position dependent kernel. 3.2 Span-based Seq2Seq pre-training We adopt span-based sequence-to-sequence pretraining as per (Raffel et al., 2019). Specifically, given an input sequence, we randomly mask spans of lengths L and replace them with a special sentinel token. The pre-training task is then to generate the masked tokens as targets. For example: Inputs: The happy cat sat [mask]. and Outputs: on the mat. 4352 3.2.1 Convolutional Seq2Seq Architecture We implement a Seq2Seq (Sutskever et al., 2014) architecture similar to (Wu et al., 2019). The key difference when compared with Transformer architectures is that we replace the multi-headed selfattention with convolutional blocks. Instead of query-key-value transforms, we use gated linear unit projections following (Wu et al., 2019). Each convolution block be written as: X1 = W IX ⊙sigmoid(W SX), X2 = ConvBlock(X1), X3 = W O(X2), where W I, W S, W O are trainable parameters. We experiment with simple lightweight convolutions, dynamic convolutions and dilated convolutions in our experiments. Following (Wu et al., 2019; Gehring et al., 2017), the encoder-decoder attention remains untouched. The convention follows the backbone Transformer model in which we wrap each submodule with layer normalization and residual connectors. Hence, each Conv block is written as: XA = LayerNorm(Conv(X)) + X, XB = LayerNorm(FFN(XA) + XA, where Conv is any of the convolution models that we explore in our experiments. FFN(.) is a two layer feed-forward network with ReLU activations in the middle. 3.2.2 Optimization The model optimizes the token-wise cross-entropy loss and is trained with teacher forcing. L = L X t=1 n X i=1 log(πt i) + (1 −yt i) log(1 −πt i), where πt i is the prediction of class i at time step t and yt i is the ground truth label of the class i at time step t. 4 Research Questions and Discussion Before we delve into our experiments, we establish a set of research questions and agenda we hope this work aims to bring clarity to. • RQ1: Do convolutions benefit from pretraining as much as Transformers? • RQ2: Are convolutional models, pre-trained or otherwise, competitive with Transformer models? When do they perform well? • RQ3: What are the benefits (if any) of using pre-trained convolution models over pretrained Transformers? Are convolutions faster alternatives to self-attention based Transformers? • RQ4: What are the failure modes, caveats and reasons to not use pre-trained convolutions? • RQ5: Are certain convolution variants better than others? 5 Experiments and Analysis This section presents our analysis and results. 5.1 Datasets Our evaluation is based on the following datasets and tasks. • Toxicity Detection - We use the CIVIL COMMENTS (Borkan et al., 2019) and WIKI TOXIC SUBTYPES dataset (Wulczyn et al., 2017). Given a piece of short text (originating from social media or wikipedia), the goal is to determine if the content is toxic, i.e., a binary classification task. For this task, we evaluate on both accuracy and F1 score. • Sentiment Classification - This is a binary classification task that determines the polarity of documents, sentences and/or tweets. We use the IMDb reviews dataset (Maas et al., 2011), Stanford Sentiment Treebank (SST2) (Socher et al., 2013) dataset, along with Twitter Sentiment140 (S140) (Go et al., 2009) dataset. • News Classification - This is a task of topic categorization for news articles. We use the AGNews dataset (Zhang et al., 2015). This is a four-way classification task. • Question Classification We use the TREC fine-grained question classification dataset (Li and Roth, 2002). This task involves classifying questions into 46 fine-grained question categories. • Semantic Parsing / Compositional Generalization Compositional generalization is the 4353 ability of models to generalize compositionally outside of the training distribution. To be specific, it needs be able to handle unseen combinations at test time. For this task, we use the COGS dataset (Kim and Linzen, 2020), a task of generating semantic representation of a given English sentence. For example, A cat smiled →cat(x1) AND smile.agent(x2, x1). All of the datasets, with the exception of the recent COGS dataset (Kim and Linzen, 2020), are Tensorflow datasets1. For each dataset, we evaluate all models with and without pre-training (details in subsequent sections). Table 1 reports the statistics of the datasets used in this paper. Dataset / Task # Train # Test # Class Civil Comments 3,820,210 205,781 2 Wiki Toxicity 561,808 234,564 2 IMDb 25,000 25,000 2 SST-2 67,000 1,800 2 S140 1,600,000 359 2 TREC 4,500 500 46 AGNews 120,000 7,600 4 COGS 24,000 3000 N/A Table 1: Statistics of datasets used in our experiments. Datasets are diverse in terms of domains, tasks and amount of labeled data. 5.2 Experimental Setup This section describes our experimental setup. 5.2.1 Models Our models are largely based on sequence to sequence models, a paradigm that has demonstrated great success made evident by models such as BART (Lewis et al., 2019) and T5(Raffel et al., 2019). We implement our models in Mesh Tensorflow (MTF) (Shazeer et al., 2018), a library for distributed and efficient parallel model training that has similar API to Tensorflow. We train models that are of base size, which corresponds to 12 layers each in the encoder and decoder, along with 3072 dimensions for the feed-forward layers, a model dimension of 768 and a total of 12 heads. Our Transformer models are largely based on T5 (Raffel et al., 2019), which is considered the current state-of-the-art Transformer model for NLP tasks and hence serves as a strong baseline. For the convolution models, our lightweight convolution 1https://www.tensorflow.org/datasets/ catalog/overview. and dynamic convolution models have a window size2 of 7 across all layers, the number of unique depth filters is 2. For dilated models, we use a filter size of [4, 4, 7, 7, 15, 15, 15, 15, 31, 31, 31] for our 12 layer convolution model. 5.2.2 Pre-training We pre-train both our convolutional and Transformer models for 524K steps with a batch size of 128. Given the input sequence length of 512, this corresponds to 65536 tokens per batch. For pre-training, we use the Colossal Cleaned CommonCrawl Corpus (C4) (Raffel et al., 2019) dataset which has demonstrated impressive results on downstream tasks. We use the span based seq2seq objective as the pre-training objective as mentioned in earlier sections. The span size is set to 3 and a corruption rate of 15% is adopted. We use the Adafactor optimizer (Shazeer and Stern, 2018) with an inverse square root learning rate scheduler. Each pre-training run is performed using 16 TPU-v3 chips and takes approximately 12 hours to complete for models of base size. 5.2.3 Downstream Fine-tuning We fine-tune the pre-trained models using the following set of hyperparameters: We use a constant learning rate which is tuned amongst {0.001, 0.0005, 0.0001}. The batch size is generally set to 64 but occasionally set to 32 for smaller datasets. Intuitively, sequence length is task dependent but generally approximately the 90th percentile for each task. We fine-tune for a maximum of 100K steps and report peak validation performance. Fine-tuning uses the same Adafactor optimizer as during training. We perform fine-tuning on similar hardware, i.e., typically 16 TPUv3 chips are used per fine-tuning job. 5.3 Experimental Results This section describes our experimental setup and results. 5.4 Results on Toxicity Detection Table 2 reports results on toxicity detection. On both toxicity detection datasets the pre-trained and no-pre-training (raw) setup, the best models are the dilated convolution models and the dynamic convolution models. In fact, all convolutional models 2We believe that tuning the hyperparameters of the convolution models can result in even better performance. However, we decided to keep these hyperparameters simple for the start. 4354 outperform Transformers on both CivilComments and WikiToxic. Before pre-training, convolutions outperform Transformers by approximately 1.5 absolute percentage points. The gap narrows after pretraining where Transformers see a better gain (e.g., +5.1% against +4.3%) from pre-training over convolutions on the CivilComments dataset. However, the converse is true on WikiToxic - the only case of performance degradation after pre-training. Overall, on this task, convolutions are competitive to Transformers and outperform them. 5.5 Results on Sentiment Classification Results on Sentiment Classification (IMDb, SST-2 and S140) can be found in Table 2. On the IMDb reviews dataset, the best non-pre-trained model is the lightweight convolution model, outperforming the Transformer model. The best pre-trained model is the Transformer model. However, all convolutional models come in close with less than a percentage point gap difference with pre-trained Transformers. On the SST-2 and S140 tasks, we observe that the best models are convolution-based, regardless of whether the model is pre-trained or not. 5.6 Results on Question Classification The best non-pre-trained model is the Lightweight Convolution model. For pre-trained models, convolutional models also outperform the pre-trained Transformer. On this task, while most models benefit significantly from pre-training, Transformers seem to benefit slightly more from pre-training. 5.7 Results on News Classification Results on news classification seems to follow similar trends as other benchmarks. Convolutional models outperform Transformers both in non-pretrained and pre-trained setups. The highest gain from pre-training is obtained from the dilated convolution model. 5.8 Results on Compositional Generalization Challenge and Semantic Parsing We conduct additional experiments on semantic parsing and compositional generalization. The task is framed as a sequence generation task. We use the recently proposed (Kim and Linzen, 2020) dataset. On the in-distribution test set, Transformers and convolutions have identical performance (95%). On the generalization or out of distribution set, Transformers perform at 77.5% while convolutions come in at 76.9. While convolutions do not exactly outperform Transformers, they come in close enough to be considered competitive. 5.9 Summary of Results On the seven tasks across a broad range of domains we find that (1) non-pre-trained convolutions are competitive and frequently outperform non-pretrained Transformers, (2) pre-trained convolutions outperform pre-trained Transformers on six out of seven tasks. This answers RQ2. We also find that convolutions are able to benefit from pre-training, in a similar fashion to self-attention-based models. Hence, the benefits achieved by pre-training are not exclusive to Transformer models. This answers RQ1. Amongst the pre-trained convolutional models, we find that dilated convolutions and dynamic convolutions are generally better than lightweight convolutions, thus answering RQ5. Finally, we observe that relative performance (i.e., rankings) do change with pre-training. This definitely shows that there is some kind of effect from composing architectures with pre-training. The direct implication of this effect is that a model that performs well (relatively) without pre-training will not necessarily perform the best when pretrained (and vice versa). Hence, aside from conflating architectures with pre-training schemes, we do also need to take note that different architectures may behave differently under pre-training. 6 Discussion and Analysis This section expands on the results via a detailed analysis and discussion. We discuss the pros/cons of pretrained convolutions, the impact of pretraining on performance and also recommendations to the broader community. 6.1 When do we expect pre-trained convolutions to fail? In our experimental section, we observed the potential upsides of convolutional models over wellestablished pre-trained Transformers and observe that we are able to get quality improvements in certain cases. However, it might be good to further understand the drawbacks of convolutions. One obvious weakness of pre-trained convolutions are their lack of cross-attention inductive bias that comes for free with self-attention in the Transformer encoder. For this reason, it is not a 4355 CIVILCOMMENT WIKITOXIC IMDb SST-2 S140 TREC News Model Acc F1 Acc F1 Acc Acc Acc Acc Acc No pre-training Trans. 77.22 85.09 91.93 95.45 84.81 78.44 58.84 78.00 84.25 Light 78.58 85.82 91.05 94.65 85.88 81.65 60.64 82.20 87.22 Dilat. 79.94 86.50 92.29 94.91 85.84 79.01 55.62 79.60 81.24 Dyna. 78.49 84.71 90.06 95.66 85.69 82.80 60.84 80.20 85.13 With pre-training Trans. 81.16 86.56 91.46 95.12 94.16 92.09 61.65 93.60 93.54 Light 81.47 87.58 93.61 96.48 93.60 92.20 61.65 93.60 93.63 Dilat. 81.67 87.78 93.84 96.21 93.92 92.09 62.85 94.20 93.26 Dyna. 81.83 87.71 93.76 96.53 93.35 91.59 62.45 92.40 93.93 Gain from pre-training Trans. +5.1% +1.7% -0.6% -0.4% +11.0% +17.4% +4.7% +20.0% +11.0% Light +3.7% +2.1% +2.8% +1.9% +9.0% +13.0% +1.7% +14.0% +7.3% Dilat. +2.1% +1.5% +1.7% +1.4% +9.4% +17.0% +13.0% +18.0% +14.8% Dyn. +4.3% +3.5% +4.1% +1.0% +8.9% +10.6% +2.6% +15.2% +10.4% Table 2: Comparison of pre-trained Convolutions and pre-trained Transformers on toxicity detection, sentiment classification, question classification and news classification. All models have approximately 230M parameters and are 12 layered seq2seq architectures. Our findings show that convolutions (1) also benefit from pretraining and (2) are consistently competitive to transformer models with and without pretraining. good idea to use pre-trained convolutions for tasks that requires modeling the relationship between two or more sequences. To verify this, we run experiments on SQuAD and MultiNLI and find that convolutions do not come close to Transformers just because of this missing inductive bias. This should be clearly distinguished when examining and evaluating models, as how the early SNLI leaderboard3 distinguished between models that used cross-attention and models that did not. Our initial evaluations on benchmarks like SQuAD/MNLI (Rajpurkar et al., 2016; Williams et al., 2017) showed that pre-trained convolutions are indeed significantly lackluster. For example, convolutions only achieve ≈75% accuracy on MultiNLI, while transformers easily achieve ≈84% accuracy. Likewise, while transformers achieve about ≈90% F1 on SQuAd, convolutions come in around ≈70%. This is entirely expected because there is no way the premise/question can interact with the hypothesis/context. (RQ4). However, our experiments show that this was only because they lack this cross-attention property. When we augment convolutions with a single layer of cross attention at the encoder, we find that pre-trained convolutions come close (a delta of 3https://nlp.stanford.edu/projects/ snli/ (≈1%)) to pre-trained Transformers on datasets such as MultiNLI (Williams et al., 2017), achieving about ≈83% accuracy. That said, we leave it to the practitioner to decide whether the cross-attention inductive bias is actually important for the problem at hand. We also like to emphasize that the pattern of concatenating sentence pairs is not necessary practical when scaling up since this requires inference on every permutation of sentence pairs. For this reason, dual encoder setups that do fast embedding space look-ups are more practical and feasible in practice (Guo et al., 2020). Given the strong performance of convolutions in a series of encoding tasks, we can expect pre-trained convolutions to do well in a dual encoder setup. 6.2 What are the benefits of pre-trained convolutions over Transformers? We observed a reasonable quality improvement from using convolutions over Transformers. This section discusses the additional benefit. 6.2.1 Convolutions are faster and scale better to long sequences Figure 1 reports training speed of convolution (LightConvs) versus transformers on a sequence to sequence task. The input lengths are varied from {64, 128, 256, 512, 1024, 2048, 4096}. We 4356 Figure 1: Effect of sequence length on processing speed (examples per second) on a seq2seq masked language modeling task. Results are benchmarked on 16 TPUv3 chips on C4 pre-training. Results are in log scale. show that convolutions are not only consistently faster (even at shorter sequences) but scale better than transformers. Convolution scales linearly while transformers are not able to scale to longer sequences. 6.2.2 Convolutions are FLOPs efficient We measure the number of FLOPs of convolutions versus transformers as we increase the sequence length. Figure 2 shows the phenomenon while varying sequence length. In general, across all sequence lengths, convolutions are more efficient in the number of floating point operations. Figure 2: Effect of sequence length on number of FLOPs (einsum ops) on a seq2seq masked language modeling task. Results are benchmarked on 16 TPUv3 chips on C4 pre-training. Results are in log scale. The overall findings that convolutions are faster both in wall clock time and in FLOPs answers RQ3. Moreover, we find that the FLOP efficiency of convolutions scales better across sequence lengths. 6.3 Are we suggesting to completely replace Transformers with convolution? While Transformers have dominated the research landscape in NLP, this paper suggests that there are commonly overlooked benefits to convolutions such as model quality, speed, FLOPs and scalability. Moreover, it is previously unknown to whether convolutions benefit from pre-training. In this paper, we showed that they are competitive on some tasks and also benefit from pre-training in similar fashion to transformer models. However, on the flip side, we also highlighted that they are unable to handle tasks that require cross-attention or when there is a need to model > 1 sentence or documents within the same sequence. We believe that practitioners have good options and it might be worthwhile to explore architectures outside the well-established transformer models. 6.4 On not conflating pre-training with architectural advances In this paper, we showed that three other (convolutional-based) architectures (e.g., lightweight, dymamic and dilated) also benefit from pre-training to the same extent as transformer models. In the current research landscape, pre-training has always be tightly coupled and associated with transformers architectures. As a result, the success of BERT, transformers and large language models seem to be pretty conflated. While it is true that, to this date, the only model that large-scale pretraining has been applied to are transformer models, we believe there might be potential in other architectures. Based on our empirical findings, we believe there is still significant room for the improving the understanding of the compositional effects of architecture and pre-training. Hence, we believe that the impact of this work extends beyond showing the competitiveness of convolution models in NLP. More concretely, the take home message is that there should be a healthy level of optimism in exploring architectural alternatives. 7 Conclusion In this paper, we conducted an extensive study of the viability and feasibility of pre-trained convolu4357 tions. Our experimental results show that convolutions can outperform Transformers in both pretrain and non-pre-trained setups. Our extensive experiments across 8 datasets spanning a diverse range of tasks, show that convolutions are able to benefit from pre-training to the same (or sometimes greater) extent than Transformers. While pre-trained transformers are the de-facto choice of architecture, our results show that they might not be the best in certain scenarios. Additionally, we discussed the caveats, trade-offs pertaining with runtime, scalability, number of FLOPS and model quality. Finally, we discussed the situations or data types that convolutions are not well equipped to handle and make an empirically informed recommendation for practitioners. References Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271. Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2019. Nuanced metrics for measuring unintended bias with real data for text classification. CoRR, abs/1903.04561. Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. Muthuraman Chidambaram, Yinfei Yang, Daniel Cer, Steve Yuan, Yun-Hsuan Sung, Brian Strope, and Ray Kurzweil. 2018. Learning cross-lingual sentence representations via a multi-task dual-encoder model. arXiv preprint arXiv:1810.12836. Franc¸ois Chollet. 2017. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251–1258. Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. 2020. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555. Andrew M Dai and Quoc V Le. 2015. Semisupervised sequence learning. arXiv preprint arXiv:1511.01432. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N Dauphin. 2017. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122. Alec Go, Richa Bhayani, and Lei Huang. 2009. Twitter sentiment classification using distant supervision. CS224N project report, Stanford, 1(12):2009. Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. 2020. Accelerating large-scale inference with anisotropic vector quantization. In International Conference on Machine Learning. Sepp Hochreiter and J¨urgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735–1780. Lukasz Kaiser, Aidan N Gomez, and Francois Chollet. 2017. Depthwise separable convolutions for neural machine translation. arXiv preprint arXiv:1706.03059. Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukcuoglu. 2016. Neural machine translation in linear time. arXiv preprint arXiv:1610.10099. Najoung Kim and Tal Linzen. 2020. Cogs: A compositional generalization challenge based on semantic interpretation. arXiv preprint arXiv:2010.05465. Yoon Kim. 2014. Convolutional neural networks for sentence classification. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1746–1751, Doha, Qatar. Association for Computational Linguistics. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461. Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics. Qi Liu, Matt J Kusner, and Phil Blunsom. 2020. A survey on contextual embeddings. arXiv preprint arXiv:2003.07278. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. 4358 Andrew L Maas, Raymond E Daly, Peter T Pham, Dan Huang, Andrew Y Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies-volume 1, pages 142–150. Association for Computational Linguistics. Bryan McCann, James Bradbury, Caiming Xiong, and Richard Socher. 2017. Learned in translation: Contextualized word vectors. In Advances in neural information processing systems, pages 6294–6305. Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. arXiv preprint arXiv:1310.4546. Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532–1543. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. arXiv preprint arXiv:1802.05365. Xipeng Qiu, Tianxiang Sun, Yige Xu, Yunfan Shao, Ning Dai, and Xuanjing Huang. 2020. Pre-trained models for natural language processing: A survey. Science China Technological Sciences, pages 1–26. Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. arXiv preprint arXiv:1606.05250. Noam Shazeer, Youlong Cheng, Niki Parmar, Dustin Tran, Ashish Vaswani, Penporn Koanantakool, Peter Hawkins, HyoukJoong Lee, Mingsheng Hong, Cliff Young, et al. 2018. Mesh-tensorflow: Deep learning for supercomputers. In Advances in Neural Information Processing Systems, pages 10414–10423. Noam Shazeer and Mitchell Stern. 2018. Adafactor: Adaptive learning rates with sublinear memory cost. arXiv preprint arXiv:1804.04235. Laurent Sifre and St´ephane Mallat. 2014. Rigidmotion scattering for image classification. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. 2019. Mass: Masked sequence to sequence pre-training for language generation. arXiv preprint arXiv:1905.02450. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. arXiv preprint arXiv:1409.3215. Yi Tay, Dara Bahri, Donald Metzler, Da-Cheng Juan, Zhe Zhao, and Che Zheng. 2020a. Synthesizer: Rethinking self-attention in transformer models. arXiv preprint arXiv:2005.00743. Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long range arena : A benchmark for efficient transformers. In International Conference on Learning Representations. Yi Tay, Mostafa Dehghani, Dara Bahri, and Donald Metzler. 2020b. Efficient transformers: A survey. arXiv preprint arXiv:2009.06732. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. Wei Wang, Bin Bi, Ming Yan, Chen Wu, Zuyi Bao, Jiangnan Xia, Liwei Peng, and Luo Si. 2019. Structbert: Incorporating language structures into pretraining for deep language understanding. arXiv preprint arXiv:1908.04577. Adina Williams, Nikita Nangia, and Samuel R Bowman. 2017. A broad-coverage challenge corpus for sentence understanding through inference. arXiv preprint arXiv:1704.05426. Felix Wu, Angela Fan, Alexei Baevski, Yann N Dauphin, and Michael Auli. 2019. Pay less attention with lightweight and dynamic convolutions. arXiv preprint arXiv:1901.10430. Ellery Wulczyn, Nithum Thain, and Lucas Dixon. 2017. Ex machina: Personal attacks seen at scale. In Proceedings of the 26th International Conference on World Wide Web, WWW ’17, pages 1391–1399, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. arXiv preprint arXiv:1906.08237. 4359 Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification.
|
2021
|
335
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4360–4369 August 1–6, 2021. ©2021 Association for Computational Linguistics 4360 PairRE: Knowledge Graph Embeddings via Paired Relation Vectors Linlin Chao, Jianshan He, Taifeng Wang, Wei Chu AntGroup {chulin.cll,yebai.hjs}@antgroup.com {taifeng.wang,wei.chu}@alibaba-inc.com Abstract Distance based knowledge graph embedding methods show promising results on link prediction task, on which two topics have been widely studied: one is the ability to handle complex relations, such as N-to-1, 1-to-N and N-to-N, the other is to encode various relation patterns, such as symmetry/antisymmetry. However, the existing methods fail to solve these two problems at the same time, which leads to unsatisfactory results. To mitigate this problem, we propose PairRE, a model with paired vectors for each relation representation. The paired vectors enable an adaptive adjustment of the margin in loss function to fit for complex relations. Besides, PairRE is capable of encoding three important relation patterns, symmetry/antisymmetry, inverse and composition. Given simple constraints on relation representations, PairRE can encode subrelation further. Experiments on link prediction benchmarks demonstrate the proposed key capabilities of PairRE. Moreover, We set a new stateof-the-art on two knowledge graph datasets of the challenging Open Graph Benchmark. 1 Introduction Knowledge graphs store huge amounts of structured data in the form of triples, with projects such as WordNet (Miller, 1995), Freebase (Bollacker et al., 2008), YAGO (Suchanek et al., 2007) and DBpedia (Lehmann et al., 2015). They have gained widespread attraction from their successful use in tasks such as question answering (Bordes et al., 2014), semantic parsing (Berant et al., 2013), and named entity disambiguation (Zheng et al., 2012) and so on. Since most knowledge graphs suffer from incompleteness, predicting missing links between entities has been a fundamental problem. This problem is named as link prediction or knowledge graph completion. Knowledge graph embedding methods, which embed all entities and relations into a low dimensional space, have been proposed for this problem. Distance based embedding methods from TransE (Bordes et al., 2013) to the recent state-of-the-art RotatE (Sun et al., 2019) have shown substantial improvements on knowledge graph completion task. Two major problems have been widely studied. The first one refers to handling of 1-toN, N-to-1, and N-to-N complex relations (Bordes et al., 2013; Lin et al., 2015). In case of the 1-toN relations, given triples like (StevenSpielberg, DirectorOf, ?), distance based models should make all the corresponding entities about film name like Jaws and JurassicPark have closer distance to entity StevenSpielberg after transformation via relation DirectorOf. The difficulty is that all these entities should have different representations. Same issue happens in cases of N-to-N and N-to-1 relations. The latter is learning and inferring relation patterns according to observed triples, as the success of knowledge graph completion heavily relies on this ability (Bordes et al., 2013; Sun et al., 2019). There are various types of relation patterns: symmetry (e.g., IsSimilarTo), antisymmetry (e.g., FatherOf), inverse (e.g., PeopleBornHere and PlaceOfBirth), composition (e.g., my mother’s father is my grandpa) and so on. Previous methods solve these two problems separately. TransH (Wang et al., 2014), TransR (Lin et al., 2015), TransD (Ji et al., 2015) all focus on ways to solve complex relations. However, these methods can only encode symmetry/antisymmetry relations. The recent state-ofthe-art RotatE shows promising results to encode symmetry/antisymmetry, inverse and composition relations. However, complex relations remain challenging to predict. 4361 Here we present PairRE, an embedding method that is capable of encoding complex relations and multiple relation patterns simultaneously. The proposed model uses two vectors for relation representation. These vectors project the corresponding head and tail entities to Euclidean space, where the distance between the projected vectors is minimized. This provides three important benefits: • The paired relation representations enable an adaptive adjustment of the margin in loss function to fit for different complex relations; • Semantic connection among relation vectors can be well captured, which enables the model to encode three important relation patterns, symmetry/antisymmetry, inverse and composition; • Adding simple constraints on relation representations, PairRE can encode subrelation further. Besides, PairRE is a highly efficient model, which contributes to large scale datasets. We evaluate PairRE on six standard knowledge graph benchmarks. The experiment results show PairRE can achieve either state-of-the-art or highly competitive performance. Further analysis also proves that PairRE can better handle complex relations and encode symmetry/antisymmetry, inverse, composition and subrelation relations. 2 Background and Notation Given a knowledge graph that is represented as a list of fact triples, knowledge graph embedding methods define scoring function to measure the plausibility of these triples. We denote a triple by (h, r, t), where h represents head entity, r represents relation and t represents tail entity. The column vectors of entities and relations are represented by bold lower case letters, which belong to set E and R respectively. We denote the set of all triples that are true in a world as T . fr(h, t) represents the scoring function. We take the definition of complex relations from (Wang et al., 2014). For each relation r, we compute average number of tails per head (tphr) and average number of heads per tail (hptr). If tphr < 1.5 and hptr < 1.5, r is treated as 1-to-1; if tphr > 1.5 and hptr > 1.5, r is treated as a N-to-N; if tphr > 1.5 and hptr < 1.5, r is treated as 1-to-N. We focus on four important relation patterns, which includes: (1) Symmetry/antisymmetry. A relation r is symmetric if ∀e1, e2 ∈E, (e1, r, e2) ∈ T ⇐⇒(e2, r, e1) ∈T and is antisymmetric if (e1, r, e2) ∈T ⇒(e2, r, e1) /∈T ; (2) Inverse. If ∀e1, e2 ∈E, (e1, r1, e2) ∈T ⇐⇒(e2, r2, e1) ∈ T , then r1 and r2 are inverse relations; (3) Composition. If ∀e1, e2, e3 ∈E, (e1, r1, e2) ∈T ∧ (e2, r2, e3) ∈T ⇒(e1, r3, e3) ∈T , then r3 can be seen as the composition of r1 and r2; (4) Subrelation (Qu and Tang, 2019). If ∀e1, e2 ∈ E, (e1, r1, e2) ∈T ⇒(e1, r2, e2) ∈T , then r2 can be seen as a subrelation of r1. 3 Related Work Distance based models. Distance based models measure plausibility of fact triples as distance between entities. TransE interprets relation as a translation vector r so that entities can be connected, i.e., h + r ≈t. TransE is efficient, though cannot model symmetry relations and have difficulty in modeling complex relations. Several models are proposed for improving TransE to deal with complex relations, including TransH, TransR, TransD, TranSparse (Ji et al., 2016) and so on. All these methods project the entities to relation specific hyperplanes or spaces first, then translate projected entities with relation vectors. By projecting entities to different spaces or hyperplanes, the ability to handle complex relations is improved. However, with the added projection parameters, these models are unable to encode inverse and composition relations. The recent state-of-the-art, RotatE, which can encode symmetry/antisymmetry, inverse and composition relation patterns, utilizes rotation based translational method in a complex space. Although expressiveness for different relation patterns, complex relations remain challenging. GC-OTE (Tang et al., 2020) proposes to improve complex relation modeling ability of RotatE by introducing graph context to entity embedding. However, the calculation of graph contexts for head and tail entities is time consuming, which is inefficient for large scale knowledge graphs, e.g. ogbl-wikikg (Hu et al., 2020). Another related work is SE (Bordes et al., 2011), which utilizes two separate relation matrices to project head and tail entities. As pointed out by (Sun et al., 2019), this model is not able to encode symmetry/antisymmetry, inverse and composition 4362 Method Score Function Performance of complex relations Relation Patterns Sym Asym Inv Comp Sub TransE −||h + r −t|| Low TransR −||Mrh + r −Mrt|| High RotatE −||h ◦r −t|| Low PairRE −||h ◦rH −t ◦rT || High * Table 1: Comparison between PairRE and some distance based embedding methods. Sym, Asym, Inv, Comp and Sub are abbreviations for symmetry, antisymmetry, inverse and subrelation respectively. * means the model can have the specific capability with some constraints. relations. Table 1 shows comparison between our method and some representative distance based methods. As the table shows, our model is the most expressive one, with the ability to handle complex relations and encode four key relation patterns. Semantic matching models. Semantic matching models exploit similarity-based scoring functions, which can be divided into bilinear models and neural network based models. As the models have been developed, such as RESCAL (Nickel et al., 2011), DistMult (Yang et al., 2014), HolE (Nickel et al., 2016), ComplEx (Trouillon et al., 2016) and QuatE (Zhang et al., 2019), the key relation encoding abilities are enriched. However, all these models have the flaw in encoding composition relations (Sun et al., 2019). RESCAL, ComplEx and SimplE (Kazemi and Poole, 2018) are all proved to be fully expressive when embedding dimensions fulfill some requirements (Wang et al., 2018; Trouillon et al., 2016; Kazemi and Poole, 2018). The fully expressiveness means these models can express all the ground truth which exists in the data, including complex relations. However, these requirements are hardly fulfilled in practical use. It is proved by (Wang et al., 2018) that, to achieve complete expressiveness, the embedding dimension should be greater than N/32, where N is the number of entities in dataset. Neural networks based methods, e.g., convolution neural networks (Dettmers et al., 2018), graph convolutional networks (Schlichtkrull et al., 2018) show promising performances. However, they are difficult to analyze as they work as a black box. Encoding Subrelation. Existing methods encode subrelation by utilizing first order logic rules. One way is to augment knowledge graphs with grounding of rules, including subrelation rules (Guo et al., 2018; Qu and Tang, 2019). The other way is adding constraints on entity and relation representations, e.g., ComplEx-NNE-AER and SimplE+. The second way enriches the models’ expressiveness with relatively low cost. In this paper, we show that PairRE can encode subrelation with constraints on relation representations while keeping the ability to encode symmetry/antisymmetry, inverse and composition relations. 4 Methodology To overcome the problem of modeling 1-to-N/Nto-1/N-to-N complex relations and enrich the capabilities for different relation patterns, we propose a model with paired vectors for each relation. Given a training triple (h, r, t), our model learns vector embeddings of entities and relation in real space. Specially, PairRE takes relation embedding as paired vectors, which is represented as [rH, rT ]. rH and rT project head entity h and tail entity t to Euclidean space respectively. The projection operation is the Hadamard product1 between these two vectors. PairRE then computes distance of the two projected vectors as plausibility of the triple . We want that h ◦rH ≈t ◦rT when (h, r, t) holds, while h◦rH should be far away from t◦rT otherwise. In this paper, we take the L1-norm to measure the distance. In order to remove scaling freedoms, we also add constraint on embeddings similar to previous distance based models (Bordes et al., 2013; Wang et al., 2014; Lin et al., 2015). And the constraint is only added on entity embeddings. We want relation embeddings to capture semantic connection among relation vectors (e.g., PeopleBornHere and PlaceOfBirth) and complex characteristic (e.g., 1-N) easily and sufficiently. For entity embedding, the L2-norm is set to be 1. The scoring function is defined as follows: fr(h, t) = −||h ◦rH −t ◦rT ||, (1) where h, rH, rT , t ∈Rd and ||h||2 = ||t||2 = 1. The model parameters are, all the entities’ embed1Hadamard product means entry-wise product. 4363 (a) TransE (b) RotatE (c) PairRE Figure 1: Illustration of TransE, RotatE and PairRE when the entities stay in a plane. For PairRE, all entities are on the unit circle. The relation vectors project entities to different locations. dings, {ej}E j=1 and all the relations’ embeddings, {rj}R j=1. Illustration of the proposed PairRE is shown in Figure 1. Compared to TransE/RotatE, PairRE enables an entity to have distributed representations when involved in different relations. We also find the paired relation vectors enable an adaptive adjustment of the margin in loss function, which alleviates the modeling problem for complex relations. Let’s take a 1-to-N relation as an example. We set the embedding dimension to one and remove the constraint on entity embeddings for better illustration. Given triples (h, r, ?), where the correct tail entities belong to set S = {t1, t2, ..., tN}, PairRE predicts tail entities by letting ||h ◦rH −ti ◦rT || < γ, where γ is a fixed margin for distance based embedding models and ti ∈S. The value of ti should stay in the following range: ti ∈ ((h ◦rH −γ)/rT , (h ◦rH + γ)/rT ), if rT > 0, ((h ◦rH + γ)/rT , (h ◦rH −γ)/rT ), if rT < 0, (−∞, +∞), otherwise. The above analysis shows PairRE can adjust the value of rT to fit the entities in S. The larger the size of S, the smaller the absolute value rT . While models like TransE or RotatE have a fixed margin for all complex relation types. When the size of S is large enough, these models will be difficult to fit the data. For N-to-1 relations, PairRE can also adjust the value of rH adaptively to fit the data. Meanwhile, not adding a relation specific translational vector enables the model to encode several key relation patterns. We show these capabilities below. Proposition 1. PairRE can encode symmetry/antisymmetry relation pattern. Proof. If (e1, r1, e2) ∈T and (e2, r1, e1) ∈T , we have e1 ◦rH 1 = e2 ◦rT 1 ∧e2 ◦rH 1 = e1 ◦rT 1 ⇒rH 1 2 = rT 1 2 (2) if (e1, r1, e2) ∈T and (e2, r1, e1) /∈T , we have e1 ◦rH 1 = e2 ◦rT 1 ∧e2 ◦rH 1 ̸= e1 ◦rT 1 ⇒rH 1 2 ̸= rT 1 2 (3) Proposition 2. PairRE can encode inverse relation pattern. Proof. If (e1, r1, e2) ∈T and (e2, r2, e1) ∈T , we have e1 ◦rH 1 = e2 ◦rT 1 ∧e2 ◦rH 2 = e1 ◦rT 2 ⇒rH 1 ◦rH 2 = rT 1 ◦rT 2 (4) Proposition 3. PairRE can encode composition relation pattern. Proof. If (e1, r1, e2) ∈T , (e2, r2, e3) ∈T and (e1, r3, e3) ∈T , we have e1 ◦rH 1 = e2 ◦rT 1 ∧e2 ◦rH 2 = e3 ◦rT 2 ∧ e1 ◦rH 3 = e3 ◦rT 3 ⇒rT 1 ◦rT 2 ◦rH 3 = rH 1 ◦rH 2 ◦rT 3 (5) Moreover, with some constraint, PairRE can also encode subrelations. For a subrelation pair, ∀h, t ∈E : (h, r1, t) →(h, r2, t), it suggests triple (h, r2, t) should be always more plausible than triple (h, r1, t). In order to encode this pattern, PairRE should have the capability to enforce fr2(h, r2, t) ≥fr1(h, r1, t). 4364 Proposition 4. PairRE can encode subrelation relation pattern using inequality constraint. Proof. Assume a subrelation pair r1 and r2 that ∀h, t ∈E: (h, r1, t)→(h, r2, t). We impose the following constraints: rH 2,i rH 1,i = rT 2,i rT 1,i = αi, |αi| ≤1, (6) where α ∈Rd. Then we can get fr2(h, t) −fr1(h, t) = ||h ◦rH 1 −t ◦rT 1 || −||h ◦rH 2 −t ◦rT 2 || = ||h ◦rH 1 −t ◦rT 1 || −||α ◦(h ◦rH 1 −t ◦rT 1 )|| ≥0. (7) When the constraints are satisfied, PairRE forces triple (h, r2, t) to be more plausible than triple (h, r1, t). Optimization. To optimize the model, we utilize the self-adversarial negative sampling loss (Sun et al., 2019) as objective for training: L = −log σ(γ −fr(h, t)) − n X i=1 p(h ′ i, r, t ′ i) log σ(fr(h ′ i, t ′ i) −γ), (8) where γ is a fixed margin and σ is the sigmoid function. (h ′ i, r, t ′ i) is the ith negative triple and p(h ′ i, r, t ′ i) represents the weight of this negative sample. p(h ′ i, r, t ′ i) is defined as follows: p((h ′ i, r, t ′ i)|(h, r, t)) = exp fr(h ′ i, t ′ i) P j exp fr(h ′ j, t ′ j). (9) 5 Experimental results 5.1 Experimental setup We evaluate the proposed method on link prediction tasks. At first, we validate the ability to deal with complex relations and symmetry/antisymmetry, inverse and composition relations on four benchmarks. Then we validate our model on two subrelation specific benchmarks. Statistics of these benchmarks are shown in Table 2. ogbl-wikikg22 (Hu et al., 2020) is extracted from Wikidata knowledge base (Vrandeˇci´c and Kr¨otzsch, 2014). One of the main challenges for this dataset is complex relations. ogbl-biokg 2ogbl-wikikg2 fixes a bug in test/validation negative samples from original ogbl-wikikg. Dataset |R| |E| Train Valid Test ogbl-wikikg2 535 2,500k 16,109k 429k 598k ogbl-biokg 51 94k 4,763k 163k 163k FB15k 13k 15k 483k 50k 59k FB15k-237 237 15k 272k 18k 20k DB100k 470 100k 598k 50k 50k Sports 4 1039 1312 307 Table 2: Number of entities, relations, and observed triples in each split for the six benchmarks. (Hu et al., 2020) contains data from a large number of biomedical data repositories. One of the main challenges for this dataset is symmetry relations. FB15k (Bordes et al., 2013) contains triples from Freebase. The main relation patterns are inverse and symmetry/antisymmetry. FB15k-237 (Toutanova and Chen, 2015) is a subset of FB15k, with inverse relations removed. The main relation patterns are antisymmetry and composition. DB100k (Ding et al., 2018) is a subset of DBpedia. The main relation patterns are composition, inverse and subrelation. Sports (Wang et al., 2015) is a subset of NELL (Mitchell et al., 2018). The main relation patterns are antisymmetry and subrelation. Evaluation protocol. Following the state-ofthe-art methods, we measure the quality of the ranking of each test triple among all possible head entity and tail entity substitutions: (h ′, r , t) and (h, r, t ′), ∀h ′, ∀t ′ ∈E. Three evaluation metrics, including Mean Rank(MR), Mean Reciprocal Rank (MRR) and Hit ratio with cut-off values n = 1, 3, 10, are utilized. MR measures the average rank of all correct entities. MRR is the average inverse rank for correct entities with higher value representing better performance. Hit@n measures the percentage of correct entities in the top n predictions. The rankings of triples are computed after removing all the other observed triples that appear in either training, validation or test set. For experiments on ogbl-wikikg2 and ogbl-biokg, we follow the evaluation protocol of these two benchmarks (Hu et al., 2020). Implementation. We utilize the official implementations of benchmarks ogbl-wikikg2 and ogblbiokg (Hu et al., 2020) for the corresponding experiments3. Only the hypeparameter γ and embedding dimension are tuned. The other settings are kept the same with baselines. For the rest experiments, we implement our models based on the implementation of RotatE (Sun et al., 2019). All hypeparam3Our code is available at: https://github.com/alipay/KnowledgeGraphEmbeddingsViaPairedRelationVectors PairRE 4365 ogbl-wikikg2 ogbl-biokg Model #Dim Test MRR Valid MRR #Dim Test MRR Valid MRR TransE 100 0.2622 ± 0.0045 0.2465 ± 0.0020 DistMult 100 0.3447 ± 0.0082 0.3150 ± 0.0088 ComplEx 50 0.3804 ± 0.0022 0.3534 ± 0.0052 RotatE 50 0.2530 ± 0.0034 0.2250 ± 0.0035 PairRE 100 0.4849 ± 0.0029 0.4941 ± 0.0035 TransE 500† 0.4256 ± 0.0030 0.4272 ± 0.0030 2000 0.7452 ± 0.0004 0.7456 ± 0.0003 DistMult 500† 0.3729 ± 0.0045 0.3506 ± 0.0042 2000 0.8043 ± 0.0003 0.8055 ± 0.0003 ComplEx 250† 0.4027 ± 0.0027 0.3759 ± 0.0016 1000 0.8095 ± 0.0007 0.8105 ± 0.0001 RotatE 250† 0.4332 ± 0.0025 0.4353 ± 0.0028 1000 0.7989 ± 0.0004 0.7997 ± 0.0002 PairRE 200 0.5208 ± 0.0027 0.5423 ± 0.0020 2000 0.8164 ± 0.0005 0.8172 ± 0.0005 Table 3: Link prediction results on ogbl-wikikg2 and ogbl-biokg. Best results are in bold. All the results except PairRE are from (Hu et al., 2020). † requires a GPU with 48GB memory. PairRE runs on a GPU with 16GB memory. FB15k FB15k-237 Model MR MRR Hit@10 Hit@3 Hit@1 MR MRR Hit@10 Hit@3 Hit@1 TransE† 0.463 0.749 0.578 0.297 357 0.294 0.465 DistMult3 42 0.798 0.893 254 0.241 0.419 0.263 0.155 HolE 0.524 0.739 0.759 0.599 ConvE 51 0.657 0.831 0.723 0.558 244 0.325 0.501 0.356 0.237 ComplEx 0.692 0.840 0.759 0.599 339 0.247 0.428 0.275 0.158 SimplE 0.727 0.838 0.773 0.660 RotatE 40 0.797 0.884 0.830 0.746 177 0.338 0.533 0.375 0.241 SeeK 0.825 0.886 0.841 0.792 OTE 0.351 0.537 0.388 0.258 GC-OTE 0.361 0.550 0.396 0.267 PairRE 37.7 0.811 0.896 0.845 0.765 160 0.351 0.544 0.387 0.256 ±0.4979 ±0.00077 ±0.00071 ±0.0011 ±0.0012 ±0.9949 ±0.00066 ±0.00093 ±0.00079 ±0.00097 Table 4: Link prediction results on FB15k and FB15k-237. Results of [†] are taken from (Nickel et al., 2016); Results of [3] are taken from (Kadlec et al., 2017). Other results are taken from the corresponding papers. GC-OTE adds graph context to OTE (Tang et al., 2020). Subrelation (h, CoachesTeam, t) →(h, PersonBelongsToOrganization, t) (h, AthleteLedSportsTeam, t) →(h, AtheletePlaysForTeam, t) Table 5: The added subrelation rules for Sports dataset. Model MRR hit@1 SimplE 0.230 0.184 SimplE+ 0.404 0.349 PairRE 0.468 ± 0.003 0.416 ± 0.005 PairRE+Rule 0.475 ± 0.003 0.432 ± 0.004 Table 6: Link prediction results on Sports dataset. Other results are taken from (Fatemi et al., 2019). eters except γ and embedding dimension are kept the same with RotatE. 5.2 Main results Comparisons for ogbl-wikikg2 and ogbl-biokg are shown in Table 3. On these two large scale datasets, PairRE achieves state-of-the-art performances. For ogbl-wikikg2 dataset, PairRE performs best on both limited embedding dimension and increased embedding dimension. With the same number of parameters to ComplEx (dimension 100), PairRE Model MRR Hit@10 Hit@3 Hit@1 TransE 0.111 0.270 0.164 0.016 DistMult 0.233 0.448 0.301 0.115 HolE 0.260 0.411 0.309 0.182 ComplEx 0.242 0.440 0.312 0.126 SeeK 0.338 0.467 0.370 0.268 ComplEx-NNE 0.298 0.426 0.330 0.229 ComplEx-NNE-AER 0.306 0.418 0.334 0.244 PairRE 0.412 0.600 0.472 0.309 ±0.0015 ±0.0006 ±0.0015 ±0.0027 PairRE+rule 0.419 0.599 0.475 0.321 ±0.0010 ±0.0008 ±0.0008 ±0.0016 Table 7: Link prediction results on DB100k. All the results are taken from the corresponding papers. improves Test MRR close to 10%. With increased dimension, all models are able to achieve higher MRR on validation and test sets. Due to the limitation of hardware, we only increase embedding dimension to 200 for PairRE. PairRE also outperforms all baselines and improves Test MRR 8.7%. Based on performances of baselines, the performance of PairRE may be improved further if embedding dimension is increased to 500. Under the same experiment setting and the same number of parameters, PairRE also outperforms all baselines on ogbl-biokg dataset. It improves Test 4366 MRR by 0.69%, which proves the superior ability to encode symmetry relations. Comparisons for FB15k and FB15k-237 datasets are shown in Table 4. Since our model shares the same hyper-parameter settings and implementation with RotatE, comparing with this state-of-the-art model is fair to show the advantage and disadvantage of the proposed model. Besides, the comparisons also include several leading methods, such as TransE (Bordes et al., 2013), DistMult (Yang et al., 2014), HolE (Nickel et al., 2016), ConvE (Dettmers et al., 2018), ComplEx (Trouillon et al., 2016), SimplE (Kazemi and Poole, 2018), SeeK (Xu et al., 2020) and OTE (Tang et al., 2020). Compared with RotatE, PairRE shows clear improvements on FB15k and FB15k-237 for all evaluation metrics. For MRR metric, the improvements are 1.4% and 1.3% respectively. Compared with the other leading methods, PairRE also shows highly competitive performances. All these comparisons prove the effectiveness of PairRE to encode inverse and composition relations. 5.3 Further experiments on subrelation We further compare our method with two of the leading methods ComplEx-NNE-AER and SimplE+, which focus on encoding subrelation. These two methods add subrelation rules to semantic matching models. We utilize these rules as constraints on relation representations for PairRE. Two ways are validated. We first test the performance of weight tying for subrelation rules on Sports dataset. The rules (r1−→r2) are added as follows: rH 2 = rH 1 ◦cosine(θ), rT 2 = rT 1 ◦cosine(θ), (10) where θ ∈Rd. The added rules are shown in Table 5. The experiments results in Table 6 show effectiveness of the proposed method. Weight tying on relation representation is a way to incorporate hard rules. The soft rules can also be incorporated into PairRE by approximate entailment constraints on relation representations. In this section, we add the same rules from ComplExNNE-AER, which includes subrelation and inverse rules. We denote by r1 λ −→r2 the approximate entailment between relations r1 and r2, with confidence level λ. The objective for training is then changed to: Lrule = L + µ X τsubrelation λ1T (rH 1 ◦rT 2 −rT 1 ◦rH 2 )2 + µ X τinverse λ1T (rH 1 ◦rH 2 −rT 1 ◦rT 2 )2, (11) where L is calculated from Equation 8, µ is loss weight for added constraints, τsubrelation and τinverse are the sets of subrelation rules and inverse rules respectively. Following (Ding et al., 2018), we take the corresponding two relations from subrelation rules as equivalence. Because τsubrelation contains both rule r1→r2 and rule r2→r1. We validate our method on DB100k dataset. The results are shown in Table 7. We can see PairRE outperforms the recent state-of-the-art SeeK and ComplEx based models with large margins on all evaluation metrics. With added constraints, the performance of PairRE is improved further. The improvements for the added rules are 0.7%, 1.2% for MRR and Hit@1 metrics respectively. 5.4 Model analysis Analysis on complex relations We analyze the performances of PairRE for complex relations. The results of PairRE on different relation categories on FB15k and ogbl-wikikg2 are summarized into Table 8. We can see PairRE performs quite well on N-to-N and N-to-1 relations. It has a significant lead over baselines. We also notice that performance of 1-to-N relations on ogblwikikg2 dataset is not as strong as the other relation categories. One of the reasons is that only 2.2% of test triples belong to the 1-to-N relation category. In order to further test the performance of paired relation vectors, we change the relation vector in RotatE to paired vectors. In the modified RotatE model, both head and tail entities are rotated with different angles based on the paired Figure 2: Performance comparison between RotatE and RotatE+PairRelation on ogbl-wikikg2 dataset. 4367 FB15k(Hits@10) ogbl-wikikg2(Hits@10) Model 1-to-1 1-to-N N-to-1 N-to-N 1-to-1 1-to-N N-to-1 N-to-N KGE2E KL(He et al., 2015) 0.925 0.813 0.802 0.715 TransE 0.887 0.822 0.766 0.895 0.074 0.063 0.400 0.220 ComplEx 0.939 0.896 0.822 0.902 0.394 0.278 0.483 0.504 RotatE 0.923 0.840 0.782 0.908 0.164 0.144 0.431 0.261 PairRE 0.785 0.899 0.872 0.940 0.262 0.270 0.594 0.587 Table 8: Experimental results on FB15k and ogbl-wikikg2 by relation category. Results on FB15k are taken from RotatE (Sun et al., 2019). The embedding dimensions for models on ogbl-wikikg2 are same to the experiments in Table 3, which is 100 for real space models and 50 for complex value based models. (a) r1 (b) rH 1 2 −rT 1 2 (c) r2 (d) rH 2 2 −rT 2 2 (e) r3 (f) rH 2 ◦rH 3 −rT 2 ◦rT 3 (g) r4 (h) r5 (i) r6 (j) rH 4 ◦rH 5 ◦rT 6 −rT 4 ◦rT 5 ◦rH 6 Figure 3: Histograms of relation embeddings for different relation patterns. r1 is relation spouse. r2 is relation /broadcast/tv station/owner. r3 is relation /broadcast/tv station owner/tv stations. r4 is relation /location/administrative division/capital/location/administrative divisioncapital relationship/capital. r5 is relation /location/hud county place/place. r6 is relation base/areas/schema/administrative area/capital. relation vectors. This model can also be seen as complex value based PairRE. We name this model as RotatE+PairRelation. The experiment results are shown in Figure 2. With the same embedding dimension (50 in the experiments), RotatE+PairRelation improves performance of RotatE with 20.8%, 27.5%, 14.4% and 39.1% on 1-to-1, 1-to-N, N-to-1 and N-to-N relation categories respectively. These significant improvements prove the superior ability of paired relation vectors to handle complex relations. Analysis on relation patterns To further verify the learned relation patterns, we visualize some examples. Histograms of the learned relation embeddings are shown in Figure 3 . Symmetry/AntiSymmetry. Figure 3a shows a symmetry relation spouse from DB100k. The embedding dimension is 500. For PairRE, symmetry relation pattern can be encoded when embedding r satisfies rH2 = rT 2. Figure 3b shows most of the paired elements in rH and rT have the same absolute value. Figure 3c shows a antisymmetry relation tv station owner, where most of the paired 4368 elements do not have the same absolute value as shown in Figure 3d. Inverse. Figure 3c and Figure 3e show an example of inverse relations from FB15k. As the histogram in Figure 3f shows these two inverse relations tv station owner (r2) and tv station owner tv stations (r3) close to satisfy rH 3 ◦rH 2 = rT 3 ◦rT 2 . Composition. Figures 3g, 3h, 3i show an example of composition relation pattern from FB15k, where the third relation r6 can be seen as the composition of the first relation r4 and the second relation r5. As Figure 3j shows these three relations close to satisfy rH 4 ◦rH 5 ◦rT 6 −rT 4 ◦rT 5 ◦rH 6 . 6 Conclusion To better handle complex relations and tackle more relation patterns, we proposed PairRE, which represents each relation with paired vectors. With a slight increase in complexity, PairRE can solve the aforementioned two problems efficiently. Beyond the symmetry/antisymmetry, inverse and composition relations, PairRE can further encode subrelation with simple constraint on relation representations. On large scale benchmark ogbl-wikikg2 an ogbl-biokg, PairRE outperforms all the state-of-theart baselines. Experiments on other well designed benchmarks also demonstrate the effectiveness of the focused key abilities. References Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. 2013. Semantic parsing on Freebase from question-answer pairs. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1533–1544, Seattle, Washington, USA. Association for Computational Linguistics. Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: a collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1247–1250. AcM. Antoine Bordes, Nicolas Usunier, Alberto GarciaDuran, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In Advances in neural information processing systems, pages 2787–2795. Antoine Bordes, Jason Weston, Ronan Collobert, and Yoshua Bengio. 2011. Learning structured embeddings of knowledge bases. In Conference on artificial intelligence, CONF. Antoine Bordes, Jason Weston, and Nicolas Usunier. 2014. Open question answering with weakly supervised embedding models. In Joint European conference on machine learning and knowledge discovery in databases, pages 165–180. Springer. Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Thirty-Second AAAI Conference on Artificial Intelligence. Boyang Ding, Quan Wang, Bin Wang, and Li Guo. 2018. Improving knowledge graph embedding using simple constraints. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 110–121, Melbourne, Australia. Association for Computational Linguistics. Bahare Fatemi, Siamak Ravanbakhsh, and David Poole. 2019. Improved knowledge graph embedding using background taxonomic information. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3526–3533. Shu Guo, Quan Wang, Lihong Wang, Bin Wang, and Li Guo. 2018. Knowledge graph embedding with iterative guidance from soft rules. In Thirty-Second AAAI Conference on Artificial Intelligence. Shizhu He, Kang Liu, Guoliang Ji, and Jun Zhao. 2015. Learning to represent knowledge graphs with gaussian embedding. In Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, pages 623–632. Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs. arXiv preprint arXiv:2005.00687. Guoliang Ji, Shizhu He, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Knowledge graph embedding via dynamic mapping matrix. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 687–696. Guoliang Ji, Kang Liu, Shizhu He, and Jun Zhao. 2016. Knowledge graph completion with adaptive sparse transfer matrix. In AAAI, volume 16, pages 985– 991. Rudolf Kadlec, Ondrej Bajgar, and Jan Kleindienst. 2017. Knowledge base completion: Baselines strike back. In Proceedings of the 2nd Workshop on Representation Learning for NLP, pages 69–74, Vancouver, Canada. Association for Computational Linguistics. Seyed Mehran Kazemi and David Poole. 2018. Simple embedding for link prediction in knowledge graphs. In Advances in Neural Information Processing Systems, pages 4284–4295. 4369 Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, S¨oren Auer, et al. 2015. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia. Semantic web, 6(2):167–195. Yankai Lin, Zhiyuan Liu, Maosong Sun, Yang Liu, and Xuan Zhu. 2015. Learning entity and relation embeddings for knowledge graph completion. In Twenty-ninth AAAI conference on artificial intelligence. George A Miller. 1995. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39– 41. Tom Mitchell, William Cohen, Estevam Hruschka, Partha Talukdar, Bishan Yang, Justin Betteridge, Andrew Carlson, Bhanava Dalvi, Matt Gardner, Bryan Kisiel, et al. 2018. Never-ending learning. Communications of the ACM, 61(5):103–115. Maximilian Nickel, Lorenzo Rosasco, and Tomaso Poggio. 2016. Holographic embeddings of knowledge graphs. In Thirtieth Aaai conference on artificial intelligence. Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. 2011. A three-way model for collective learning on multi-relational data. In ICML, volume 11, pages 809–816. Meng Qu and Jian Tang. 2019. Probabilistic logic neural networks for reasoning. In Advances in Neural Information Processing Systems, pages 7710–7720. Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. In European Semantic Web Conference, pages 593–607. Springer. Fabian M Suchanek, Gjergji Kasneci, and Gerhard Weikum. 2007. Yago: a core of semantic knowledge. In Proceedings of the 16th international conference on World Wide Web, pages 697–706. ACM. Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations. Yun Tang, Jing Huang, Guangtao Wang, Xiaodong He, and Bowen Zhou. 2020. Orthogonal relation transforms with graph context modeling for knowledge graph embedding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2713–2722, Online. Association for Computational Linguistics. Kristina Toutanova and Danqi Chen. 2015. Observed versus latent features for knowledge base and text inference. In Proceedings of the 3rd Workshop on Continuous Vector Space Models and their Compositionality, pages 57–66. Th´eo Trouillon, Johannes Welbl, Sebastian Riedel, ´Eric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In International Conference on Machine Learning, pages 2071–2080. Denny Vrandeˇci´c and Markus Kr¨otzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78–85. Quan Wang, Bin Wang, and Li Guo. 2015. Knowledge base completion using embeddings and rules. In Twenty-Fourth International Joint Conference on Artificial Intelligence. Yanjie Wang, Rainer Gemulla, and Hui Li. 2018. On multi-relational link prediction with bilinear models. Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. In Twenty-Eighth AAAI conference on artificial intelligence. Wentao Xu, Shun Zheng, Liang He, Bin Shao, Jian Yin, and Tie-Yan Liu. 2020. SEEK: Segmented embedding of knowledge graphs. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3888–3897, Online. Association for Computational Linguistics. Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2014. Embedding entities and relations for learning and inference in knowledge bases. arXiv preprint arXiv:1412.6575. Shuai Zhang, Yi Tay, Lina Yao, and Qi Liu. 2019. Quaternion knowledge graph embedding. arXiv preprint arXiv:1904.10281. Zhicheng Zheng, Xiance Si, Fangtao Li, Edward Y Chang, and Xiaoyan Zhu. 2012. Entity disambiguation with freebase. In Proceedings of the The 2012 IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent TechnologyVolume 01, pages 82–89.
|
2021
|
336
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4370–4379 August 1–6, 2021. ©2021 Association for Computational Linguistics 4370 Hierarchy-aware Label Semantics Matching Network for Hierarchical Text Classification Haibin Chen, Qianli Ma*, Zhenxi Lin, Jiangyue Yan School of Computer Science and Engineering, South China University of Technology, Guangzhou, China haibin [email protected] [email protected]∗ Abstract Hierarchical text classification is an important yet challenging task due to the complex structure of the label hierarchy. Existing methods ignore the semantic relationship between text and labels, so they cannot make full use of the hierarchical information. To this end, we formulate the text-label semantics relationship as a semantic matching problem and thus propose a hierarchy-aware label semantics matching network (HiMatch). First, we project text semantics and label semantics into a joint embedding space. We then introduce a joint embedding loss and a matching learning loss to model the matching relationship between the text semantics and the label semantics. Our model captures the text-label semantics matching relationship among coarse-grained labels and fine-grained labels in a hierarchy-aware manner. The experimental results on various benchmark datasets verify that our model achieves state-of-the-art results. 1 Introduction Hierarchical text classification (HTC) is widely used in Natural Language Processing (NLP), such as news categorization (Lewis et al., 2004) and scientific paper classification (Kowsari et al., 2017). HTC is a particular multi-label text classification problem, which introduces hierarchies to organize label structure. As depicted in Figure 1, HTC models predict multiple labels in a given label hierarchy, which generally construct one or multiple paths from coarse-grained labels to fine-grained labels in a top-down manner (Aixin Sun and Ee-Peng Lim, 2001). Generally speaking, fine-grained labels are the most appropriate labels for describing the input text. Coarse-grained labels are generally the parent nodes of coarse- or fine-grained labels, expressing a more general concept. The key challenges of ∗*Corresponding author HTC are to model the large-scale, imbalanced, and structured label hierarchy (Mao et al., 2019). Root Economics Debt Revenue Society Coarse-grained Labels Fine-grained Labels Input Text: "Global debt is set to reach $200 trillion ..." Label Hierarchy Figure 1: An hierarchical text classification example tagged with labels Economics and Debt from coarsegrained label to fine-grained label. Existing work in HTC has introduced various methods to use hierarchical information in a holistic way. To capture the holistic label correlation features, some researchers proposed a hierarchyaware global model to exploit the prior probability of label dependencies through Graph Convolution Networks (GCN) and TreeLSTM (Zhou et al., 2020). Some researchers also introduced more label correlation features such as label semantic similarity and label co-occurrence (Lu et al., 2020). They followed the traditional way to transform HTC into multiple binary classifiers for every label (F¨urnkranz et al., 2008). However, they ignored the interaction between text semantics and label semantics (F¨urnkranz et al., 2008; Wang et al., 2019), which is highly useful for classification (Chen et al., 2020). Hence, their models may not be sufficient to model complex label dependencies and provide comparable text-label classification scores (Wang et al., 2019). A natural strategy for modeling the interaction between text semantics and label semantics is to introduce a text-label joint embedding by label attention (Xiao et al., 2019) or autoencoders (Yeh et al., 2017). Label attention-based methods adopted a 4371 self-attention mechanism to identify label-specific information (Xiao et al., 2019). Autoencoder-based methods extended the vanilla Canonical Correlated Autoencoder (Yeh et al., 2017) to a ranking-based autoencoder architecture to produce comparable text-label scores (Wang et al., 2019). However, these methods assume all the labels are independent without fully considering the correlation between coarse-grained labels and fine-grained labels, which cannot be simply transferred to HTC models (Zhou et al., 2020). In this paper, we formulate the interaction between text and label as a semantic matching problem and propose a Hierarchy-aware Label Semantics Matching Network (HiMatch). The principal idea is that the text representations should be semantically similar to the target label representations (especially fine-grained labels), while they should be semantically far away from the incorrect label representations. First, we adopt a text encoder and a label encoder (shown in Figure 2) to extract textual semantics and label semantics, respectively. Second, inspired by the methods of learning common embeddings (Wang et al., 2019), we project both textual semantics and label semantics into a text-label joint embedding space where correlations between text and labels are exploited. In this joint embedding space, we introduce a joint embedding loss between text semantics and target label semantics to learn a text-label joint embedding. After that, we apply a matching learning loss to capture text-label matching relationships in a hierarchy-aware manner. In this way, the finegrained labels are semantically closest to the text semantics, followed by the coarse-grained labels, while the incorrect labels should be semantically far away from the text semantics. Hence, we propose a hierarchy-aware matching learning method to capture different matching relationships through different penalty margins on semantic distances. Finally, we employ the textual representations guided by the joint embedding loss and matching learning loss to perform the hierarchical text classification. The major contributions of this paper are: 1. By considering the text-label semantics matching relationship, we are the first to formulate HTC as a semantic matching problem rather than merely multiple binary classification tasks. 2. We propose a hierarchy-aware label semantics matching network (HiMatch), in which we introduce a joint embedding loss and a matching learning loss to learn the text-label semantics matching relationship in a hierarchy-aware manner. 3. Extensive experiments (with/without BERT) on various datasets show that our model achieves state-of-the-art results. 2 Related Work 2.1 Hierarchical Text Classification Hierarchical text classification is a particular multilabel text classification problem, where the classification results are assigned to one or more nodes of a taxonomic hierarchy. Existing state-of-the-art methods focus on encoding hierarchy constraint in a global view such as directed graph and tree structure. Zhou et al. (2020) proposed a hierarchyaware global model to exploit the prior probability of label dependencies. Lu et al. (2020) introduced three kinds of label knowledge graphs, i.e., taxonomy graph, semantic similarity graph, and cooccurrence graph to benefit hierarchical text classification. They regarded hierarchical text classification as multiple binary classification tasks (F¨urnkranz et al., 2008). The limitation is that these models did not consider the interaction of label semantics and text semantics. Therefore, they failed to capture complex label dependencies and can not provide comparable text-label classification scores (Wang et al., 2019), which leads to restricted performance (Chen et al., 2020). Hence, it is crucial to exploit the relationship between text and label semantics, and help the model distinguish target labels from incorrect labels in a comparable and hierarchy-aware manner. We perform matching learning in a joint embedding of text and label to solve these problems in this work. 2.2 Exploit Joint Embedding of Text and Label To determine the correlation between text and label, researchers proposed various methods to exploit a text-label joint embedding such as (Xiao et al., 2019) or Autoencoder (Yeh et al., 2017). In the field of multi-label text classification, Xiao et al. (2019) proposed a Label-Specific Attention Network (LSAN) to learn a text-label joint embedding by label semantic and document semantic. Wang et al. (2019) extended vanilla Canonical Correlated AutoEncoder (Yeh et al., 2017) to a ranking-based autoencoder architecture to produce comparable label scores. However, they did not fully consider label semantics and holistic label correlation 4372 Global debt is set to reach... Input text Text Encoder Classification Layer Classification Loss Economics(Coarse -grained Target Label) Debt(Fine-grained Target Label) Revenue(Incorrect Sibling Label) Society(Other Incorrect Label) Matching Learning Loss Joint Embedding Loss CNN+pooling Bi-GRU Label Encoder Graph Convolution Text Representations Label Representations Classification Learning MLP Root Economics Debt Revenue Society MLP Label Set Text Representations Label Representations Feature Propagation Text-label Joint Embedding d1 d2 d3 d4 d1 < d2 < d3 < d4 Minimize(Text, Target Labels) Maximize(Text, Incorrect Labels) Joint Embedding Learning Hierarchy-aware Matching Learning Coarse-grained Labels Fine-grained Labels Figure 2: The overall architecture of the proposed model. Firstly, the text encoder and label encoder extract the text semantics and label semantics, respectively. Then text semantics and label semantics are projected into a joint embedding space. Joint embedding loss encourages the text semantics to be similar to the target label semantics. By introducing matching learning loss, fine-grained labels semantics (Debt) is semantically closest to the text semantics, followed by coarse-grained labels (Economics), while other incorrect labels semantics is semantically far away from text semantics (Revenue, Society). The relative order is d1 < d2 < d3 < d4, where d represents the metric distances in joint embedding. among fine-grained labels, coarse-grained labels, and incorrect labels. In addition, we can not simply transfer these multi-label classification methods to HTC due to the constraint of hierarchy (Zhou et al., 2020). 3 Proposed Method In this section, we will describe the details about our Hierarchy-aware Label Semantics Matching Network. Figure 2 shows the overall architecture of our proposed model. 3.1 Text Encoder In the HTC task, given the input sequence xseq = {x1, ..., xn}, the model will predict the label y = {y1, ..., yk} where n is the number of words and k is the number of label sets. The label with a probability higher than a fixed threshold (0.5) will be regarded as the prediction result. The sequence of token embeddings is firstly fed into a bidirectional GRU layer to extract contextual feature H = {h1, ..., hn}. Then, CNN layers with top-k max-pooling are adopted for generating key n-gram features T ∈Rk×dcnn where dcnn indicates the output dimension of the CNN layer. Following the previous work (Zhou et al., 2020), we further introduce a hierarchy-aware text feature propagation module to encode label hierarchy information. We define a hierarchy label structure as a directed graph G = Vt, ←− E , −→ E , where Vt indicates the set of hierarchy structure nodes. ←− E are built from the top-down hierarchy paths representing the prior statistical probability from parent nodes to children nodes. −→ E are built from the bottom-up hierarchy paths representing the connection relationship from children nodes to parent nodes. The feature size of graph adjacency matrix ←E and →E is ∈Rk×k, where k is the number of label sets. Text feature propagation module firstly projects text features T to node inputs Vt by a linear transformation Wproj ∈Rk×dcnn×dt, where dt represents the hierarchy structure node dimension from text feature. Then a Graph Convolution Network (GCN) is adopted to explicitly combine text semantics with prior hierarchical information ←− E and −→ E : St = σ ←− E · Vt · Wg1 + −→ E · Vt · Wg2 (1) where σ is the activation function ReLU. Wg1, Wg2 ∈Rdt×dt are the weight matrix of GCN. St is the text representation aware of prior hierarchy paths. 3.2 Label Encoder In the HTC task, the hierarchical label structure can be regarded as a directed graph G = Vl, ←− E , −→ E , 4373 where Vl indicates the set of hierarchy structure nodes with label representation. The graph G in label encoder shares the same structure ←− E and −→ E with the graph in text encoder. Given the total label set y = {y1, ..., yk} as input, we create label embeddings Vl ∈Rdl by averaging of pre-trained label embeddings first. Then GCN could be utilized as label encoder: Sl = σ ←− E · Vl · Wg3 + −→ E · Vl · Wg4 (2) where σ is the activation function ReLU. Wg3, Wg4 ∈Rdl×dl are the weight matrix of GCN. Sl is the label representation aware of prior hierarchy paths. It must be noted that the weight matrix and input representation of the label encoder are different with those in the text encoder. 3.3 Label Semantics Matching 3.3.1 Joint Embedding Learning In this section, we will introduce the methods of learning a text-label joint embedding and hierarchyaware matching relationship. For joint embedding learning, firstly, we project text semantics St and label semantics Sl into a common latent space as follows: Φt = FFNt (St) , (3) Φl = FFNl (Sl) (4) where FFNt and FFNl are independent two-layer feedforward neural networks. Φt, Φl ∈Rdϕ represent text semantics and label semantics in joint embedding space, respectively. dϕ indicates the dimension of joint embedding. In order to align the two independent semantic representations in the latent space, we employ the mean squared loss between text semantics and target labels semantics: Ljoint = X p∈P(y) Φt −Φp l 2 2 (5) where P(y) is target label sets. Ljoint aims to minimize the common embedding loss between input text and target labels. 3.3.2 Hierarchy-aware Matching Learning Based on the text-label joint embedding loss, the model only captures the correlations between text semantics and target labels semantics, while correlations among different granular labels are ignored. Economics (Coarse-grained Target Label) Debt (Fine-grained Target Label) Revenue (Incorrect Sibling Label) Society (Other Incorrect Label) Large Semantic Distance Small Semantic Distance Root d4 d3 d1 Text: "Global Debt is set to..." Matching Matching Matching Matching d2 Large Penalty Margin Small Penalty Margin γ4 γ3 γ1 γ2 Figure 3: Illustration of hierarchy-aware margin. Target labels are colored yellow. Each colored line represent the matching operation between text and different labels. The two vertical axes for semantic matching distance and penalty margin are on the right. The semantic matching distance can be sorted by the order of d1 (fine-grained target labels) < d2 (coarse-grained target labels) < d3 (incorrect sibling labels) < d4 (other incorrect labels). We introduce penalty margins γ to model the relative matching relationships. In the HTC task, it is expected that the matching relationship between text semantics and fine-grained labels should be the closest, followed by coarsegrained labels. Text semantics and incorrect labels semantics should not be related. Insight of these, we propose a hierarchy-aware matching loss Lmatch to incorporate the correlations among text semantics and different labels semantics. Lmatch aims to penalize the small semantic distance between text semantics and incorrect labels semantics with a margin γ: Lmatch = max 0, D Φt, Φp l −D (Φt, Φn l ) + γ (6) where Φp l represents target labels semantics and Φn l represents incorrect labels semantics. We use L2-normalized euclidean distance for metric D and γ is a margin constant for margin-based triplet loss. We take the average of all the losses between every label pairs as the margin loss. Hierarchy-aware Margin Due to the large label sets in the HTC task, it is time-consuming to calculate every label’s matching loss. Therefore, we propose hierarchy-aware sampling to alleviate the problem. Specifically, we sample all parent labels (coarse-grained labels), one sibling label, and one random incorrect label for every fine-grained label to obtain its negative label sets n ∈N(y). It is also unreasonable to assign the same margin for different label pairs since the label semantics similarity is quite different in a large structured label hierarchy. Our basic idea is that the semantics relationship should be closer if two labels are closer 4374 in the hierarchical structure. Firstly, the text semantics should match fine-grained labels the most, which is exploited in joint embedding learning. Then we regard the pair with the smallest semantic distance (d1) as a positive pair and regard other textlabel matching pairs as negative pairs. As depicted in the schema figure 3, compared with the positive pair, the semantics matching distance between text and coarse-grained target labels (d2) should be larger. The incorrect sibling labels have a certain semantic relationship with the target labels. Hence, the semantics matching distance between text and the incorrect sibling labels of fine-grained labels (d3) should be further larger, while the semantics matching distance between text and other incorrect labels (d4) should be the largest. We introduce hierarchy-aware penalty margins γ1, γ2, γ3, γ4 to model the comparable relationship. The penalty margin is smaller if we expect the semantic matching distance to be smaller. We neglect γ1 because the matching relationships between text semantics and fine-grained labels are exploited in joint embedding learning. γ2, γ3, γ4 are penalty margins compared with the matching relationships between text semantics and fine-grained labels semantics. We introduce two hyperparameters α, β to measure different matching relationships of γ: γ2 = αγ; γ3 = βγ; γ4 = γ (7) where 0 < α < β < 1. The proposed loss captures the relative semantics similarity rankings among target labels and incorrect labels in a hierarchyaware manner. 3.4 Classification Learning and Objective Function We find that it is easier to overfit for classification learning if we perform classification learning in the text-label joint embedding directly. Hence, we use the text semantics representation St guided by joint embedding loss and matching learning loss to perform classification learning. St is fed into a fully connected layer to get the label probability ˆy for prediction. The overall objective function includes a crossentropy category loss, joint embedding loss and hierarchy-aware matching loss: L = Lcls(y, ˆy) + λ1Ljoint + λ2Lmatch (8) where y and ˆy are the ground-truth label and output probability, respectively. λ1, λ2 are the hyperparameters for balancing the joint embedding loss and Dataset |L| Depth Avg(|Li|) Train V al Test RCV1-V2 103 4 3.24 20833 2316 781265 WOS 141 2 2 30070 7518 9397 EURLEX-57K 4271 5 5 45000 6000 6000 Table 1: Statistics of three datasets for hierarchical multi-label text classification. |L|: Number of target classes. Depth: Maximum level of hierarchy. Avg(|Li|): Average Number of classes per sample. Train/V al/Test: Size of train/validation/test set. matching learning loss. We minimize the above function by gradient descent during training. 4 Experiment 4.1 Experiment Setup Datasets To evaluate the effectiveness of our model, we conduct experiments on three widelystudied datasets for hierarchical multi-label text classification. Statistics of these datasets are listed in Table 1. RCV1-V2 (Lewis et al., 2004) is a news categorization corpora, and WOS (Kowsari et al., 2017) includes abstracts of published papers from Web of Science. EURLEX57K is a large hierarchical multi-label text classification (LMTC) dataset that contains 57k English EU legislative documents, and is tagged with about 4.3k labels from the European Vocabulary (Chalkidis et al., 2019). The label sets are split into zero-shot labels, few-shot labels, and frequent labels. Few-shot labels are labels whose frequencies in the training set are less than or equal to 50. Frequent labels are labels whose frequencies in the training set are more than 50. The label setting is the same as previous work (Lu et al., 2020). In EURLEX57K, the corpora are only tagged with fine-grained labels, and the parent labels of fine-grained labels are not tagged as the target labels. Evaluation Metric On RCV1-V2 and WOS datasets, we measure the experimental results by Micro-F1 and Macro-F1. Micro-F1 takes the overall precision and recall of all the instances into account, while Macro-F1 equals the average F1score of labels. We report the results of two ranking metrics on large hierarchical multi-label text classification dataset EURLEX-57K, including Recall@5 and nDCG@5. The ranking metrics are preferable for EURLEX-57K since it does not introduce a significant bias towards frequent labels (Lu et al., 2020). Implementation Details We initialize the word embeddings with 300D pre-trained GloVe vectors 4375 (Pennington et al., 2014). Then we use a one-layer BiGRU with hidden dimension 100 and used 100 filters with kernel size [2,3,4] to setup the CNNs. The dimension of the text propagation feature and graph convolution weight matrix are both 300. The hidden size of joint embedding is 200. The matching margin γ is set to 0.2 on RCV1-V2 and WOS datasets, and set to 0.5 on EURLEX-57K dataset. We set the value of hierarchy-aware penalty hyperparameters α, β to 0.01 and 0.5, respectively. The loss balancing factor λ1, λ2 are set to 1. For fair comparisons with previous work (Lu et al., 2020; Chalkidis et al., 2019) on EURLEX-57K dataset, firstly, we do not use CNN layer and text feature propagation module. Secondly, to adapt to the zeroshot settings, the prediction is generated by the dot product similarity between text semantics and label semantics. Our model is optimized by Adam with a learning rate of 1e-4. For pretrained language model BERT (Devlin et al., 2018), we use the top-level representation hCLS of BERT’s special CLS token to perform classification. To combine our model with BERT, we replace the text encoder of HiMatch with BERT, and the label representations are initiated by pretrained BERT embedding. The batch size is set to 16, and the learning rate is 2e-5. Comparison Models On RCV1-V2 and WOS datasets, we compare our model with three types of strong baselines: 1) Text classification baselines: TextRCNN (Lai et al., 2015), TextRCNN with label attention (TextRCNN-LA) (Zhou et al., 2020), and SGM (Yang et al., 2018). 2) Hierarchy-aware models: HE-AGCRCNN (Peng et al., 2019), HMCN (Mao et al., 2019), Htrans (Banerjee et al., 2019), HiLAP-RL (Mao et al., 2019) which introduced reinforcement learning to simulate the assignment process, HiAGM (Zhou et al., 2020) which exploited the prior probability of label dependecies through Graph Convolution Network and TreeLSTM. 3) Pretrained language model: a more powerful pretrained language model BERT (Devlin et al., 2018) than tradition text classification models when fine-tuned on downstream tasks. On EURLEX-57K dataset, we compare our model with strong baselines with/without zeroshot settings such as BIGRU-ATT, BIGRU-LWAN (Chalkidis et al., 2019) which introduced labelwise attention. The models starting with “ZERO” make predictions by calculating similarity scores between text and label semantics for zero-shot settings. AGRU-KAMG (Lu et al., 2020) is a stateof-the-art model which introduced various label knowledge. 4.2 Experiment Results Models Micro Macro Baselines TextRCNN (Zhou et al., 2020) 81.57 59.25 TextRCNN-LA (Zhou et al., 2020) 81.88 59.85 SGM (Zhou et al., 2020) 77.30 47.49 Hierarchy-Aware Models HE-AGCRCNN (Peng et al., 2019) 77.80 51.30 HMCN (Mao et al., 2019) 80.80 54.60 Htrans (Banerjee et al., 2019) 80.51 58.49 HiLAP-RL (Mao et al., 2019) 83.30 60.10 HiAGM (Zhou et al., 2020) 83.96 63.35 HiMatch 84.73 64.11 Pretrained Language Models BERT (Devlin et al., 2018) 86.26 67.35 BERT+HiMatch 86.33 68.66 Table 2: The experimental results comparing to other state-of-the-art models on RCV1-V2 dataset. Models Micro Macro Baselines TextRNN (Zhou et al., 2020) 77.94 69.65 TextCNN (Zhou et al., 2020) 82.00 76.18 TextRCNN (Zhou et al., 2020) 83.55 76.99 Hierarchy-Aware Models HiAGM (Zhou et al., 2020) 85.82 80.28 HiMatch 86.20 80.53 Pretrained Language Models BERT (Devlin et al., 2018) 86.26 80.58 BERT+HiMatch 86.70 81.06 Table 3: The experimental results comparing to other state-of-the-art models on Web-of-Science dataset. Table 2, 3 and 4 report the performance of our approaches against other methods. HiAGM is an effective baseline on RCV1-V2 and WOS due to the introduction of holistic label information. However, they ignored the semantic relationship between text and labels. Our model achieves the best results by capturing the matching relationships among text and labels in a hierarchy-aware manner, which achieves stronger performances especially on Macro-F1. The improvements show that our model can make better use of structural information to help imbalanced HTC classification. The pretrained language model BERT is an effective method when fine-tuned on downstream tasks. Compared with the results regarding HTC 4376 Frequent Few Zero Overall R@5 nDCG@5 R@5 nDCG@5 R@5 nDCG@5 R@5 nDCG@5 BIGRU-ATT (Chalkidis et al., 2019) 0.740 0.813 0.596 0.580 0.051 0.027 0.675 0.789 BIGRU-LWAN (Chalkidis et al., 2019) 0.755 0.819 0.661 0.618 0.029 0.019 0.692 0.796 ZERO-CNN-LWAN (Chalkidis et al., 2019) 0.683 0.745 0.494 0.454 0.321 0.264 0.617 0.717 ZERO-BIGRU-LWAN (Chalkidis et al., 2019) 0.716 0.780 0.560 0.510 0.438 0.345 0.648 0.752 AGRU-KAMG (Lu et al., 2020) 0.731 0.795 0.563 0.518 0.528 0.414 0.661 0.766 HiMatch 0.769 0.830 0.697 0.648 0.399 0.372 0.705 0.807 Table 4: The experimental results comparing to other state-of-the-art models on EURLEX-57K dataset. as multiple binary classifiers, our results show that the full use of structured label hierarchy can bring great improvements to BERT model on RCV1-V2 and WOS datasets. On EURLEX57K dataset, our model achieves the best results on different matrics except for zeroshot labels. The largest improvements come from few-shot labels. AGRU-KAMG achieves the best results on zero-shot labels by fusing various knowledge such as label semantics similarities and label co-occurrence. However, our model performs semantics matching among seen labels based on training corpora, which is not designed for a specific zero-shot learning task. 4.3 Analysis 4.3.1 Ablation Study In this section, we investigate to study the independent effect of each component in our proposed model. Firstly, we validate the influence of two proposed losses, and the hierarchy-aware sampling. The results are reported in Table 5. The results show that F1 will decrease with removing joint embedding loss or matching learning loss. Joint embedding loss has a great influence since label semantics matching relies on the joint embedding. Besides, in the hierarchy-aware margin subsection, we perform hierarchy-aware sampling by sampling coarse-grained labels, incorrect sibling labels, and other incorrect labels as negative label sets. When we remove hierarchy-aware sampling and replace it with random sampling, the results will decrease, which shows the effectiveness of hierarchy-aware sampling. 4.3.2 Hyperparameters Study To study the influence of the hyperparameters γ, α, and β, we conduct seven experiments on RCV1V2 dataset. The results are reported in Table 6. The first experiment is the best hyperparameters of our model. Then we fine-tune the matching learning margin γ in experiments two and three. We Ablation Models Micro Macro TextRCNN 81.57 59.25 HiMatch 84.73 64.11 - w/o Joint Embedding Loss 84.49 62.57 - w/o Matching Learning Loss 84.46 63.58 - w/o Hierarchy-aware Sampling 84.67 63.45 Table 5: Ablation study on RCV1-V2 dataset. No. γ α β Micro Macro HiMatch x 0.2 0.01 0.5 84.73 64.11 Fine-tuning γ y 0.02 0.01 0.5 84.51 63.26 z 2 0.01 0.5 84.69 63.55 Fine-tuning α, β { 0.2 0.5 0.01 84.52 63.35 | 0.2 1 1 84.37 63.45 } 0.2 0.01 0.01 84.49 63.20 ~ 0.2 0.5 0.5 84.47 64.02 Table 6: Hyperparameter study on RCV1-V2 dataset. find that a proper margin γ = 0.2 is beneficial for matching learning compared with a large or small margin. Furthermore, we validate the effectiveness of the hierarchy-aware margin. In experiment four, the performance will decrease if we violate the hierarchical structure by setting a large penalty margin for coarse-grained labels, and setting a small penalty margin for incorrect sibling labels. In experiment five, the performance has a relatively larger decrease if we set α = 1 and β = 1, which ignores hierarchical structure completely. We speculate that the penalty margin that violates the hierarchical structure will affect the results, since the semantics relationship should be closer if the labels are closer in the hierarchical structure. Moreover, we validate the effectiveness of different penalty margins among different granular labels. In experiments six and seven, the results will degrade if we ignore the relationships between coarse-grained target labels and incorrect sibling labels, by setting the same margin for α and 4377 a) Label Hierarchy . . . C17: funding/captial C172: bonds/debt issues b) Joint Embedding Loss x Label Representations Text Representations C171: share captial GWELF: welfare/social services E61: housing starts GWELF E61 C171 C17 C172 GWELF E61 C171 C172 c) Joint Embedding Loss and Matching Learning Loss Figure 4: Figure a) is a part of the hierarchical label structure. Figure b) is the T-SNE visualization of text representations and label representations of the labels in Figure a) by introducing joint embedding loss. Figure c) is the T-SNE visualization with both joint embedding loss and matching learning loss. Figure 5: Performance study on label granularity based on hierarchical levels. β. Therefore, it is necessary to set a small penalty margin for coarse-grained target labels, and a larger penalty margin for incorrect sibling labels. 4.3.3 T-SNE Visualization of Joint Embedding We plot the T-SNE projection of the text representations and label representations in the joint embedding in Figure 4. Figure a) is a part of the hierarchical label structure in RCV1-V2. Label C171 and C172 are fine-grained labels, and label C17 is coarse-grained label of C171 and C172. GWELF and E61 are other labels with different semantics with C17, C171 and C172. In Figure b), by introducing joint embedding loss, we can see that the text representations are close to their corresponding label representations. Furthermore, the text representations of labels C171 and C172 are close to the label representation of their coarse-grained label C17. However, the text representations of different labels may overlap, since the matching relationships among different labels are ignored. In Figure c), by introducing both joint embedding loss and matching learning loss, the text representations of different labels are more separable. Other unrelated text representations and label representations such as labels GWELF, E61 are far away from C17, C171, C172. Besides, the text representations of semantically similar labels (C171 and C172) are far away relatively compared with Figure b). The T-SNE visualization shows that our model can capture the semantics relationship among texts, coarsegrained labels, fine-grained labels and unrelated labels. 4.3.4 Performance Study on Label Granularity We analyze the performance with different label granularity based on their hierarchical levels. We compute level-based Micro-F1 and Macro-F1 scores of the RCV1-V2 dataset on TextRCNN, HiAGM, and our model in Figure 5. On RCV1-V2 dataset, both the second and third hierarchical levels contain fine-grained labels (leaf nodes). The second level has the largest number of labels and contains confusing labels with similar concepts, so its Micro-F1 is relatively low. Both the second and third levels contain some long-tailed labels, so their Macro-F1 are relatively low. Figure 5 shows that our model achieves a better performance than other models on all levels, especially among deep levels. The results demonstrate that our model has a better ability to capture the hierarchical label semantic, especially on fine-grained labels with a complex hierarchical structure. 4.3.5 Computational Complexity In this part, we compare the computational complexity between HiAGM and our model. For time complexity, the training time of HiMatch is 1.11 times that of HiAGM with batch size 64. For space complexity during training, HiMatch has 37.4M parameters, while HiAGM has 27.8M. The increase mainly comes from the label encoder with large 4378 label sets. However, during testing, the time and space complexity of HiMatch is the same as HiAGM. The reason is that only the classification results are needed, and we can remove the joint embedding. HiMatch achieves new state-of-the-art results, and we believe that the increase of computational complexity is acceptable. 5 Conclusion Here we present a novel hierarchical text classification model called HiMatch that can capture semantic relationships among texts and labels at different abstraction levels. Instead of treating HTC as multiple binary classification tasks, we consider the text-label semantics matching relationship and formulate it as a semantic matching problem. We learn a joint semantic embedding between text and labels. Finally, we propose a hierarchy-aware matching strategy to model different matching relationships among coarse-grained labels, fine-grained labels and incorrect labels. In future work, we plan to extend our model to the zero-shot learning scenario. Acknowledgments We thank the anonymous reviewers for their helpful feedbacks. The work described in this paper was partially funded by the National Natural Science Foundation of China (Grant No. 61502174, and 61872148), the Natural Science Foundation of Guangdong Province (Grant No. 2017A030313355, 2019A1515010768 and 2021A1515011496), the Guangzhou Science and Technology Planning Project (Grant No. 201704030051, and 201902010020), the Key R&D Program of Guangdong Province (No. 2018B010107002) and the Fundamental Research Funds for the Central Universities. References Aixin Sun and Ee-Peng Lim. 2001. Hierarchical text classification and evaluation. In Proceedings 2001 IEEE International Conference on Data Mining, pages 521–528. Siddhartha Banerjee, Cem Akkaya, Francisco PerezSorrosal, and Kostas Tsioutsiouliklis. 2019. Hierarchical transfer learning for multi-label text classification. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6295–6300, Florence, Italy. Association for Computational Linguistics. Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, and Ion Androutsopoulos. 2019. Large-scale multi-label text classification on EU legislation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6314–6322, Florence, Italy. Association for Computational Linguistics. Boli Chen, Xin Huang, Lin Xiao, Zixin Cai, and Liping Jing. 2020. Hyperbolic interaction model for hierarchical multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 7496–7503. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. Johannes F¨urnkranz, Eyke H¨ullermeier, Eneldo Loza Menc´ıa, and Klaus Brinker. 2008. Multilabel classification via calibrated label ranking. Mach. Learn., 73(2):133–153. K. Kowsari, D. E. Brown, M. Heidarysafa, K. Jafari Meimandi, M. S. Gerber, and L. E. Barnes. 2017. Hdltex: Hierarchical deep learning for text classification. In 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), pages 364–371. Siwei Lai, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Recurrent convolutional neural networks for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 29. David D. Lewis, Yiming Yang, Tony G. Rose, and Fan Li. 2004. Rcv1: A new benchmark collection for text categorization research. 5:361–397. Jueqing Lu, Lan Du, Ming Liu, and Joanna Dipnall. 2020. Multi-label few/zero-shot learning with knowledge aggregated from multiple label graphs. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2935–2943, Online. Association for Computational Linguistics. Yuning Mao, Jingjing Tian, Jiawei Han, and Xiang Ren. 2019. Hierarchical text classification with reinforced label assignment. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 445–455, Hong Kong, China. Association for Computational Linguistics. Hao Peng, Jianxin Li, Qiran Gong, Senzhang Wang, Lifang He, Bo Li, Lihong Wang, and Philip S. Yu. 2019. Hierarchical taxonomy-aware and attentional graph capsule rcnns for large-scale multi-label text classification. CoRR, abs/1906.04898. Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference 4379 on empirical methods in natural language processing (EMNLP), pages 1532–1543. Bingyu Wang, Li Chen, Wei Sun, Kechen Qin, Kefeng Li, and Hui Zhou. 2019. Ranking-based autoencoder for extreme multi-label classification. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2820–2830, Minneapolis, Minnesota. Association for Computational Linguistics. Lin Xiao, Xin Huang, Boli Chen, and Liping Jing. 2019. Label-specific document representation for multi-label text classification. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 466–475, Hong Kong, China. Association for Computational Linguistics. Pengcheng Yang, Xu Sun, Wei Li, Shuming Ma, Wei Wu, and Houfeng Wang. 2018. SGM: Sequence generation model for multi-label classification. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3915–3926, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Chih-Kuan Yeh, Wei-Chieh Wu, Wei-Jen Ko, and YuChiang Frank Wang. 2017. Learning deep latent space for multi-label classification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 31. Jie Zhou, Chunping Ma, Dingkun Long, Guangwei Xu, Ning Ding, Haoyu Zhang, Pengjun Xie, and Gongshen Liu. 2020. Hierarchy-aware global model for hierarchical text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1106–1117, Online. Association for Computational Linguistics.
|
2021
|
337
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4380–4390 August 1–6, 2021. ©2021 Association for Computational Linguistics 4380 HiddenCut: Simple Data Augmentation for Natural Language Understanding with Better Generalization Jiaao Chen, Dinghan Shen1, Weizhu Chen1, Diyi Yang Georgia Institute of Technology, 1Microsoft Dynamics 365 AI {jchen896,dyang888}@gatech.edu {dishen,wzchen}@microsoft.com Abstract Fine-tuning large pre-trained models with taskspecific data has achieved great success in NLP. However, it has been demonstrated that the majority of information within the selfattention networks is redundant and not utilized effectively during the fine-tuning stage. This leads to inferior results when generalizing the obtained models to out-of-domain distributions. To this end, we propose a simple yet effective data augmentation technique, HiddenCut, to better regularize the model and encourage it to learn more generalizable features. Specifically, contiguous spans within the hidden space are dynamically and strategically dropped during training. Experiments show that our HiddenCut method outperforms the state-of-the-art augmentation methods on the GLUE benchmark, and consistently exhibits superior generalization performances on out-of-distribution and challenging counterexamples. We have publicly released our code at https://github.com/ GT-SALT/HiddenCut. 1 Introduction Fine-tuning large-scale pre-trained language models (PLMs) has become a dominant paradigm in the natural language processing community, achieving state-of-the-art performances in a wide range of natural language processing tasks (Devlin et al., 2019; Liu et al., 2019; Yang et al., 2019a; Joshi et al., 2019; Sun et al., 2019; Clark et al., 2019; Lewis et al., 2020; Bao et al., 2020; He et al., 2020; Raffel et al., 2020). Despite the great success, due to the huge gap between the number of model parameters and that of task-specific data available, the majority of the information within the multi-layer self-attention networks is typically redundant and ineffectively utilized for downstream tasks (Guo et al., 2020; Gordon et al., 2020; Dalvi et al., 2020). As a result, after task-specific fine-tuning, models are very likely to overfit and make predictions based on spurious patterns (Tu et al., 2020; Kaushik et al., 2020), making them less generalizable to outof-domain distributions (Zhu et al., 2019; Jiang et al., 2019; Aghajanyan et al., 2020). In order to improve the generalization abilities of over-parameterized models with limited amount of task-specific data, various regularization approaches have been proposed, such as adversarial training that injects label-preserving perturbations in the input space (Zhu et al., 2019; Liu et al., 2020; Jiang et al., 2019), generating augmented data via carefully-designed rules (McCoy et al., 2019; Xie et al., 2020; Andreas, 2020; Shen et al., 2020), and annotating counterfactual examples (Goyal et al., 2019; Kaushik et al., 2020). Despite substantial improvements, these methods often require significant computational and memory overhead (Zhu et al., 2019; Liu et al., 2020; Jiang et al., 2019; Xie et al., 2020) or human annotations (Goyal et al., 2019; Kaushik et al., 2020). In this work, to alleviate the above issues, we rethink the simple and commonly-used regularization technique—dropout (Srivastava et al., 2014)— in pre-trained transformer models (Vaswani et al., 2017). With multiple self-attention heads in transformers, dropout converts some hidden units to zeros in a random and independent manner. Although PLMs have already been equipped with the dropout regularization, they still suffer from inferior performances when it comes to out-of-distribution cases (Tu et al., 2020; Kaushik et al., 2020). The underlying reasons are two-fold: (1) the linguistic relations among words in a sentence is ignored while dropping the hidden units randomly. In reality, these masked features could be easily inferred from surrounding unmasked hidden units with the self-attention networks. Therefore, redundant information still exists and gets passed to the upper 4381 layers. (2) The standard dropout assumes that every hidden unit is equally important with the random sampling procedure, failing to characterize the different roles these features play in distinct tasks. As a result, the learned representations are not generalized enough while applied to other data and tasks. To drop the information more effectively, Shen et al. (2020) recently introduce Cutoff to remove tokens/features/spans in the input space. Even though models will not see the removed information during training, examples with large noise may be generated when key clues for predictions are completely removed from the input. To overcome these limitations, we propose a simple yet effective data augmentation method, HiddenCut, to regularize PLMs during the fine-tuning stage. Specifically, the approach is based on the linguistic intuition that hidden representations of adjacent words are more likely to contain similar and redundant information. HiddenCut drops hidden units more structurally by masking the whole hidden information of contiguous spans of tokens after every encoding layer. This would encourage models to fully utilize all the task-related information, instead of learning spurious patterns during training. To make the dropping process more efficient, we dynamically and strategically select the informative spans to drop by introducing an attentionbased mechanism. By performing HiddenCut in the hidden space, the impact of dropped information is only mitigated rather than completely removed, avoiding injecting too much noise to the input. We further apply a Jensen-Shannon Divergence consistency regularization between the original and these augmented examples to model the consistent relations between them. To demonstrate the effectiveness of our methods, we conduct experiments to compare our HiddenCut with previous state-of-the-art data augmentation method on 8 natural language understanding tasks from the GLUE (Wang et al., 2018) benchmark for in-distribution evaluations, and 5 challenging datasets that cover single-sentence tasks, similarity and paraphrase tasks and inference tasks for out-ofdistribution evaluations. We further perform ablation studies to investigate the impact of different selecting strategies on HiddenCut’s effectiveness. Results show that our method consistently outperforms baselines, especially on out-of-distribution and challenging counterexamples. To sum up, our contributions are: • We propose a simple data augmentation method, HiddenCut, to regularize PLMs during fine-tuning by cutting contiguous spans of representations in the hidden space. • We explore and design different strategic sampling techniques to dynamically and adaptively construct the set of spans to be cut. • We demonstrate the effectiveness of HiddenCut through extensive experiments on both indistribution and out-of-distribution datasets. 2 Related Work 2.1 Adversarial Training Adversarial training methods usually regularize models through applying perturbations to the input or hidden space (Szegedy et al., 2013; Goodfellow et al., 2014; Madry et al., 2017) with additional forward-backward passes, which influence the model’s predictions and confidence without changing human judgements. Adversarial-based approaches have been actively applied to various NLP tasks in order to improve models’ robustness and generalization abilities, such as sentence classification (Miyato et al., 2017), machine reading comprehension (MRC) (Wang and Bansal, 2018) and natural language inference (NLI) tasks (Nie et al., 2020). Despite its success, adversarial training often requires extensive computation overhead to calculate the perturbation directions (Shafahi et al., 2019; Zhang et al., 2019a). In contrast, our HiddenCut adds perturbations in the hidden space in a more efficient way that does not require extra computations as the designed perturbations can be directly derived from self-attentions. 2.2 Data Augmentation Another line of work to improve the model robustness is to directly design data augmentation methods to enrich the original training set such as creating syntactically-rich examples (McCoy et al., 2019; Min et al., 2020) with specific rules, crowdsourcing counterfactual augmentation to avoid learning spurious features (Goyal et al., 2019; Kaushik et al., 2020), or combining examples in the dataset to increase compositional generalizabilities (Jia and Liang, 2016; Andreas, 2020; Chen et al., 2020b,a). However, they either require careful design (McCoy et al., 2019; Andreas, 2020) to infer labels for generated data or extensive human annotations (Goyal et al., 2019; Kaushik et al., 2020), 4382 which makes them hard to generalize to different tasks/datasets. Recently Shen et al. (2020) introduce a set of cutoff augmentation which directly creates partial views to augment the training in a more task-agnostic way. Inspired by these prior work, our HiddenCut aims at improving models’ generalization abilities to out-of-distribution via linguistic-informed strategically dropping spans of hidden information in transformers. 2.3 Dropout-based Regularization Variations of dropout (Srivastava et al., 2014) have been proposed to regularize neural models by injecting noise through dropping certain information so that models do not overfit training data. However, the major efforts have been put to convolutional neural networks and trimmed for structures in images recently such as DropPath (Larsson et al., 2017), DropBlock (Ghiasi et al., 2018), DropCluster (Chen et al., 2020c) and AutoDropout (Pham and Le, 2021). In contrast, our work takes a closer look at transformer-based models and introduces HiddenCut for natural language understanding tasks. HiddenCut is closely related to DropBlock (Ghiasi et al., 2018), which drops contiguous regions from a feature map. However, different from images, hidden dimensions in PLMs that contain syntactic/semantic information for NLP tasks are more closely related (e.g., NER and POS information), and simply dropping spans of features in certain hidden dimensions might still lead to information redundancy. 3 HiddenCut Approach To regularize transformer models in a more structural and efficient manner, in this section, we introduce a simple yet effective data augmentation technique, HiddenCut, that reforms dropout to cutting contiguous spans of hidden representations after each transformer layer (Section 3.1). Intuitively, the proposed approach encourages the models to fully utilize all the hidden information within the self-attention networks. Furthermore, we propose an attention-based mechanism to strategically and judiciously determine the specific spans to cut (Section 3.2). The schematic diagram of HiddenCut, applied to the transformer architecture (and its comparison to dropout) are shown in Figure 1. 3.1 HiddenCut For an input sequence s = {w0, w1, ..., wL} with L tokens associated with a label y, we employ a pre-trained transformer model f1:M(·) with M layers like RoBERTa (Liu et al., 2019) to encode the text into hidden representations. Thereafter, an inference network g(·) is learned on top of the pretrained models to predict the corresponding labels. In the hidden space, after layer m, every word wi in the input sequence is encoded into a D dimensional vector hm i ∈RD and the whole sequence could be viewed as a hidden matrix Hm ∈RL×D. With multiple self-attention heads in the transformer layers, it is found that there is extensive redundant information across hm i ∈H that are linguistically related (Dalvi et al., 2020) (e.g., words that share similar semantic meanings). As a result, the removed information from the standard dropout operation may be easily inferred from the remaining unmasked hidden units. The resulting model might easily overfit to certain high-frequency features without utilizing all the important task-related information in the hidden space (especially when task-related data is limited). Moreover, the model also suffers from poor generalization ability while being applied to out-of-distribution cases. Inspired by Ghiasi et al. (2018); Shen et al. (2020), we propose to improve the dropout regularization in transformer models by creating augmented training examples through HiddenCut, which drops a contiguous span of hidden information encoded in every layer, as shown in Figure 1 (c). Mathematically, in every layer m, a span of hidden vectors, S ∈Rl×D, with length l = αL in the hidden matrix Hm ∈RL×D are converted to 0, and the corresponding attention masks are adjusted to 0, where α is a pre-defined hyper-parameter indicating the dropping extent of HiddenCut. After being encoded and hiddencut through all the hidden layers in pre-trained encoders, augmented training data fHiddenCut(s) is created for learning the inference network g(·) to predict task labels. 3.2 Strategic Sampling Different tasks rely on learning distinct sets of information from the input to predict the corresponding task labels. Performing HiddenCut randomly might be inefficient especially when most of the dropping happens at task-unrelated spans, which fails to effectively regularize model to take advantage of all the task-related features. To this end, we 4383 Figure 1: Illustration of the differences between Dropout (a) and HiddenCut (b), and the position of HiddenCut in transformer layers (c). A sentence in the hidden space can be viewed as a L × D matrix where L is the length of the sentence and D is the number of hidden dimensions. The cells in blue represent that they are masked. Dropout masks random independent units in the matrix while our HiddenCut selects and masks a whole span of hidden representations based on attention weights received in the current layer. In our experiments, we perform HiddenCut after the feed-forward network in every transformer layer. propose to select the spans to be cut dynamically and strategically in every layer. In other words, we mask the most informative span of hidden representations in one layer to force models to discover other useful clues to make predictions instead of relying on a small set of spurious patterns. Attention-based Sampling Strategy The most direct way is to define the set of tokens to be cut by utilizing attention weights assigned to tokens in the self-attention layers (Kovaleva et al., 2019). Intuitively, we can drop the spans of hidden representations that are assigned high attentions by the transformer layers. As a result, the information redundancy is alleviated and models would be encourage to attend to other important information. Specifically, we first derive the average attention for each token, ai, from the attention weights matrix A ∈RP×L×L after self-attention layers, where P is the number of attention heads and L is the sequence length: ai = PP j (PL k A[j][k][i]) P . We then sample the start token hi for HiddenCut from the set that contains top βL tokens with higher average attention weights (β is a pre-defined parameter). Then HiddenCut is performed to mask the hidden representations between hi and hi+l. Note that the salient sets are different across different layers and updated throughout the training. Other Sampling Strategies We also explore other widely used word importance discovery methods to find a set of tokens to be strategically cut by HiddenCut, including: • Random: All spans of tokens are viewed as equally important, thus are randomly cut. • LIME (Ribeiro et al., 2016) defines the importance of tokens by examining the locally faithfulness where weights of tokens are assigned by classifiers trained with sentences whose words are randomly removed. We utilized LIME on top of a SVM classifier to pre-define a fixed set of tokens to be cut. • GEM (Yang et al., 2019b) utilizes orthogonal basis to calculate the novelty scores that measure the new semantic meaning in tokens, significance scores that estimate the alignment between the semantic meaning of tokens and the sentence-level meaning, and the uniqueness scores that examine the uniqueness of the semantic meaning of tokens. We compute the GEM scores using the hidden representations at every layer to generate the set of tokens to be cut, which are updated during training. • Gradient (Baehrens et al., 2010): We define the set of tokens to be cut based on the rankings of the absolute values of gradients they received at every layer in the backward-passing. This set would be updated during training. 4384 3.3 Objectives During training, for an input text sequence s with a label y, we generate N augmented examples {fHiddenCut 1 (s), ..., fHiddenCut N (s)} through performing HiddenCut in pre-trained encoder f(·). The whole model g(f(·)) is then trained though several objectives including general classification loss (Lori and Laug) on data-label pairs and consistency regularization (Ljs) (Miyato et al., 2017, 2018; Clark et al., 2018; Xie et al., 2019; Shen et al., 2020) across different augmentations: Lori = CE(g(f(s)), y) Laug = X N CE(g(fHiddenCut i (s)), y) Ljs = X N KL[p(y|g(fHiddenCut i (s))||pavg] where CE and KL represent the cross-entropy loss and KL-divergence respectively. pavg stands for the average predictions across the original text and all the augmented examples. Combining these three losses, our overall objective function is: L = Lori + γLaug + ηLjs where γ and η are the weights used to balance the contributions of learning from the original data and augmented data. 4 Experiments 4.1 Datasets We conducted experiments on both in-distribution datasets and out-of-distribution datasets to demonstrate the effectiveness of our proposed HiddenCut. In-Distribution Datasets We mainly trained and evaluated our methods on the widely-used GLUE benchmark (Wang et al., 2018) which covers a wide range of natural language understanding tasks: single-sentence tasks including: (i) Stanford Sentiment Treebank (SST-2) which predict the sentiment of movie reviews to be positive or negative, and (ii) Corpus of Linguistic Acceptability (CoLA) which predict whether a sentence is linguistically acceptable or not; similarity and paraphrase tasks including (i) Quora Question Pairs (QQP) which predict whether two question are paraphrases, (ii) Semantic Textual Similarity Benchmark (STS-B) which predict the similarity ratings between two sentences, and (iii) Microsoft Research Paraphrase Corpus (MRPC) which predict whether two given sentences are semantically equivalent; inference tasks including (i) Multi-Genre Natural Language Inference (MNLI) which classified the relationships between two sentences into entailment, contradiction, or neutral, (ii) Question Natural Language Inference (QNLI) which predict whether a given sentence is the correct answer to a given question, and (iii) Recognizing Textual Entailment (RTE) which predict whether the entailment relation holds between two sentences. Accuracy was used as the evaluation metric for most of the datasets except that Matthews correlation was used for CoLA and Spearman correlation was utilized for STS-B. Out-Of-Distribution Datasets To demonstrate the generalization abilities of our proposed methods, we directly evaluated on 5 different out-ofdistribution challenging sets, using the models that are fine-tuned on GLUE benchmark datasets: • Single Sentence Tasks: Models fine-tuned from SST-2 are directly evaluated on two recent challenging sentiment classification datasets: IMDB Contrast Set (Gardner et al., 2020) including 588 examples and IMDB Counterfactually Augmented Dataset (Kaushik et al., 2020) including 733 examples. Both of them were constructed by asking NLP researchers (Gardner et al., 2020) or Amazon Mechanical Turkers (Kaushik et al., 2020) to make minor edits to examples in the original IMDB dataset (Maas et al., 2011) so that the sentiment labels change while the major contents keep the same. • Similarity and Paraphrase Tasks: Models fine-tuned from QQP are directly evaluated on the recently introduced challenging paraphrase dataset PAWS-QQP (Zhang et al., 2019b) that has 669 test cases. PAWS-QQP contains sentence pairs with high word overlap but different semantic meanings created via word-swapping and back-translation from the original QQP dataset. • Inference Tasks: Models fine-tuned from MNLI are directly evaluated on two challenging NLI sets: HANS (McCoy et al., 2019) with 30,000 test cases and Adversarial NLI (A1 dev sets) (Nie et al., 2020) including 1,000 test cases. The former one was constructed by using syntactic rules (lexical overlap, subsequence and constituent) to generate 4385 Method MNLI QNLI QQP RTE SST-2 MRPC CoLA STS-B Avg RoBERTa-base 87.6 92.8 91.9 78.7 94.8 89.5 63.6 91.2 86.3 ALUM 88.1 93.1 92.0 80.2 95.3 90.9 63.6 91.1 86.8 Token Cutoff 88.2 93.1 91.9 81.2 95.1 91.1 64.1 91.2 87.0 Feature Cutoff 88.2 93.3 92.0 81.6 95.3 90.7 63.6 91.2 87.0 Span Cutoff 88.4 93.4 92.0 82.3 95.4 91.1 64.7 91.2 87.3 HiddenCut † 88.2 93.7 92.0 83.4 95.8 92.0 66.2 91.3 87.8 Table 1: In-distribution evaluation results on the dev sets of the GLUE benchmark. † means our proposed method. Method Single-Sentence Similarity&Paraphrase Inference IMDB-Cont. IMDB-CAD PAWS-QQP HANS AdvNLI (A1) RoBERTa-base 84.6 88.4 38.4 67.8 31.2 Span Cutoff 85.5 89.2 38.8 68.4 31.1 HiddenCut † 87.8 90.4 41.5 71.2 32.8 Table 2: Out-of-distribution evaluation results on 5 different challenging sets. † means our proposed method. For all the datasets, we did not use their training sets to further fine-tune the derived models from GLUE. non-entailment examples with high premisehypothesis overlap from MNLI. The latter one was created by adversarial human-and-modelin-the-loop framework (Nie et al., 2020) to create hard examples based on BERT-Large models(Devlin et al., 2019) pre-trained on SNLI (Bowman et al., 2015) and MNLI. 4.2 Baselines We compare our methods with several baselines: • RoBERTa (Liu et al., 2019) is used as our base model. Note that RoBERTa is regularized with dropout during fine-tuning. • ALUM (Liu et al., 2020) is the state-of-theart adversarial training method for neural language models, which regularizes fine-tuning via perturbations in the embedding space. • Cutoff (Shen et al., 2020) is a recent data augmentation for natural language understanding tasks by removing information in the input space, including three variations: token cutoff, feature cutoff, and span cutoff. 4.3 Implementation Details We used the RoBERTa-base model (Liu et al., 2019) to initialize all the methods. Note that HiddenCut is agnostic to different types of pre-trained models. We followed Liu et al. (2019) to set the linear decay scheduler with a warmup ratio of 0.06 for training. The maximum learning rate was selected from {5e−6, 8e−6, 1e−5, 2e−5} and the max number of training epochs was set to be either 5 or 10. All these hyper-parameters are shared for all the models. The HiddenCut ratio α was set 0.1 after a grid search from {0.05, 0.1, 0.2, 0.3, 0.4}. The selecting ratio β in the important sets sampling process was set 0.4 after a grid search from {0.1, 0.2, 0.4, 0.6}. The weights γ and η in our objective function were both 1. All the experiments were performed using a GeForce RTX 2080Ti. 4.4 Results on In-Distribution Datasets Based on Table 1, we observed that, compared to RoBERTa-base with only dropout regularization, ALUM with perturbations in the embedding space through adversarial training has better results on most of these GLUE tasks. However, the extra additional backward passes to determine the perturbation directions in ALUM can bring in significantly more computational and memory overhead. By masking different types of input during training, Cutoff increased the performances while being more computationally efficient. In contrast to Span Cutoff, HiddenCut not only introduced zero additional computation cost, but also demonstrated stronger performances on 7 out of 8 GLUE tasks, especially when the size of training set is small (e.g., an increase of 1.1 on RTE and 1.5 on CoLA). Moreover, HiddenCut achieved the best average result compared to previous stateof-the-art baselines. These in-distribution improvements indicated that, by strategically dropping contiguous spans in the hidden space, HiddenCut not 4386 only helps pre-trained models utilize hidden information in a more effective way, but also injects less noise during the augmentation process compared to cutoff, e.g., Span Cutoff might bring in additional noises for CoLA (which aims to judge whether input sentences being linguistically acceptable or not) when one span in the input is removed, since it might change the labels. 4.5 Results on Out-Of-Distribution Datasets To validate the better generalizability of HiddenCut, we tested our models trained on SST-2, QQP and MNLI directly on 5 out-of-distribution/outof-domain challenging sets in zero-shot settings. As mentioned earlier, these out-of-distribution sets were either constructed with in-domain/out-ofdomain data and further edited by human to make them harder, or generated by rules that exploited spurious correlations such as lexical overlap, which made them challenging to most existing models. As shown in Table 2, Span Cutoff slightly improved the performances compared to RoBERTa by adding extra regularizations through creating restricted input. HiddenCut significantly outperformed both RoBERTa and Span Cutoff. For example, it outperformed Span Cutoff. by 2.3%(87.8% vs. 85.5%) on IMDB-Conts, 2.7%(41.5% vs. 38.8%) on PAWS-QQP, and 2.8%(71.2% vs 68.4%) on HANS consistently. These superior results demonstrated that, by dynamically and strategically dropping contiguous span of hidden representations, HiddenCut was able to better utilize all the important task-related information which improved the model generalization to out-of-distribution and challenging adversary examples. 4.6 Ablation Studies This section presents our ablation studies on different sampling strategies and the effect of important hyper-parameters in HiddenCut. 4.6.1 Sampling Strategies in HiddenCut We compared different ways to cut hidden representations (DropBlock (Ghiasi et al., 2018) which randomly dropped spans in certain random hidden dimensions instead of the whole hidden space) and different sampling strategies for HiddenCut described in Section 3.2 (including Random, LIME (Ribeiro et al., 2016), GEM (Yang et al., 2019b), Gradient (Yeh et al., 2019), Attention) based on the performances on SST-2 and QNLI. For these strategies, we also experimented with a reverse set Strategy SST-2 QNLI RoBERTa 94.8 92.8 DropBlock 95.4 93.2 Random 95.4 93.5 LIME 95.2 93.1 LIME-R 95.3 93.2 GEM 95.5 93.4 GEM-R 95.1 93.2 Gradient 95.6 93.6 Gradient-R 95.1 93.4 Attention 95.8 93.7 Attention-R 94.6 93.4 Table 3: The performances on SST-2 and QNLI with different strategies when dropping information in the hidden space. Different sampling strategies combined with HiddenCut are presented. “-R” means sampling outside the set to be cut given by these strategies. denoted by “-R” where we sampled outside the important set given by above strategies. From Table 3, we observed that (i) sampling from important sets resulted in better performances than random sampling. Sampling outside the defined importance sets usually led to inferior performances. These highlights the importance of strategically selecting spans to drop. (ii) Sampling from dynamic sets sampled by their probabilities often outperformed sampling from predefined fixed sets (LIME), indicating the effectiveness of dynamically adjusting the sampling sets during training. (iii) The attention-based strategy outperformed all other sampling strategies, demonstrating the effectiveness of our proposed sampling strategies for HiddenCut. (iv) Completely dropping out the spans of hidden representations generated better results than only removing certain dimensions in the hidden space, which further validated the benefit of HiddenCut over DropBlock in natural language understanding tasks. 4.6.2 The Effect of HiddenCut Ratios The length of spans that are dropped by HiddenCut is an important hyper-parameter, which is controlled by the HiddenCut ratio α and the length of input sentences. α could also be interpreted as the extent of perturbations added to the hidden space. We presented the results of HiddenCut on MNLI with a set of different α including {0.05, 0.1, 0.2, 0.3, 0.4} in Table 5. HiddenCut achieved the best performance with α = 0.1, and 4387 Method Original and Counterfactual Sentences Prediction RoBERTa <s> I would rate 8 stars out of 10 </s> Positive HiddenCut <s> I would rate 8 stars out of 10 </s> Positive RoBERTa <s> The movie became more and more intriguing </s> Positive HiddenCut <s> The movie became more and more intriguing </s> Positive RoBERTa <s> I would rate 8 stars out of 20 </s> Positive HiddenCut <s> I would rate 8 stars out of 20 </s> Negative RoBERTa <s> The movie became only slightly more intriguing </s> Positive HiddenCut <s> The movie became only slightly more intriguing </s> Negative Table 4: Visualization of the attention weights at the last layer in models. The sentences in the first section are from IMDB with positive labels and the sentences in the second section is constructed by changing ratings or diminishing via qualifiers (Kaushik et al., 2020) to flip their corresponding labels. Deeper blue represents that those tokens receive higher attention weights. α 0.05 0.1 0.2 0.3 0.4 MNLI 88.07 88.23 88.13 88.07 87.64 Table 5: Performances on MNLI with different HiddenCut ratio α, which controls the length of span to cut in the hidden space. the performance gradually decreased with higher α since larger noise might be introduced when dropping more hidden information. This suggested the importance of balancing the trade-off between applying proper perturbations to regularize models and injecting potential noises. 4.6.3 The Effect of Sampling Ratios The number of words that are considered important and selected by HiddenCut is also an influential hyper-parameter controlled by the sampling ratio β and the length of input sentences. As shown in Table 6, we compared the performances on SST-2 by adopting different β including {0.1, 0.2, 0.4, 0.6}. When β is too small, the number of words in the important sets is limited, which might lead HiddenCut to consistently drop certain hidden spans during the entire training process. The low diversities reduce the improvements over baselines. When β is too large, the important sets might cover all the words except stop words in sentences. As a result, the Attention-based Strategy actually became Random Sampling, which led to lower gains over baselines. The best performance was achieved when β = 0.4, indicating a reasonable trade-off between diversities and efficiencies. 4.7 Visualization of Attentions To further demonstrate the effectiveness of HiddenCut, we visualize the attention weights that the special start token (“<s>”) assigns to other tokens at the last layer, via several examples and their counβ 0.1 0.2 0.4 0.6 SST-2 95.18 95.30 95.76 95.46 Table 6: Performances on SST-2 with different sampling ratio β, which controls the size of important token set from which HiddenCut would sample. terfactual examples in Table 4. We observed that RoBERTa only assigned higher attention weights on certain tokens such as “8 stars”, “intriguing” and especially the end special token “</s>”, while largely ignored other context tokens that were also important to make the correct predictions such as scale descriptions (e.g., “out of 10”) and qualifier words (e.g., “more and more”). This was probably because words like “8 stars” and “intriguing” were highly correlated with positive label and RoBERTa might overfit such patterns without probable regularization. As a result, when the scale of ratings (e.g., from “10” to “20”) or the qualifier words changed (e.g., from “more and more” to “only slightly more”), RoBERTa still predicted the label as positive even when the groundtruth is negative. With HiddenCut, models mitigated the impact of tokens with higher attention weights and were encouraged to utilize all the related information. So the attention weights in HiddenCut were more uniformly distributed, which helped models make the correct predictions for out-of-distribution counterfactual examples. Taken together, HiddenCut helps improve model’s generalizability by facilitating it to learn from more task-related information. 5 Conclusion In this work, we introduced a simple yet effective data augmentation technique, HiddenCut, to improve model robustness on a wide range of natural language understanding tasks by drop4388 ping contiguous spans of hidden representations in the hidden space directed by strategic attentionbased sampling strategies. Through HiddenCut, transformer models are encouraged to make use of all the task-related information during training rather than only relying on certain spurious clues. Through extensive experiments on indistribution datasets (GLUE benchmarks) and outof-distribution datasets (challenging counterexamples), HiddenCut consistently and significantly outperformed state-of-the-art baselines, and demonstrated superior generalization performances. Acknowledgment We would like to thank the anonymous reviewers, and the members of Georgia Tech SALT group for their feedback. This work is supported in part by grants from Amazon and Salesforce. References Armen Aghajanyan, Akshat Shrivastava, Anchit Gupta, Naman Goyal, Luke Zettlemoyer, and Sonal Gupta. 2020. Better fine-tuning by reducing representational collapse. arXiv preprint arXiv:2008.03156. Jacob Andreas. 2020. Good-enough compositional data augmentation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7556–7566, Online. Association for Computational Linguistics. David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and KlausRobert Müller. 2010. How to explain individual classification decisions. Journal of Machine Learning Research, 11(61):1803–1831. Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Liu, Yu Wang, Songhao Piao, Jianfeng Gao, Ming Zhou, et al. 2020. Unilmv2: Pseudo-masked language models for unified language model pre-training. arXiv preprint arXiv:2002.12804. Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632–642, Lisbon, Portugal. Association for Computational Linguistics. Jiaao Chen, Zhenghui Wang, Ran Tian, Zichao Yang, and Diyi Yang. 2020a. Local additivity based data augmentation for semi-supervised ner. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1241–1251. Jiaao Chen, Zichao Yang, and Diyi Yang. 2020b. MixText: Linguistically-informed interpolation of hidden space for semi-supervised text classification. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2147– 2157, Online. Association for Computational Linguistics. Liyan Chen, P. Gautier, and Sergül Aydöre. 2020c. Dropcluster: A structured dropout for convolutional networks. ArXiv, abs/2002.02997. Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. 2019. Electra: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations. Kevin Clark, Minh-Thang Luong, Christopher D. Manning, and Quoc V. Le. 2018. Semi-supervised sequence modeling with cross-view training. In EMNLP. Fahim Dalvi, Hassan Sajjad, Nadir Durrani, and Yonatan Belinkov. 2020. Analyzing redundancy in pretrained transformer models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4908– 4926, Online. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT. Matt Gardner, Yoav Artzi, Victoria Basmov, Jonathan Berant, Ben Bogin, Sihao Chen, Pradeep Dasigi, Dheeru Dua, Yanai Elazar, Ananth Gottumukkala, Nitish Gupta, Hannaneh Hajishirzi, Gabriel Ilharco, Daniel Khashabi, Kevin Lin, Jiangming Liu, Nelson F. Liu, Phoebe Mulcaire, Qiang Ning, Sameer Singh, Noah A. Smith, Sanjay Subramanian, Reut Tsarfaty, Eric Wallace, Ally Zhang, and Ben Zhou. 2020. Evaluating models’ local decision boundaries via contrast sets. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1307–1323, Online. Association for Computational Linguistics. G. Ghiasi, Tsung-Yi Lin, and Quoc V. Le. 2018. Dropblock: A regularization method for convolutional networks. In NeurIPS. Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572. Mitchell Gordon, Kevin Duh, and Nicholas Andrews. 2020. Compressing BERT: Studying the effects of weight pruning on transfer learning. In Proceedings of the 5th Workshop on Representation Learning for NLP, pages 143–155, Online. Association for Computational Linguistics. 4389 Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Counterfactual visual explanations. In ICML, pages 2376–2384. Demi Guo, Alexander M. Rush, and Yoon Kim. 2020. Parameter-efficient transfer learning with diff pruning. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654. Robin Jia and Percy Liang. 2016. Data recombination for neural semantic parsing. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12–22, Berlin, Germany. Association for Computational Linguistics. Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Tuo Zhao. 2019. Smart: Robust and efficient fine-tuning for pretrained natural language models through principled regularized optimization. arXiv preprint arXiv:1911.03437. Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2019. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64–77. Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. 2020. Learning the difference that makes a difference with counterfactually-augmented data. In International Conference on Learning Representations. Olga Kovaleva, Alexey Romanov, Anna Rogers, and Anna Rumshisky. 2019. Revealing the dark secrets of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4365–4374, Hong Kong, China. Association for Computational Linguistics. Gustav Larsson, M. Maire, and Gregory Shakhnarovich. 2017. Fractalnet: Ultra-deep neural networks without residuals. ArXiv, abs/1605.07648. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2020. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. SCL. Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142–150, Portland, Oregon, USA. Association for Computational Linguistics. Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. 2017. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448, Florence, Italy. Association for Computational Linguistics. Junghyun Min, R. Thomas McCoy, Dipanjan Das, Emily Pitler, and Tal Linzen. 2020. Syntactic data augmentation increases robustness to inference heuristics. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2339–2352, Online. Association for Computational Linguistics. Takeru Miyato, Andrew M. Dai, and Ian J. Goodfellow. 2017. Adversarial training methods for semi-supervised text classification. arXiv: Machine Learning. Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. 2018. Virtual adversarial training: a regularization method for supervised and semisupervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979– 1993. Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885–4901, Online. Association for Computational Linguistics. Hieu Pham and Quoc V. Le. 2021. Autodropout: Learning dropout patterns to regularize deep networks. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference 4390 on Knowledge Discovery and Data Mining, KDD ’16, page 1135–1144, New York, NY, USA. Association for Computing Machinery. Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. 2019. Adversarial training for free! In Advances in Neural Information Processing Systems, pages 3358–3369. Dinghan Shen, M. Zheng, Y. Shen, Yanru Qu, and W. Chen. 2020. A simple but tough-to-beat data augmentation approach for natural language understanding and generation. ArXiv, abs/2009.13818. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958. Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223. Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199. Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Transactions of the Association for Computational Linguistics, 8:621–633. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, L. Kaiser, and Illia Polosukhin. 2017. Attention is all you need. ArXiv, abs/1706.03762. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In BlackboxNLP@EMNLP. Yicheng Wang and Mohit Bansal. 2018. Robust machine comprehension models via adversarial training. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 575–581, New Orleans, Louisiana. Association for Computational Linguistics. Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. 2019. Unsupervised data augmentation for consistency training. arXiv preprint arXiv:1904.12848. Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. 2020. Unsupervised data augmentation for consistency training. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019a. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pages 5754– 5764. Ziyi Yang, Chenguang Zhu, and Weizhu Chen. 2019b. Parameter-free sentence embedding via orthogonal basis. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 638–648, Hong Kong, China. Association for Computational Linguistics. Chih-Kuan Yeh, Cheng-Yu Hsieh, Arun Sai Suggala, David I. Inouye, and Pradeep Ravikumar. 2019. On the (in)fidelity and sensitivity of explanations. In NeurIPS. Dinghuai Zhang, Tianyuan Zhang, Yiping Lu, Zhanxing Zhu, and Bin Dong. 2019a. You only propagate once: Painless adversarial training using maximal principle. arXiv preprint arXiv:1905.00877, 2(3). Yuan Zhang, Jason Baldridge, and Luheng He. 2019b. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298–1308, Minneapolis, Minnesota. Association for Computational Linguistics. Chen Zhu, Yu Cheng, Zhe Gan, Siqi Sun, Tom Goldstein, and Jingjing Liu. 2019. Freelb: Enhanced adversarial training for natural language understanding. In International Conference on Learning Representations.
|
2021
|
338
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4391–4401 August 1–6, 2021. ©2021 Association for Computational Linguistics 4391 Neural Stylistic Response Generation with Disentangled Latent Variables Qingfu Zhu♯, Weinan Zhang♯∗, Ting Liu♯, William Yang Wang♭ ♯Harbin Institute of Technology, Harbin, China ♭University of California, Santa Barbara, USA {qfzhu, wnzhang, tliu}@ir.hit.edu.cn [email protected] Abstract Generating open-domain conversational responses in the desired style usually suffers from the lack of parallel data in the style. Meanwhile, using monolingual stylistic data to increase style intensity often leads to the expense of decreasing content relevance. In this paper, we propose to disentangle the content and style in latent space by diluting sentence-level information in style representations. Combining the desired style representation and a response content representation will then obtain a stylistic response. Our approach achieves a higher BERT-based style intensity score and comparable BLEU scores, compared with baselines. Human evaluation results show that our approach significantly improves style intensity and maintains content relevance. 1 Introduction Linguistic style is an essential aspect of natural language interaction and provides particular ways of using language to engage with the audiences (Kabbara and Cheung, 2016). In human-bot conversations, it is crucial to generate stylistic responses for increasing user engagement to conversational systems (Gan et al., 2017). Currently, most of the existing parallel datasets are not stylistically consistent. Samples in these datasets are usually contributed by a variety of users, resulting in an averaging effect across style characteristics (Zhang et al., 2018a). Meanwhile, constructing a parallel stylistic dataset for training the open-domain conversational agents is both labor-intensive and time-consuming. Recent studies show the effect of stylizing responses using a monolingual dataset in the desired style and a conventional conversational dataset (Niu and Bansal, 2018; Gao et al., 2019b). However, increasing style intensity often leads to ∗Corresponding author. Dialogue History: A: Hello, this is <name> apartment office, what can I do for you? B: I want to rent an apartment. A: Do you want the whole lease or a shared lease? Content Relevance Style Intensity S2S: I just want to rent a room. Style Fusion: I hope I can share. Ours: I should prefer having a partner to being alone. S2S+LM: My friend had a considerable share in clearing the matter up. Figure 1: An example of responses generated by S2S, S2S+LM (Niu and Bansal, 2018), Style Fusion (Gao et al., 2019b), and our approach, targeting the Holmes style, which is quite formal and polite. the expense of decreasing content relevance between dialogue history and response. As an example in Figure 1 shows, Niu and Bansal (2018) independently train a response generation model and a stylistic language model and subsequently interpolates them in the inference phase. Lacking the interaction between the stylistic language model and response generation encoder, it usually yields a trade-off between style intensity and content relevance. Gao et al. (2019a,b) fuse a structured latent space where the direction denotes the diversity, and the distance denotes style intensity and content relevance. The main issue is that style intensity and content relevance are contradictory in measurement but are coupling to the same “distance” metric of the latent space. To sum up, the key issue of the above studies is the improper entanglement of style and content. To address the issue, we propose to disentangle the style and content of a response. The disentanglement is conducted on the structured latent space, where each sentence (dialogue history, response, 4392 and stylistic sentence) is projected into a vector representation. We further split the representation into two components: style and content representations. The former is a corpus-level feature since sentences within a dataset have the same style. In contrast, the content representation is a sentence-level feature decided by a sentence itself. We thus disentangle the content and style by diluting sentence-level information in the style representation. This encourages the encoding of content information into the content representation. Otherwise, the content information will be corrupted in the style representation, making it hard to reconstruct the original content in the subsequent decoding process. We conduct experiments on DailyDialogue conversational dataset (Li et al., 2017) and Holmes monolingual stylistic dataset (Gao et al., 2019b). Experimental results show that our proposed approach improves style intensity and maintains content relevance. Our contributions are listed below: • We propose a unified framework to simultaneously improve style intensity and maintain content relevance for neural stylistic response generation. • We introduce a scheme of learning latent variables by a diluting strategy to disentangle the style and content. • Experimental results show that our approach achieves higher performance in style intensity without decreasing content relevance, compared with previous approaches. 2 Method 2.1 Task Definition The task of stylistic response generation is defined as follows: given a monolingual stylistic dataset S = {S1, ..., SN}1 and a conversational dataset C = {(X1, Y1), ..., (XM, YM)}, where Si, Xi, and Yi denote a stylistic sentence, dialogue history, and a response respectively, the goal is to learn a generation model P( ˆY |X), where ˆY is a generated response expected to be in the style of S (called the desired style in the following sections). We will first briefly review the concept of structured latent space and then introduce our disentanglement approach. 1Throughout the paper, we use bold letters to denote vectors, i.e., V = {V1, V2, ..., VN}. ^ 1 2 ZS2S(Xi) ZAE(Yi) ZAE(Yi) Z(Yi) ZAE(Sj) Structured Latent Space Xi : We are going hiking this weekend. Do you want to join us? Yi : Yes, of course. Yi : I don’t like hiking. Yi : I would like to join. Sj : Thanks for your help. I would like to book. 1 2 ^ Figure 2: An example of a dialogue in the structured latent space. The center point corresponds to the dialogue history representation ZS2S(Xi). The k-th response representation ZAE(Y k i ) (denoted by a black point) is optimized to be distributed around ZS2S(Xi). The red point ZAE(Sj) and the purple point Z( ˆYi) are representations of a monolingual stylistic sentence and a stylistic response, respectively. 2.2 Background: Structured Latent Space Overview The structured latent space is constructed by two main mechanisms: (i) sharing a decoder between a sequence-to-sequence (S2S) model and an auto-encoder (AE), and (ii) fusion and smoothness objectives. As an example in Figure 2 shows, a response representation ZAE(Yi) is regularized by the two mechanisms to be distributed around its dialogue history representation ZS2S(Xi). The notations ZAE(·) and ZS2S(·) denote the representations computed by AE encoder and S2S encoder, respectively. Such a latent space makes it possible to predict a response ˆY by sampling nearby the dialogue history representation. Based on that, Gao et al. (2019b) further align stylistic sentence representations into the latent space, which improves the style intensity of generated responses. In summary, the construction of the structred latent space is a process of aligning the three spaces (ZS2S(Xi), ZAE(Yi), and ZAE(Sj)) by two mechanisms (sharing the decoder, and fusion and smoothness objectives). Fusion Objective cross-aligns sentences of different spaces. Since Xi and Yi are paired, we align them by minimizing their pair-wise dissimilarity: dconv = X i∈batch dE(ZS2S(Xi), ZAE(Yi)) n √ l , (1) where dE denotes the Euclidean distance, n is the batch size, and l is the dimensionality of the latent space. In contrast, the pair-wise dissimilarity can4393 not be applied to stylistic sentences since they are not paired with conversational data. To this end, the fusion objective instead optimizes the nearest neighbor distance between the two datasets: dstyle =1 2dcross NN ({ZS2S(Xi)}, {ZAE(Sj)}) +1 2dcross NN ({ZAE(Sj)}, {ZS2S(Xi)}), (2) where dcross NN ({ai}, {bj}) denotes the batch average distance between ai and its nearest neighbor in the set {bj}. To further encourage the representations spread-out the latent space, a inner-distance loss is introduced: dspread-out = min{dinner NN (ZS2S(Xi)), dinner NN (ZAE(Yi)), dinner NN (ZAE(Sj))}, (3) where dinner NN ({ai}) denotes the batch average distance between ai and its nearest neighbor in the set {ai}. The final fusion objective is defined as: Lfuse = dconv + dstyle −dspread-out. (4) Smoothness Objective aims to make the structured latent space a continuous space, where each point can decode a natural sentence. Given three discrete points ZS2S(Xi), ZAE(Yi), and ZAE(Sj), the objective encourages points in the area between ZS2S(Xi) and ZAE(Yi) to generate Yi: Zconv = UZS2S(Xi) + (1 −U)ZAE(Yi) + ϵ, Lsmooth,conv = −log P(Yi|Zconv), (5) where ϵ ∼N(0, σ2I), and U ∼U(0, 1). Meanwhile, as a point moves from ZAE(Yi) to ZAE(Sj), the corresponding generation is expected to gradually move from Yi to Sj: Zstyle = UZAE(Yi) + (1 −U)ZAE(Sj) + ϵ Lsmooth,style = −U log P(Yi|Zstyle) −(1 −U) log P(Sj|Zstyle). (6) The smoothness objective Lsmooth is the sum of Lsmooth,conv and Lsmooth,style, and is added to the final loss function along with the fusion objective and response generation loss of S2S. 2.3 Our Method Despite aligning monolingual stylistic sentences into the structured latent space helps stylize generated responses, their style intensity is still limited. We conjecture this is due to the coupling of the style and the content in sentence representations. To this end, we propose to disentangle the two aspects in the structured latent space. In our proposed approach, a sentence representation Z ∈Rl in the latent space consists of two components: content representation Zc ∈Rlc and style representation Zs ∈Rls, where l is the dimensionality of latent space and lc + ls = l. Zs encodes all the style information of a sentence. It is a corpus-level feature because Zs for different sentences in the same corpus should be similar. In contrast, Zc can be seen as a sentence-level feature which only decided by the content of its corresponding sentence. Figure 3 shows an example of our approach, where Zc and Zs can be seen as two “containers”. Colored squares represent the content and style information. We encourage the disentanglement of the two types of information by diluting sentence-level content information in Zs. As an example in Figure 3 (a) shows, the content and style information may be mixed in both Zc and Zs. During the decoding process of a sentence, i.e., Yi, we replace its style representation Zs AE(Yi) with its batch average style representation ¯ Zs AE(Yi) = 1 n P j∈batch Zs AE(Yj). In this way, its sentence-level content information will be diluted since it greatly varies from other sentences’ content information, which introduces extra noise. In contrast, its corpus-level style information, which is similar to that of other sentences within the batch, will remain unaffected. As the training processes, the content information will be encouraged to be encoded into Zc where it can remain unchanged, as an example in Figure 3 (b) shows. Otherwise, the content information will be corrupted in Zs, making it hard to recover the content of Yi. As a result, the encoding process will be punished by the response generation loss of S2S and the reconstruction loss of AE, as shown in Figure 3 (a). Based on that, we update the response generation process by replacing its style representation Zs with the corresponding batch average style representation ¯ Zs: LS2S = −log P(Yi|[Zc S2S(Xi) : ¯ Zs S2S(Xi)]), (7) where the bracket [:] denotes concatenation. The decoding process in the smoothness objective is updated similarly. Note that when we move from 4394 Content Information of Sentence #1 (S1) Content Information of Sentence #2 (S2) Style Information Content Representation Zc Style Representation ZS S1: Could you please tell me how I can go job-hunting in the web? Encoding S1: Could you please tell me how I can go job-hunting in the web? Encoding Average Style Representation Average Style Representation Decoding Decoding S1: Could you please tell me how I can put my bags? S1: Could you please tell me how I can go job-hunting in the web? S2: Sorry, sir, where shall I put my bags? S2: Sorry, sir, where shall I put my bags? S2: Sorry, sir, where shall I put my bags? S2: Sorry, how I can put my bags? Figure 3: An example of disentangling content and style. The purple block is the content information of the first sentence. The yellow block is the content information of the second sentence. Style information in both two sentences is denoted by red blocks as it is a corpus-level feature shared among samples within the corpus. (a): A negative example whose content and style information is mixed in Zc and Zs. Its content information is corrupted after averaging Zs within the batch and fails to recover the input content. (b): A positive example. Content information in Zc and style information in Zs will not be affected after averaging Zs. Yi to Sj, and from Xi to Yi, we only interpolate their content representations Zc in the latent space: Zc conv = UZc S2S(Xi) + (1 −U)Zc AE(Yi) + ϵ, Zc style = UZc AE(Yi) + (1 −U)Zc AE(Sj) + ϵ. (8) The batch average style representation ¯ Zs remains consistent with the target, i.e., being ¯ Zs AE(Sj) when the target is Sj. The updated smoothness objective is as follows: Lsmooth,conv = −log P(Yi|[Zc conv : ¯ Zs AE(Yi)]), Lsmooth,style = −U log P(Yi|[Zc style : ¯ Zs AE(Yi)]) −(1 −U) log P(Sj|[Zc style : ¯ Zs AE(Sj)]). (9) The final training loss is the sum of the response generation loss, fusion objective, and smoothness objective: L =LS2S + Lfuse + Lsmooth. (10) Here, we do not employ pre-training models, i.e., DialoGPT (Zhang et al., 2020b) and OpenAI GPT2 (Radford et al., 2019). This is because the disentanglement is usually conducted on a sentence representation. While most of the pre-training models depend on the attention mechanism, and there is no static global sentence representation during the decoding process. 2.4 Inference To generate a stylistic response ˆYi given dialogue history Xi during the inference process, we first obtain Zc S2S(Xi) by S2S encoder and subsequently sample Zc( ˆYi) from the hypersphere of Zc S2S(Xi) with a mannually tuned radius r. After that, we generate ˆYi by concatenating Zc( ˆYi) and ¯ Zs AE(Sj), which is the batch average style representation of randomly sampled stylistic sentences. Considering the discrepancy between training and inference that content and style representations in different corpora have never been concatenated for generation, we propose a soft combination approach to introduce the desired style by interpolating Zs S2S(Xi) and ¯ Zs AE(Sj): Zs soft = Zs S2S(Xi) + α ∗¯ Zs AE(Sj), (11) where α is the weight of the desired style. After that, ˆYi is generated by the decoder whose hidden state is set to [Zc( ˆYi) : Zs soft]. To further balance style intensity and content relevance, we also employ the re-ranking strategy following Gao et al. (2019b). It samples Ny candidate responses and re-ranks them by: sr = γ ∗PS2S( ˆYi|Xi)+(1−γ)∗Pstyle( ˆYi), (12) where PS2S( ˆYi|Xi) is the generation probability under a S2S model measuring the relevance. Pstyle( ˆYi) is the probability that ˆYi has the desired style. It is a interpolation between the probabilities of a neural-based classifier and a n-gram classifier: Pstyle( ˆYi) =η ∗Pneural( ˆYi) + (1 −η) ∗ N X n=1 wn ∗Pn-gram( ˆYi), (13) 4395 Training Dialogues 11,118 Validation Dialogues 1,000 Test Dialogues 1,000 Average Tokens Per Dialogue 114.7 Average Tokens Per Utterance 14.6 Table 1: Statistics of the DailyDialog dataset. where wn is a weight which is set to the accuracy of the corresponding classifier. 3 Experiments 3.1 Data Conversational Dataset We employ DailyDialog2 (Li et al., 2017) as our conversational dataset C. It is a human-written multi-turn dataset covering various topics of daily life. Table 1 shows some statistics of its training, validation, and test set. We split dialogue of K utterances into K-1 samples. Each sample consists of at most three continuous utterances. The last utterance of a sample is regarded as the response. The previous utterances of the response are concatenated as its dialogue history. Here, Reddit dataset is not employed as Gao et al. (2019b) because the post-reply format data collected from social networks is noisy and different from real conversations (Li et al., 2017). Monolingual Stylistic Dataset Following Gao et al. (2019b), we use Holmes3 as the stylistic dataset S. It is collected from the Sherlock Holmes novel series and consists of roughly 38k sentences. We do not use the arXiv dataset as it contains too many special tokens, i.e., equations, and incomplete sentences, such as “is concerned” and “exactly identical restrictions”. 3.2 Baselines We compare the proposed approach with the following baselines: • S2S, the sequence-to-sequence response generation model (Shang et al., 2015). • S2S+LM, a S2S trained on C and a stylistic language model trained on S (Niu and Bansal, 2018). During the inference process, it generates a stylistic response by interpolating outputs of the two models. 2http://yanran.li/dailydialog 3https://github.com/golsun/StyleFusion Model Time (s) # of parameters S2S 4.55 63M Style Fusion 4.60 75M Ours 4.60 75M Table 2: The average running time (in seconds per batch) and the number of parameters. • Style Fusion, a multi-task learning based model whose latent space fuses dialogue history, responses, and stylistic sentences with a specific structure (Gao et al., 2019b). Note that we do not consider the Label-FineTuning model and Polite Reinforcement Learning model (Niu and Bansal, 2018), because they require some training samples in the conversational dataset to have the desired style (Gao et al., 2019b). 3.3 Experiment Settings We implement the proposed approach based on the released code of Style Fusion model4. The vocabulary table consists of the most frequent 20,000 words. S2S encoder, AE encoder, and the shared decoder are two-layer LSTMs. The number of their hidden units is 1000, which is also the size of the structured latent space. The dimension of Zc and Zs is 950 and 50, respectively. The maximum length is set to 90 for the dialogue history and 30 for the response. During the training process, we use the ADAM optimizer, whose learning rate is 0.0003. σ2 for sampling ϵ in Equation 8 is 0.12. Table 2 shows the average running time on a single TITAN X (Pascal) GPU. During the inference process, the weights γ and η for re-ranking are set to 0.5. The weight (accuracy) of n-gram classifier is 0.93, 0.87, 0.77, and 0.65 for n from 1 to 4. The number of candidate responses, Ny, is set to 10. The radius r is set to 3. 4 Results 4.1 Evaluation Metrics Automatic Evaluation Considering that it is unfair to evaluate a response by the classifiers that are used for selecting the response (Song et al., 2020), we fine-tune a BERT (Devlin et al., 2019) to measure style intensity. Concretely, positive samples are the stylistic sentences. Negative samples are 4https://github.com/golsun/StyleFusion 4396 Model SI(%) Dist-1 Dist-2 BLEU-3 BLEU-4 Mean S2S (Shang et al., 2015) 6.32 0.035 0.227 0.70 0.20 0.10 S2S+LM (Niu and Bansal, 2018) 32.79 0.015 0.086 0.55 0.08 0.13 Style Fusion (Gao et al., 2019b) 10.58 0.043 0.280 0.82 0.22 0.14 Ours (α=0.25) 11.91 0.041 0.275 0.79 0.23 0.16 Ours (α=0.50) 20.67 0.040 0.275 0.64 0.17 0.19 Ours (α=0.75) 34.85 0.038 0.285 0.47 0.10 0.16 Table 3: Automatic evaluation results of SI, Dist-1, Dist-2, and BLEU. The last column is the harmonic mean of SI and BLEU-4 measuring the overall performance of style intensity and content relevance. 0 0.1 0.2 0.3 0.4 0 0.25 0.5 0.75 SI BLEU-4 Mean Figure 4: The trade-off between style intensity measured by SI and content relevance measured by BLEU4. The x-axis corresponds to α. The harmonic mean achieves the maximum around α=0.5. randomly selected from DailyDialog’s responses, which are of the same amount of sentences as the positive samples. Given the fine-tuned BERT classifier (whose accuracy achieves 0.96 on the validation set), we report the average probability of responses being positive as a measurement of the style intensity. For brevity, we denote this metric as SI. The content relevance is evaluated by BLEU. Since it may correlate weakly with human judgments of quality in a single reference setting (Liu et al., 2016), we employ the expanded responses in multi-reference DailyDialog test set (Gupta et al., 2019) as references to alleviate the problem. Meanwhile, we evaluate the diversity by Dist-k (Li et al., 2016), which is the number of distinct k-grams normalized by the total number of words of responses. Human Evaluation We randomly sample 200 messages from the test set of C to conduct the human evaluation from two aspects: style intensity and content relevance. Each aspect is independently evaluated by five Amazon Mechanical Turk (AMT)5 workers whose approval rate is greater than 95%, and the number of approved is greater than 500. Given dialogue history and two responses generated by a baseline and our approach, the workers are asked to give a preference of which one is 5https://www.mturk.com Content Relevance Style Intensity Win Lose Win Lose vs. S2S 40.21 39.79 49.37 36.84 vs. S2S+LM 65.00 20.00 53.30 32.50 vs. Style Fusion 43.32 42.67 48.77 36.68 Table 4: Pair-wise human evaluation results of content relevance and style intensity. better (ties are also permitted). 4.2 Results Figure 4 shows the trade-off between style intensity and content relevance in our approach. There is an improvement in SI and a decrease in BLEU associated with the increase of α in Equation 11. To assess the overall performance, we also compute their harmonic mean, whose maximum lies around α = 0.5. We thus conduct the human evaluation and analysis in this parameter setting. We report the human evaluation results in Table 4. Our approach is clearly preferred in style intensity because the percentage of Win is significantly higher than that of Lose (p <0.001, T-test). In terms of content relevance, the ratios of Win in “vs. S2S” and “vs. Style Fusion” are similar to those of Lose. This suggests that our approach can significantly improve the style intensity without decreasing the content relevance. In contrast, S2S+LM loses in most of the cases in the content relevance. Following Zhou et al. (2018) and Ke et al. (2018), we evaluate the agreement of annotators via inter-rater consistency. The percentage of samples that at least three annotators have the same preference (3/5 agreement) is 81.80%. And the percentage for 4/5 agreement is 32.15%. Table 3 shows the results of the automatic evaluation. Our approach has the highest mean score, which indicates that it achieves the best overall performance. S2S+LM has a high SI score, but its BLEU scores are not as good as others, i.e., S2S. 4397 SI BLEU-3 BLEU-4 Mean Full Model 11.71 0.67 0.17 0.14 -Disentangle 7.52 0.68 0.17 0.11 -Lfuse 6.46 0.59 0.15 0.09 -Lsmooth 6.02 0.63 0.17 0.09 Table 5: Results of the ablation study. Style Fusion Ours Stylistic Samples Conversational Samples Figure 5: MDS visualization of Zs (black) and three continuous sub-sequences extracted from the head (yellow), middle (red), and tail (blue) of Zc. This is in line with our human evaluation results and Niu and Bansal (2018)’s observation that biasing a decoder with a stylistic language model may harm the content relevance. In contrast, our approach (α = 0.25) significantly outperforms S2S and is comparable to Style Fusion. By increasing α to 0.5, the BLEU score drops slightly but is comparable to baselines (evidenced by the human evaluation results). Meanwhile, there is a significant improvement (up to 95.37%) in SI comparing with Style Fusion. This verifies the effectiveness of our disentanglement approach in improving the style intensity and maintaining the content relevance. Besides, the Dist-k results in Table 3 also indicate that the diversity of our approach is comparable to the best-performed Style Fusion. 4.3 Ablation Study We conduct ablation studies to investigate the contributions of the fusion objective, smoothness objective, and our disentanglement approach. To focus on their effects on the generation process, in this section, we sample a single response without using the re-ranking strategy (Equation 12). Table 5 shows the results of the ablation study. There is a significant decline in SI and a slight change in BLEU-3 and BLEU-4 after removing each component. This indicates that a multi-task learning architecture without the three components [Zc : Zs] Zs Style Fusion 0.83 0.72 (-13.02%) Ours 0.88 0.86 (-1.71%) Table 6: Style classification accuracy of the full latent variable ([Zc : Zs]) and Zs. can achieve a good content relevance performance but fails to stylize a response. By removing the disentanglement component, our approach degenerates into Style Fusion. In this case, the SI score decreases significantly while BLEU scores are nearly unchanged, which demonstrates the disentanglement could improve the style intensity and maintain the relevance at the same time. The decreases in SI after removing the fusion objective and smoothness objective are more significant than that after removing the disentanglement. This is because the two objectives are bottom components for constructing the structured latent space, where our approach and Style Fusion are built upon. 4.4 Analysis In this section, we analyze whether style information is disentangled into Zs. To achieve this goal, we train style classifiers taking as input a latent variable and use the validation accuracy as an indicator. Taking our approach as an instance, we first freeze the parameters of our well-trained model. Then we independently learn two style classifiers whose inputs are the full latent variable ([Zc : Zs]) and Zs respectively. Note that Zc and Zs in Style Fusion are a simple partition of its latent variable. There are not any disentanglement approaches applied to obtain the two representations. As shown in Table 6, Style Fusion achieves 0.83 validation accuracy training on its full latent variable. And the accuracy decreases by 13.02% when the classification is only based on Zs. In contrast, the decrease of our approach is only 1.71%, indicating that most of the style information is disentangled into Zs. We show a visualization of the disentanglement of the latent variable by MDS (Borg and Groenen, 2005) in Figure 5. Each figure consists of Zs (black) and three continuous sub-sequences extracted from the head (yellow), middle (red), and tail (blue) of Zc. The sub-sequences are of the same length with Zs. For both stylistic and conversational samples, all the sub-sequences and Zs are mixed in Style Fusion. In contrast, there is a clear separation between Zs and the sub-sequences 4398 Dialogue – Yes , after my graduation, History I worked in a trade company in Macao for one year. – What section did you work in? – The export business section. S2S – Have you received any other honors? S2S+LM – Did you ever find a job that you did not? Style Fusion – What was your major in? Ours – I ’m glad that you have been doing great. Dialogue – I will always have you come History in for a discussion if there is a major problem. – Will I be able to read the results on the website and understand them? – Yes, if you go there, you can see what each test is about. S2S – OK, I will ask the manager to take care of that. S2S+LM – I’m not sure I’ll have you to do that, Style Fusion – What are my parents for me? Ours – I am looking forward to getting the information about that. Table 7: An example of responses generated by baselines and our approach. in our approach. This is because most of the style information is disentangled into Zs in our approach, making its distribution different from sub-sequences of Zc. 4.5 Case Study Table 7 shows some examples of generated responses. There is no significant Holmes style in responses of S2S. Similarly, the style intensity of responses in Style Fusion is also limited. The semantics of S2S+LM’s response in the first example is not very clear, making it less relevant to the dialogue history than other responses. We believe this is also due to the lack of interaction between the response generation encoder and the stylistic language model. In contrast, our approach not only achieves a good content relevance performance but also has a significant Holmes style, which is quite polite and formal. 5 Related Work 5.1 Text Style Transfer without Parallel Data The task of text style transfer aims at transferring the style of a sentence while preserving its meaning. One way is to disentangle the content and style, and subsequently combine the content with the desired style. The disentanglement can be achieved by adversarial learning (Shen et al., 2017; Hu et al., 2017; Fu et al., 2018; Yang et al., 2018; Logeswaran et al., 2018), reinforcement learning (Jain et al., 2019), back-translation (Prabhumoye et al., 2018; Nogueira dos Santos et al., 2018), multi-task learning (John et al., 2019), and removing stylistic phrases (Li et al., 2018; Xu et al., 2018; Zhang et al., 2018b). The other way transfers the style without disentangled representations, for example using generator-evaluator architecture (Gong et al., 2019), cycle reconstruction (Dai et al., 2019), parameter sharing (Wang et al., 2020), and data augmentation (Zhang et al., 2020a). The main difference between our task and text style transfer lies in two aspects. First, all the content to be generated is available in the input in text style transfer, while our task needs to create new (response) content. And the key is content relevance to the dialogue history, rather than content preservation of the input. Second, the data for text style transfer is isomorphic. Data in different styles are in the same free-text format. However, our conversational data are context-response pairs while the stylistic data are free-texts, which is heterogeneous and requires more sophisticated structures, i.e., the structured latent space (Gao et al., 2019b). 5.2 Stylistic Response Generation without Parallel Stylistic Data Niu and Bansal(2018) propose three weaksupervised models based on reinforcement learning, conditional text generation, and language model. Gao et al. (2019b) fuses the latent spaces of a response generation model and a stylistic autoencoder to improve the style intensity of sampled responses. Yang et al. (2020) inject the style information by introducing a word-level KL loss and a sentence-level style classifier to the fine-turning process of DialoGPT (Zhang et al., 2020b). Distinct from previous work, we explicitly disentangle the style and content in the latent space and employ a unified architecture to jointly optimize the style intensity and content relevance. 6 Conclusion We propose a uniform framework to simultaneously improve the style intensity and maintain the content relevance for neural stylistic response generation. In contrast to existing approaches, our approach 4399 disentangles the style and the content in the latent space by a diluting strategy. Experiments show that our approach improves the style intensity of generated responses and maintains the content relevance at the same time, which demonstrates the effectiveness of this approach. Acknowledgments The authors would like to thank all the anonymous reviewers for their insightful comments. The authors from HIT are supported by the National Natural Science Foundation of China (No. 62076081, No. 61772153, and No. 61936010) and Science and Technology Innovation 2030 Major Project of China (No. 2020AAA0108605). The author from UCSB is not supported by any of the projects above. Ethical Statement This paper honors the ACL Code of Ethics. Stylistic response generation intends to improve the engagement of a dialogue system in human-bot conversations. It responds to users with the desired style, i.e., being polite, humorous, or romantic, rather than imitating any specific person. Meanwhile, style is a linguistic aspect of natural language interaction. There is not any identity characteristic being used as a variable. References Ingwer Borg and Patrick JF Groenen. 2005. Modern multidimensional scaling: Theory and applications. Springer Science & Business Media. Ning Dai, Jianze Liang, Xipeng Qiu, and Xuanjing Huang. 2019. Style transformer: Unpaired text style transfer without disentangled latent representation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5997–6007. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. 2018. Style transfer in text: Exploration and evaluation. In Thirty-Second AAAI Conference on Artificial Intelligence. Chuang Gan, Zhe Gan, Xiaodong He, Jianfeng Gao, and Li Deng. 2017. Stylenet: Generating attractive visual captions with styles. In Proceedings of CVPR, pages 955–964. IEEE. Xiang Gao, Sungjin Lee, Yizhe Zhang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2019a. Jointly optimizing diversity and relevance in neural response generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1229–1238, Minneapolis, Minnesota. Association for Computational Linguistics. Xiang Gao, Yizhe Zhang, Sungjin Lee, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2019b. Structuring latent spaces for stylized response generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 1814–1823, Hong Kong, China. Association for Computational Linguistics. Hongyu Gong, Suma Bhat, Lingfei Wu, JinJun Xiong, and Wen-mei Hwu. 2019. Reinforcement learning based text style transfer without parallel training corpus. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pages 3168–3180. Prakhar Gupta, Shikib Mehri, Tiancheng Zhao, Amy Pavel, Maxine Eskenazi, and Jeffrey Bigham. 2019. Investigating evaluation of open-domain dialogue systems with human generated multiple references. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pages 379–391, Stockholm, Sweden. Association for Computational Linguistics. Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P Xing. 2017. Toward controlled generation of text. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1587–1596. JMLR. org. Parag Jain, Abhijit Mishra, Amar Prakash Azad, and Karthik Sankaranarayanan. 2019. Unsupervised controllable text formalization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 6554–6561. Vineet John, Lili Mou, Hareesh Bahuleyan, and Olga Vechtomova. 2019. Disentangled representation learning for non-parallel text style transfer. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 424–434. Jad Kabbara and Jackie Chi Kit Cheung. 2016. Stylistic transfer in natural language generation systems using recurrent neural networks. In Proceedings of the Workshop on Uphill Battles in Language Processing: Scaling Early Achievements to Robust Methods, pages 43–47. 4400 Pei Ke, Jian Guan, Minlie Huang, and Xiaoyan Zhu. 2018. Generating informative responses with controlled sentence function. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1), pages 1499–1508. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119. Juncen Li, Robin Jia, He He, and Percy Liang. 2018. Delete, retrieve, generate: a simple approach to sentiment and style transfer. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics,Volume 1, pages 1865–1874. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986–995. Chia-Wei Liu, Ryan Lowe, Iulian Serban, Mike Noseworthy, Laurent Charlin, and Joelle Pineau. 2016. How NOT to evaluate your dialogue system: An empirical study of unsupervised evaluation metrics for dialogue response generation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2122–2132, Austin, Texas. Association for Computational Linguistics. Lajanugen Logeswaran, Honglak Lee, and Samy Bengio. 2018. Content preserving text generation with attribute controls. In Advances in Neural Information Processing Systems, pages 5103–5113. Tong Niu and Mohit Bansal. 2018. Polite dialogue generation without parallel data. Transactions of the Association for Computational Linguistics, 6:373–389. Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhutdinov, and Alan W Black. 2018. Style transfer through back-translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1), pages 866–876. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Cicero Nogueira dos Santos, Igor Melnyk, and Inkit Padhi. 2018. Fighting offensive language on social media with unsupervised text style transfer. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2), pages 189–194, Melbourne, Australia. Lifeng Shang, Zhengdong Lu, and Hang Li. 2015. Neural responding machine for short-text conversation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1), pages 1577–1586. Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2017. Style transfer from non-parallel text by cross-alignment. In Advances in neural information processing systems, pages 6830–6841. Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating persona consistent dialogues by exploiting natural language inference. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8878–8885. Yunli Wang, Yu Wu, Lili Mou, Zhoujun Li, and Wenhan Chao. 2020. Formality style transfer with shared latent space. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2236–2249. Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xuancheng Ren, Houfeng Wang, and Wenjie Li. 2018. Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, Volume 1, pages 979– 988. Ze Yang, Wei Wu, Can Xu, Xinnian Liang, Jiaqi Bai, Liran Wang, Wei Wang, and Zhoujun Li. 2020. StyleDGPT: Stylized response generation with pretrained language models. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1548–1559, Online. Association for Computational Linguistics. Zichao Yang, Zhiting Hu, Chris Dyer, Eric P Xing, and Taylor Berg-Kirkpatrick. 2018. Unsupervised text style transfer using language models as discriminators. In Advances in Neural Information Processing Systems, pages 7287–7298. Ye Zhang, Nan Ding, and Radu Soricut. 2018a. SHAPED: Shared-private encoder-decoder for text style adaptation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pages 1528–1538. Yi Zhang, Tao Ge, and Xu Sun. 2020a. Parallel data augmentation for formality style transfer. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3221– 3228, Online. Association for Computational Linguistics. Yi Zhang, Jingjing Xu, Pengcheng Yang, and Xu Sun. 2018b. Learning sentiment memories for sentiment modification without parallel data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1103–1108. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing 4401 Liu, and Bill Dolan. 2020b. DIALOGPT : Largescale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270– 278, Online. Association for Computational Linguistics. Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018. Commonsense knowledge aware conversation generation with graph attention. In the 27th International Joint Conference on Artificial Intelligence and the 23rd European Conference on Artificial Intelligence, pages 4623–4629.
|
2021
|
339
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 404–414 August 1–6, 2021. ©2021 Association for Computational Linguistics 404 A Training-free and Reference-free Summarization Evaluation Metric via Centrality-weighted Relevance and Self-referenced Redundancy Wang Chen1∗ Piji Li2 Irwin King1 1Department of Computer Science and Engineering, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong 2Tencent AI Lab 1{wchen, king}@cse.cuhk.edu.hk [email protected] Abstract In recent years, reference-based and supervised summarization evaluation metrics have been widely explored. However, collecting human-annotated references and ratings are costly and time-consuming. To avoid these limitations, we propose a training-free and reference-free summarization evaluation metric. Our metric consists of a centralityweighted relevance score and a self-referenced redundancy score. The relevance score is computed between the pseudo reference built from the source document and the given summary, where the pseudo reference content is weighted by the sentence centrality to provide importance guidance. Besides an F1-based relevance score, we also design an Fβ-based variant that pays more attention to the recall score. As for the redundancy score of the summary, we compute a self-masked similarity score with the summary itself to evaluate the redundant information in the summary. Finally, we combine the relevance and redundancy scores to produce the final evaluation score of the given summary. Extensive experiments show that our methods can significantly outperform existing methods on both multi-document and single-document summarization evaluation. The source code is released at https://github.com/Chen-WangCUHK/Training-Free-and-Ref-Free-SummEvaluation. 1 Introduction Text summarization systems have been developed rapidly due to the appearance of sequence-tosequence frameworks (Sutskever et al., 2014; Bahdanau et al., 2015; See et al., 2017; Chan et al., 2020), transformer architectures (Vaswani et al., 2017) and large-scale pre-training models (Devlin et al., 2019; Liu et al., 2019). How to accurately ∗This work was mainly done when Wang Chen was an intern at Tencent AI Lab. evaluate the summaries generated from these systems also attracts more and more attention in this research area. One of the most accurate evaluation methods is human evaluation. However, human evaluation is expensive, time-consuming, and nonreproducible. Thus, it is necessary to develop automatic evaluation metrics for text summarization systems. Existing automatic summarization evaluation metrics can be roughly categorized into two groups: reference-based metrics and reference-free metrics. In this work, we focus on reference-free metrics. Reference-free summarization evaluation metrics have been developed in parallel in multidocument summarization and single-document summarization. The SOTA reference-free method for multi-document summarization evaluation, SUPERT (Gao et al., 2020), predicts a relevance score for each (document, summary) pair to estimate the informativeness of the summary and then averages all the scores from multiple documents as the final evaluation score. For each pair, SUPERT employs the top-ranked sentences which are ranked by the position or centrality as a pseudo reference of the document and then applies BERTScore (Zhang et al., 2020) to produce a relevance score between the pseudo reference and the given summary. The SOTA single-document summarization referencefree evaluation metric, LS Score (Wu et al., 2020), combines a learned linguistic scorer for the summary and a cosine similarity scorer for the (document, summary) pair to produce the final score. Although SUPERT and LS Score achieve the SOTA performance on their own areas respectively, they still have several drawbacks. For example, SUPERT only considers the relevance score between the document and the summary while ignoring the other aspects such as how much redundant information is contained in the summary. Besides, SUPERT assumes that all pseudo reference sen405 tences are equally-important. However, in the real world, the key information of a document is unevenly distributed over sentences. Therefore, such an assumption may introduce extra noise for the evaluation. Note that although SUPERT may employ sentence centrality to select document sentences as a pseudo reference, they ignore the sentence centrality after the selection and still treat the selected sentences equally-important. As for LS Score, although it does not require a reference during the evaluation of a summary, it requires a large-scale training dataset with reference summaries to train the linguistic scorer. Besides the intrinsic drawbacks in these SOTA methods, to our best knowledge, there is no reference-free evaluation metric showing that it can achieve the SOTA performance on both multi-document and singledocument summarization. To solve the above limitations, based on SUPERT, we propose a novel training-free and reference-free metric for both multiple and single document summarization evaluation. Our metric is composed of a centrality-weighted relevance score and a self-referenced redundancy score. For the relevance score which is employed to estimate the informativeness of the summary, we incorporate the following new features. First, unlike previous work which only utilizes the tokenlevel representations, motivated by Clark et al. (2019), we engage a hybrid way that contains both token-level representations and sentence-level representations to encode the document and the summary. The purpose of the hybrid representation is to enable our method to consider richer mapping styles (i.e., token-to-token, sentence-to-token, and sentence-to-sentence) and help to produce a more comprehensive evaluation score. Second, we utilize the sentence centrality computed from sentence-level representations of the source document to produce the importance weights of the pseudo reference sentences and tokens. Based on the weights, we compute a weighted relevance score that is more precise by considering the relative importance. Third, besides the F1 version of our relevance score, we also propose an adaptive Fβ version where recall is considered β times as important as precision. β is computed based on the length ratio between the pseudo reference and the given summary. The motivation is to punish the short summary that can easily get high precision while covering very limited important information in the pseudo reference (i.e., low recall). To measure the redundancy of a summary, we design a simple but effective self-referenced similarity score. If a summary contains much redundant information, there must exist plenty of semantically similar tokens or sentences. Based on this assumption, we use the summary itself as the reference and input a (summary, summary) pair into a selfmasked BERTScore to produce a redundancy score that evaluates the averaged degree of semantic similarity of each token or sentence with other tokens or sentences. After obtaining the centrality-weighted relevance score and the self-referenced redundancy score, we combine them to predict the final evaluation score. Depending on either F1 or Fβ is applied in our relevance score, we propose two variants of our method: the F1-based version and the Fβ-based version. Extensive experiments are conducted on both multi-document and single-document summarization datasets. The results show that our F1based method already outperforms all the SOTA baselines on all datasets. Moreover, our Fβ-based method can further improve the performance on multi-document summarization datasets. Our contributions are summarized as follows: (1) A novel training-free and reference-free summarization evaluation metric which considers both relevance and redundancy; (2) A centrality-weighted relevance score that effectively utilizes the sentence centrality of the documents to provide importance guidance for the pseudo reference tokens and sentences. Besides the F1 version, we also develop an Fβ based relevance score which pays more attention to recall; (3) A self-referenced redundancy score that utilizes a self-masked BERTScore to detect the duplicated information of the given summary; (4) To the best of our knowledge, we are the first evaluation metric that can achieve SOTA performance on both multiple and single document summarization under the reference-free setting. 2 Preliminary Notations. We denote vectors as bold lowercase characters and matrices as bold uppercase characters. The characters that are not bold are used to denote scalars. Calligraphy uppercase characters are utilized to represent sets. Problem Definition. We formally define the reference-free summarization evaluation problem as follows. Give a set of documents D = 406 Summary 𝑥 Document 𝑑# BERT 𝐰% & … 𝐰' & 𝐬% & … 𝐬) & 𝐰% * … 𝐰+ * 𝐬% * … 𝐬, * Centrality-based Sentence Selection Centrality-weighted BERTScore (𝐹% or 𝐹.) Self-masked BERTScore (𝐹%) Sentence Weights Merge Final Score Averaged Relevance Score Redundancy Score 𝐾 Pseudo Reference 𝑟 Figure 1: Overall framework of our method. w and s are the token-level and sentence-level representations. n and N (m and M) are the token number and the sentence number of the summary (pseudo reference). For multidocument summary (i.e., K > 1), we compute relevance scores between the summary x and each document dk, and then average them as the final relevance score. {d1, d2, ..., dK} and a generated summary x, the goal is to predict a score to represent the overall quality of the summary. K = 1 and K > 1 indicate single-document and multi-document summarization respectively. 3 Our Methodology The overall framework is illustrated in Figure 1. Our final evaluation score of a summary consists of an averaged centrality-weighted relevance score and a self-referenced redundancy score. Both scores are calculated on a semantic-level instead of utilizing n-gram overlapping. The averaged relevance score is computed from the relevance score between the summary and each document in the document set. The redundancy score is calculated based on the summary itself. 3.1 Centrality-weighted Relevance Score Our relevance score aims to estimate the informativeness of the given summary. We first encode each document in the document set and the summary into hidden representations. Then, for each document, we select essential sentences by centrality to build a pseudo reference. Next, we compute a centrality-weighted relevance score between the summary and each pseudo reference. Finally, we average all the relevance scores as the final relevance score of the summary. We use the k-th document dk and a summary x as an example to show the workflow. Encoding. Following SUPERT (Gao et al., 2020), we first split the document dk and the summary x into sentences. Then, the pre-trained SBERT1 is employed to encode the tokens of each sentence into token-level contextual hidden representations. We also apply max-pooling on all the tokens of a sentence to obtain the sentence-level hidden representation. Following previous work, when utilizing the token-level representations to compute the relevance and redundancy scores, we will filter out the non-informative tokens such as stop-words to improve the efficiency. Building Pseudo Reference. We do not choose all the document sentences of dk to evaluate the relevance of the summary. Because the whole document usually contains plenty of unimportant sentences which may introduce extra noise for the relevance evaluation. Thus, we select important document sentences to build a pseudo reference r for the evaluation. The sentence selection is based on the centrality of each sentence, which is computed by the unsupervised algorithm, PacSum (Zheng and Lapata, 2019), using the sentence-level representation. After obtaining the centrality scores of all sentences of the document, we choose the top-M2 sentences as the pseudo reference. Besides, we normalize the centrality scores to [0, 1] and denote the normalized centrality scores of the selected sen1bert-large-nli-stsb-mean-tokens 2In experiments, we follow the default configuration of SUPERT and set M as 12 for all the datasets. 407 tences as ¯as = [¯as 1, ¯as 2, ..., ¯as M] where ¯as i ∈[0, 1] and the superscript s means sentence-level. We denote the pseudo reference building process as PacSumTopM. Computing Relevance Score with One Pseudo Reference. Instead of only using token-level representations, we also leverage the sentence-level representations to provide multi-level information. The hybrid representations of the summary x and the pseudo reference r are denoted as follows: X = [wx 1, ..., wx n, sx 1, ..., sx N], (1) Rk = [wr 1, ..., wr m, sr 1, ..., sr M], (2) where n and N (m and M) are the token number and sentence number of the summary (pseudo reference). w and s represent the token and sentence hidden representations respectively. Besides the hybrid representations, we also introduce a centrality weighting scheme to weight the tokens and sentences of the pseudo reference, which is different from previous work that either treats them equally or uses the surface statistics like IDF as the weights. Based on the centrality scores of the selected pseudo reference sentences i.e., ¯as = [¯as 1, ¯as 2, ..., ¯as M], we assign the weights of the pseudo reference tokens as follows: ¯aw = [¯aw 1 , ¯aw 2 , ..., ¯aw m], (3) ¯aw j = ¯as i:wj∈si, (4) where ¯ai:wj∈si indicates the token wj inherits the centrality score from its sentence si. Since we have already removed the non-informative tokens in the token-level representations of each sentence, the remaining tokens capture the key information of the sentence and consequently it is reasonable to perform such a weight inheritance. Next, we combine token weights ¯aw and sentence weights ¯as to get the final normalized centrality-based weights of the hybrid representations: a = [aw 1 , ..., aw m, as 1, ..., as M], (5) aw j = ¯aw j /sum([¯aw; ¯as]), (6) as i = ¯as i/sum([¯aw; ¯as]), (7) where “[·; ·]” represents concatenation. Based on the hybrid representations (i.e., X and Rk) and the centrality-based weights of the pseudo reference tokens and sentences (i.e., a), we compute the relevance score between the summary and the pseudo reference by a weighted BERTScore (Zhang et al., 2020). For brevity, we denote the j-th element of X as xj, the i-th element of Rk as ri, and the i-th element of a as ai: Recall = P i ai maxj Sim(ri, xj) P i ai , (8) Precision = P j maxi Sim(ri, xj) |X| , (9) F1 = 2 ∗Recall ∗Precision Recall + Precision , (10) where “Sim” denotes the cosine similarity and |X| equals to n + N. Recall, Precision, and F1 are in the range of [-1, 1]. Besides the F1 version, we also propose an adaptive Fβ version of relevance score as follows: Fβ = (1 + β2) ∗Recall ∗Precision Recall + β2 ∗Precision , (11) β2 = 1, if ( |Rk| |X| )1/γ ≤1 2, if ( |Rk| |X| )1/γ ≥2 ( |Rk| |X| )1/γ, otherwise , (12) where |Rk| = m+M, |X| = n+N, and γ is a positive integer hyper-parameter. In our experiments, γ is set as 2 after fine-tuning on the validation dataset and is fixed for all the testing datasets. The physical meaning of β is that the Recall score is considered β times as important as the Precision score. In summarization evaluation, the coverage of the key information is always the most important quality indicator of the summary. Thus, we set the lower bound of β as 1. On the other hand, the metric should not only evaluate the key information coverage, containing less unimportant content in the summary should also be considered. Therefore, we set the upper bound of β as √ 2. As shown in Eq.12, within the range of [1, √ 2], β adaptively changes according to the ratio between |Rk| and |X|. The intuition comes from that a longer pseudo reference implies more key information needs to be covered by the summary. Besides, a shorter summary can easily get high precision but covers very limited important information in the pseudo reference. Thus, we give Recall a higher weight to punish such short summaries when the pseudo reference is long. Final Averaged Relevance Score. After computing the centrality-weighted relevance score between the summary and the pseudo reference of each source document, we employ the average as 408 the final relevance score of the summary: scorerel = mean([F 1 ∗, ..., F k ∗, ..., F K ∗]), (13) where * is 1 for the F1 variant and β for the Fβ variant. The superscript k indicates the F∗score is computed with the k-th document. Note that scorerel ∈[−1, 1] and higher is better. 3.2 Self-referenced Redundancy Score In this section, we introduce our self-referenced redundancy score. We engage the summary itself as the reference to evaluate the degree of the semantic similarity between each summary token or sentence with the other tokens or sentences. The averaged semantic similarity degree is used as the redundancy score. The computation is based on a self-masked BERTScore as follows: scorered = P i maxj:i̸=j Sim(xj, xi) |X| , (14) where “j : i ̸= j” means we do not consider the similarity between xi and itself, i.e, self-masked. Because of the symmetric property, the F1, precision, and recall scores are equal with each other. This is also the reason that we use precision in Eq.14 as the final redundancy score. Note that scorered ∈[−1, 1] and lower is better. 3.3 Final Evaluation Score After obtaining the relevance score and the redundancy score, we apply a linear combination to produce the final evaluation score of the summary based on the document set: score = scorerel −λ ∗scorered 1 + λ , (15) where 0 < λ ≤1 is a hyper-parameter to scale the redundancy score and score ∈[−1, 1]. Higher score means better summary quality. In our experiments, after fine-tuning on the validation set, λ is set as 0.6 and is fixed for all the testing datasets. We denote the variants of our final method as Ours(Fβ)-PacSumTopM and Ours(F1)-PacSumTopM depending on whether the adaptive Fβ is employed. 4 Experiment Setup 4.1 Datasets For comprehensively investigating our summarization evaluation methods, we test our methods on both multi-document and single-document summarization datasets. We leverage TAC3 datasets 3https://tac.nist.gov/ Dataset |Topic| Document Summary |Set| Ave.S Ave.T |Systems| Ave.S Ave.T Valid. TAC-2010 46 10 23.2 651.8 43 4.3 118.9 Test. TAC-2011 44 10 20.1 560.5 50 4.3 120.9 TAC-2009 44 10 24.9 705.8 55 4.1 117.6 TAC-2008 48 10 23.3 660.0 58 4.2 119.6 CNNDM 499 1 36.0 921.1 4 3.5 73.2 Table 1: Statistics of datasets. “Valid.” and “Test.” indicate the dataset is used for validation and testing, respectively. “|Topic|” is the number of topics. Under each topic, a set of documents is given and summaries are from different systems associating with humanannotated quality scores. “|Set|” is the number of documents in the document set. “Ave.S” and “Ave.T” represent the averaged sentence number and token number per document or summary. Note that the token number is counted after the tokenization. “|Systems|” denotes the number of summarization systems in the dataset. for multi-document summarization evaluation testing. We choose TAC-2010 as the validation dataset and TAC-2008/TAC-2009/TAC-2011 as the testing datasets. Following previous work, we only utilize the initial summaries in TAC datasets, i.e., the summaries for the document set A. For the singledocument summarization evaluation, we employ CNNDM4 (Chaganty et al., 2018) as the testing dataset. The statistics of these datasets are shown in Table 1. Note that the hyper-parameters of our methods are fine-tuned on TAC-2010 and then fixed for all the testing datasets. For TAC datasets, we compute correlation coefficients between predicted scores of an evaluation method and the annotated Pyramid scores of summaries to measure the effectiveness of the method. Following Gao et al. (2020), a correlation is computed for each topic. Then, the averaged correlation from all the topics is engaged as the final correlation of the method with human ratings. For CNNDM dataset, correlations are calculated with the human scores in three dimensions including Overall, Grammar, and Redundancy. Following Wu et al. (2020), the correlation is computed between predicted scores of the 499 × 4 = 1996 (document, summary) pairs with corresponding human ratings. 4.2 Baselines In this section, we briefly introduce our baselines. We choose TF-IDF, JS (Louis and Nenkova, 2013), and REPEAR (Rioux et al., 2014) as traditional reference-free baselines. All these traditional baselines do not build pseudo references and 4https://bit.ly/price-of-debiasing 409 Method TAC-2011 TAC-2009 TAC-2008 r ρ τ r ρ τ r ρ τ TF-IDF 0.313 0.294 0.209 0.372 0.382 0.279 0.375 0.341 0.243 JS 0.377 0.333 0.240 0.376 0.381 0.279 0.385 0.338 0.242 REAPER 0.377 0.334 0.237 0.358 0.357 0.256 0.287 0.261 0.187 Ours(F1)-All 0.495 0.451 0.329 0.478 0.476 0.353 0.466 0.426 0.310 Ours(Fβ)-All 0.498 0.455 0.332 0.480 0.471 0.348 0.462 0.423 0.307 ROUGE-1-PacSumTopM 0.436 0.377 0.274 0.418 0.406 0.301 0.397 0.348 0.252 ROUGE-2-PacSumTopM 0.429 0.388 0.287 0.380 0.419 0.314 0.410 0.355 0.259 ROUGE-L-PacSumTopM 0.436 0.370 0.272 0.427 0.415 0.306 0.385 0.336 0.245 MoverScore-PacSumTopM 0.521 0.475 0.351 0.483 0.485 0.362 0.479 0.440 0.323 S+WMS-PacSumTopM 0.291 0.292 0.211 0.350 0.358 0.264 0.364 0.358 0.260 C-ELMO-PacSumTopM 0.386 0.302 0.217 0.317 0.235 0.167 0.210 0.162 0.114 C-SBERT-PacSumTopM 0.332 0.293 0.207 0.314 0.277 0.197 0.183 0.196 0.143 SUPERT-PacSumTopM 0.511 0.481 0.357 0.486 0.494 0.368 0.493 0.457 0.334 SUPERT-IDF-PacSumTopM 0.507 0.476 0.353 0.485 0.492 0.367 0.489 0.450 0.328 Ours(F1)-PacSumTopM 0.531 0.493 0.365 0.502 0.506 0.381 0.495 0.461 0.337 Ours(Fβ)-PacSumTopM 0.541 0.505 0.374 0.507 0.508 0.380 0.500 0.465 0.339 Table 2: Main results on multi-document summarization datasets. Pearson’s r, Spearman’s ρ, and Kendall’s τ with human scores are reported. The best results are bold and the second-best results are underlined. Method Overall Grammar Redundancy r ρ τ r ρ τ r ρ τ TF-IDF 0.264 0.249 0.187 0.186 0.170 0.127 0.281 0.253 0.187 JS 0.265 0.232 0.174 0.210 0.180 0.136 0.317 0.278 0.208 REAPER 0.036 0.032 0.024 0.004 -0.006 -0.005 -0.020 -0.031 -0.024 LS Score (Wu et al., 2020) − 0.334 − − 0.266 − − 0.288 − Ours(F1)-All 0.390 0.370 0.281 0.306 0.306 0.232 0.413 0.381 0.287 Ours(Fβ)-All 0.361 0.337 0.255 0.273 0.270 0.204 0.395 0.356 0.268 ROUGE-1-PacSumTopM 0.224 0.215 0.159 0.126 0.114 0.084 0.289 0.254 0.186 ROUGE-2-PacSumTopM 0.347 0.335 0.253 0.254 0.240 0.181 0.398 0.369 0.274 ROUGE-L-PacSumTopM 0.235 0.224 0.166 0.135 0.122 0.090 0.300 0.264 0.193 MoverScore-PacSumTopM 0.373 0.341 0.259 0.264 0.240 0.181 0.411 0.359 0.267 S+WMS-PacSumTopM 0.324 0.353 0.267 0.240 0.256 0.193 0.360 0.385 0.286 C-ELMO-PacSumTopM 0.355 0.297 0.223 0.232 0.201 0.151 0.425 0.354 0.262 C-SBERT-PacSumTopM 0.405 0.378 0.286 0.295 0.299 0.225 0.415 0.373 0.279 SUPERT-PacSumTopM 0.384 0.374 0.284 0.318 0.317 0.240 0.381 0.369 0.277 SUPERT-IDF-PacSumTopM 0.382 0.373 0.283 0.316 0.314 0.238 0.377 0.365 0.274 Ours(F1)-PacSumTopM 0.416 0.404 0.308 0.341 0.341 0.259 0.428 0.408 0.308 Ours(Fβ)-PacSumTopM 0.400 0.381 0.290 0.314 0.311 0.235 0.427 0.395 0.298 Table 3: Main results on single-document summarization dataset (CNNDM). Pearson’s r, Spearman’s ρ, and Kendall’s τ with human scores are reported. The best results are bold and the second-best results are underlined. directly utilize the full content of the documents. For fairness, we also show the performance of our methods without building pseudo reference. We denote them as Ours(F1)-All and Ours(Fβ)-All since they use the whole document as a reference. We also extend several popular referencebased methods as baselines. We adapt ROUGE1/2/L (Lin, 2004), MoverScore (Zhao et al., 2019), and S+WMS (Clark et al., 2019) into the referencefree scenario via building the pseudo reference with the PacSumTopM method. We add the suffix “PacSumTopM” to these baseline names to indicate the pseudo reference building process. Besides, the SOTA reference-free summary evaluation metrics are also selected as our strong baselines, including C-ELMO/C-SBERT (Sun and Nenkova, 2019), SUPERT/SUPERT-IDF (Gao et al., 2020), and LS Score (Wu et al., 2020). CELMO (C-SBERT) encodes the document and the summary using the pre-trained ELMO (SBERT) and then computes their cosine similarity. SUPERTIDF is an extension of SUPERT, which utilizes the inverse document frequency (IDF) as the importance weight of each token. For fair comparisons, we also apply the same pseudo reference building process i.e., PacSumTopM, to C-ELMO/CSBERT/SUPERT/SUPERT-IDF and add the suffix “-PacSumTopM” to the their names. 5 Results and Analysis 5.1 Main Results The main experimental results on multi-document summarization datasets are shown in Table 2. We find that our F1 version (i.e., Ours(F1)PacSumTopM) already consistently outperforms all the baselines, which indicates the effectiveness of our centrality-weighted relevance score and our self-referenced redundancy score. The results also 410 1 2 3 5 7 9 all Different |Set| −0.02 −0.01 0.00 0.01 0.02 Ours(Fβ)'s ρ - Ours(F1)'s ρ 0.004 0.011 0.007 0.01 0.008 0.009 0.012 4 8 12 16 20 28 36 44 all Different |Systems| −0.02 −0.01 0.00 0.01 0.02 Ours(Fβ)'s ρ - Ours(F1)'s ρ -0.024 -0.007 0.007 0.014 0.018 0.018 0.014 0.014 0.012 Figure 2: The gap of Spearman’s ρ between Ours(Fβ) and Ours(F1) on TAC-2011 for different |Set| and |Systems|. Positive gaps mean our Fβ can improve the performance while negative gaps indicate our Fβ degrades the performance. When changing one of them, the other is fixed. “all” means the full size is applied, i.e., 10 for |Set| and 50 for |Systems|. demonstrate that our Fβ version can further improve the performance of multi-document summarization evaluation. By comparing Ours(Fβ)PacSumTopM and Ours(Fβ)-All, we see that the pseudo reference building process can significantly improve the performance. This is also the reason why we apply the same pseudo reference building process into SOTA baselines for fair comparisons. In the remaining part of this paper, we omit the suffix “-PacSumTopM” for simplicity when we mention a method. We also test our methods on the single-document summarization dataset without further fine-tuning the hyper-parameters. The main results are displayed in Table 3. We note that our F1 version still outperforms all the baselines, which manifests the high generalization ability of our F1-based method. One interesting finding is that the performance significantly drops after incorporating the Fβ score. To study the reason for the performance degradation on CNNDM after incorporating Fβ, we compare CNNDM and TAC datasets first. From Table 1, we note the main differences between them are the size of the document set for each topic (i.e., |Set|) and the number of the summarization systems (i.e., |Systems|). CNNDM has much smaller |Set| and |Systems|. We use the TAC-2011 dataset as an example to investigate whether our Fβ is unsuitable for smaller |Set| and |Systems|. We change |Set| and |Systems| respectively and report the gap of Spearman’s ρ between Ours(Fβ) and Ours(F1) in Figure 2. From the results, we observe that our Fβ Method TAC CNNDM 2011 2009 2008 Overall Grammar Redundancy Ours(F1) 0.493 0.506 0.461 0.404 0.341 0.408 Ours(Fβ) 0.505 0.508 0.465 0.381 0.311 0.395 MoverScore 0.475 0.485 0.440 0.341 0.240 0.359 +CentralityW. 0.472 0.467 0.431 0.350 0.257 0.364 +Redundancy 0.237 0.202 0.221 0.448 0.326 0.546 +Both 0.261 0.220 0.241 0.455 0.341 0.545 Table 4: Spearman’s ρ of incorporating the centrality weighting and redundancy score into MoverScore based framework. “+Both” means these two features are simultaneously applied. can consistently improve the performance for different |Set|. For the single-document summarization setting, i.e., |Set|=1, it still obtains a positive gap. Nevertheless, when the |Systems| is small such as 4, applying our Fβ leads to a dramatic performance dropping. From Table 1, we also see that CNNDM and TAC-2011 have different summary lengths (73.2 for CNNDM and 120.9 for TAC2011). However, when we limit the |Systems| of TAC-2011 to smaller numbers, the average length of generated summaries is still around 120, which indicates the performance degeneration is indeed from the change of system numbers. Therefore, we suggest using Ours(Fβ) when |Systems| is large like 12 and employing Ours(F1) when |Systems| is small like 4. 5.2 Ablation Study For better understanding the contributions of our proposed components, we conduct ablation studies on the best-performed method on each dataset, i.e., Ours(Fβ) for the multi-document summarization datasets and Ours(F1) for the single-document summarization dataset. We display results of the rank-based Spearman’s ρ in Figure 3. As shown in the figure, after removing one of the three components (i.e., the centrality weighting, the hybrid representation, and the redundancy score), the performance of our methods become worse in most cases. This finding demonstrates the effectiveness of our proposed components. Besides, we also note that removing the redundancy score significantly degrades the performance on the redundancy evaluation on CNNDM, which indicates our redundancy score effectively captures the redundancy degree of the summaries. 5.3 Apply Centrality Weighting and Redundancy Score into MoverScore Besides basing on BERTScore, we also study whether our key features i.e., the centrality weighting and redundancy score, can work well in a 411 TAC2011 TAC2009 TAC2008 0.30 0.35 0.40 0.45 0.50 0.55 0.60 Spearman's ρ 0.505 0.508 0.465 0.501 0.51 0.46 0.502 0.509 0.468 0.498 0.496 0.466 Ablation Study on TAC Datasets Ours(Fβ) -CentralityW. -HybridR. -Redundancy Overall Grammar Redundancy 0.20 0.25 0.30 0.35 0.40 0.45 0.50 Spearman's ρ 0.404 0.341 0.408 0.399 0.332 0.406 0.4 0.337 0.406 0.382 0.33 0.374 Ablation Study on CNNDM Dataset Ours(F1) -CentralityW. -HybridR. -Redundancy Figure 3: Ablation studies for Ours(Fβ) on TAC datasets and Ours(F1) on CNNDM. “-CentralityW.” means that we remove the centrality weighting when computing relevance scores. “-HybridR.” represents we only utilize the token-level representations when calculating relevance and redundancy scores. “Redundancy” indicates we omit the redundancy score. MoverScore based framework (i.e., the relevance and redundancy scores are computed using MoverScore). Note that our Fβ is not applicable to MoverScore since it is not an F-measure. The results are listed in Table 4. We find that these two features significantly improve the performance of the original MoverScore on single-document summarization evaluation while degrading the performance dramatically on multi-document summarization evaluation. On CNNDM, the enhanced MoverScore even outperforms Ours(F1) on the “Overall” and “Redundancy” aspects, which indicates MoverScore is a promising basis for our proposed new features. We leave solving the performance dropping of the enhanced MoverScore on multi-document setting as future work. 5.4 Robustness Analysis We investigate the robustness of our method on the following factors and report the experimental results on the validation dataset (i.e., TAC-2010) in Figure 4: (1) the hyper-parameter λ for scaling the redundancy score; (2) the hyper-parameter γ in Fβ; (3) the number of selected sentences for pseudo reference i.e., M; (4) different pre-trained contextual encoding models including BERT-base5, BERTlarge6, RoBERTa-base7, and RoBERTa-large8. 5bert-base-nli-stsb-mean-tokens 6bert-large-nli-stsb-mean-tokens 7roberta-base-nli-stsb-mean-tokens 8roberta-large-nli-stsb-mean-tokens 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Different λ 0.550 0.575 0.600 0.625 Correlations r ρ 1 2 3 4 5 6 Different γ 0.550 0.575 0.600 0.625 Correlations r ρ 3 6 9 12 15 18 21 Different M 0.550 0.575 0.600 0.625 Correlations r ρ BERT-base BERT-large RoBERTa-base RoBERTa-large 0.550 0.575 0.600 0.625 Correlations r ρ Figure 4: The performance of Ours(Fβ) on TAC-2010 under different λ, γ, M, and encoding models. When we change one of them, the others are fixed. The Pearson’s r and Spearman’s ρ are reported. Since both Spearman’s ρ and Kendall’s τ are rank-based correlation coefficients, we omit Kendall’s τ for simplicity. From this figure, we observe that the performance of our method is relatively stable for different λ and γ. We also find that a small M leads to lower correlations because much important information may be abandoned when building the pseudo references. But a large M will also degenerate the correlations since more noises are introduced. Thus, a moderate M is better. As for encoding models, we note that large encoding models obtain better performance than base encoding models. However, large models need more computation resources and time to encode the input text. Note that for our final method, we only fine-tune λ and γ on the TAC-2010 and set them as 0.6 and 2. As for M and encoding models, following the configuration of SUPERT (Gao et al., 2020), we directly set M as 12 and employ the BERT-large as the encoding model. All these factors are fixed for all testing datasets. 5.5 Performance on Bad/Good Summaries In this section, we evaluate the ability of our method to distinguish bad and good summaries. The bad and good summaries are selected by human ratings. We use TAC-2011 as an example and choose SUPERT as a strong baseline. The corresponding distributions of the reversed rank for bad and good summaries are illustrated in Figure 5. A smaller (larger) reversed rank represents the summary is assigned with a lower (higher) score. From the figure, we find that compared with SUPERT, Our(Fβ) has a better ability to assign bad sum412 SUPERT Ours(Fβ) 0 10 20 30 40 Reversed Rank Reversed Rank of Bad Summaries SUPERT Ours(Fβ) 0 10 20 30 40 Reversed Rank Reversed Rank of Good Summaries Figure 5: Distributions of the reversed rank from SUPERT and Ours(Fβ) for bad and good summaries on TAC-2011. The bar in the middle indicates the median. maries lower scores and good summaries higher scores, which demonstrates the effectiveness of our method again. Moreover, we also note that both SUPERT and Ours(Fβ) are good at giving bad summaries lower scores while having difficulty in assigning good summaries higher scores. We leave solving this problem as another future work under the reference-free setting. 6 Related Work Reference-based Evaluation Metrics mainly measure the relevance between the humanannotated references and the system-generated text, which are widely adopted in text summarization (Lin, 2004; Zhao et al., 2019), machine translation (Papineni et al., 2002; Zhang et al., 2020), and dialogue systems (Papineni et al., 2002; Gao et al., 2021; Xiang et al., 2021). For example, ROUGE (Lin, 2004) evaluates the token sequence overlapping. BERTScore (Zhang et al., 2020), S+WMS (Clark et al., 2019), and MoverScore (Zhao et al., 2019) measure the semantic similarity between the references and the summary via a greedy or optimized minimum Earth Mover’s Distance. Reference-free Evaluation Metrics have been developed to avoid the dependency on humanannotated references, which obtain more and more attention in recent years (B¨ohm et al., 2019; Gao et al., 2020; Wu et al., 2020; Chan et al., 2021). Some of them need to train a scorer (Peyrard and Gurevych, 2018; Xenouleas et al., 2019; Scialom et al., 2019; B¨ohm et al., 2019). For example, LS Score (Wu et al., 2020) designs a metric which combines a linguistic quality scorer trained from the built positive and negative summaries, and a relevance scorer based on cosine similarity. The others do not require training (Louis and Nenkova, 2013; Rioux et al., 2014; Peyrard, 2019; Sun and Nenkova, 2019). For instance, SUPERT (Gao et al., 2020) builds the pseudo references from the source document first and then engages BERTScore to compute the relevance score between the pseudo reference and the summary. 7 Conclusion In this paper, we propose a novel training-free and reference-free summarization evaluation metric consisting of a relevance score and a redundancy score. Experiments on multi-document and single-document summarization settings show the effectiveness of our methods. One promising future direction is to solve the performance dropping issue after applying our key features into MoverScore and the other is to tackle the problem that current metrics struggle to assign higher scores for good summaries. Acknowledgements The work described in this paper was partially supported by the Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410021, Research Impact Fund (RIF), R5034-18). References Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Florian B¨ohm, Yang Gao, Christian M. Meyer, Ori Shapira, Ido Dagan, and Iryna Gurevych. 2019. Better rewards yield better summaries: Learning to summarise without references. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3110–3120, Hong Kong, China. Association for Computational Linguistics. Arun Chaganty, Stephen Mussmann, and Percy Liang. 2018. The price of debiasing automatic metrics in natural language evalaution. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 413 pages 643–653, Melbourne, Australia. Association for Computational Linguistics. Hou Pong Chan, Wang Chen, and Irwin King. 2020. A unified dual-view model for review summarization and sentiment classification with inconsistency loss. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, SIGIR 2020, Virtual Event, China, July 25-30, 2020, pages 1191–1200. ACM. Zhangming Chan, Lemao Liu, Juntao Li, Haisong Zhang, Dongyan Zhao, Shuming Shi, and Rui Yan. 2021. Enhancing the open-domain dialogue evaluation in latent space. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics: Findings. Elizabeth Clark, Asli C¸ elikyilmaz, and Noah A. Smith. 2019. Sentence mover’s similarity: Automatic evaluation for multi-sentence texts. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 2748–2760. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Jun Gao, Wei Bi, Ruifeng Xu, and Shuming Shi. 2021. Ream#: An enhancement approach to referencebased evaluation metrics for open-domain dialog generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics: Findings. Yang Gao, Wei Zhao, and Steffen Eger. 2020. SUPERT: towards new frontiers in unsupervised evaluation metrics for multi-document summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 1347–1354. Association for Computational Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692. Annie Louis and Ani Nenkova. 2013. Automatically assessing machine summary content without a gold standard. Comput. Linguistics, 39(2):267–300. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Maxime Peyrard. 2019. A simple theoretical model of importance for summarization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1059–1073, Florence, Italy. Association for Computational Linguistics. Maxime Peyrard and Iryna Gurevych. 2018. Objective function learning to match human judgements for optimization-based summarization. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACLHLT, New Orleans, Louisiana, USA, June 1-6, 2018, Volume 2 (Short Papers), pages 654–660. Association for Computational Linguistics. Cody Rioux, Sadid A. Hasan, and Yllias Chali. 2014. Fear the REAPER: A system for automatic multidocument summarization with reinforcement learning. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 681–690, Doha, Qatar. Association for Computational Linguistics. Thomas Scialom, Sylvain Lamprier, Benjamin Piwowarski, and Jacopo Staiano. 2019. Answers unite! unsupervised metrics for reinforced summarization models. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 3246–3256, Hong Kong, China. Association for Computational Linguistics. Abigail See, Peter J. Liu, and Christopher D. Manning. 2017. Get to the point: Summarization with pointergenerator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073– 1083, Vancouver, Canada. Association for Computational Linguistics. Simeng Sun and Ani Nenkova. 2019. The feasibility of embedding based automatic evaluation for single document summarization. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1216–1221, Hong Kong, China. Association for Computational Linguistics. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 3104–3112. 414 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pages 5998–6008. Hanlu Wu, Tengfei Ma, Lingfei Wu, Tariro Manyumwa, and Shouling Ji. 2020. Unsupervised reference-free summary quality evaluation via contrastive learning. CoRR, abs/2010.01781. Stratos Xenouleas, Prodromos Malakasiotis, Marianna Apidianaki, and Ion Androutsopoulos. 2019. SUMQE: a BERT-based summary quality estimation model. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6005–6011, Hong Kong, China. Association for Computational Linguistics. Jiannan Xiang, Yahui Liu, Deng Cai, Huayang Li, Defu Lian, and Lemao Liu. 2021. Assessing dialogue systems with distribution distances. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics: Findings. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. 2019. MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563–578, Hong Kong, China. Association for Computational Linguistics. Hao Zheng and Mirella Lapata. 2019. Sentence centrality revisited for unsupervised summarization. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 6236–6247. Association for Computational Linguistics.
|
2021
|
34
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4402–4417 August 1–6, 2021. ©2021 Association for Computational Linguistics 4402 Intent Classification and Slot Filling for Privacy Policies Wasi Uddin Ahmad†∗, Jianfeng Chi‡∗, Tu Le‡ Thomas Norton§, Yuan Tian‡, Kai-Wei Chang† †University of California, Los Angeles, ‡University of Virginia, §Fordham University †{wasiahmad, kwchang}@cs.ucla.edu ‡{jc6ub,tnl6wk,yuant}@virginia.edu §[email protected] Abstract Understanding privacy policies is crucial for users as it empowers them to learn about the information that matters to them. Sentences written in a privacy policy document explain privacy practices, and the constituent text spans convey further specific information about that practice. We refer to predicting the privacy practice explained in a sentence as intent classification and identifying the text spans sharing specific information as slot filling. In this work, we propose PolicyIE, an English corpus consisting of 5,250 intent and 11,788 slot annotations spanning 31 privacy policies of websites and mobile applications. PolicyIE corpus is a challenging real-world benchmark with limited labeled examples reflecting the cost of collecting large-scale annotations from domain experts. We present two alternative neural approaches as baselines, (1) intent classification and slot filling as a joint sequence tagging and (2) modeling them as a sequence-tosequence (Seq2Seq) learning task. The experiment results show that both approaches perform comparably in intent classification, while the Seq2Seq method outperforms the sequence tagging approach in slot filling by a large margin. We perform a detailed error analysis to reveal the challenges of the proposed corpus. 1 Introduction Privacy policies inform users about how a service provider collects, uses, and maintains the users’ information. The service providers collect the users’ data via their websites or mobile applications and analyze them for various purposes. The users’ data often contain sensitive information; therefore, the users must know how their information will be used, maintained, and protected from unauthorized and unlawful use. Privacy policies are meant to explain all these use cases in detail. This makes ∗Equal contribution. Listed by alphabetical order. privacy policies often very long, complicated, and confusing (McDonald and Cranor, 2008; Reidenberg et al., 2016). As a result, users do not tend to read privacy policies (Commission et al., 2012; Gluck et al.; Marotta-Wurgler, 2015), leading to undesirable consequences. For example, users might not be aware of their data being sold to third-party advertisers even if they have given their consent to the service providers to use their services in return. Therefore, automating information extraction from verbose privacy policies can help users understand their rights and make informed decisions. In recent years, we have seen substantial efforts to utilize natural language processing (NLP) techniques to automate privacy policy analysis. In literature, information extraction from policy documents is formulated as text classification (Wilson et al., 2016a; Harkous et al., 2018; Zimmeck et al., 2019), text alignment (Liu et al., 2014; Ramanath et al., 2014), and question answering (QA) (Shvartzshanider et al., 2018; Harkous et al., 2018; Ravichander et al., 2019; Ahmad et al., 2020). Although these approaches effectively identify the sentences or segments in a policy document relevant to a privacy practice, they lack in extracting fine-grained structured information. As shown in the first example in Table 1, the privacy practice label “Data Collection/Usage” informs the user how, why, and what types of user information will be collected by the service provider. The policy also specifies that users’ “username” and “icon or profile photo” will be used for “marketing purposes”. This informs the user precisely what and why the service provider will use users’ information. The challenge in training models to extract finegrained information is the lack of labeled examples. Annotating privacy policy documents is expensive as they can be thousands of words long and requires domain experts (e.g., law students). Therefore, prior works annotate privacy policies at the 4403 [We]Data Collector: First Party Entity may also [use]Action or display [your]Data Provider: user [username]Data Collected: User Online Activities/Profiles and [icon or profile photo]Data Collected: User Online Activities/Profiles on [marketing purpose or press releases]Purpose: Advertising/Marketing. Privacy Practice. Data Collection/Usage [We]Data Sharer: First Party Entity do [not]Polarity: Negation [sell]Action [your]Data Provider: user [personal information]Data Shared: General Data to [third parties]Data Receiver: Third Party Entity. Privacy Practice. Data Sharing/Disclosure Table 1: Annotation examples from PolicyIE Corpus. Best viewed in color. sentence level, without further utilizing the constituent text spans to convey specific information. Sentences written in a policy document explain privacy practices, which we refer to as intent classification and identifying the constituent text spans that share further specific information as slot filling. Table 1 shows a couple of examples. This formulation of information extraction lifts users’ burden to comprehend relevant segments in a policy document and identify the details, such as how and why users’ data are collected and shared with others. To facilitate fine-grained information extraction, we present PolicyIE, an English corpus consisting of 5,250 intent and 11,788 slot annotations over 31 privacy policies of websites and mobile applications. We perform experiments using sequence tagging and sequence-to-sequence (Seq2Seq) learning models to jointly model intent classification and slot filling. The results show that both modeling approaches perform comparably in intent classification, while Seq2Seq models outperform the sequence tagging models in slot filling by a large margin. We conduct a thorough error analysis and categorize the errors into seven types. We observe that sequence tagging approaches miss more slots while Seq2Seq models predict more spurious slots. We further discuss the error cases by considering other factors to help guide future work. We release the code and data to facilitate research.1 2 Construction of PolicyIE Corpus 2.1 Privacy Policies Selection The scope of privacy policies primarily depends on how service providers function. For example, service providers primarily relying on mobile applications (e.g., Viber, Whatsapp) or websites and applications (e.g., Amazon, Walmart) have different privacy practices detailed in their privacy policies. 1https://github.com/wasiahmad/ PolicyIE In PolicyIE, we want to achieve broad coverage across privacy practices exercised by the service providers such that the corpus can serve a wide variety of use cases. Therefore, we go through the following steps to select the policy documents. Initial Collection Ramanath et al. (2014) introduced a corpus of 1,010 privacy policies of the top websites ranked on Alexa.com. We crawled those websites’ privacy policies in November 2019 since the released privacy policies are outdated. For mobile application privacy policies, we scrape application information from Google Play Store using play-scraper public API2 and crawl their privacy policy. We ended up with 7,500 mobile applications’ privacy policies. Filtering First, we filter out the privacy policies written in a non-English language and the mobile applications’ privacy policies with the app review rating of less than 4.5. Then we filter out privacy policies that are too short (< 2,500 words) or too long (> 6,000 words). Finally, we randomly select 200 websites and mobile application privacy policies each (400 documents in total).3 Post-processing We ask a domain expert (working in the security and privacy domain for more than three years) to examine the selected 400 privacy policies. The goal for the examination is to ensure the policy documents cover the four privacy practices: (1) Data Collection/Usage, (2) Data Sharing/Disclosure, (3) Data Storage/Retention, and (4) Data Security/Protection. These four practices cover how a service provider processes users’ data in general and are included in the General Data Protection Regulation (GDPR). Finally, we shortlist 50 policy documents for annotation, 25 in each category (websites and mobile applications). 2https://github.com/danieliu/ play-scraper 3We ensure the mobile applications span different application categories on the Play Store. 4404 2.2 Data Annotation Annotation Schema To annotate sentences in a policy document, we consider the first four privacy practices from the annotation schema suggested by Wilson et al. (2016a). Therefore, we perform sentence categorization under five intent classes that are described below. (1) Data Collection/Usage: What, why and how user information is collected; (2) Data Sharing/Disclosure: What, why and how user information is shared with or collected by third parties; (3) Data Storage/Retention: How long and where user information will be stored; (4) Data Security/Protection: Protection measures for user information; (5) Other: Other privacy practices that do not fall into the above four categories. Apart from annotating sentences with privacy practices, we aim to identify the text spans in sentences that explain specific details about the practices. For example, in the sentence “we collect personal information in order to provide users with a personalized experience”, the underlined text span conveys the purpose of data collection. In our annotation schema, we refer to the identification of such text spans as slot filling. There are 18 slot labels in our annotation schema (provided in Appendix). We group the slots into two categories: type-I and typeII based on their role in privacy practices. While the type-I slots include participants of privacy practices, such as Data Provider, Data Receiver, type-II slots include purposes, conditions that characterize more details of privacy practices. Note that type-I and type-II slots may overlap, e.g., in the previous example, the underlined text span is the purpose of data collection, and the span “user” is the Data Provider (whose data is collected). In general, typeII slots are longer (consisting of more words) and less frequent than type-I slots. In total, there are 14 type-I and 4 type-II slots in our annotation schema. These slots are associated with a list of attributes, e.g., Data Collected and Data Shared have the attributes Contact Data, Location Data, Demographic Data, etc. Table 1 illustrates a couple of examples. We detail the slots and their attributes in the Appendix. Annotation Procedure General crowdworkers such as Amazon Mechanical Turkers are not suitable to annotate policy documents as it requires specialized domain knowledge (McDonald and CraDataset Train Test # Policies 25 6 # Sentences 4,209 1,041 # Type-I slots 7,327 1,704 # Type-II slots 2,263 494 Avg. sentence length 23.73 26.62 Avg. # type-I slot / sent. 4.48 4.75 Avg. # type-II slot / sent. 1.38 1.38 Avg. type-I slot length 2.01 2.15 Avg. type-II slot length 8.70 10.70 Table 2: Statistics of the PolicyIE Corpus. nor, 2008; Reidenberg et al., 2016). We hire two law students to perform the annotation. We use the web-based annotation tool, BRAT (Stenetorp et al., 2012) to conduct the annotation. We write a detailed annotation guideline and pretest them through multiple rounds of pilot studies. The guideline is further updated with notes to resolve complex or corner cases during the annotation process.4 The annotation process is closely monitored by a domain expert and a legal scholar and is granted IRB exempt by the Institutional Review Board (IRB). The annotators are presented with one segment from a policy document at a time and asked to perform annotation following the guideline. We manually segment the policy documents such that a segment discusses similar issues to reduce ambiguity at the annotator end. The annotators worked 10 weeks, with an average of 10 hours per week, and completed annotations for 31 policy documents. Each annotator is paid $15 per hour. Post-editing and Quality Control We compute an inter-annotator agreement for each annotated segment of policy documents using Krippendorff’s Alpha (αK) (Klaus, 1980). The annotators are asked to discuss their annotations and re-annotate those sections with token-level αK falling below 0.75. An αK value within the range of 0.67 to 0.8 is allowed for tentative conclusions (Artstein and Poesio, 2008; Reidsma and Carletta, 2008). After the re-annotation process, we calculate the agreement for the two categories of slots individually. The inter-annotator agreement is 0.87 and 0.84 for type-I and type-II slots, respectively. Then the adjudicators discuss and finalize the annotations. The adjudication process involves one of the annotators, the legal scholar, and the domain expert. 4We release the guideline as supplementary material. 4405 Joint intent and slot tagging Input: [CLS] We may also use or display your username and icon or profile photo on marketing purpose or press releases . Type-I slot tagging output Data-Collection-Usage B-DC.FPE O O B-Action O O B-DP.U B-DC.UOAP O B-DC.UOAP I-DC.UOAP I-DC.UOAP I-DC.UOAP O O O O O O O Type-II slot tagging output Data-Collection-Usage O O O O O O O O O O O O O O B-P.AM I-P.AM I-P.AM I-P.AM I-P.AM O Sequence-to-sequence (Seq2Seq) learning Input: We may also use or display your username and icon or profile photo on marketing purpose or press releases . Output: [IN:Data-Collection-Usage [SL:DC.FPE We] [SL:Action use] [SL:DP.U your] [SL:DC.UOAP username] [SL:DC.UOAP icon or profile photo] [SL:P.AM marketing purpose or press releases]] Table 3: An example of input / output used to train the two types of models on PolicyIE. For brevity, we replaced part of label strings with symbols: DP.U, DC.FPE, DC.UOAP, P.AM represents Data-Provider.User, DataCollector.First-Party-Entity, Data-Collected.User-Online-Activities-Profiles, and Purpose.Advertising-Marketing. Data Statistics & Format Table 2 presents the statistics of the PolicyIE corpus. The corpus consists of 15 and 16 privacy policies of websites and mobile applications, respectively. We release the annotated policy documents split into sentences.5 Each sentence is associated with an intent label, and the constituent words are associated with a slot label (following the BIO tagging scheme). 3 Model & Setup PolicyIE provides annotations of privacy practices and corresponding text spans in privacy policies. We refer to privacy practice prediction for a sentence as intent classification and identifying the text spans as slot filling. We present two alternative approaches; the first approach jointly models intent classification and slot tagging (Chen et al., 2019), and the second modeling approach casts the problem as a sequence-to-sequence learning task (Rongali et al., 2020; Li et al., 2021). 3.1 Sequence Tagging Following Chen et al. (2019), given a sentence s = w1, . . . , wl from a privacy policy document D, a special token (w0 = [CLS]) is prepended to form the input sequence that is fed to an encoder. The encoder produces contextual representations of the input tokens h0, h1, . . . , hl where h0 and h1, . . . , hl are fed to separate softmax classifiers 5We split the policy documents into sentences using UDPipe (Straka et al., 2016). to predict the target intent and slot labels. yi = softmax(W T i h0 + bi), ys n = softmax(W T s hn + bs), n ∈1, . . . l, where Wi ∈Rd×I, Ws ∈Rd×S, br ∈RI and bi ∈RI, bs ∈RS are parameters, and I, S are the total number of intent and slot types, respectively. The sequence tagging model (composed of an encoder and a classifier) learns to maximize the following conditional probability to perform intent classification and slot filling jointly. P(yi, ys∣s) = p(yi∣s) l ∏ n=1 p(ys n∣s). We train the models end-to-end by minimizing the cross-entropy loss. Table 3 shows an example of input and output to train the joint intent and slot tagging models. Since type-I and type-II slots have different characteristics as discussed in § 2.2 and overlap, we train two separate sequential tagging models for type-I and type-II slots to keep the baseline models simple.6 We use BiLSTM (Liu and Lane, 2016; Zhang and Wang, 2016), Transformer (Vaswani et al., 2017), BERT (Vaswani et al., 2017), and RoBERTa (Liu et al., 2019) as encoder to form the sequence tagging models. Besides, we consider an embedding based baseline where the input word embeddings are fed to the softmax classifiers. The special token (w0 = 6Span enumeration based techniques (Wadden et al., 2019; Luan et al., 2019) can be utilized to perform tagging both types of slots jointly, and we leave this as future work. 4406 Model # param Intent F1 Type-I Type-II (in millions) Slot F1 EM Slot F1 EM Human 96.5 84.3 56.6 62.3 55.6 Embedding 1.7 50.9±27.3 19.1±0.3 0.8±0.3 0.0±0.0 0.0±0.0 BiLSTM 8 75.9±1.1 40.8±0.9 7.6±0.9 3.9±3.0 10.0±2.7 Transformer 34.8 80.1±0.6 41.0±3.5 6.5±2.8 3.5±1.0 13.1±2.4 BERT 110 84.7±0.7 55.5±1.1 17.0±1.1 29.6±2.4 24.2±4.2 RoBERTa 124 84.5±0.7 54.2±1.9 14.3±2.4 29.8±1.7 24.8±1.4 Embedding w/ CRF 1.7 67.9±0.6 26.0±1.5 1.20±0.3 5.7±4.6 3.1±0.6 BiLSTM w/ CRF 8 76.7±1.4 45.1±1.2 9.2±0.9 26.8±2.2 18.1±2.0 Transformer w/ CRF 34.8 77.9±2.7 43.7±2.3 8.9±3.0 5.7±0.9 11.0±2.1 BERT w/ CRF 110 82.1±2.0 56.0±0.8 19.2±1.1 31.7±1.9 19.7±2.6 RoBERTa w/ CRF 124 83.3±1.6 57.0±0.6 18.2±1.2 34.5±1.3 27.7±3.9 Table 4: Test set performance of the sequence tagging models on PolicyIE corpus. We individually train and evaluate the models on intent classification and type-I and type-II slots tagging and report average intent F1 score. [CLS]) embedding is formed by applying average pooling over the input word embeddings. We train WordPiece embeddings with a 30,000 token vocabulary (Devlin et al., 2019) using fastText (Bojanowski et al., 2017) based on a corpus of 130,000 privacy policies collected from apps on the Google Play Store (Harkous et al., 2018). We use the hidden state corresponding to the first WordPiece of a token to predict the target slot labels. Conditional Random Field (CRF) helps structure prediction tasks, such as semantic role labeling (Zhou and Xu, 2015) and named entity recognition (Cotterell and Duh, 2017). Therefore, we model slot labeling jointly using a conditional random field (CRF) (Lafferty et al., 2001) (only interactions between two successive labels are considered). We refer the readers to Ma and Hovy (2016) for details. 3.2 Sequence-to-Sequence Learning Recent works in semantic parsing (Rongali et al., 2020; Zhu et al., 2020; Li et al., 2021) formulate the task as sequence-to-sequence (Seq2Seq) learning. Taking this as a motivation, we investigate the scope of Seq2Seq learning for joint intent classification and slot filling for privacy policy sentences. In Table 3, we show an example of encoder input and decoder output used in Seq2Seq learning. We form the target sequences by following the template: [IN:LABEL [SL:LABEL w1, . . . , wm] . . . ]. During inference, we use greedy decoding and parse the decoded sequence to extract intent and slot labels. Note that we only consider text spans in the decoded sequences that are surrounded by “[]”; the rest are discarded. Since our proposed PolicyIE corpus consists of a few thousand examples, instead of training Seq2Seq models from scratch, we finetune pre-trained models as the baselines. Specifically, we consider five state-of-the-art models: MiniLM (Wang et al., 2020), UniLM (Dong et al., 2019), UniLMv2 (Bao et al., 2020), MASS (Song et al.), and BART (Lewis et al., 2020). 3.3 Setup Implementation We use the implementation of BERT and RoBERTa from transformers API (Wolf et al., 2020). For the Seq2Seq learning baselines, we use their public implementations.7,8,9 We train BiLSTM, Transformer baseline models and fine-tune all the other baselines for 20 epochs and choose the best checkpoint based on validation performance. From 4,209 training examples, we use 4,000 examples for training (∼95%) and 209 examples for validation (∼5%). We tune the learning rate in [1e-3, 5e-4, 1e-4, 5e-5, 1e-5] and set the batch size to 16 in all our experiments (to fit in one GeForce GTX 1080 GPU with 11gb memory). We train (or fine-tune) all the models five times with different seeds and report average performances. Evaluation Metrics To evaluate the baseline approaches, we compute the F1 score for intent classification and slot filling tasks.10 We also compute an exact match (EM) accuracy (if the predicted intent matches the reference intent and slot F1 = 1.0). 7https://github.com/microsoft/unilm 8https://github.com/microsoft/MASS 9https://github.com/pytorch/fairseq/ tree/master/examples/bart 10We use a micro average for intent classification. 4407 Model # param Intent F1 Type-I Type-II (in millions) Slot F1 EM Slot F1 EM Human 96.5 84.3 56.6 62.3 55.6 MiniLM 33 83.9±0.3 52.4±1.5 19.8±1.6 40.4±0.4 27.9±1.6 UniLM 110 83.6±0.5 58.2±0.7 28.6±1.2 53.5±1.4 35.4±1.9 UniLMv2 110 84.7±0.5 61.4±0.9 29.9±1.2 53.5±1.5 33.5±1.5 MASS 123 81.8±1.2 54.1±2.5 21.3±2.0 44.9±1.2 25.3±1.3 BART 140 83.3±1.1 53.6±1.7 10.6±1.7 52.4±2.7 27.5±2.2 400 83.6±1.3 63.7±1.3 23.0±1.3 55.2±1.0 31.6±2.0 Table 5: Test set performance of the Seq2Seq models on PolicyIE corpus. Human Performance is computed by considering each annotator’s annotations as predictions and the adjudicated annotations as the reference. The final score is an average across all annotators. 4 Experiment Results & Analysis We aim to address the following questions. 1. How do the two modeling approaches perform on our proposed dataset (§ 4.1)? 2. How do they perform on different intent and slot types (§ 4.2)? 3. What type of errors do the best performing models make (§ 4.3)? 4.1 Main Results Sequence Tagging The overall performances of the sequence tagging models are presented in Table 4. The pre-trained models, BERT and RoBERTa, outperform other baselines by a large margin. Using conditional random field (CRF), the models boost the slot tagging performance with a slight degradation in intent classification performance. For example, RoBERTa + CRF model improves over RoBERTa by 2.8% and 3.9% in terms of typeI slot F1 and EM with a 0.5% drop in intent F1 score. The results indicate that predicting type-II slots is difficult compared to type-I slots as they differ in length (type-I slots are mostly phrases, while type-II slots are clauses) and are less frequent in the training examples. However, the EM accuracy for type-I slots is lower than type-II slots due to more type-I slots (∼4.75) than type-II slots (∼1.38) on average per sentence. Note that if models fail to predict one of the slots, EM will be zero. Seq2Seq Learning Seq2Seq models predict the intent and slots by generating the labels and spans following a template. Then we extract the intent and slot labels from the generated sequences. The experiment results are presented in Table 5. To our surprise, we observe that all the models perform well in predicting intent and slot labels. The best performing model is BART (according to slot F1 score) with 400 million parameters, outperforming its smaller variant by 10.1% and 2.8% in terms of slot F1 for type-I and type-II slots, respectively. Sequence Tagging vs. Seq2Seq Learning It is evident from the experiment results that Seq2Seq models outperform the sequence tagging models in slot filling by a large margin, while in intent classification, they are competitive. However, both the modeling approaches perform poorly in predicting all the slots in a sentence correctly, resulting in a lower EM score. One interesting factor is, the Seq2Seq models significantly outperform sequence tagging models in predicting type-II slots. Note that type-II slots are longer and less frequent, and we suspect conditional text generation helps Seq2Seq models predict them accurately. In comparison, we suspect that due to fewer labeled examples of type-II slots, the sequence tagging models perform poorly on that category (as noted before, we train the sequence tagging models for the type-I and type-II slots individually). Next, we break down RoBERTa (w/ CRF) and BART’s performances, the best performing models in their respective model categories, followed by an error analysis to shed light on the error types. 4.2 Performance Breakdown Intent Classification In the PolicyIE corpus, 38% of the sentences fall into the first four categories: Data Collection, Data Sharing, Data Storage, Data Security, and the remaining belong to the Other category. Therefore, we investigate how much the models are confused in predicting the accurate intent label. We provide the confusion matrix of the models in Appendix. Due to an imbalanced distribution of labels, BART makes many 4408 Intent labels Intent F1 Slot F1 Type-I Type-II RoBERTa Data Collection 74.1±1.1 59.8±0.8 28.9±2.7 Data Sharing 67.2±2.0 53.6±5.7 34.4±3.4 Data Storage 61.7±3.6 40.1±3.7 31.6±3.1 Data Security 68.9±2.9 53.9±4.9 21.9±2.5 BART Data Collection 73.5±2.3 67.0±4.2 56.2±2.8 Data Sharing 70.4±2.7 61.2±1.6 53.5±3.4 Data Storage 63.1±4.7 56.2±8.2 64.9±2.5 Data Security 67.2±3.9 66.0±2.2 32.8±1.3 Table 6: Test performance of the RoBERTa and BART model for each intent type. incorrect predictions. We notice that BART is confused most between Data Collection and Data Storage labels. Our manual analysis reveals that BART is confused between slot labels {“Data Collector”, “Data Holder”} and {“Data Retained”, “Data Collected”} as they are often associated with the same text span. We suspect this leads to BART’s confusion. Table 6 presents the performance breakdown across intent labels. Slot Filling We breakdown the models’ performances in slot filling under two settings. First, Table 6 shows slot filling performance under different intent categories. Among the four classes, the models perform worst on slots associated with the “Data Security” intent class as PolicyIE has the lowest amount of annotations for that intent category. Second, we demonstrate the models’ performances on different slot types in Figure 1. RoBERTa’s recall score for “polarity”, “protectagainst”, “protection-method” and “storage-place” slot types is zero. This is because these slot types have the lowest amount of training examples in PolicyIE. On the other hand, BART achieves a higher recall score, specially for the “polarity” label as their corresponding spans are short. We also study the models’ performances on slots of different lengths. The results show that BART outperforms RoBERTa by a larger margin on longer slots (see Figure 2), corroborating our hypothesis that conditional text generation results in more accurate predictions for longer spans. 4.3 Error Analysis We analyze the incorrect intent and slot predictions by RoBERTa and BART. We categorize the errors 0.0 0.2 0.4 0.6 0.8 action condition data-collected data-collector data-holder data-protected data-protector data-provider data-receiver data-retained data-shared data-sharer polarity protect-against protection-method purpose retention-period storage-place RoBERTa BART Figure 1: Test set performance (Recall score) on PolicyIE for the eighteen slot types. 0.0 0.2 0.4 0.6 0.8 2 3 4 5 6 7 8 9 10 [11-20] [21-30] [31-40] 50+ RoBERTa BART Figure 2: Test set performance (Recall score) on PolicyIE for slots with different length. into seven types. Note that a predicted slot is considered correct if its’ label and span both match (exact match) one of the references. We characterize the error types as follows. 1. Wrong Intent (WI): The predicted intent label does not match the reference intent label. 2. Missing Slot (MS): None of the predicted slots exactly match a reference slot. 3. Spurious Slot (SS): Label of a predicted slot does not match any of the references. 4. Wrong Split (WSp): Two or more predicted slot spans with the same label could be merged to match one of the reference slots. A merged span and a reference span may only differ in punctuations or stopwords (e.g., and). 5. Wrong Boundary (WB): A predicted slot span is a sub-string of the reference span or vice versa. The slot label must exactly match. 4409 + [IN:data-collection-usage [SL:data-provider.third-party-entity third parties] [SL:action collect] [SL:dataprovider.user your] [SL:data-collected.data-general information] [SL:data-collector.first-party-entity us]] −[IN:data-sharing-disclosure [SL:data-receiver.third-party-entity third parties] [SL:action share] [SL:data-provider.user your] [SL:data-shared.data-general information] [SL:data-sharer.first-party-entity us] [SL:condition where applicable] [SL:condition based on their own privacy policies]] Error types: Wrong Intent (WI), Wrong Label (WL), Wrong Slot (WS), Spurious Slot (SS) + [.. .[SL:data-provider.third-party-entity third parties] [SL:condition it is allowed by applicable law or according to your agreement with third parties]] −[. . . [SL:condition allowed by applicable law or according to your agreement with third parties]] Error types: Wrong Boundary (WB), Missing Slot (MS) + [. . . [SL:data-receiver.third-party-entity social media and other similar platforms] ...] −[. . . [SL:data-receiver.third-party-entity social media] [SL:data-receiver.third-party-entity other similar platforms] . . . ] Error types: Wrong Split (WSp) Table 7: Three examples showing different error types appeared in BART’s predictions. + and −indicates the reference and predicted sequences, respectively. Best viewed in color. Error RoBERTa BART Wrong Intent 161 178 Spurious Slot 472 723 Missing Slot 867 517 Wrong Boundary 130 160 Wrong Slot 103 143 Wrong Split 32 27 Wrong Label 18 19 Total Slots 2,198 2,198 Correct Prediction 1,064 1,361 Total Errors 1,622 1,589 Total Predictions 2,686 2,950 Table 8: Counts for each error type on the test set of PolicyIE using RoBERTa and BART models. 6. Wrong Label (WL): A predicted slot span matches a reference, but the label does not. 7. Wrong Slot (WS): All other types of errors fall into this category. We provide one example of each error type in Table 7. In Table 8, we present the counts for each error type made by RoBERTa and BART models. The two most frequent error types are SS and MS. While BART makes more SS errors, RoBERTa suffers from MS errors. While both the models are similar in terms of total errors, BART makes more correct predictions resulting in a higher Recall score, as discussed before. One possible way to reduce SS errors is by penalizing more on wrong slot label prediction than slot span. On the other hand, reducing MS errors is more challenging as many missing slots have fewer annotations than others. We provide more qualitative examples in Appendix (see Table 11 and 12) . In the error analysis, we exclude the test examples (sentences) with the intent label “Other” and no slots. Out of 1,041 test instances in PolicyIE, there are 682 instances with the intent label “Other”. We analyze RoBERTa and BART’s predictions on those examples separately to check if the models predict slots as we consider them as spurious slots. While RoBERTa meets our expectation of performing highly accurate (correct prediction for 621 out of 682), BART also correctly predicts 594 out of 682 by precisely generating “[IN:Other]”. Overall the error analysis aligns with our anticipation that the Seq2Seq modeling technique has promise and should be further explored in future works. 5 Related Work Automated Privacy Policy Analysis Automating privacy policy analysis has drawn researchers’ attention as it enables the users to know their rights and act accordingly. Therefore, significant research efforts have been devoted to understanding privacy policies. Earlier approaches (Costante et al., 2012) designed rule-based pattern matching techniques to extract specific types of information. Under the Usable Privacy Project (Sadeh et al., 2013), several works have been done (Bhatia and Breaux, 2015; Wilson et al., 2016a,b; Sathyendra et al., 2016; Bhatia et al., 2016; Hosseini et al., 2016; Mysore Sathyendra et al., 2017; Zimmeck et al., 2019; Bannihatti Kumar et al., 2020). No4410 table works leveraging NLP techniques include text alignment (Liu et al., 2014; Ramanath et al., 2014), text classification (Wilson et al., 2016a; Harkous et al., 2018; Zimmeck et al., 2019), and question answering (QA) (Shvartzshanider et al., 2018; Harkous et al., 2018; Ravichander et al., 2019; Ahmad et al., 2020). Bokaie Hosseini et al. (2020) is the most closest to our work that used named entity recognition (NER) modeling technique to extract third party entities mentioned in policy documents. Our proposed PolicyIE corpus is distinct from the previous privacy policies benchmarks: OPP115 (Wilson et al., 2016a) uses a hierarchical annotation scheme to annotate text segments with a set of data practices and it has been used for multilabel classification (Wilson et al., 2016a; Harkous et al., 2018) and question answering (Harkous et al., 2018; Ahmad et al., 2020); PrivacyQA (Ravichander et al., 2019) frame the QA task as identifying a list of relevant sentences from policy documents. Recently, Bui et al. (2021) created a dataset by tagging documents from OPP-115 for privacy practices and uses NER models to extract them. In contrast, PolicyIE is developed by following semantic parsing benchmarks, and we model the task following the NLP literature. Intent Classification and Slot Filling Voice assistants and chat-bots frame the task of natural language understanding via classifying intents and filling slots given user utterances. Several benchmarks have been proposed in literature covering several domains, and languages (Hemphill et al., 1990; Coucke et al., 2018; Gupta et al., 2018; Upadhyay et al., 2018; Schuster et al., 2019; Xu et al., 2020; Li et al., 2021). Our proposed PolicyIE corpus is a new addition to the literature within the security and privacy domain. PolicyIE enables us to build conversational solutions that users can interact with and learn about privacy policies. 6 Conclusion This work aims to stimulate research on automating information extraction from privacy policies and reconcile it with users’ understanding of their rights. We present PolicyIE, an intent classification and slot filling benchmark on privacy policies with two alternative neural approaches as baselines. We perform a thorough error analysis to shed light on the limitations of the two baseline approaches. We hope this contribution would call for research efforts in the specialized privacy domain from both privacy and NLP communities. Acknowledgments The authors acknowledge the law students Michael Rasmussen and Martyna Glaz at Fordham University who worked as annotators to make the development of this corpus possible. This work was supported in part by National Science Foundation Grant OAC 1920462. Any opinions, findings, conclusions, or recommendations expressed herein are those of the authors, and do not necessarily reflect those of the US Government or NSF. Broader Impact Privacy and data breaches have a significant impact on individuals. In general, security breaches expose the users to different risks such as financial loss (due to losing employment or business opportunities), physical risks to safety, and identity theft. Identity theft is among the most severe and fastest-growing crimes. However, the risks due to data breaches can be minimized if the users know their rights and how they can exercise them to protect their privacy. This requires the users to read the privacy policies of websites they visit or the mobile applications they use. As reading privacy policies is a tedious task, automating privacy policy analysis reduces the burden of users. Automating information extraction from privacy policies empowers users to be aware of their data collected and analyzed by service providers for different purposes. Service providers collect consumer data at a massive scale and often fail to protect them, resulting in data breaches that have led to increased attention towards data privacy and related risks. Reading privacy policies to understand users’ rights can help take informed and timely decisions on safeguarding data privacy to mitigate the risks. Developing an automated solution to facilitate policy document analysis requires labeled examples, and the PolicyIE corpus adds a new dimension to the available datasets in the security and privacy domain. While PolicyIE enables us to train models to extract fine-grained information from privacy policies, the corpus can be coupled with other existing benchmarks to build a comprehensive system. For example, PrivacyQA corpus (Ravichander et al., 2019) combined with PolicyIE can facilitate building QA systems that can answer questions with fine-grained details. We believe our experiments and analysis will help direct future research. 4411 References Wasi Ahmad, Jianfeng Chi, Yuan Tian, and Kai-Wei Chang. 2020. PolicyQA: A reading comprehension dataset for privacy policies. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 743–749, Online. Association for Computational Linguistics. Ron Artstein and Massimo Poesio. 2008. Survey article: Inter-coder agreement for computational linguistics. Computational Linguistics, 34(4):555– 596. Vinayshekhar Bannihatti Kumar, Roger Iyengar, Namita Nisal, Yuanyuan Feng, Hana Habib, Peter Story, Sushain Cherivirala, Margaret Hagan, Lorrie Cranor, Shomir Wilson, et al. 2020. Finding a choice in a haystack: Automatic extraction of optout statements from privacy policy text. In Proceedings of The Web Conference 2020, pages 1943– 1954. Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Liu, Yu Wang, Jianfeng Gao, Songhao Piao, Ming Zhou, et al. 2020. Unilmv2: Pseudomasked language models for unified language model pre-training. In International Conference on Machine Learning, pages 642–652. PMLR. Jaspreet Bhatia and Travis D Breaux. 2015. Towards an information type lexicon for privacy policies. In 2015 IEEE eighth international workshop on requirements engineering and law (RELAW), pages 19–24. IEEE. Jaspreet Bhatia, Morgan C Evans, Sudarshan Wadkar, and Travis D Breaux. 2016. Automated extraction of regulated information types using hyponymy relations. In 2016 IEEE 24th International Requirements Engineering Conference Workshops (REW), pages 19–25. IEEE. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. Transactions of the Association for Computational Linguistics, 5:135–146. Mitra Bokaie Hosseini, Pragyan K C, Irwin Reyes, and Serge Egelman. 2020. Identifying and classifying third-party entities in natural language privacy policies. In Proceedings of the Second Workshop on Privacy in NLP, pages 18–27, Online. Association for Computational Linguistics. Duc Bui, Kang G Shin, Jong-Min Choi, and Junbum Shin. 2021. Automated extraction and presentation of data practices in privacy policies. Proceedings on Privacy Enhancing Technologies, 2021(2):88–110. Qian Chen, Zhu Zhuo, and Wen Wang. 2019. Bert for joint intent classification and slot filling. arXiv preprint arXiv:1902.10909. Federal Trade Commission et al. 2012. Protecting consumer privacy in an era of rapid change. FTC report. Elisa Costante, Jerry den Hartog, and Milan Petkovi´c. 2012. What websites know about you. In Data Privacy Management and Autonomous Spontaneous Security, pages 146–159. Springer. Ryan Cotterell and Kevin Duh. 2017. Lowresource named entity recognition with crosslingual, character-level neural conditional random fields. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 91–96, Taipei, Taiwan. Asian Federation of Natural Language Processing. Alice Coucke, Alaa Saade, Adrien Ball, Th´eodore Bluche, Alexandre Caulier, David Leroy, Cl´ement Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, et al. 2018. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. arXiv preprint arXiv:1805.10190. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In Advances in Neural Information Processing Systems, pages 13063–13075. Joshua Gluck, Florian Schaub, Amy Friedman, Hana Habib, Norman Sadeh, Lorrie Faith Cranor, and Yuvraj Agarwal. How short is too short? implications of length and framing on the effectiveness of privacy notices. In Twelfth Symposium on Usable Privacy and Security ({SOUPS} 2016). Abhirut Gupta, Anupama Ray, Gargi Dasgupta, Gautam Singh, Pooja Aggarwal, and Prateeti Mohapatra. 2018. Semantic parsing for technical support questions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3251– 3259, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Hamza Harkous, Kassem Fawaz, R´emi Lebret, Florian Schaub, Kang G Shin, and Karl Aberer. 2018. Polisis: Automated analysis and presentation of privacy policies using deep learning. In 27th {USENIX} Security Symposium ({USENIX} Security 18), pages 531–548. Charles T Hemphill, John J Godfrey, and George R Doddington. 1990. The atis spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, Pennsylvania, June 24-27, 1990. 4412 Mitra Bokaei Hosseini, Sudarshan Wadkar, Travis D Breaux, and Jianwei Niu. 2016. Lexical similarity of information type hypernyms, meronyms and synonyms in privacy policies. In 2016 AAAI Fall Symposium Series. Krippendorff Klaus. 1980. Content analysis: An introduction to its methodology. John Lafferty, Andrew McCallum, and Fernando CN Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Haoran Li, Abhinav Arora, Shuohui Chen, Anchit Gupta, Sonal Gupta, and Yashar Mehdad. 2021. MTOP: A comprehensive multilingual task-oriented semantic parsing benchmark. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2950–2962, Online. Association for Computational Linguistics. Bing Liu and Ian Lane. 2016. Attention-based recurrent neural network models for joint intent detection and slot filling. In Interspeech 2016, 17th Annual Conference of the International Speech Communication Association, pages 685–689. Fei Liu, Rohan Ramanath, Norman Sadeh, and Noah A. Smith. 2014. A step towards usable privacy policy: Automatic alignment of privacy statements. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 884–894, Dublin, Ireland. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Yi Luan, Dave Wadden, Luheng He, Amy Shah, Mari Ostendorf, and Hannaneh Hajishirzi. 2019. A general framework for information extraction using dynamic span graphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3036–3046, Minneapolis, Minnesota. Association for Computational Linguistics. Xuezhe Ma and Eduard Hovy. 2016. End-to-end sequence labeling via bi-directional LSTM-CNNsCRF. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1064–1074, Berlin, Germany. Association for Computational Linguistics. Florencia Marotta-Wurgler. 2015. Does “notice and choice” disclosure regulation work? an empirical study of privacy policies,”. In Michigan Law: Law and Economics Workshop. Aleecia M McDonald and Lorrie Faith Cranor. 2008. The cost of reading privacy policies. Isjlp, 4:543. Kanthashree Mysore Sathyendra, Shomir Wilson, Florian Schaub, Sebastian Zimmeck, and Norman Sadeh. 2017. Identifying the provision of choices in privacy policy text. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. Rohan Ramanath, Fei Liu, Norman Sadeh, and Noah A Smith. 2014. Unsupervised alignment of privacy policies using hidden markov models. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 605–610. Abhilasha Ravichander, Alan W Black, Shomir Wilson, Thomas Norton, and Norman Sadeh. 2019. Question answering for privacy policies: Combining computational and legal perspectives. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4947–4958. Joel R Reidenberg, Jaspreet Bhatia, Travis D Breaux, and Thomas B Norton. 2016. Ambiguity in privacy policies and the impact of regulation. The Journal of Legal Studies, 45(S2):S163–S190. Dennis Reidsma and Jean Carletta. 2008. Reliability measurement without limits. Computational Linguistics, 34(3):319–326. Subendhu Rongali, Luca Soldaini, Emilio Monti, and Wael Hamza. 2020. Don’t parse, generate! a sequence to sequence architecture for task-oriented semantic parsing. In Proceedings of The Web Conference 2020, pages 2962–2968. Norman Sadeh, Alessandro Acquisti, Travis D Breaux, Lorrie Faith Cranor, Aleecia M McDonald, Joel R Reidenberg, Noah A Smith, Fei Liu, N Cameron Russell, Florian Schaub, et al. 2013. The usable privacy policy project. Technical report, Technical Report, CMU-ISR-13-119. Kanthashree Mysore Sathyendra, Florian Schaub, Shomir Wilson, and Norman Sadeh. 2016. Automatic extraction of opt-out choices from privacy policies. In 2016 AAAI Fall Symposium Series. Sebastian Schuster, Sonal Gupta, Rushin Shah, and Mike Lewis. 2019. Cross-lingual transfer learning for multilingual task oriented dialog. In Proceedings of the 2019 Conference of the North American 4413 Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3795–3805. Yan Shvartzshanider, Ananth Balashankar, Thomas Wies, and Lakshminarayanan Subramanian. 2018. RECIPE: Applying open domain question answering to privacy policies. In Proceedings of the Workshop on Machine Reading for Question Answering, pages 71–77, Melbourne, Australia. Association for Computational Linguistics. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and TieYan Liu. Mass: Masked sequence to sequence pretraining for language generation. In International Conference on Machine Learning. Pontus Stenetorp, Sampo Pyysalo, Goran Topi´c, Tomoko Ohta, Sophia Ananiadou, and Jun’ichi Tsujii. 2012. Brat: a web-based tool for nlp-assisted text annotation. In Proceedings of the Demonstrations at the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 102–107. Milan Straka, Jan Hajic, and Jana Strakov´a. 2016. Udpipe: trainable pipeline for processing conll-u files performing tokenization, morphological analysis, pos tagging and parsing. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 4290– 4297. Shyam Upadhyay, Manaal Faruqui, Gokhan T¨ur, Hakkani-T¨ur Dilek, and Larry Heck. 2018. (almost) zero-shot cross-lingual spoken language understanding. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6034–6038. IEEE. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30, pages 5998–6008. Curran Associates, Inc. David Wadden, Ulme Wennberg, Yi Luan, and Hannaneh Hajishirzi. 2019. Entity, relation, and event extraction with contextualized span representations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5784– 5789, Hong Kong, China. Association for Computational Linguistics. Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep selfattention distillation for task-agnostic compression of pre-trained transformers. In Advances in Neural Information Processing Systems. Shomir Wilson, Florian Schaub, Aswarth Abhilash Dara, Frederick Liu, Sushain Cherivirala, Pedro Giovanni Leon, Mads Schaarup Andersen, Sebastian Zimmeck, Kanthashree Mysore Sathyendra, N. Cameron Russell, Thomas B. Norton, Eduard Hovy, Joel Reidenberg, and Norman Sadeh. 2016a. The creation and analysis of a website privacy policy corpus. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1330–1340. Shomir Wilson, Florian Schaub, Rohan Ramanath, Norman Sadeh, Fei Liu, Noah A Smith, and Frederick Liu. 2016b. Crowdsourcing annotations for websites’ privacy policies: Can it really work? In Proceedings of the 25th International Conference on World Wide Web, pages 133–143. International World Wide Web Conferences Steering Committee. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45, Online. Association for Computational Linguistics. Weijia Xu, Batool Haider, and Saab Mansour. 2020. End-to-end slot alignment and recognition for crosslingual NLU. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5052–5063, Online. Association for Computational Linguistics. Xiaodong Zhang and Houfeng Wang. 2016. A joint model of intent determination and slot filling for spoken language understanding. In IJCAI, volume 16, pages 2993–2999. Jie Zhou and Wei Xu. 2015. End-to-end learning of semantic role labeling using recurrent neural networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1127–1137, Beijing, China. Association for Computational Linguistics. Qile Zhu, Haidar Khan, Saleh Soltan, Stephen Rawls, and Wael Hamza. 2020. Don’t parse, insert: Multilingual semantic parsing with insertion based decoding. In Proceedings of the 24th Conference on Computational Natural Language Learning, pages 496– 506, Online. Association for Computational Linguistics. Sebastian Zimmeck, Peter Story, Daniel Smullen, Abhilasha Ravichander, Ziqi Wang, Joel Reidenberg, N Cameron Russell, and Norman Sadeh. 2019. Maps: Scaling privacy compliance analysis to a million apps. Proceedings on Privacy Enhancing Technologies, 2019(3):66–86. 4414 Type-I slots Attributes Action None Data Provider (1) User (2) Third party entity Data Collector (1) First party entity Data Collected (1) General Data (2) Aggregated/Non-identifiable data (3) Contact data (4) Financial data (5) Location data (6) Demographic data (7) Cookies, web beacons and other technologies (8) Computer/Device data (9) User online activities/profiles (10) Other data Data Sharer (1) First party entity Data Shared (1) General Data (2) Aggregated/Non-identifiable data (3) Contact data (4) Financial data (5) Location data (6) Demographic data (7) Cookies, web beacons and other technologies (8) Computer/Device data (9) User online activities/profiles (10) Other data Data Receiver (1) Third party entity Data Holder (1) First party entity (2) Third party entity Data Retained (1) General Data (2) Aggregated/Non-identifiable data (3) Contact data (4) Financial data (5) Location data (6) Demographic data (7) Cookies, web beacons and other technologies (8) Computer/Device data (9) User online activities/profiles (10) Other data Storage Place None Retention Period None Data Protector (1) First party entity (2) Third party entity Data Protected (1) General Data (2) Aggregated/Non-identifiable data (3) Contact data (4) Financial data (5) Location data (6) Demographic data (7) Cookies, web beacons and other technologies (8) Computer/Device data (9) User online activities/profiles (10) Other data Protect Against Security threat Type-II slots Attributes Purpose (1) Basic service/feature (2) Advertising/Marketing (3) Legal requirement (4) Service operation and security (5) Personalization/customization (6) Analytics/research (7) Communications (8 Merge/Acquisition (9) Other purpose Condition None Polarity (1) Negation Protection Method (1) General safeguard method (2) User authentication (3) Access limitation (5) Encryptions (6) Other protection method Table 9: Slots and their associated attributes. “None” indicates there are no attributes for the those slots. 4415 Privacy Practices Data Data Data Data Collection/Usage Sharing/Disclosure Storage/Retention Security/Protection Type-I slots Action 750 / 169 344 / 70 198 / 57 102 / 31 Data Provider 784 / 172 247 / 54 139 / 44 65 / 20 Data Collector 653 / 151 Data Collected 1833 / 361 Data Sharer 288 / 54 Data Shared 541 / 110 Data Receiver 456 / 115 Data Holder 192 / 59 Data Retained 291 / 119 Storage Place 70 / 21 Retention Period 101 / 17 Data Protector 105 / 31 Data Protected 119 / 34 Protect Against 49 / 15 Type-II slots Purpose 894 / 193 327 / 65 168 / 40 5 / 0 Condition 337 / 81 154 / 26 81 / 25 43 / 7 Polarity 50 / 15 21 / 1 22 / 1 18 / 5 Protection Method 143 / 35 # of slots 5301 / 1142 2378 / 495 1262 / 383 649 / 178 # of sequences 919 / 186 380 / 83 232 / 61 103 / 29 Table 10: Privacy practices and the associated slots with their distributions. “X / Y” indicates there are X instances in the train set and Y instances in the test set. Other Data Collection Data Sharing Data Storage Data Security Other Data Collection Data Sharing Data Storage Data Security 0.92 0.05 0.01 0.01 0.01 0.12 0.77 0.05 0.06 0.00 0.24 0.07 0.65 0.02 0.02 0.10 0.25 0.06 0.59 0.00 0.17 0.02 0.10 0.03 0.67 0.2 0.4 0.6 0.8 Figure 3: Confusion matrix for intent classification using the RoBERTa model. Other Data Collection Data Sharing Data Storage Data Security Other Data Collection Data Sharing Data Storage Data Security 0.88 0.08 0.02 0.01 0.01 0.06 0.81 0.05 0.08 0.01 0.15 0.09 0.72 0.01 0.02 0.08 0.21 0.06 0.66 0.00 0.16 0.03 0.05 0.03 0.73 0.00 0.15 0.30 0.45 0.60 0.75 Figure 4: Confusion matrix for intent classification using the BART model. 4416 Label Text Ground truth data-holder.first-party-entity We action keep data-retained.data-general records retention-period.retention-period a period of no more than 6 years RoBERTa (P:1.0, R: 0.75) data-holder.first-party-entity We action keep retention-period.retention-period a period of no more than 6 years BART (P:1.0, R: 1.0) data-holder.first-party-entity We action keep data-retained.data-general records retention-period.retention-period a period of no more than 6 years Ground truth data-collector.first-party-entity We action access data-collected.data-general information RoBERTa (P:0.0, R: 0.0) data-sharer.first-party-entity We data-shared.data-general information BART (P:0.0, R: 0.0) data-sharer.first-party-entity We action disclose data-shared.data-general information Ground truth data-sharer.first-party-entity Marco Polo data-receiver.third-party-entity third party data-shared.data-general Personal Information data-provider.user users action transferred RoBERTa (P:0.6, R: 0.6) data-receiver.third-party-entity Marco data-sharer.first-party-entity our data-receiver.third-party-entity third party data-shared.data-general Personal Information action transferred BART (P:0.83, R: 1.0) data-sharer.first-party-entity Marco Polo data-receiver.third-party-entity third party data-shared.data-general Personal Information data-sharer.first-party-entity us data-provider.user users action transferred Ground truth data-sharer.first-party-entity We data-receiver.third-party-entity third parties action provide data-shared.data-general information RoBERTa (P:1.0, R: 1.0) data-sharer.first-party-entity We data-receiver.third-party-entity third parties action provide data-shared.data-general information BART (P:0.25, R: 0.25) data-collector.first-party-entity We data-provider.third-party-entity third parties action provide data-collected.data-general information Table 11: Sample RoBERTa and BART predictions of Type-I slots. () and () indicates correct and incorrect predictions, respectively. Precision (P) and recall (R) score is reported for each example in the left column. 4417 Ground truth [Label] condition [Text] you use our product and service or view the content provided by us RoBERTa (P:1.0, R: 1.0) [Label] condition [Text] you use our product and service or view the content provided by us BART (P:1.0, R: 1.0) [Label] condition [Text] you use our product and service or view the content provided by us Ground truth [Label] purpose.other [Text] their own purposes [Label] purpose.advertising-marketing [Text ] inform advertising related services provided to other clients RoBERTa (P:0.0, R: 0.0) [Label] None [Text] None BART (P:1.0, R: 1.0) [Label] purpose.other [Text] their own purposes [Label] purpose.advertising-marketing [Text] inform advertising related services provided to other clients Ground truth [Label] purpose.personalization-customization [Text] provide more tailored services and user experiences [Label] purpose.basic-service-feature [Text] remembering your account identity [Label] purpose.service-operation-and-security [Text] analyzing your account ’s security [Label] purpose.analytics-research [Text] analyzing your usage of our product and service [Label] purpose.advertising-marketing [Text] advertisement optimization ( helping us to provide you with more targeted advertisements instead of general advertisements based on your information ) RoBERTa (P:0.17, R: 0.2) [Label] purpose.basic-service-feature [Text] provide [Label] purpose.other [Text] purposes [Label] purpose.analytics-research [Text] remembering your account identity [Label] purpose.analytics-research [Text] analyzing your account ’s security [Label] purpose.analytics-research [Text] analyzing your usage of our product and service [Label] purpose.advertising-marketing [Text] advertisement optimization BART (P:0.43, R: 0.6) [Label] purpose.personalization-customization [Text] provide more tailored services and user experiences [Label] purpose.service-operation-and-security [Text] remembering your account identity [Label] purpose.service-operation-and-security [Text] analyzing your account ’s security [Label] purpose.analytics-research [Text] analyzing your usage of our product and service [Label] purpose.advertising-marketing [Text] advertisement optimization [Label] purpose.advertising-marketing [Text] provide you with more targeted advertisements instead of general advertisements [Label] purpose.advertising-marketing [Text] based on your information Table 12: Sample RoBERTa and BART predictions of Type-II slots. () and () indicates correct and incorrect predictions, respectively. Precision (P) and recall (R) score is reported for each example in the left column.
|
2021
|
340
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4418–4429 August 1–6, 2021. ©2021 Association for Computational Linguistics 4418 RADDLE: An Evaluation Benchmark and Analysis Platform for Robust Task-oriented Dialog Systems Baolin Peng1, Chunyuan Li1, Zhu Zhang12† , Chenguang Zhu1, Jinchao Li1, Jianfeng Gao1 1Microsoft Research, Redmond, WA 2Iowa State University / Ames, IA {bapeng,chunyl,chezhu,jincli,jfgao}@microsoft.com [email protected] Abstract For task-oriented dialog systems to be maximally useful, it must be able to process conversations in a way that is (1) generalizable with a small number of training examples for new task domains, and (2) robust to user input in various styles, modalities, or domains. In pursuit of these goals, we introduce the RADDLE1 benchmark 2, a collection of corpora and tools for evaluating the performance of models across a diverse set of domains. By including tasks with limited training data, RADDLE is designed to favor and encourage models with a strong generalization ability. RADDLE also includes a diagnostic checklist that facilitates detailed robustness analysis in aspects such as language variations, speech errors, unseen entities, and out-of-domain utterances. We evaluate recent state-of-the-art systems based on pre-training and fine-tuning, and find that grounded pre-training on heterogeneous dialog corpora performs better than training a separate model per domain. Adversarial training is also proposed to improve model robustness against noisy inputs. Overall, existing models are less than satisfactory in robustness evaluation, which suggests opportunities for future improvement. 1 Introduction Dialogs constitute a crucial communication channel in completing a broad range of tasks, such as weather query, flight and restaurant booking, movie booking, IT help desk, etc. Comparing to chitchat systems that are usually modeled with singleturn context-response pairs, task-oriented dialog systems involve retrieving information from knowledge bases and reasoning over multiple dialog turns. This makes it especially important for a system to †Work was done when Zhu Zhang was visiting MSR 1Robust tAsk-orienteD DiaLog systems Evaluation 2Benchmark link: http://aka.ms/raddle be able to produce response that are grounded on tasks goals and user intents. In a bid to support human-computer interactions, task-oriented dialog systems have been built to allow users to converse with a computer system using natural language, such as Siri, Google Assistant, Amazon Alexa, Microsoft XiaoIce (Zhou et al., 2020). Traditionally, a task-oriented dialog system uses a modularized pipeline with four modules that execute sequentially (Gao et al., 2019). A natural language understanding (NLU) module identifies user intents and extracts associated information such as slots and corresponding values from user input. A dialog state tracker (DST) infers the belief state (or user goal) from dialog history. The belief state is often used to query a task-specific database (DB) to obtain the DB state, such as the number of entities that match the user goal. The dialog state and DB state are then passed to a dialog policy (POL) module to select the next system action. A natural language generation (NLG) module converts the action to a natural language response. The human ability to converse is general, flexible, and robust. In contrast, most popular tools for dialog system development adopting the above modular systems are designed for specific tasks and struggle with out-of-scope data. If we aspire to develop models beyond extensively handcrafted rules and annotated data for each single domain/task, it is critical to develop a more unified, efficient and robust model that can more quickly learn to execute a range of tasks in different domains. To fuel research in this direction, we present the RADDLE benchmark. It includes a collection of task-oriented dialog tasks in diverse domains (e.g. end-to-end modeling, dialog state tracking). The benchmark also has a companion online platform for model evaluation, comparison, and robustness analysis. Importantly, RADDLE exhibits two 4419 unique advantages that pave the way for building more pragmatic dialog systems: (i) Limited data setting is the major focus of RADDLE, to evaluate the generalization ability of models. It aims at simulating the real-world application scenarios where only very limited amount of labelled data is available for new domains. Given this focus, RADDLE is therefore a favorable benchmark to evaluate recent models in the pre-training and finetuning paradigm, which learn to represent linguistic knowledge in a way that facilitates sample-efficient learning and effective knowledge transfer. (ii) Robustness analysis is introduced to study model performance in various challenging scenarios, where models are evaluated with anomalous user input such as language variations, speech errors, unseen entities and out-of-domain utterances. Failing to handle these inputs often produce inappropriate responses leading to frustrating user experience. These scenarios are common for deployed systems in the real world, but are largely ignored in existing dialog benchmarks. To the best of our knowledge, RADDLE presents the first work to fill this gap. To better understand the challenges posed by RADDLE, we conduct experiments with simple baselines and state-of-the-art task-oriented dialog models. We find that grounded pre-trained models with a unified multi-task learning objective outperform models separately trained on each domain. Moreover, even the best performing model (SOLOIST (Peng et al., 2020a)) in our evaluation achieves a fairly low score in robustness analysis. This suggests that our baseline models can handle common inputs with strong regularities, but struggle with anomalous inputs that require deeper reasoning. In summary, our key contributions are: (i) A novel dialog benchmark with an emphasis on limited data and multiple domains/tasks, which formally creates a scenario to evaluate the grounding and generalization ability of pre-trained models. (ii) A crowd-sourced diagnostic evaluation dataset to cover a broad range of real-world sophistication to study model robustness. (iii) An online evaluation platform and leaderboard to track research progress, with human evaluation services to be granted to top-ranked submissions on a bi-monthly basis. (iv) Baseline results for major existing approaches to task-oriented dialogs are reported. An adversarially robust model is proposed to improve the generalization ability in noisy environments. Starter codes, pre-trained models, and scripts to reproduce the results will be provided together with the benchmark. 2 Related Work 2.1 Dialog Benchmarks To drive the progress of building dialog systems using data-driven approaches, a number of conversational corpora have been released. They are roughly grouped into two categories: (i) Corpora with structured semantic labels (Wen et al., 2017; Shah et al., 2018). These datasets are often specifically annotated, and used to study an individual module in the dialog pipeline. For example, DialoGLUE (Mehri et al., 2020) is a recently proposed benchmark with a focus on NLU and DST tasks. (ii) Corpora with an implicit user goal (Lowe et al., 2015). These datasets are often without semantic labels but can be used in end-to-end (E2E) dialog modeling (Li et al., 2016; Zhu, 2020; Wu et al., 2019; Zhu et al., 2019a; Lee et al., 2019; Zhu et al., 2020). MultiWOZ (Budzianowski et al., 2018) is the most related work to RADDLE. It is a large-scale multi-turn conversational corpus across several domains. It can be used to develop individual dialog modules as separate tasks for existing modularbased methods, or serves as a benchmark for E2E dialog modeling methods. RADDLE inherits the advantages of MultiWOZ in its flexibility for separate/joint task modeling and its comprehensiveness in multi-domain data coverage, but differs significantly in two aspects: an emphasis on limited data settings and an unique robustness checklist. Both are essential qualities in building task bots at scale. Further, RADDLE provides an online platform for model evaluation and fair comparison based on privately-held test data, inspired by GLUE (Wang et al., 2018). To the best of our knowledge, RADDLE is the first online platform for DST and E2E tasks in the dialog community. This can reduce the inconsistency caused by different researchers/teams using varying processing/evaluation scripts to dilute where the gain comes from. 2.2 Evaluation of Pre-Trained Models Pre-trained language models (PLMs) have substantially advanced the state of the art across a variety of language understanding and generation tasks (Peters et al., 2018; Devlin et al., 2019; Yang et al., 2019; Liu et al., 2019; Radford et al., 2019; 4420 Standard Language Variations / Speech Errors Unseen OOD Domain Attraction Train Hotel Restaurant Attraction Train Hotel Restaurant Reminder Attraction #Train 50 50 50 50 50 50 #Test 100 200 200 200 100 200 200 200 400 800 Task Dialog State Tracking / End-to-End Modeling DST / IC DST / OOD Metrics Joint Goal Accuracy / Combined Score JGA / Acc. JGA / F1 Table 1: Dataset descriptions and statistics. DST is short for Dialog State Tracking, E2E denotes End-to-End modeling, and IC stands for Intent Classification. Joint Goal Accuracy (JGA) is used for DST and Combined score is used for E2E. Keskar et al., 2019; Dong et al., 2019; Peng et al., 2020b,c; Li et al., 2020a). PLMs are often trained to predict words based on their context on massive text data, and the learned models can be fine-tuned to quickly adapt to various downstream tasks, exhibiting strong generalization capacity even with just a few in-domain training examples. Building task bots at scale requires the model to deal with the limited data problem for each domain, which can be used as a testbed to evaluate the generalization ability of PLMs. To this end, we limit the number of task-specific training examples in RADDLE to evaluate the sample-efficiency of models. Meanwhile, task-oriented dialogs pose a unique set of challenges for PLMs (Gao et al., 2020): a dialog is intrinsically goal-driven, multi-turn and often informal/noisy. Indeed, dialog-specific PLMs are proposed (Wu et al., 2020a; Peng et al., 2020a). However, the robustness of PLMs to linguistic perturbations often occurring in dialog settings (See Section 4 for details) is largely unexplored. Note that our notion of robustness emphasizes natural language variations, which is different from adversarial examples/training that aim to fool a trained model (Nie et al., 2019). From this perspective, RADDLE provides an unique benchmark for assessing PLMs with a robustness orientation. 3 Tasks RADDLE is centered on five English dialog scenarios in daily life, which cover a broad range of data collection schemes, task types and complexities. As our first goal of RADDLE is to spur development of generalizable dialog systems, we design the benchmark such that a good performance requires a model to leverage substantial knowledge (e.g., pretrained parameters) learned from its previous life cycle, while still maintaining some task-specific components (Coope et al., 2020; Henderson et al., 2020; Peng et al., 2020a; Wu et al., 2020b). Specifically, we deliberately keep a small number of training examples for each scenario. This is consistent with the common practice that only limited labelled data is provided when deploying a dialog system to new domains. Table 1 shows the data statistics. Four domains in the standard-setting are sampled from MultiWOZ 2.0 (Budzianowski et al., 2018). Reminder is intentionally only utilized for unseen entity tracking. Because it is a humanmachine corpus with a relatively smaller action space meaning that the impact of policy learning on models is largely alleviated. Therefore, the performance of models on this corpus will mostly reflect its capability of unseen entity tracking. Note that the number of training examples is limited to 50, an accepted scale that users can provide. Though it is possible to train a single model for each task from scratch without outside sources of knowledge, we expect that our focus on data-scarce settings will render this approach uncompetitive. Furthermore, a typical task-oriented dialog system uses a modularized pipeline that has four modules and executes sequentially. Recent research has shown promising results on parameterizing the modularized pipeline using a single neural autoregressive model, and training it in an end-to-end manner (Peng et al., 2020a; Ham et al., 2020; Hosseini-Asl et al., 2020). In fact, a single autoregressive model can significantly ease the workflow of training and deploying dialog systems for new tasks, compared to existing modularized tools and methods. Therefore, we design the benchmark to allow evaluations on end-to-end dialog modeling, in addition to the modularized evaluation on dialog state tracking. To reveal the gap between the complexity of dialogs in lab environments and that in real scenarios, we construct a suite of tasks to study the robustness of models. We describe these tasks below and in Table 1. On the evaluation front, we concentrate on 4421 simulation-based methodologies, in order to facilitate automation. Though we only offer human evaluations (Gao et al., 2019) to top-ranked submissions at this point, we emphasize realistic scenarios in pursuit of system robustness (see Section 4). Task 1: Dialog State Tracking A robust NLU and DST is the first step towards building a reliable dialog system. The dialog state is a summary of the entire conversation till the current turn. In a task-oriented system, it is represented in the form of slot-value pairs, where slot indicates the category/attribute of the user goal expressed in the utterance, and value is the corresponding information. For the evaluation metric, we report joint goal accuracy, which indicates the proportion of dialog turns where all the user’s search goal constraints are correctly identified (Mrksic et al., 2017). To specially study the NLU performance, we consider intent classification, which aims to automatically extract meaning from a natural language utterance in order to understand user’s goal (Hemphill et al., 1990; Zhu et al., 2019b). Task 2: End-to-End Modeling The end-to-end (E2E) dialog models consider dialog history as input, and produce the natural language response. It jointly implements the dialog management (including DST and POL) and response generation (i.e., NLG) components. Following Budzianowski et al. (2018), Inform, Success, and BLEU scores are reported. The first two metrics evaluate dialog task completion: Inform measures if the system provides a correct entity (inform rate), meanwhile Success measures the exact matching of answering all the requested information (success rate), and if the answered information matches users’ goal. BLEU evaluates how fluent the generated responses are compared to human-written responses. A combined score (Combined) is also reported using Combined = (Inform + Success) × 0.5 + BLEU as an overall quality measure, as suggested in (Budzianowski et al., 2018). 4 Robustness Diagnostic Checklist Existing benchmarks assume a world of a “perfect” user who always provides precise, concise, and semantically unambiguous utterances. These goal-oriented dialog datasets are largely collected by crowd-sourcing, where a crowd-sourced worker enacts the part of a real user by following a set of template instructions provided for the task. This method results in a dataset where most user utterances are straight-forward, stick to the goal and tend to leave out the variation/errors commonly found in real-world conversational data. To this end, we collect a suite of language variations to reveal the dialog sophistication in the real world, and measure the robustness of dialog models. 4.1 Checklist Tasks Language Variations It is well-known that humans communicate using language with fairly large variations such as different ways of expressions or personalized styles (Sacks et al., 1978), while template-based crowd-sourcing fails in covering the linguistic variations (Schegloff et al., 1977; Moore and Arar, 2019). Specifically, we consider four types of variations in RADDLE: (i) Paraphrase widely exists among different users, who may present restatements of the meaning of a text or message using other words. (ii) Verbosity describes a quality that users may express their intents using more words than needed. (iii) Simplification is a quality that users express their intents using fewer words to be concise. (iv) Typos often result from mistakes made in the typing. In Figure 1(b)-(e), we provide examples to illustrate these language variations. Speech Errors It is desirable that dialog systems can leverage automatic speech recognition (ASR) techniques to serve the speech modality, as in Amazon Alexa. However, almost all dialog systems have typically assumed that the user input is written text, and hoped that the system would seamlessly integrate with speech inputs. Recently, it has been empirically shown in Gopalakrishnan et al. (2020) that dialog systems trained on written data is very sensitive to various types of synthetic and actual ASR hypotheses in the dialog history. To bring attention to this gap, RADDLE promotes speech robustness as an evaluation criterion. For example in Figure 1(f), “what’s available” can be transcribed as “once available” due to ASR deficiency, and a robust dialog system is expected to still correctly perceive user intents. Unseen Entities Most existing DST methods are not designed to handle slot values that are not known to the tracker. The assumption that a predefined ontology exists for the dialog and one can enumerate all possible values for each slot is often not valid in real-world scenarios. Even if such lists or dictionaries exist, they can be very large in size 4422 User : I want to tour a college in the center of town. what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Christ's College is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : i want to tour college in the center of town once available System : I have several different options available . did you want one with free entrance ? User : entrance phi is not an issue. System : Christ's College is free to enter. User : Can i have the postcard and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : i know that's it uh thanks System : thank you , goodbye . User : I want to tour a park in the Bellevue downtown . what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Downtown Park is a nice one and it is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : I want to tour a college in the center of town . what's available? System : I have several different options available . did you want one with free entrance ? User : today is a sunny day we should go outside for a walk. System : Sorry I do not understand. Can you paraphrase? By the way, Christ's College is a good place and free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . Paraphrase Verbose User : I want to take a tour in a college which lies in the center of town. Any recommendation? System : I have several different options available . did you want one with free entrance ? User : I don't have a problem paying the entrance fee. System : Christ's College is free to enter. User : Will you please give me postcode and address of that college? System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, perfect. I appreciate that. System : thank you , goodbye . User : I want to tour a colleg center area is awesome. System : I have several diffe User : I do not have any pref something extra? System : Christ's College is f User : Okay, sounds good! I w you have its address and pos System : Yes, it's located at information you need? User : no, thanks. System : thank you , goodby User : Any college available in the center of town? System : I have several different options available . did you want one with free entrance ? User : do not care about fee System : Christ's College is free to enter. User : Give me address and postcode. System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thanks. System : thank you , goodbye . User : want to tour a college System : I have several diffe User : I don't have a problem System : Christ's College is f User : Plz tell me addr and p System : Yes, it's located at information you need? User : no, thats all. thx System : thank you , goodby (a) Standard dialog session (b) Paraphrase Verbose er of town. Any want one with free college? cb23bu. Is there any other User : I want to tour a college. what do you have available? I know some college in the center area is awesome. System : I have several different options available . did you want one with free entrance ? User : I do not have any preference on the fee but does paying the fee give me something extra? System : Christ's College is free to enter. User : Okay, sounds good! I would like to visit that college to enjoy this afternoon, do you have its address and postcode? System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thanks. System : thank you , goodbye . want one with free cb23bu. Is there any other User : want to tour a college in center. whats available System : I have several different options available . did you want one with free entrance ? User : I don't have a problem paying the entrance fee. System : Christ's College is free to enter. User : Plz tell me addr and postcode of that college. System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thats all. thx System : thank you , goodbye . Paraphrase Verbose User : I want to take a tour in a college which lies in the center of town. Any recommendation? System : I have several different options available . did you want one with free entrance ? User : I don't have a problem paying the entrance fee. System : Christ's College is free to enter. User : Will you please give me postcode and address of that college? System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, perfect. I appreciate that. System : thank you , goodbye . User : I want to tour a colleg center area is awesome. System : I have several diffe User : I do not have any pre something extra? System : Christ's College is User : Okay, sounds good! I you have its address and pos System : Yes, it's located at information you need? User : no, thanks. System : thank you , goodb User : Any college available in the center of town? System : I have several different options available . did you want one with free entrance ? User : do not care about fee System : Christ's College is free to enter. User : Give me address and postcode. System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thanks. System : thank you , goodbye . User : want to tour a college System : I have several diffe User : I don't have a problem System : Christ's College is User : Plz tell me addr and p System : Yes, it's located at information you need? User : no, thats all. thx System : thank you , goodb (c) Verbosity (d) Simplification Verbose ter of town. Any want one with free college? cb23bu. Is there any other User : I want to tour a college. what do you have available? I know some college in the center area is awesome. System : I have several different options available . did you want one with free entrance ? User : I do not have any preference on the fee but does paying the fee give me something extra? System : Christ's College is free to enter. User : Okay, sounds good! I would like to visit that college to enjoy this afternoon, do you have its address and postcode? System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thanks. System : thank you , goodbye . want one with free cb23bu. Is there any other User : want to tour a college in center. whats available System : I have several different options available . did you want one with free entrance ? User : I don't have a problem paying the entrance fee. System : Christ's College is free to enter. User : Plz tell me addr and postcode of that college. System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, thats all. thx System : thank you , goodbye . User : I want to tour a college in the center of town. what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Christ's College is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : i want to tour college in the center of town once available System : I have several different options available . did you want one with free entrance ? User : entrance phi is not an issue. System : Christ's College is free to enter. User : Can i have the postcard and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : i know that's it uh thanks System : thank you , goodbye . User : I want to tour a park in the Bellevue downtown . what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Downtown Park is a nice one and it is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : I want to tour a college in the center of town . what's available? System : I have several different options available . did you want one with free entrance ? User : today is a sunny day we should go outside for a walk. System : Sorry I do not understand. Can you paraphrase? By the way, Christ's College is a good place and free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . (e) Typos (f) Speech errors User : I want to tour a college in the center of town. what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Christ's College is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : i want to tour college in the center of town once available System : I have several different options available . did you want one with free entrance ? User : entrance phi is not an issue. System : Christ's College is free to enter. User : Can i have the postcard and address for that college?, System : Yes, it's located at saint andrew's street, postcode cb23bu. Is there any other information you need? User : i know that's it uh thanks System : thank you , goodbye . User : I want to tour a park in the Bellevue downtown . what's available? System : I have several different options available . did you want one with free entrance ? User : Entrance fee is not an issue., System : Downtown Park is a nice one and it is free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : I want to tour a college in the center of town . what's available? System : I have several different options available . did you want one with free entrance ? User : today is a sunny day we should go outside for a walk. System : Sorry I do not understand. Can you paraphrase? By the way, Christ's College is a good place and free to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . User : I want to tour a college in the center of town. what's available?, System : I have several different options available . did you want one with free entrance ? User : today is a sunny day we should go outside for a walk. System : Sorry I do not understand. Can you paraphrase? By the way, Christ’s College is a good place and fee to enter. User : Can i have the postcode and address for that college?, System : Yes, it's located at 10201 NE 4th St, postcode 98004. Is there any other information you need? User : no, that's it. thanks. System : thank you , goodbye . (g) Unseen entities (h) Out-of-domain utterance Figure 1: Illustration of different language perturbations in the robustness diagnostic checklist. The standard dialog example is shown in (a). Based on it, (b)-(e) are four types of language variations , (f) shows speech error , (e) shows unseen entities , and (h) shows out-of-domain utterance . In each case, some representative examples are highlighted in red text. and highly dynamic (Xu and Hu, 2018). Therefore, unseen entities are common in dialogs, i.e., entities that are not observed during training, but appear in the testing stage. In Figure 1(g), the entity Bellevue downtown is in the knowledge base but never appears in model training, a robust DST should be able to recognize it as a city/place, via generalizing from other similar entities learned during training. Out-of-Domain Utterances Most deployed task-oriented dialog systems are built for a closed set of target domains. Thus, they are fragile when dealing with out-of-domain (OOD) utterances (Lee and Shalyminov, 2019). Failure to detect OOD utterances often prevents the model from responding with an appropriate fallback action, hence leading to frustrating user experience. Therefore, it is important to endow task bots with the ability to detect OOD utterances for special handling (Larson et al., 2019). For example, in Figure 1(h), the user suggests an excursion to a task bot trained in college consulting, which is out of the bot’s scope. The bot is expected to raise a flag to label the utterance as an outlier, and guides the user to focus on the 4423 current domain. 4.2 Collection Protocols The standard setting is sampled from MultiWOZ 2.0 (Budzianowski et al., 2018) but re-purposed in a few-shot learning setting. The language variations corpus is created by workers on Amazon Mechanical Turks based on the standard corpus. To maximize the quality, we require workers in US locale and have a minimal previous approval rate of 90%. Assignments are constructed at the turn level. Given a user utterance and associated dialog history, workers are required to answer four questions, what are the paraphrase, typos, verbose, and simplified versions of the user utterance. Moreover, in each assignment, the workers are instructed to exactly mention the slot values in the answers if the given user utterance has them. We pay Turks 0.5$ per assignment and each assignment can be finished in one to two minutes. For the speech recognition errors setting, we employ the audio-level error simulation (Gopalakrishnan et al., 2020), which generates audio signals from texts, adds noise into the audio, and then decodes the audio with an ASR model to obtain hypotheses. In particular, we employ Microsoft Cognition text-to-speech service to synthesize audio signals. After injecting background noise into the audio signals, we use the speech recognition service to obtain a corpus of Word Error Rate (WER) of 30%. For the reminder domain that is applied for unseen entity evaluation, we firstly simulate several dialogs as seed scenarios using an agenda-based simulator and then randomly replace the slots in the dialogs with new values. Similar to constructing the language variations corpus, we then hire workers to rewrite the corpus as diverse and realistic as possible. Finally, the out-of-domain corpus is developed following Lee and Shalyminov (2019). We randomly choose 50% utterances in DSTC (Henderson et al., 2014) for the Attraction domain as the training set. For the test set, besides utterance from DSTC, we also introduce utterance from a diverse set of domains like Stanford (Eric and Manning, 2017), Reddit, Twitter (Sordoni et al., 2015) to evaluate the capability of handling different out-of-domain utterances. A board of data researchers reviews all the collected data to ensure no ethical concerns in it. 5 Methods 5.1 Competitive Baselines For baselines, we consider three representative methods, holding state-of-the-art positions on existing benchmarks such as MultiWoZ (Budzianowski et al., 2018). DAMD (Zhang et al., 2020) is a state-of-theart modular system, where each dialog module is implemented using a neural network, and the whole system is trained in an end-to-end manner. GPT-2 represents a single multi-task learning model with impressive results on general language understanding and generation tasks. GPT-2 is an auto-regressive language model that leverages 12-24 layers of masked, multi-head self-attention Transformers. GPT-2 is pre-trained on extremely massive text data OpenWebText (Radford et al., 2019). It has demonstrated superior performance on characterizing human language data distribution and knowledge transfer. Given text prompts, GPT-2 can often generate fluent sentences. Its ancestral work GPT (with a smaller model size and less training data) has shown impressive results on language understanding tasks. In this paper, we consider GPT-2FT as the approach of directly finetuning the pre-trained GPT-2 on a specific domain. Hence, GPT-2FT can be viewed as SOLOIST without grounded pre-training, and serve as a strong baseline for both DST and E2E task. SOLOIST represents recent model variants (Ham et al., 2020; Hosseini-Asl et al., 2020) to parameterize dialog system as a single auto-regressive model. SOLOIST subsumes different dialog modules (e.g. state tracker, dialog policy, response generator) into a single Transformer model. It has the similar capability with GPT-2 in understanding and generating natural language sentences but is pre-trained on large heterogeneous dialog corpora to gain additional capability of grounding text response in user goals and real-world knowledge for task completion (Peng et al., 2020a; Gao et al., 2020). For detailed description, please see Section A in Appendix. 5.2 Adversarially Robust SOLOIST It is known that adversarial training can improve a model’s adversarial robustness, which refers to a model’s invariance to small (often imperceptible) perturbations of its inputs (i.e., clean exam4424 Standard Para. Simp. Typos Verbo. Speech ERR Unseen OOD Model Avg. Avg.C JGA ↑ C ↑ JGA ↑ C ↑ JGA ↑ C ↑ JGA ↑ C ↑ JGA ↑ C ↑ JGA ↑ C ↑ JGA ↑ IC ↑ JGA ↑ F1 ↑ DAMD 14.18 48.99 6.75 44.13 5.78 42.93 5.33 42.58 7.08 42.56 9.1 45.94 GPT-2FT 47.46 46.53 40.52 67.36 31.36 62.72 28.82 59.44 22.31 54.15 30.40 54.16 31.41 65.95 28.28 51.29 47.37 83.86 SOLOIST 59.09 58.30 53.17 76.13 40.27 64.89 37.18 63.61 22.73 57.77 38.21 65.71 36.81 70.48 69.05 96.98 56.28 96.18 SOLOISTAdv 61.03 60.14 55.47 79.06 42.11 71.13 38.28 69.89 23.30 63.17 40.02 69.36 39.02 72.33 69.56 98.79 55.03 89.94 Table 2: Overall results of baselines across all RADDLE tasks. C indicates the Combined metric, IC denotes intent classification accuracy. Avg. is averaged over all the tasks while Avg.C is averaged over all the roubust checklist tasks. Para., Simp., Verbo. are short for Paraphrase, Simplification, and Verbosity. Note that it is not straightforward to directly apply DAMD to Unseen and OOD tasks since it requires extra annotations. As such, we omit results of DAMD on these two tasks. ples) (Madry et al., 2017; Miyato et al., 2018; Liu et al., 2020; Li et al., 2020b). Adversarial examples are produced by adding perturbations on clean examples to fool the predictions of a trained model the most. Though fundamentally different, one may view adversarial examples as resembling the variations in natural language to some extent. Inspired by this idea, we propose an adversarially robust SOLOIST model, denoted as SOLOISTAdv. Specifically, for a dialog turn x drawn from the training dataset D, and a neural model SOLOIST parameterized by θ, the standard training minimizes the empirical risk: minθ Ex∼DLθ(x), where Lθ(x) is the SOLOIST learning objective defined in Appendix Section A. The key idea of adversarial training is to modify the objective by applying small perturbation δ to input word embeddings that maximize the adversarial loss: minθ Ex∼D maxδ Lθ(x+δ), where the inner maximization can be solved by running a number of projected gradient descent steps (Goodfellow et al., 2014; Bubeck, 2014). SOLOISTAdv is trained in a hybrid manner that combines standard training and adversarial training. It augments the training dataset with adversarial examples that add perturbations in the word embedding space of original dialog turns, which improve the model’s robustness against noisy inputs that arguably covers language variations. In our experiments, SOLOISTAdv employs adversarial training in both task-specific pre-training and fine-tuning stages. 5.3 Submission Details Training We leverage the pre-trained checkpoints from the corresponding work, and fine-tune them on RADDLE. For SOLOISTAdv, We apply 100k steps of adversarial training to the pre-trained checkpoints. Each domain is trained separately. We train our models with Adam with initial learning rate 5e-5 and batch size 1 for 20 epochs. We encourage subsequent submissions systems to devote the same computation efforts in fine-tuning stage, e.g., up to one hour GPU time, for each model to ensure fair comparisons. Evaluation The RADDLE benchmark follows the same evaluation model as GLUE (Wang et al., 2018) or Kaggle3. To evaluate a system on the benchmark, one must run the system on the provided test data for the tasks, then upload the results to the website http://aka.ms/raddle for scoring. The benchmark site shows per-task scores and a macro-average of those scores to determine a system’s position on the leaderboard. The website also provides fine- and coarse-grained results on the robustness diagnostic datasets. We will provide human evaluation services for top-ranked submissions on a quarterly basis. The human evaluation protocol follows Peng et al. (2020a) and Li et al. (2020c). 6 Benchmark Results 6.1 Overall Results We first present the results of baseline methods across all tasks on the RADDLE benchmark in Table 2. As shown, GPT-2FT fine-tuned with domainspecific dialog corpora outperforms the strong modular-based method DAMD. This highlights the efficacy of pre-trained language models. SOLOIST improves upon GPT-2FT over 10 points in terms of average score, and consistently performs better than GPT-2FT across all the tasks. These strong results indicate that large-scale task-specific pretraining on dialog corpora is crucial for effective and robust task adaptation. However, the performance of SOLOIST drops on robust checklist tasks. Benefiting from adversarial training, SOLOISTAdv outperforms SOLOIST about 2 points. 3https://www.kaggle.com/ 4425 6.2 Robustness Diagnostic Checklist Results Table 2 shows the overall performance of DST and E2E modeling under different variation settings. Language Variations It is noticeable that all the models incur significant performance drops under each type of variation. Among all variation types, Typos has the most substantial impact on both JGA and Combined score resulting in 10 to 20 points of drop in performance. This is expected as misspelled keywords pose significant challenges for state tracking. The influence of other three types of variations are also prominent. The results reveal that existing SoTA dialog models trained on limited task-specific examples are not robust enough to handle various types of user utterances. Adversarial training improves robustness to language variations, boosting performance across all the language variations tasks. Speech Errors We observe a clear degradation in all metrics for all models. This shows that during inference, models trained on textual data are sensitive and not robust to actual ASR hypotheses introduced in dialog history. Unseen Entities Without task-specific pretraining, GPT-2FT only achieves less than 30% of JGA and 51.20 of dialog act accuracy even on a simple domain with most of the common entity values. SOLOIST performs significantly better than GPT-2FT by achieving 69.05% JGA and 96.98 dialog act accuracy but remains imperfect. SOLOISTAdv performs similar to SOLOIST, which is expected as adversarial training does not provides additional knowledge. These results imply that task-specific pre-training can improve the generalization capability of models but is still far from enough for production environments. Out-of-Domain Utterances It is non-trivial for conventional modular-based dialog systems to handle OOD detection. It often requires an additional component to classify whether a user utterance as in-domain or not. As such, we omit the result of DAMD in our experiments. GPT-2FT achieves 83.96 F1 score while SOLOIST has 96.18 F1 score, which shows that task-specific pre-training can improve robustness of models to OOD utterances. It is interesting to observe that adversarial training hurts model’s performance on OOD detection. We conjecture that adversarial training enable models to tolerate disturbances on the inputs and thus yield 1 2 3 4 5 Team ID 0.5 0.6 0.7 0.8 0.9 Success Rate Corpus E. Human E. Non-Pre-trained Models Pre-trained Models 1 2 3 4 5 Team ID 0.5 0.6 0.7 0.8 0.9 Success Rate DSTC-8 1 2 3 4 5 Team ID 0.5 0.6 0.7 0.8 0.9 Success Rate DSTC-9 (a) DSTC8 (b) DSTC9 Figure 2: Corpus and human evaluation for different models in two recent Multi-domain Dialog Challenges: (a) DSTC8 and (b) DSTC9. The regions indicate the gap between human and corpus evaluations for different types of models. We observe that (i) In DSTC8, Team 5 is the winner, and the only submission adopting pre-trained GPT-2 models; The performance discrepancy between the corpus and human evaluation is significantly smaller than other teams using modularbased methods without pre-training. (ii) a general trend shifting from modular based systems to pre-trained endto-end systems. (iii) a substantial drop in performance which indicates that pre-trained methods remain sensitive to noisy inputs. more false positive predictions on this task. Finally, it is worth pointing out some important trends in the dialog research community, based on the DSTC challenge (Kim et al., 2019; Gunasekara et al., 2020) in the last 2 years (Figure 2). In DSTC8 (Kim et al., 2019), the winning submission by Team 5 is the only one that uses pretrained models (GPT-2). When moving from corpus evaluation to human evaluation, it exhibits the least performance drop relative to other submissions, which is strong evidence to demonstrate robustness of pre-trained models. By the time of DSTC9 (Gunasekara et al., 2020), the community have witnessed a general trend shift from modular systems to pre-trained end-to-end architectures. However, the significant performance gap between corpus evaluation and human evaluation indicates that pre-trained methods remain sensitive to noisy inputs. Such observations underscore the importance of robustness-oriented design and evaluation, for which RADDLE fills a major void. 7 Conclusion We introduce RADDLE, a platform and collection of resources for evaluating and analyzing taskoriented dialog systems. We confirm (1) the utility of grounded pre-training and transfer learning methods in dialog systems: pre-training improves 4426 generalization in a limited data setting, and (2) adversarial training improves robustness, but still leaves room for improvement. When evaluating these models on our diagnostic dataset, we find that they fail (often spectacularly) on many robustness test cases, suggesting possible avenues for future work. In summary, the question of how to design unified, efficient, robust models remains largely unexplored, and we believe that RADDLE can provide fertile soil for addressing this challenge. Acknowledgement We gratefully acknowledge the entire Project Philly team inside Microsoft, who provided the computing platform for our research. We also thank the anonymous reviewers whose suggestions helped clarify this work. Ethical Considerations The collection of our RADDLE dataset is consistent with the terms of use of any sources and the original authors’ intellectual property and privacy rights. The dataset is collected with Amazon mechanical Turks, and each HIT requires up to two minutes to complete. The requested inputs are general language variations, and no privacy-related information is collected during data collection. Each HIT was paid 0.5 USD, with the hourly pay being 15% higher than the minimum wage requirements in our area. A board of data researchers has reviewed all the collected data to ensure no ethical concerns e.g., toxic language and hate speech. References S´ebastien Bubeck. 2014. Convex optimization: Algorithms and complexity. arXiv preprint arXiv:1405.4980. Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, I˜nigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gasic. 2018. Multiwoz-a largescale multi-domain wizard-of-oz dataset for taskoriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016–5026. Sam Coope, Tyler Farghly, Daniela Gerz, Ivan Vulic, and Matthew Henderson. 2020. Span-convert: Fewshot span extraction for dialog with pretrained conversational representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 510, 2020, pages 107–121. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre-training for natural language understanding and generation. In Advances in Neural Information Processing Systems, pages 13042–13054. Mihail Eric and Christopher D Manning. 2017. Keyvalue retrieval networks for task-oriented dialogue. arXiv preprint arXiv:1705.05414. Jianfeng Gao, Michel Galley, and Lihong Li. 2019. Neural approaches to conversational ai. Foundations and Trends R⃝in Information Retrieval, 13(23):127–298. Jianfeng Gao, Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and HeungYeung Shum. 2020. Robust conversational ai with grounded text generation. arXiv preprint arXiv:2009.03457. Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572. Karthik Gopalakrishnan, Behnam Hedayatnia, Longshaokan Wang, Yang Liu, and Dilek HakkaniTur. 2020. Are neural open-domain dialog systems robust to speech recognition errors in the dialog history? an empirical study. arXiv preprint arXiv:2008.07683. Chulaka Gunasekara, Seokhwan Kim, Luis Fernando D’Haro, Abhinav Rastogi, Yun-Nung Chen, Mihail Eric, Behnam Hedayatnia, Karthik Gopalakrishnan, Yang Liu, Chao-Wei Huang, et al. 2020. Overview of the ninth dialog system technology challenge: Dstc9. arXiv preprint arXiv:2011.06486. Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using gpt-2. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 583–592. Charles T. Hemphill, John J. Godfrey, and George R. Doddington. 1990. The ATIS spoken language systems pilot corpus. In Speech and Natural Language: Proceedings of a Workshop Held at Hidden Valley, Pennsylvania, June 24-27,1990. Matthew Henderson, I˜nigo Casanueva, Nikola Mrksic, Pei-Hao Su, Tsung-Hsien Wen, and Ivan Vulic. 4427 2020. Convert: Efficient and accurate conversational representations from transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, EMNLP 2020, Online Event, 16-20 November 2020, pages 2161–2174. Association for Computational Linguistics. Matthew Henderson, Blaise Thomson, and Jason D Williams. 2014. The second dialog state tracking challenge. In Proceedings of the 15th annual meeting of the special interest group on discourse and dialogue (SIGDIAL), pages 263–272. Ehsan Hosseini-Asl, Bryan McCann, Chien-Sheng Wu, Semih Yavuz, and Richard Socher. 2020. A simple language model for task-oriented dialogue. arXiv preprint arXiv:2005.00796. Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. 2019. Ctrl: A conditional transformer language model for controllable generation. arXiv preprint arXiv:1909.05858. Seokhwan Kim, Michel Galley, Chulaka Gunasekara, Sungjin Lee, Adam Atkinson, Baolin Peng, Hannes Schulz, Jianfeng Gao, Jinchao Li, Mahmoud Adada, et al. 2019. The eighth dialog system technology challenge. arXiv preprint arXiv:1911.06394. Stefan Larson, Anish Mahendran, Joseph J. Peper, Christopher Clarke, Andrew Lee, Parker Hill, Jonathan K. Kummerfeld, Kevin Leach, Michael A. Laurenzano, Lingjia Tang, and Jason Mars. 2019. An evaluation dataset for intent classification and out-of-scope prediction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1311–1316, Hong Kong, China. Association for Computational Linguistics. Sungjin Lee and Igor Shalyminov. 2019. Contextual out-of-domain utterance handling with counterfeit data augmentation. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7205–7209. IEEE. Sungjin Lee, Qi Zhu, Ryuichi Takanobu, Zheng Zhang, Yaoqin Zhang, Xiang Li, Jinchao Li, Baolin Peng, Xiujun Li, Minlie Huang, et al. 2019. Convlab: Multi-domain end-to-end dialog system platform. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 64–69. Chunyuan Li, Xiang Gao, Yuan Li, Baolin Peng, Xiujun Li, Yizhe Zhang, and Jianfeng Gao. 2020a. Optimus: Organizing sentences via pre-trained modeling of a latent space. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4678–4699, Online. Association for Computational Linguistics. Chunyuan Li, Xiujun Li, Lei Zhang, Baolin Peng, Mingyuan Zhou, and Jianfeng Gao. 2020b. Selfsupervised pre-training with hard examples improves visual representations. arXiv preprint arXiv:2012.13493. Jinchao Li, Baolin Peng, Sungjin Lee, Jianfeng Gao, Ryuichi Takanobu, Qi Zhu, Minlie Huang, Hannes Schulz, Adam Atkinson, and Mahmoud Adada. 2020c. Results of the multi-domain task-completion dialog challenge. In Proceedings of the 34th AAAI Conference on Artificial Intelligence, Eighth Dialog System Technology Challenge Workshop. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119. Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle Pineau. 2015. The ubuntu dialogue corpus: A large dataset for research in unstructured multi-turn dialogue systems. arXiv preprint arXiv:1506.08909. Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. 2017. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083. Shikib Mehri, Mihail Eric, and Dilek Hakkani-Tur. 2020. DialoGLUE: A natural language understanding benchmark for task-oriented dialogue. arXiv preprint arXiv:2009.13570. Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. 2018. Virtual adversarial training: a regularization method for supervised and semisupervised learning. T-PAMI. Robert J Moore and Raphael Arar. 2019. Conversational UX Design: A Practitioner’s Guide to the Natural Conversation Framework. ACM. Nikola Mrksic, Diarmuid ´O S´eaghdha, Tsung-Hsien Wen, Blaise Thomson, and Steve J Young. 2017. Neural belief tracker: Data-driven dialogue state tracking. In ACL (1). Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2019. Adversarial NLI: A new benchmark for natural language understanding. arXiv preprint arXiv:1910.14599. 4428 Baolin Peng, Chunyuan Li, Jinchao Li, Shahin Shayandeh, Lars Liden, and Jianfeng Gao. 2020a. SOLOIST: few-shot task-oriented dialog with A single pre-trained auto-regressive model. CoRR, abs/2005.05298. Baolin Peng, Chenguang Zhu, Chunyuan Li, Xiujun Li, Jinchao Li, Michael Zeng, and Jianfeng Gao. 2020b. Few-shot natural language generation for task-oriented dialog. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 172–182, Online. Association for Computational Linguistics. Baolin Peng, Chenguang Zhu, Michael Zeng, and Jianfeng Gao. 2020c. Data augmentation for spoken language understanding via pretrained models. arXiv preprint arXiv:2004.13952. Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. arXiv preprint arXiv:1802.05365. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. 1978. A simplest systematics for the organization of turn taking for conversation. In Studies in the organization of conversational interaction. Elsevier. Emanuel A Schegloff, Gail Jefferson, and Harvey Sacks. 1977. The preference for self-correction in the organization of repair in conversation. Language. Pararth Shah, Dilek Hakkani-T¨ur, Gokhan T¨ur, Abhinav Rastogi, Ankur Bapna, Neha Nayak, and Larry Heck. 2018. Building a conversational agent overnight with dialogue self-play. arXiv preprint arXiv:1801.04871. Alessandro Sordoni, Michel Galley, Michael Auli, Chris Brockett, Yangfeng Ji, Margaret Mitchell, Jian-Yun Nie, Jianfeng Gao, and Bill Dolan. 2015. A neural network approach to context-sensitive generation of conversational responses. arXiv preprint arXiv:1506.06714. Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461. Tsung-Hsien Wen, David Vandyke, Nikola Mrkˇsi´c, Milica Gasic, Lina M Rojas Barahona, Pei-Hao Su, Stefan Ultes, and Steve Young. 2017. A networkbased end-to-end trainable task-oriented dialogue system. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 438–449. Chien-Sheng Wu, Steven Hoi, Richard Socher, and Caiming Xiong. 2020a. Tod-bert: Pre-trained natural language understanding for task-oriented dialogues. arXiv preprint arXiv:2004.06871. Chien-Sheng Wu, Steven CH Hoi, Richard Socher, and Caiming Xiong. 2020b. Tod-bert: Pre-trained natural language understanding for task-oriented dialogue. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 917–929. Qingyang Wu, Yichi Zhang, Yu Li, and Zhou Yu. 2019. Alternating recurrent dialog model with largescale pre-trained language models. arXiv preprint arXiv:1910.03756. Puyang Xu and Qi Hu. 2018. An end-to-end approach for handling unknown slot values in dialogue state tracking. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1448–1457, Melbourne, Australia. Association for Computational Linguistics. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. NeurIPS. Yichi Zhang, Zhijian Ou, and Zhou Yu. 2020. Taskoriented dialog systems that consider multiple appropriate responses under the same context. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9604–9611. Li Zhou, Jianfeng Gao, Di Li, and Heung-Yeung Shum. 2020. The design and implementation of xiaoice, an empathetic social chatbot. Computational Linguistics, 46(1):53–93. Chenguang Zhu. 2020. Boosting naturalness of language in task-oriented dialogues via adversarial training. arXiv preprint arXiv:2004.14565. Chenguang Zhu, Michael Zeng, and Xuedong Huang. 2019a. Multi-task learning for natural language generation in task-oriented dialogue. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1261–1266. Chenguang Zhu, Michael Zeng, and Xuedong Huang. 2019b. Sim: A slot-independent neural model for dialogue state tracking. arXiv preprint arXiv:1909.11833. Qi Zhu, Zheng Zhang, Yan Fang, Xiang Li, Ryuichi Takanobu, Jinchao Li, Baolin Peng, Jianfeng Gao, Xiaoyan Zhu, and Minlie Huang. 2020. ConvLab2: An open-source toolkit for building, evaluating, and diagnosing dialogue systems. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 142–149, Online. Association for Computational Linguistics. 4429 A Background on SOLOIST We review the SOLOIST (Peng et al., 2020a) for completeness. Each dialog turn is represented as: x = (s, b, c, r), (1) where s is the entire dialog history up to the current dialog turn, b is the dialog belief state acquired from human annotation, c is the DB state automatically retrieved from a database using b, and r is the delexicalized dialog response, from which the system response in natural language can be easily obtained with some automatic post-processing. In sum, each item in x is by itself a sequence of tokens, the entire dialog turn can be viewed as a long sequence. SOLOIST is a neural model parameterized by θ to characterize the sequence generation probability pθ(x). It is pre-trained using publicly available heterogeneous dialog corpora with labels of belief states and DB states. The pre-trained model can be fine-tuned to any new task to generate responses grounded in task-specific user goals and a database. The pre-training and fine-tuning share the same multi-task objective for learning θ: Lθ = LB + LR + LC , (2) where each task is described as follows: Task 1: Belief Prediction For a belief state sequence of length Tb, we define the objective of predicting the belief state as: LB = log p(b|s) = Tb X t=1 log pθ(bt|b<t, s), (3) where b<t indicates all tokens before t. Task 2: Grounded Response Generation A delexicalized response of length Tr, r = [r1, · · · , rTr], is generated by our model token-bytoken from left to right, grounded in dialog history c, belief state b and DB state s. The corresponding training objective is defined as LR = log p(r|c, b, s) (4) = Tr X t=1 log pθ(rt|r<t, c, b, s). Task 3: Contrastive Objective A contrastive objective is employed to promote the matched items (y = 1 for positive samples x) while driving down the mismatched items (y = 0 for negative samples x′). Since the the special token [EOS] attends all tokens in the sequence, the output feature on [EOS] is the fused representation of all items. We apply a binary classifier on top of the feature LC =y log(pθ(x)) + (1−y) log(1 −pθ(x′)). (5) Please refer (Peng et al., 2020a) for more details.
|
2021
|
341
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4430–4445 August 1–6, 2021. ©2021 Association for Computational Linguistics 4430 Semantic Representation for Dialogue Modeling Xuefeng Bai♠♥, Yulong Chen♠♥, Linfeng Song♣, Yue Zhang♥♦ ♠Zhejiang University, China ♥School of Engineering, Westlake University, China ♣Tencent AI Lab, Bellevue, WA, USA ♦Institute of Advanced Technology, Westlake Institute for Advanced Study, China Abstract Although neural models have achieved competitive results in dialogue systems, they have shown limited ability in representing core semantics, such as ignoring important entities. To this end, we exploit Abstract Meaning Representation (AMR) to help dialogue modeling. Compared with the textual input, AMR explicitly provides core semantic knowledge and reduces data sparsity. We develop an algorithm to construct dialogue-level AMR graphs from sentence-level AMRs and explore two ways to incorporate AMRs into dialogue systems. Experimental results on both dialogue understanding and response generation tasks show the superiority of our model. To our knowledge, we are the first to leverage a formal semantic representation into neural dialogue modeling. 1 Introduction Dialogue systems have received increasing research attention (Wen et al., 2015; Serban et al., 2017; Bao et al., 2020), with much recent work focusing on social chats (Ritter et al., 2011; Li et al., 2017) and task-oriented dialogues (Wen et al., 2017; Dinan et al., 2019). There are two salient subtasks in dialogue modeling, namely dialogue understanding (Choi et al., 2018; Reddy et al., 2019; Yu et al., 2020) and response generation (Li et al., 2017; Budzianowski et al., 2018). The former refers to understanding of semantic and discourse details in a dialogue history, and the latter concerns making a fluent, novel and coherent utterance. The current state-of-the-art methods employ neural networks and end-to-end training (Sutskever et al., 2014; Bahdanau et al., 2015) for dialogue modeling. For instance, sequence-to-sequence models have been used to encode a dialogue history, before directly synthesizing the next utterance (Vinyals and Le, 2015; Wen et al., 2017; Bao et al., Dialogue History: … SPEAKER-1 : Recently, I’ve been obsessed with horror films. SPEAKER-2 : Oh, how can you be infatuated with horror films? They’re so scary . SPEAKER-1 : Yeah, you are right I used to not watch horror films, but after seeing Silence of the Lamb with Mike last month, I fell in love with them. SPEAKER-2 : It’s amazing. But if I were you, I wouldn't have the courage to watch the first one. SPEAKER-1 : But it's really exciting . Ground-Truth: Maybe, but I would rather watch romance, science fiction, crime or even disaster movie instead of a horror picture… Transformer: Great. I’m looking forward to it. I just can’t keep away from the food that I saw. Figure 1: A conversation from DailyDialog. Some important contents are marked with squares. 2020). Despite giving strong empirical results, neural models can suffer from spurious feature associations in their neural semantic representation (Poliak et al., 2018; Kaushik et al., 2020), which can lead to weak robustness, inducing irrelevant dialogue states (Xu and Sarikaya, 2014; Sharma et al., 2019; Rastogi et al., 2019) and generating unfaithful or irrelevant text (Maynez et al., 2020; Niu and Bansal, 2020). As shown in Figure 1, the baseline Transformer model pays attention to the word “lamb” but ignores its surrounding context, which has important contents (marked with squares) that indicate its true meaning, thereby giving an irrelevant response that is related to food. Intuitively, such issues can be alleviated by having a structural representation of semantic information, which treats entities as nodes and builds structural relations between nodes, making it easy to find the most salient context. Explicit structures are also more interpretable compared to 4431 neural representation and have been shown useful for information extraction (Strubell et al., 2018; Sun et al., 2019; Li et al., 2020; Bai et al., 2021; Sachan et al., 2021), summarization (Liu et al., 2015; Hardy and Vlachos, 2018; Liao et al., 2018) and machine translation (Marcheggiani et al., 2018; Song et al., 2019a). We explore AMR (Banarescu et al., 2013) as a semantic representation for dialogue histories in order to better represent conversations. As shown in the central block of Figure 2, AMR is one type of sentential semantic representations, which models a sentence using a rooted directed acyclic graph, highlighting its main concepts (e.g. “mistake”) and semantic relations (e.g., “ARG0”1), while abstracting away function words. It can thus potentially offer core concepts and explicit structures needed for aggregating the main content in dialogue. In addition, AMR can also be useful for reducing the negative influence of variances in surface forms with the same meaning, which adds to data sparsity. Existing work on AMR parsing focuses on the sentence level. However, as the left block of Figure 2 shows, the semantic structure of a dialogue history can consist of rich cross-utterance coreference links (marked with squares) and multiple speaker interactions. To this end, we propose an algorithm to automatically derive dialogue-level AMRs from utterance-level AMRs, by adding cross-utterance links that indicate speakers, identical mentions and co-reference links. One example is shown in the right block of Figure 2, where newly added edges are in color. We consider two main approaches of making use of such dialogue-level AMR structures. For the first method, we merge an AMR with tokens in its corresponding sentence via AMR-to-text alignments, before encoding the resulting structure using a graph Transformer (Zhu et al., 2019). For the second method, we separately encode an AMR and its corresponding sentence, before leveraging both representations via feature fusion (Mangai et al., 2010) or dual attention (Calixto et al., 2017). We verify the effectiveness of the proposed framework on a dialogue relation extraction task (Yu et al., 2020) and a response generation task (Li et al., 2017). Experimental results show that the proposed framework outperforms previous 1Please refer to PropBank (Kingsbury and Palmer, 2002; Palmer et al., 2005) for more details. methods (Vaswani et al., 2017; Bao et al., 2020; Yu et al., 2020), achieving the new state-of-the-art results on both benchmarks. Deep analysis and human evaluation suggest that semantic information introduced by AMR can help our model to better understand long dialogues and improve the coherence of dialogue generation. One more advantage is that AMR is helpful to enhance the robustness and has a potential to improve the interpretability of neural models. To our knowledge, this is the first attempt to leverage the AMR semantic representation into neural networks for dialogue understanding and generation. Our code is available at https://github.com/muyeby/AMR-Dialogue. 2 Constructing Dialogue AMRs Figure 2 illustrates our method for constructing a dialogue-level AMR graph from multiple utterancelevel AMRs. Given a dialogue consisting multiple utterances, we adopt a pretrained AMR parser (Cai and Lam, 2020) to obtain an AMR graph for each utterance. For utterances containing multiple sentences, we parse them into multiple AMR graphs, and mark them belonging to the same utterance. We construct each dialogue AMR graph by making connections between utterance AMRs. In particular, we take three strategies according to speaker, identical concept and co-reference information. Speaker We add a dummy node and connect it to all root nodes of utterance AMRs. We add speaker tags (e.g., SPEAKER1 and SPEAKER2) to the edges to distinguish different speakers. The dummy node ensures that all utterance AMRs are connected so that information can be exchanged during graph encoding. Besides, it serves as the global root node to represent the whole dialogue. Identical Concept There can be identical mentions in different utterances (e.g. “possible” in the first and the forth utterances in Figure 2), resulting in repeated concept nodes in utterance AMRs. We connect nodes corresponding to the same nonpronoun concepts by edges labeled with SAME2. This type of connection can further enhance crosssentence information exchange. Inter-sentence Co-reference A major challenge for dialogues understanding is posed by pronouns, 2Compared with co-reference, identical concept relations can connect different words which share the same meaning e.g.⟨could, might⟩, ⟨fear, afraid⟩. 4432 Raw Utterance Texts (b) Utterance AMR Graphs (c) Dialogue AMR Graph AMR Parsing Graph Merge S1 S2 Could I have my bill, please? S2 Certainly, sir. S1 I’m afraid there has been a mistake. What could it be? :ARG0 possible have interrogative :mode I I :poss bill :ARG1 :ARG1 say I certain sir :ARG1 :ARG2 fear I mistake :ARG1 :ARG1 :ARG0 :ARG0 possible :ARG1 it :mode unknown possible possible :ARG0 have interrogative :mode I I :poss bill :ARG1 say I :ARG0 certain sir :ARG1 :ARG2 fear I mistake :ARG0 :ARG1 Dummy :speaker1 :speaker2 :speaker1 :speaker2 :coref :coref :ARG1 :ARG1 it :same unknown :mode :coref Figure 2: Dialogue AMR graph construction process. Step 1: parse raw-text utterance into utterance AMR graphs; Step 2: connect utterance AMR graphs into a dialogue AMR graph. which are frequent in conversations (Grosz et al., 1995; Newman et al., 2008; Quan et al., 2019). We conduct co-reference resolution on dialogue text using an off-to-shelf model3 in order to identify concept nodes in utterance AMRs that refer to the same entity. For example, in Figure 2, “I” in the first utterance, and “sir” in the second utterance refer to the same entity, SPEAKR1. We add edges labeled with COREF between them, starting from later nodes to earlier nodes (later and earlier here refer to the temporal order of ongoing conversation), to indicate their relation4. 3 Baseline System We adopt a standard Transformer (Vaswani et al., 2017) for dialogue history encoding. Typically, a Transformer encoder consists of L layers, taking a sequence of tokens (i.e., dialogue history) S = {w1, w2, ..., wN}, where wi is the i-th token and N is the sequence length, as input and produces vectorized word representations {hl 1, hl 2, ..., hl N} iteratively, l ∈[1, ..., L]. Overall, a Transformer encoder can be written as: H = SeqEncoder(emb(S)), (1) where H = {hL 1 , hL 2 , ..., hL n}, and emb denotes a function that maps a sequence of tokens into the corresponding embeddings. Each Transformer layer consists of two sub-layers: a self-attention sub-layer and a position-wise feed forward network. The former calculates a set of attention scores: αij = Attn(hi, hj). (2) 3https://github.com/huggingface/neuralcoref 4For simplicity, we omit the coreference links between the second and third utterance for display. which are used to update the hidden state of wi: hl i = XN j=1 αij(W V hl−1 j ), (3) where W V is a parameter matrix. The position-wise feed-forward (FFN) layer consists of two linear transformations: FFN(h) = W2ReLU(W1h + b1) + b2, (4) where W1, W2, b1, b2 are model parameters. 3.1 Dialogue Understanding Task We take the dialogue relation extraction task (Yu et al., 2020) as an example. Given a dialogue history S and an argument (or entity) pair (a1, a2), the goal is to predict the corresponding relation type r ∈R between a1 and a2. We follow a previous dialogue relation extraction model (Chen et al., 2020) to feed the hidden states of a1 and a2 (denoted as ha1, ha2) into a classifier to obtain the probability of each relation types: Prel = softmax(W3[ha1; ha2] + b3), (5) where W3 and b3 are model parameters. The k-th value of Prel is the conditional probability of k-th relation in R. Given a training instance ⟨S, a1, a2, r⟩, the local loss is: ℓ= −logP(r|S, a1, a2; θ), (6) where θ denotes the set of model parameters. In practice, we use BERT (Devlin et al., 2019) for calculating ha1 and ha2, which can be regarded as pre-trained initialization of the Transformer encoder. 4433 𝑤𝑤1 𝑤𝑤2 𝑤𝑤3 𝑤𝑤4 𝑤𝑤5 Transformer Graph Transformer 𝑒𝑒1 𝑒𝑒3 𝑒𝑒2 Projected AMR edges Text ℎ1 ̂𝑠𝑠 ℎ2 ̂𝑠𝑠 ℎ3 ̂𝑠𝑠 ℎ4 ̂𝑠𝑠 ℎ5 ̂𝑠𝑠 ℎ1 𝑆𝑆 ℎ2 𝑆𝑆 ℎ3 𝑆𝑆 ℎ4 𝑆𝑆 ℎ5 𝑆𝑆 (a) 𝑛𝑛1 𝑛𝑛2 𝑛𝑛3 𝑛𝑛4 𝑒𝑒3 𝑒𝑒1 𝑒𝑒2 Graph Encoder Sequence Encoder 𝑤𝑤1 𝑤𝑤2 𝑤𝑤3 𝑤𝑤4 𝑤𝑤5 Feature Fusion ℎ1 ℎ2 ℎ3 ℎ4 ℎ5 (b) 𝑛𝑛1 𝑛𝑛2 𝑛𝑛3 𝑛𝑛4 𝑒𝑒3 𝑒𝑒1 𝑒𝑒2 Graph Encoder Sequence Encoder 𝑤𝑤1 𝑤𝑤2 𝑤𝑤3 𝑤𝑤4 𝑤𝑤5 Dual Attention 𝑦𝑦𝑡𝑡 𝑠𝑠𝑡𝑡 𝑐𝑐𝑡𝑡 𝑦𝑦𝑡𝑡+1 𝑠𝑠𝑡𝑡+1 … … (c) Figure 3: AMR for dialogue modeling. (a) Using AMR to enrich text representation. (b,c) Using AMR independently. 3.2 Dialogue Response Generation Task Given a dialogue history S, we use a standard autoregressive Transformer decoder (Vaswani et al., 2017) to generate a response Y = {y1, y2, ..., y|Y|}. At time step t, the previous output word yt−1 is firstly transformed into a hidden state st by a selfattention layer as Equations 2 and 3. Then an encoder-decoder attention mechanism is applied to obtain a context vector from encoder output hidden states{hL 1 , hL 2 , . . . , hL N}: ˆαti = Attn(st, hL i ), ct = XN i=1 ˆαtihL i , (7) The obtained context vector ct is then used to calculate the output probability distribution for the next word yt over the target vocabulary5: Pvoc = softmax(W4ct + b4), (8) where W4, b4 are trainable model parameters. The k-th value of Pvoc is the conditional probability of k-th word in vocabulary given a dialogue. Given a dialogue history-response pair {S, Y}, the model minimizes a cross-entropy loss: ℓ= − |Y | X t=1 logPvoc(yt|yt−1, ..., y1, S; θ), (9) where θ denotes all model parameters. 4 Proposed Model Our model takes a dialogue history S and the corresponding dialogue AMR as input. Formally, 5Similar to the encoder, there is also multi-head attention, a position-wise feed-forward layer and residual connections, which we omit in the equations. an AMR is a directed acyclic graph G = ⟨V, E⟩, where V denotes a set of nodes (i.e. AMR concepts) and E (i.e. AMR relations) denotes a set of labeled edges. An edge can be further represented by a triple ⟨ni, rij, nj⟩, meaning that the edge is from node ni to nj with label rij. We consider two main ways of making use of dialogue-level AMRs. The first method (Figure 3(a)) uses AMR semantic relations to enrich a textual representation of the dialogue history. We project AMR nodes onto the corresponding tokens, extending Transformer by encoding semantic relations between words. For the second approach, we separately encode an AMR and its sentence, and use either feature fusion (Figure 3(b)) or dual attention (Figure 3(c)) to incorporate their embeddings. 4.1 Graph Encoding We adopt a Graph Transformer (Zhu et al., 2019) to encode an AMR graph, which extends the standard Transformer (Vaswani et al., 2017) for modeling structural input. A L-layer graph Transformer takes a set of node embeddings {n1, n2, ..., nM} and a set of edge embeddings {rij|i ∈[1, ..., M], j ∈ [1, ..., M]} as input6 and produces more abstract node features {hl 1, hl 2, ..., hl M} iteratively, where l ∈[1, ..., L]. The key difference between a graph Transformer and a standard Transformer is the graph attention layer. Compared with selfattention layer (Equation 2), the graph attention layer explicitly considers graph edges when updating node hidden states. For example, give an edge ⟨ni, rij, nj⟩, the attention score ˆαij is calculated 6If there is no relation between ni and nj, rij=“None” 4434 as: ˆαij = exp(ˆeij) PM m=1 exp (ˆeim) , ˆeij = (W Qhl−1 i )T (W Khl−1 j + W Rrij) √ d , (10) where W R is a transformation matrix, rij is the embedding of relation rij, d is hidden state size, and {h0 1, h0 2, ..., h0 M} = {n1, n2, ..., nM}. The hidden state of ni is then updated as: hl i = XM j=1 αij(W V hl−1 j + W Rrij), (11) where W V is a parameter matrix. Overall, given an input AMR graph G = ⟨V, E⟩, the graph Transformer encoder can be written as H = GraphEncoder(emb(V), emb(E)), (12) where H = {hL 1 , hL 2 , ..., hL M} denotes top-layer graph encoder hidden states. 4.2 Enriching Text Representation We first use the JAMR aligner (Flanigan et al., 2014) to obtain a node-to-word alignment, then adopt the alignment to project the AMR edges onto text with following rules: ˆrij = ri′j′, if A(ni′) = wi, A(nj′) = wj, Self, if i = j, None, otherwise, (13) where A is a one-to-K alignment (K ∈ [0, . . . , N]). In this way, we obtain a projected graph G′ = ⟨V′, E′⟩, where V′ represents the set of input words {w1, w2, ..., wN} and E′ denotes a set of word-to-word semantic relations. Inspired by previous work on AMR graph modeling (Guo et al., 2019; Song et al., 2019b; Sun et al., 2019), we adopt a hierarchical encoder that stacks a sequence encoder and a graph encoder. A sequence encoder (SeqEncoder) transforms a dialogue history into a set of hidden states: HS = SeqEncoder(emb(S)). (14) A graph encoder incorporates the projected relations features into HS: H ˆS = GraphEncoder(HS, emb(E′)), (15) In addition, we add a residual connection between graph adapter and sequence encoder to fuse word representations before and after refinement (as shown in Figure 3(b)): HF = LayerNorm(HS + H ˆS). (16) where LayerNorm denotes the layer normalization (Ba et al., 2016). We name the hierarchical encoder as Hier, which can be used for both dialogue understanding and dialogue response generation. 4.3 Leveraging both Text and Structure Cues We consider integrating both text cues and AMR structure cues for dialogue understanding and response generation, using a dual-encoder network. First, a sequence encoder is used to transform a dialogue history S into a text memory (denoted as HS = {hS 1 , hS 2 , ..., hS N}) using Equation 1. Second, the AMR graph G is encoded into graph memory (denoted as HG = {hG 1 , hG 2 , ..., hG M}) by a graph Transformer encoder using Equation 12. For dialogue understanding (Figure 3(b)) and dialogue response generation (Figure 3(c)), slightly different methods of feature integration are used due to their different nature of outputs. Dialogue Understanding. Similar to Section 4.2, we first use the JAMR aligner to obtain a node-toword alignment A. Then we fuse the word and AMR node representations as follows: ˆhi = ( f(hS i , hG j ), if ∃j, A(nj) = wi, f(hS i , h∅), otherwise, (17) where h∅is the vector representation of the dummy node (see Figure 2), f is defined as: h = LayerNorm(h1 + h2). (18) The fused word representations are then fed into a classifier for relation prediction (Equation 5). Dialogue Response Generation. We replace the standard encoder-decoder attention (Equation 7) with a dual-attention mechanism (Song et al., 2019a). In particular, given a decoder hidden state st at time step t, the dual-attention mechanism calculates a graph context vector cS t and a text context vector cG t , simultaneously: ˆαti = Attn(st, hS i ), ˆαtj = Attn(st, hG j ), cS t = XN i=1 ˆαtihS i , cG t = XM j=1 ˆαtjhG j , (19) 4435 Model data-v1 data-v2 dev test dev test F1(δ) F1c(δ) F1(δ) F1c(δ) F1(δ) F1c(δ) F1(δ) F1c(δ) AGGCN† 46.6(-) 40.5(-) 46.2(-) 39.5 (-) LSR† 44.5(-) 44.4(-) DHGAT† 57.7(-) 52.7(-) 56.1(-) 50.7(-) BERT 60.6(1.2) 55.4(0.9) 58.5(2.0) 53.2(1.6) 59.4 (0.7) 54.7(0.8) 57.9(1.0) 53.1(0.7) BERTs 63.0(1.5) 57.3(1.2) 61.2(0.9) 55.4(0.9) 62.2(1.3) 57.0(1.0) 59.5(2.1) 54.2(1.4) BERTc 66.8(0.9) 60.9(1.0) 66.1(1.1) 60.2(0.8) 66.2(0.9) 60.5(1.1) 65.1(0.8) 59.8(1.2) Hier 68.2(0.8) 62.2(0.7) 67.0(0.9) 61.3(0.6) 68.0(0.6) 62.2(0.4) 66.7(0.3) 61.0(0.4) Dual 68.3(0.6) 62.2(0.2) 67.3(0.4) 61.4(0.2) 68.2(0.5) 62.3(0.4) 67.1(0.4) 61.1(0.5) Table 1: Performance on DialogRE, where δ denotes the standard deviation computed from 5 runs, and † indicates results reported by Chen et al. (2020). and the final context vector ˆct is calculated as: ct = W c[cS t ; cG t ] + bc, (20) where W c and bc are model parameters. We name the dual-encoder model as Dual. 5 Dialogue Understanding Experiments We evaluate our model on DialogRE (Yu et al., 2020), which contains totally 1,788 dialogues, 10,168 relational triples and 36 relation types in total. On average, a dialogue in DialogRE contains 4.5 relational triples and 12.9 turns. We report experimental results on both original (v1) and updated (v2) English version.7 5.1 Settings We adopt the same input format and hyperparameter settings as Yu et al. (2020) for the proposed model and baselines. In particular, the input sequence is constructed as [CLS]d[SEP]a1[SEP]a2[SEP], where d denotes the dialogue, and a1 and a2 are the two associated arguments. In the BERT model of Yu et al. (2020), only the hidden state of the [CLS] token is fed into a classifier for prediction, while our baseline (BERTc) additionally takes the hidden states of a1 and a2. All hyperparameters are selected by prediction accuracy on validation dataset (See Table 6 for detailed hyperparameters). Metrics Following previous work on DialogRE, we report macro F1 score on relations in both the standard (F1) and conversational settings (F1c; Yu et al., 2020). F1c is computed over the first few turns of a dialogue where two arguments are first mentioned. 7https://dataset.org/dialogre/ 5.2 Main Results Table 1 shows the results of different systems on DialogRE. We compare the proposed model with two BERT-based approches, BERT and BERTs. Based on BERT, BERTs (Yu et al., 2020) highlights speaker information by replacing speaker arguments with special tokens. For completeness, we also include recent methods, such as AGGCN (Guo et al., 2019), LSR (Nan et al., 2020) and DHGAT (Chen et al., 2020). BERTc and Hier, Dual represent our baseline and the proposed models, respectively. By incorporating speaker information, BERTs gives the best performance among the previous system. Our BERTc baseline outperforms BERTs by a large margin, as BERTc additionally considers argument representations for classification. Hier significantly (p < 0.01)8 outperforms BERTc in all settings, with 1.4 points of improvement in terms of F1 score on average. A similar trend is observed under F1c. This shows that semantic information in AMR is beneficial to dialogue relation extraction, since AMR highlights core entities and semantic relations between them. Dual obtains slightly better results than Hier, which shows effect of separately encoding a semantic structure. Finally, the standard deviation values of both Dual and Hier are lower than the baselines. This indicates that our approaches are more robust regarding model initialization. 5.3 Impact of Argument Distance We split the dialogues of the DialogRE (v2) devset into five groups by the utterance-based distance between two arguments. As shown in Figure 4, Dual gives better results than BERTc except when 8We use pair-wised t-test. 4436 <5 [5,10) [10, 15) [15, 20] >20 Argument Distance (# Utterances) 64 66 68 70 72 F1 Baseline Ours Figure 4: The performance of BERTc (Baseline) and Dual (Ours) regarding argument distances. the argument distance is less than 5. In particular, Dual surpasses BERTc by a large margin when the arguments distance is greater than 20. The comparison indicates that AMR can help a model to better handle long-term dependencies by improving the entity recall. In addition to utterance distance, we also consider word distance and observe a similar trend (as shown in Appendix 7). 5.4 Case Study Figure 5 shows a conversation between a manager and an employee who might have taken a leave. The baseline model incorrectly predicts that the relation between two interlocutors is parent and child. It might be influenced by the last sentence in the conversation, assuming that it is a dialogue between family members. However, the proposed model successful predicts the interlocutors’ relation, suggesting it can extract global semantic information in the dialogue from a comprehensive perspective. 6 Response Generation Experiments We conduct experiments on the DailyDialog benchmark (Li et al., 2017), which contains 13,119 daily multi-turn conversations. On average, the number of turns for each dialogue is 7.9, and each utterance has 14.6 tokens. 6.1 Settings We take Transformer as a baseline. Our hyperparameters are selected by word prediction accuracy on validation dataset. The detailed hyperparameters are given in Appendix (See Table 6). Metric We set the decoding beam size as 5 and adopt BLEU-1/2/3/4 (Papineni et al., 2002) and Distinct-1/2 (Li et al., 2016) as automatic evaluation metrics. The former measures the ngram overlap between generated response and Dialogue : SPEAKER-1: A new place for a new Ross. I'm gonna have you and all the guys from work over once it's y'know, furnished. SPEAKER-2: I must say it's nice to see you back on your feet. SPEAKER-1: Well I am that. And that whole rage thing is definitely behind me. SPEAKER-2: I wonder if its time for you to rejoin our team at the museum? SPEAKER-1: Oh Donald that-that would be great. I am totally ready to come back to work. I…What? No! Wh-What are you doing?!! GET OFF MY SISTER!!!!!!!!!!!!! Ground-Truth: per:boss(S1, S2) Baseline: per:parent(S1, S2) Ours: per:boss(S1, S2) Figure 5: Case study for dialogue relation extraction. Model BLEU-1/2/3/4 Distinct-1/2 Seq2Seq† 33.6/26.8/-/3.0/12.8 iVAEMI 30.9/24.9/-/2.9/25.0 PLATO w/o L†♭ 40.5/32.2/-/4.6/24.6 PLATO†♭ 39.7/31.1/-/5.3/29.1 Transformer 38.3/31.7/29.1/27.8 5.8/30.5 Hier 41.3/35.4/33.2/32.1 6.5/32.3 Dual 40.8/35.0/32.7/31.5 6.6/33.0 Table 2: Performance on DailyDialog. Results marked with † are from Bao et al. (2020). Models marked with ♭requires external corpus for pretraining. the target response while the latter assesses the generation diversity, which is defined as the number of distinct uni- or bi-grams divided by the total amount of generated words. In addition, we also conduct human evaluation. Following Bao et al. (2020), we ask annotators who study linguistics to evaluate model outputs from four aspects, which are fluency, coherence, informativeness and overall performance. The scores are in a scale of {0, 1, 2}. The higher, the better. 6.2 Automatic Evaluation Results Table 2 reports the performances of the previous state-of-the-art methods and proposed models on the DailyDialog testset. For the previous methods, PLATO and PLATO w/o L are both Transformer models pre-trained on large-scale conversational data (8.3 million samples) and finetuned on DailyDialog. For completeness, we also report other systems including Seq2Seq (Vinyals and Le, 2015) and iVAEMI (Fang et al., 2019). 4437 Model Fluency Coherence Inf. Overall Transformer 1.76 0.86 1.40 0.66 Hier 1.86 1.04 1.48 0.82 Dual 1.88 1.04 1.52 0.84 Table 3: Human evaluation results on DailyDialog. Inf. stands for Informativeness. Among the previous systems, PLATO and PLATO w/o L report the best performances. Our Transformer baseline is highly competitive in terms of BLEU and Distinct scores. Compared with the Transformer baseline, both Dual and Hier show better numbers regarding BLEU and Distinct, and the gains of both models are significant (p < 0.01). This indicates that semantic information in AMR graphs is useful for dialogue response generation. In particular, the gains come from better recall of the important entities and their relations in a dialogue history, which can leads to generating a more detailed response. 6.3 Human Evaluation Results We conduct human evaluation on randomly selected 50 dialogues and corresponding generated responses of the baseline and our models. As shown in Table 3, the Transformer baseline gives the lowest scores, while Dual sees the highest scores from all aspects. Our main advantage is on the Coherence, meaning that AMRs are effective on recalling important concepts and relations. As the result, it makes it easier for our models to generate coherent replies. Examples are shown in Figure 8 in Appendix. Comparatively, all systems achieve high scores regarding Fluency, suggesting that this aspect is not the current bottleneck for response generation. 7 Analysis This section contains analysis concerning the effects of graph features, dialogue length and model robustness. We use Dual model for experiments since it gives slightly better results than Hier. 7.1 Ablation on AMR graph Table 4 shows the results of our best performing models on the two datasets regarding different configurations on the dialogue AMR graphs. We report the average F1 score for DialogRE and the BLEU1/Distinct-1 score for DailyDialog. First, using utterance-level AMR improves the text baseline by 1.2 points and 1.5 points with regard to F1 and Setting DialogRE (v2) DailyDialog Dialog-AMR(Dual) 68.2 38.2/5.9 -Speaker 67.5 37.7/5.7 -Ident. concept 68.0 37.9/5.8 -Coref 67.8 37.4/5.6 Utter-AMR 67.4 36.9/5.6 Text 66.2 35.4/5.5 Table 4: Ablation study on the development sets of both DialogRE (v2) and DailyDialog. <4 [4,8) [8,12) [12, 16] >16 Dialogue Length (# Utterances) 30 40 50 60 70 Baseline(DU) Ours(DU) Baseline(RG) Ours(RG) Figure 6: Devset performance against dialogue lengths. BLEU-1 scores, respectively. This indicates that the semantic knowledge in formal AMR is helpful for dialogue modeling. Second, our manually added relations (in Section 2) also leads to improvements, ranging from 0.5 to 1.0 in BLEU-1 score. The speaker relation is the most important for dialogue relation extraction, a possible reason is that DialogRE dataset mainly focus on person entities. Also, co-reference relations help the most in dialogue response generation. The identical concept relations give least improvements among three relations. Finally, combining all relations to build a Dialog-AMR graph achieves best performance on both datasets. 7.2 Impact of Dialogue Length We group the devset of DialogRE (v2) and DailyDialog into five groups according to the number of utterances in a dialogue. Figure 6 summarizes the performance of the baseline and the proposed model on dialogue understanding (DU) and response generation (RG) tasks. In dialogue understanding, our model gives slightly better F1 scores than the baseline when a dialogue has smaller than 12 utterance. The performance improvement is more significant when modeling a long dialogue. This confirms our motivation that AMR can help to understand long dialogues. In dialogue response generation, our model consistently outperforms the Transformer baseline by a large margin on 4438 Model Original Paraphrased Baseline 100 94.50 Ours 100 98.50 Table 5: F1 on original and paraphrased testsets. dialogues of different lengths, still with more improvements on larger dialogues. Overall, these results are consistent with Table 1 and 2, showing that AMR can provide useful semantic information and alleviate the issue of long-range dependency. 7.3 Robustness Against Input Recent studies show that neural network-based dialog models lack robustness (Shalyminov and Lee, 2018; Einolghozati et al., 2019). We select 100 instances from the testset of DialogRE (v2) where both baseline and our model gives true prediction, before paraphrasing the source dialogues manually (see appendix B.3 for paraphrasing guidelines.). Results on the paraphrased dataset are given in Table 5. The performance of baseline model drop from 100 to 94.5 on paraphrased dataset. By contrast, the result of our model reaches 98.5, 4 points higher than baseline. This confirms our assumption that AMR can reduce data sparsity, thus improve the robustness of neural models. 8 Related Work Semantic Parsing for Dialogue Some previous work builds domain-specified semantic schema for task-oriented dialogues. For example, in the PEGASUS (Zue et al., 1994) system, a sentence is first transformed into a semantic frame and then used for travel planing. Wirsching et al. (2012) use semantic features to help a dialogue system perform certain database operations. Gupta et al. (2018) represent task-oriented conversations as semantic trees where intents and slots are tree nodes. They solve intent classification and slot-filling task via semantic parsing. Cheng et al. (2020) design a rooted semantic graph that integrates domains, verbs, operators and slots in order to perform dialogue state tracking. All these structures are designed for specified task only. In contrast, we investigate a general semantic representation for the modeling of everyday conversations. Constructing AMRs beyond Sentence Level There are a few attempts to construct AMRs beyond the sentence level. Liu et al. (2015) construct document-level AMRs by merging identical concepts of sentence-level AMRs for abstractive summerization, and Liao et al. (2018) further extend this approach to multi-document summerization. O’Gorman et al. (2018) manually annotate co-reference information across sentence AMRs. We focus on creating conversation-level AMRs to facilitate information exchange more effectively for dialogue modeling. Bonial et al. (2020) adapt AMRs on dialogues by enriching the standard AMR schema with dialogue acts, tense and aspect, and they construct a dataset consisting of 340 dialogue AMRs. However, they propose theoretical changes in the schema for annotating AMRs, while we explore empirical solutions that leverage existing AMRs of the standard schema on dialogues. AMR Parsing and Encoding Our work is also related to AMR parsing (Flanigan et al., 2014; Konstas et al., 2017a; Lyu and Titov, 2018; Guo and Lu, 2018; Zhang et al., 2019; Cai and Lam, 2020) and AMR encoding (Konstas et al., 2017b; Song et al., 2018; Zhu et al., 2019; Song et al., 2020; Zhao et al., 2020; Bai et al., 2020). The former task makes it possible to use automatically-generated AMRs for downstream applications, while the latter helps to effectively exploit structural information in AMRs. In this work, we investigate AMRs for dialogue representation and combine AMRs with text for dialogue modeling. 9 Conclusion We investigated the feasibility of using AMRs for dialogue modeling, describing an algorithm to construct dialogue-level AMRs automatically and exploiting two ways to incorporate AMRs into neural dialogue systems. Experiments on two benchmarks show advantages of using AMR semantic representations model on both dialogue understanding and dialogue response generation. Acknowledgments Yue Zhang is the corresponding author. We would like to thank the anonymous reviewers for their insightful comments and Jinhao Jiang for his help for data preparation. This work has been supported by Tencent AI Lab Rhino-Bird Focused Research Program. It also receives support from the Westlake University and Bright Dream Joint Institute for Intelligent Robotics, and a research grant from Rxhui Inc. 4439 References Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. 2016. Layer normalization. CoRR, abs/1607.06450. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. Xuefeng Bai, Pengbo Liu, and Yue Zhang. 2021. Investigating typed syntactic dependencies for targeted sentiment classification using graph attention neural network. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:503–514. Xuefeng Bai, Linfeng Song, and Yue Zhang. 2020. Online back-parsing for AMR-to-text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1206–1219, Online. Association for Computational Linguistics. Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob, Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. 2013. Abstract meaning representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse. Siqi Bao, Huang He, Fan Wang, Hua Wu, and Haifeng Wang. 2020. PLATO: Pre-trained dialogue generation model with discrete latent variable. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 85–96, Online. Association for Computational Linguistics. Claire Bonial, Lucia Donatelli, Mitchell Abrams, Stephanie M. Lukin, Stephen Tratz, Matthew Marge, Ron Artstein, David Traum, and Clare Voss. 2020. Dialogue-AMR: Abstract Meaning Representation for dialogue. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 684– 695, Marseille, France. European Language Resources Association. Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, I˜nigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gaˇsi´c. 2018. MultiWOZ - a large-scale multi-domain Wizard-of-Oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016–5026, Brussels, Belgium. Association for Computational Linguistics. Deng Cai and Wai Lam. 2020. AMR parsing via graph-sequence iterative inference. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1290–1301, Online. Association for Computational Linguistics. Iacer Calixto, Qun Liu, and Nick Campbell. 2017. Doubly-attentive decoder for multi-modal neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1913– 1924, Vancouver, Canada. Association for Computational Linguistics. Hui Chen, Pengfei Hong, Wei Han, Navonil Majumder, and Soujanya Poria. 2020. Dialogue relation extraction with document-level heterogeneous graph attention networks. CoRR, abs/2009.05092. Jianpeng Cheng, Devang Agrawal, H´ector Mart´ınez Alonso, Shruti Bhargava, Joris Driesen, Federico Flego, Dain Kaplan, Dimitri Kartsaklis, Lin Li, Dhivya Piraviperumal, Jason D. Williams, Hong Yu, Diarmuid ´O S´eaghdha, and Anders Johannsen. 2020. Conversational semantic parsing for dialog state tracking. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8107–8117, Online. Association for Computational Linguistics. Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. 2018. QuAC: Question answering in context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2174–2184, Brussels, Belgium. Association for Computational Linguistics. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net. Arash Einolghozati, Sonal Gupta, Mrinal Mohit, and Rushin Shah. 2019. Improving robustness of task oriented dialog systems. CoRR, abs/1911.05153. Le Fang, Chunyuan Li, Jianfeng Gao, Wen Dong, and Changyou Chen. 2019. Implicit deep latent variable models for text generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3946–3956, Hong Kong, China. Association for Computational Linguistics. Jeffrey Flanigan, Sam Thomson, Jaime Carbonell, Chris Dyer, and Noah A. Smith. 2014. A discriminative graph-based parser for the Abstract Meaning 4440 Representation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1426– 1436, Baltimore, Maryland. Association for Computational Linguistics. Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. 1995. Centering: A framework for modeling the local coherence of discourse. Computational Linguistics, 21(2):203–225. Zhijiang Guo and Wei Lu. 2018. Better transitionbased AMR parsing with a refined search space. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1712–1722, Brussels, Belgium. Association for Computational Linguistics. Zhijiang Guo, Yan Zhang, and Wei Lu. 2019. Attention guided graph convolutional networks for relation extraction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 241–251, Florence, Italy. Association for Computational Linguistics. Sonal Gupta, Rushin Shah, Mrinal Mohit, Anuj Kumar, and Mike Lewis. 2018. Semantic parsing for task oriented dialog using hierarchical representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2787–2792, Brussels, Belgium. Association for Computational Linguistics. Hardy Hardy and Andreas Vlachos. 2018. Guided neural language generation for abstractive summarization using Abstract Meaning Representation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 768–773, Brussels, Belgium. Association for Computational Linguistics. Divyansh Kaushik, Eduard H. Hovy, and Zachary Chase Lipton. 2020. Learning the difference that makes A difference with counterfactually-augmented data. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. Paul R. Kingsbury and Martha Palmer. 2002. From treebank to propbank. In Proceedings of the Third International Conference on Language Resources and Evaluation, LREC 2002, May 29-31, 2002, Las Palmas, Canary Islands, Spain. European Language Resources Association. Ioannis Konstas, Srinivasan Iyer, Mark Yatskar, Yejin Choi, and Luke Zettlemoyer. 2017a. Neural AMR: sequence-to-sequence models for parsing and generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, ACL 2017, Vancouver, Canada, July 30 - August 4, Volume 1: Long Papers, pages 146–157. Association for Computational Linguistics. Ioannis Konstas, Srinivasan Iyer, Mark Yatskar, Yejin Choi, and Luke Zettlemoyer. 2017b. Neural AMR: Sequence-to-sequence models for parsing and generation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 146–157, Vancouver, Canada. Association for Computational Linguistics. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California. Association for Computational Linguistics. Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986–995, Taipei, Taiwan. Asian Federation of Natural Language Processing. Zhongli Li, Qingyu Zhou, Chao Li, Ke Xu, and Yunbo Cao. 2020. Improving BERT with syntax-aware local attention. CoRR, abs/2012.15150. Kexin Liao, Logan Lebanoff, and Fei Liu. 2018. Abstract Meaning Representation for multi-document summarization. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1178–1190, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Fei Liu, Jeffrey Flanigan, Sam Thomson, Norman Sadeh, and Noah A. Smith. 2015. Toward abstractive summarization using semantic representations. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1077–1086, Denver, Colorado. Association for Computational Linguistics. Chunchuan Lyu and Ivan Titov. 2018. AMR parsing as graph prediction with latent alignment. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 397–407. Association for Computational Linguistics. Utthara Gosa Mangai, Suranjana Samanta, Sukhendu Das, and Pinaki Roy Chowdhury. 2010. A survey of decision fusion and feature fusion strategies for pattern classification. IETE Technical Review, 27(4):293–307. Diego Marcheggiani, Jasmijn Bastings, and Ivan Titov. 2018. Exploiting semantics in neural machine translation with graph convolutional networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 486–492, New 4441 Orleans, Louisiana. Association for Computational Linguistics. Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906–1919, Online. Association for Computational Linguistics. Guoshun Nan, Zhijiang Guo, Ivan Sekulic, and Wei Lu. 2020. Reasoning with latent structure refinement for document-level relation extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1546–1557, Online. Association for Computational Linguistics. Matthew L Newman, Carla J Groom, Lori D Handelman, and James W Pennebaker. 2008. Gender differences in language use: An analysis of 14,000 text samples. Discourse Processes: A Multidisciplinary Journal, 45(3):211–236. Tong Niu and Mohit Bansal. 2020. Avgout: A simple output-probability measure to eliminate dull responses. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8560–8567. Tim O’Gorman, Michael Regan, Kira Griffitt, Ulf Hermjakob, Kevin Knight, and Martha Palmer. 2018. AMR beyond the sentence: the multi-sentence AMR corpus. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3693–3702, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Martha Palmer, Paul R. Kingsbury, and Daniel Gildea. 2005. The proposition bank: An annotated corpus of semantic roles. Comput. Linguistics, 31(1):71–106. Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. Adam Poliak, Jason Naradowsky, Aparajita Haldar, Rachel Rudinger, and Benjamin Van Durme. 2018. Hypothesis only baselines in natural language inference. In Proceedings of the Seventh Joint Conference on Lexical and Computational Semantics, pages 180–191, New Orleans, Louisiana. Association for Computational Linguistics. Jun Quan, Deyi Xiong, Bonnie Webber, and Changjian Hu. 2019. GECOR: An end-to-end generative ellipsis and co-reference resolution model for taskoriented dialogue. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLPIJCNLP), pages 4547–4557, Hong Kong, China. Association for Computational Linguistics. Pushpendre Rastogi, Arpit Gupta, Tongfei Chen, and Mathias Lambert. 2019. Scaling multi-domain dialogue state tracking via query reformulation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Industry Papers), pages 97–105, Minneapolis, Minnesota. Association for Computational Linguistics. Siva Reddy, Danqi Chen, and Christopher D. Manning. 2019. CoQA: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249–266. Alan Ritter, Colin Cherry, and William B. Dolan. 2011. Data-driven response generation in social media. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 583–593, Edinburgh, Scotland, UK. Association for Computational Linguistics. Devendra Singh Sachan, Yuhao Zhang, Peng Qi, and William L. Hamilton. 2021. Do syntax trees help pre-trained transformers extract information? In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, EACL 2021, Online, April 19 - 23, 2021, pages 2647–2661. Association for Computational Linguistics. Iulian Vlad Serban, Alessandro Sordoni, Ryan Lowe, Laurent Charlin, Joelle Pineau, Aaron C. Courville, and Yoshua Bengio. 2017. A hierarchical latent variable encoder-decoder model for generating dialogues. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 3295– 3301. AAAI Press. Igor Shalyminov and Sungjin Lee. 2018. Improving robustness of neural dialog systems in a data-efficient way with turn dropout. In The Thirty-second Annual Conference on Neural Information Processing Systems (NIPS) 2018, workshop on Conversational AI: “Today’s Practice and Tomorrow’s Potential. Sanuj Sharma, Prafulla Kumar Choubey, and Ruihong Huang. 2019. Improving dialogue state tracking by discerning the relevant context. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 576–581, Minneapolis, Minnesota. Association for Computational Linguistics. Linfeng Song, Daniel Gildea, Yue Zhang, Zhiguo Wang, and Jinsong Su. 2019a. Semantic neural machine translation using AMR. Transactions of the Association for Computational Linguistics, 7:19–31. Linfeng Song, Ante Wang, Jinsong Su, Yue Zhang, Kun Xu, Yubin Ge, and Dong Yu. 2020. Structural information preserving for graph-to-text generation. 4442 In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7987–7998, Online. Association for Computational Linguistics. Linfeng Song, Yue Zhang, Daniel Gildea, Mo Yu, Zhiguo Wang, and Jinsong Su. 2019b. Leveraging dependency forest for neural medical relation extraction. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 208–218, Hong Kong, China. Association for Computational Linguistics. Linfeng Song, Yue Zhang, Zhiguo Wang, and Daniel Gildea. 2018. A graph-to-sequence model for AMRto-text generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1616– 1626, Melbourne, Australia. Association for Computational Linguistics. Emma Strubell, Patrick Verga, Daniel Andor, David Weiss, and Andrew McCallum. 2018. Linguistically-informed self-attention for semantic role labeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, October 31 November 4, 2018, pages 5027–5038. Association for Computational Linguistics. Kai Sun, Richong Zhang, Samuel Mensah, Yongyi Mao, and Xudong Liu. 2019. Aspect-level sentiment analysis via convolution over dependency tree. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5679– 5688, Hong Kong, China. Association for Computational Linguistics. Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pages 3104–3112. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 49, 2017, Long Beach, CA, USA, pages 5998–6008. Oriol Vinyals and Quoc V. Le. 2015. A neural conversational model. CoRR, abs/1506.05869. Tsung-Hsien Wen, Milica Gaˇsi´c, Nikola Mrkˇsi´c, PeiHao Su, David Vandyke, and Steve Young. 2015. Semantically conditioned LSTM-based natural language generation for spoken dialogue systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1711–1721, Lisbon, Portugal. Association for Computational Linguistics. Tsung-Hsien Wen, David Vandyke, Nikola Mrkˇsi´c, Milica Gaˇsi´c, Lina M. Rojas-Barahona, Pei-Hao Su, Stefan Ultes, and Steve Young. 2017. A networkbased end-to-end trainable task-oriented dialogue system. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 438–449, Valencia, Spain. Association for Computational Linguistics. G¨unther Wirsching, Markus Huber, Christian K¨olbl, Robert Lorenz, and Ronald R¨omer. 2012. Semantic dialogue modeling. In Anna Esposito, Antonietta M. Esposito, Alessandro Vinciarelli, R¨udiger Hoffmann, and Vincent C. M¨uller, editors, Cognitive behavioural systems: COST 2102 International Training School, Dresden, Germany, February 2126, 2011, volume 7403. P. Xu and R. Sarikaya. 2014. Contextual domain classification in spoken language understanding systems using recurrent neural network. In 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 136–140. Dian Yu, Kai Sun, Claire Cardie, and Dong Yu. 2020. Dialogue-based relation extraction. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4927–4940, Online. Association for Computational Linguistics. Sheng Zhang, Xutai Ma, Kevin Duh, and Benjamin Van Durme. 2019. AMR parsing as sequence-tograph transduction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 80–94, Florence, Italy. Association for Computational Linguistics. Yanbin Zhao, Lu Chen, Zhi Chen, Ruisheng Cao, Su Zhu, and Kai Yu. 2020. Line graph enhanced AMR-to-text generation with mix-order graph attention networks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 732–741, Online. Association for Computational Linguistics. Jie Zhu, Junhui Li, Muhua Zhu, Longhua Qian, Min Zhang, and Guodong Zhou. 2019. Modeling graph structure in transformer for better AMR-to-text generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5459–5468, Hong Kong, China. Association for Computational Linguistics. Victor Zue, Stephanie Seneff, Joseph Polifroni, Michael Phillips, Christine Pao, David Goddeau, James Glass, and Eric Brill. 1994. PEGASUS: A spoken language interface for on-line air travel 4443 planning. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994. 4444 <30 [30,60) [60,90) [90,120] >120 Argument Distance (# tokens) 62 64 66 68 70 F1 Baseline Ours Figure 7: Performance against argument word distance. A Model parameters Table 6 lists all model hyperparameters used for experiments. In particular, we share the word vocabulary of encoder and decoder for response generation. We implement our baselines and proposed model based on Pytorch. The preprocessed data and source code will be released at https: //github.com/muyeby/AMR-Dialogue. B More Experimental Results B.1 Impact of Argument Distance In addition to utterance distance used in Figure 4, we also consider word-based distance as a metric to measure argument distance. Figure 7 shows F1 scores of baseline and our model on 5 groups of test instances. It can be seen that our model gives better results than baseline system among all distances longer than 30. In particular, our model surpass baseline by 8 points when argument distance is longer than 120. Dialogue History: … SPEAKER-1 : We have new room rates, sir. Will that be acceptable to you? SPEAKER-2 : Well , it depends on the price, of course. What is it? SPEAKER-2 : It's $ 308 a night. SPEAKER-1 : I have no problem with that. SPEAKER-2 : Great! Would you prefer smoking or nonsmoking? SPEAKER-1 : Definitely nonsmoking. I can't handle that smell. Ground-Truth: Now, is a queen-size bed okay? Transformer: I’m sorry, sir. I’ll be fine. Ours: That’ll be nonsmoking. Now, do you prefer a single queen-size bed? Figure 8: Case study for dialogue response generation. B.2 Case Study for Dialogue Response Generation Figure 8 represents a conversation between a hotel service and a guest who wants to book a room, along with its ground-truth response and model-generated responses. We can observe that Transformer’s output is general and not consistent with dialogue history. While proposed models’ outputs can capture the core information “room” from the history, and are more relevant to the topic. Besides, the output given by proposed model is semantically similar to the ground-truth output, but using novel words to response, indicating that the model not only captures the simple dependency between input and output sentences, but also learns deep semantic information of the dialogue history. B.3 Paraphrasing Guidelines We ask annotators to paraphrase the dialogues following 3 guidelines: • do not change the original meaning. • paraphrase the sentence by using different lexicon and syntax structures. • paraphrase the dialogue as much as they can. We also ask a judge to evaluate whether the paraphrased dialogue (sentences) convey the same meaning of the original ones. 4445 Setting DialogRE DailyDialog Sequence Encoder Dropout 0.1 0.1 Encoder Layers 12 4 Attention Heads 12 8 Embedding Size 768 512 Hidden Layer size 768 512 Word Vocabulary size 31k 16k Feed-Forward Layer size 3072 1024 Number of parameters 110M 38M Graph Encoder (Hier) Dropout 0.1 0.1 Encoder Layers 2 2 Attention Heads 8 8 Hidden Layer size 512 512 Relation Embedding size 64 64 Feed-Forward Layer size 1024 1024 Number of parameters 4M 4M Graph Encoder (Dual) Dropout 0.1 0.1 Encoder Layers 3 4 Attention Heads 8 8 Hidden Layer Size 512 512 Relation Embedding Size 64 64 Concept Vocabulary Size 5.2k 10k Feed-Forward Layer Size 1024 1024 Number of parameters 11M 20M Others Optimizer Adam Adam Batch Size 48 20 Learning Rate 3e-5 1e-4 Training Epoch 30 200 Decoder Layers 4 Training Device Tesla V100 Tesla V100 Training Time 120min 48h Table 6: Hyperparameters of our models on DialogRE and DailyDialog.
|
2021
|
342
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4446–4457 August 1–6, 2021. ©2021 Association for Computational Linguistics 4446 A Pre-training Strategy for Zero-Resource Response Selection in Knowledge-Grounded Conversations Chongyang Tao1∗, Changyu Chen2∗, Jiazhan Feng1, Jirong Wen2,3 and Rui Yan2,3† 1Peking University, Beijing, China 2Gaoling School of Artificial Intelligence, Renmin University of China 3Beijing Academy of Artificial Intelligence 1{chongyangtao,fengjiazhan}@pku.edu.cn 2{chen.changyu,jrwen,ruiyan}@ruc.edu.cn Abstract Recently, many studies are emerging towards building a retrieval-based dialogue system that is able to effectively leverage background knowledge (e.g., documents) when conversing with humans. However, it is non-trivial to collect large-scale dialogues that are naturally grounded on the background documents, which hinders the effective and adequate training of knowledge selection and response matching. To overcome the challenge, we consider decomposing the training of the knowledge-grounded response selection into three tasks including: 1) query-passage matching task; 2) query-dialogue history matching task; 3) multi-turn response matching task, and joint learning all these tasks in a unified pre-trained language model. The former two tasks could help the model in knowledge selection and comprehension, while the last task is designed for matching the proper response with the given query and background knowledge (dialogue history). By this means, the model can be learned to select relevant knowledge and distinguish proper response, with the help of ad-hoc retrieval corpora and a large number of ungrounded multi-turn dialogues. Experimental results on two benchmarks of knowledge-grounded response selection indicate that our model can achieve comparable performance with several existing methods that rely on crowd-sourced data for training. 1 Introduction Along with the very recent prosperity of artificial intelligence empowered conversation systems in the spotlight, many studies have been focused on building human-computer dialogue systems (Wen et al., 2017; Zhang et al., 2020) with either retrievalbased methods (Wang et al., 2013; Wu et al., 2017; ∗Equal Contribution. †Corresponding author: Rui Yan ([email protected]). Whang et al., 2020) or generation-based methods (Li et al., 2016; Serban et al., 2016; Zhang et al., 2020), which both predict the response with only the given context. In fact, unlike a person who may associate the conversation with the background knowledge in his or her mind, the machine can only capture limited information from the query message itself. As a result, it is difficult for a machine to properly comprehend the query, and to predict a proper response to make it more engaging. To bridge the gap of the knowledge between the human and the machine, researchers have begun to simulating this motivation by grounding dialogue agents with background knowledge (Zhang et al., 2018; Dinan et al., 2019; Li et al., 2020), and lots of impressive results have been obtained. In this paper, we consider the response selection problem in knowledge-grounded conversion and specify the background knowledge as unstructured documents that are common sources in practice. The task is that given a conversation context and a set of knowledge entries, one is required 1): to select proper knowledge and grasp a good comprehension of the selected document materials (knowledge selection); 2): to distinguish the true response from a candidate pool that is relevant and consistent with both the conversation context and the background documents (knowledge matching). While there exists a number of knowledge documents on the Web, it is non-trivial to collect large-scale dialogues that are naturally grounded on the documents for training a neural response selection model, which hinders the effective and adequate training of knowledge selection and response matching. Although some benchmarks built upon crowd-sourcing have been released by recent works (Zhang et al., 2018; Dinan et al., 2019), the relatively small training size makes it hard for the dialogue models to generalize on other domains or topics (Zhao et al., 2020). Thus, in this work, we 4447 focus on a more challenging and practical scenario, learning a knowledge-grounded conversation agent without any knowledge-grounded dialogue data, which is known as zero-resource settings. Since knowledge-grounded dialogues are unavailable in training, it raises greater challenges for learning the grounded response selection model. Fortunately, there exists a large number of unstructured knowledge (e.g., web pages or wiki articles), passage search datasets (e.g., query-passage pairs coming from ad-hoc retrieval tasks) (Khattab and Zaharia, 2020) and multi-turn dialogues (e.g., context-response pairs collected from Reddit) (Henderson et al., 2019), which might be beneficial to the learning of knowledge comprehension, knowledge selection and response prediction respectively. Besides, in multi-turn dialogues, the background knowledge and conversation history (excluding the latest query) are symmetric in terms of the information they convey, and we assume that the dialogue history can be regarded as another format of background knowledge for response prediction. Based on the above intuition, in this paper, we consider decomposing the training of the grounded response selection task into several sub-tasks, and joint learning all those tasks in a unified model. To take advantage of the recent breakthrough on pretraining for natural language tasks, we build the grounded response matching models on the basis of a pre-trained language model (PLMs) (Devlin et al., 2019; Yang et al., 2019), which are trained with large-scale unstructured documents from the web. On this basis, we further train the PLMs with query-passage matching task, query-dialogue history matching task, and multi-turn response matching task jointly. The former two tasks could help the model not only in knowledge selection but also in knowledge (and dialogue history) comprehension, while the last task is designed for matching the proper response with the given query and background knowledge (dialogue history). By this means, the model can be learned to select relevant knowledge and distinguish proper responses, with the help of a large number of ungrounded dialogues and ad-hoc retrieval corpora. During the testing stage, we first utilize the trained model to select proper knowledge, and then feed the query, dialogue history, selected knowledge, and the response candidate into our model to calculate the final matching degree. Particularly, we design two strategies to compute the final matching score. In the first strategy, we directly concatenate the selected knowledge and dialogue history as a long sequence of background knowledge and feed into the model. In the second strategy, we first compute the matching degree between each queryknowledge and the response candidates, and then integrate all matching scores. We conduct experiments with benchmarks of knowledge-grounded dialogue that are constructed by crowd-sourcing, such as the Wizard-ofWikipedia Corpus (Dinan et al., 2019) and the CMU DoG Corpus (Zhou et al., 2018a). Evaluation results indicate that our model achieves comparable performance on knowledge selection and response selection with several existing models trained on crowd-sourced benchmarks. Our contributions are summarized as follows: • To the best of our knowledge, this is the first exploration of knowledge-grounded response selection under the zero-resource setting. • We propose decomposing the training of the grounded response selection models into several sub-tasks, so as to empower the model through these tasks in knowledge selection and response matching. • We achieve a comparable performance of response selection with several existing models learned from crowd-sourced training sets. 2 Related Work Early studies of retrieval-based dialogue focus on single-turn response selection where the input of a matching model is a message-response pair (Wang et al., 2013; Ji et al., 2014; Wang et al., 2015). Recently, researchers pay more attention to multiturn context-response matching and usually adopt the representation-matching-aggregation paradigm to build the model. Representative methods include the dual-LSTM model (Lowe et al., 2015), the sequential matching network (SMN) (Wu et al., 2017), the deep attention matching network (DAM) (Zhou et al., 2018b), interaction-overinteraction network (IoI) (Tao et al., 2019) and multi-hop selector network (MSN) (Yuan et al., 2019). More recently, pre-trained language models (Devlin et al., 2019; Yang et al., 2019) have shown significant benefits for various NLP tasks, and some researchers have tried to apply them on multi-turn response selection. Vig and Ramea (2019) exploit BERT to represent each utteranceresponse pair and fuse these representations to 4448 calculate the matching score; Whang et al. (2020) and Xu et al. (2020) treat the context as a long sequence and conduct context-response matching with BERT. Besides, Gu et al. (2020a) integrate speaker embeddings into BERT to improve the utterance representation in multi-turn dialogue. To bridge the gap of the knowledge between the human and the machine, researchers have investigated into grounding dialogue agents with unstructured background knowledge (Ghazvininejad et al., 2018; Zhang et al., 2018; Dinan et al., 2019). For example, Zhang et al. (2018) build a persona-based conversation data set that employs the interlocutor’s profile as the background knowledge; Zhou et al. (2018a) publish a data where conversations are grounded in articles about popular movies; Dinan et al. (2019) release another documentgrounded data with Wiki articles covering a wide range of topics. Meanwhile, several retrievalbased knowledge-grounded dialogue models are proposed, such as document-grounded matching network (DGMN) (Zhao et al., 2019) and dually interactive matching network (DIM) (Gu et al., 2019) which let the dialogue context and all knowledge entries interact with the response candidate respectively via the cross-attention mechanism. Gu et al. (2020b) further propose to pre-filter the context and the knowledge and then use the filtered context and knowledge to perform the matching with the response. Besides, with the help of gold knowledge index annotated by human wizards, Dinan et al. (2019) consider joint learning the knowledge selection and response matching in a multi-task manner or training a two-stage model. 3 Model In this section, we first formalize the knowledgegrounded response matching problem and then introduce our method from preliminary to response matching with PLMs to details of three pre-training tasks. 3.1 Problem Formalization We first describe a standard knowledge-grounded response selection task such as Wizard-ofWikipedia. Suppose that we have a knowledgegrounded dialogue data set D = {ki, ci, ri, yi}N i=1 where ki = {p1, p2, . . . , plk} represents a collection of knowledge with pj the j-th knowledge entry (a.k.a., passage) and lk is the number of entries; ci = {u1, u2, . . . , ulc} denotes multi-turn dialogue context with uj the j-th turn and lc is the number of dialogue turns. It should be noted that in this paper we denote the latest turn ulc as dialogue query qi, and dialogue context except for query is denoted as hi = ci/{qi}. ri stands for a candidate response. yi = 1 indicates that ri is a proper response for ci and ki, otherwise yi = 0. N is the number of samples in data set. The goal knowledge-grounded dialogue is to learn a matching model g(k, c, r) from D, and thus for any new (k, c, r), g(k, c, r) returns the matching degree between r and (k, c). Finally, one can collect the matching scores of a series of candidate responses and conduct response ranking. Zero-resource grounded response selection then is formally defined as follows. There is a standard multi-turn dialogue dataset Dc = {qi, hi, ri}N i=1 and an ad-hoc retrieval dataset Dp = {qi, pi, zi}M i=1 where qi is a query and pi stands a candidate passage, zi = 1 indicates that pi is a relevant passage for qi, otherwise zi = 0. Our goal is to learn a model g(k, h, q, r) from Dc and Dp, and thus for any new input (k, h, q, r), our model can select proper knowledge ˆk from k and calculate the matching degree between r and (ˆk, q, h). 3.2 Preliminary: Response Matching with PLMs Pre-trained language models have been widely used in many NLP tasks due to the strong ability of language representation and understanding. In this work, we consider building a knowledge-grounded response matching model with BERT. Specifically, given a query q, a dialogue history h = {u1, u2, ..., unh} where ui is the i-th turn in the history, a response candidate r = {r1, r2, ..., rlr} with lr words, we concatenate all sequences as a single consecutive tokens sequence with special tokens, which can be represented as x = {[CLS], u1, [SEP], . . . , [SEP], ulh, [SEP], q, [SEP], r, [SEP]}. [CLS] and [SEP] are classification symbol and segment separation symbol respectively. For each token in x, BERT uses a summation of three kinds of embeddings, including WordPiece embedding (Wu et al., 2016), segment embedding, and position embedding. Then, the embedding sequence of x is fed into BERT, giving us the contextualized embedding sequence {E[CLS], E2, . . . , Elx}. E[CLS] is an aggregated representation vector that contains the 4449 Input ··· Dialogue History or Knowledge Response Pre-trained Language Model (BERT) 𝑔𝑞, 𝑘, 𝑟 Output Layer MLP Token Embeddings Position Embeddings Segment Embeddings ··· ··· ··· ··· Response Matching Task Query Query-Dialogue History Matching Task Query-Passage Matching Task ··· [Background Knowledge] [Response] [Query] ··· ··· ··· ··· ··· ··· 𝐸!"# 𝑬$! 𝐸#%& 𝑬$" 𝐸#%& 𝑬$# 𝐸#%& 𝑬' 𝐸#%& 𝐸(! 𝐸($% 𝐸#%& ··· ··· ··· [CLS] 𝑢! [SEP] 𝑢" [SEP] 𝑢# [SEP] 𝑞 [SEP] 𝑟! 𝑟$! [SEP] 𝑤",! 𝑤",$" 𝑤&,! 𝑤&,$# Figure 1: The overall architecture of our model. semantic interaction information between the query, history, and response candidate. Finaly, E[CLS] is fed into a non-linear layer to calculate the final matching score, which is formulated as: g(h, q, r) = σ(W2 · tanh(W1E[CLS] + b1) + b2) (1) where W{1,2} and b{1,2} is training parameters for response selection task, σ is a sigmoid function. In knowledge-grounded dialogue, each dialogue is associated with a large collection of knowledge entries k = {p1, p2, . . . , plk}1. The model is required to select m(m ≥1) knowledge entries based on semantic relevance between the query and each knowledge, and then performs the response matching with the query, dialogue history and the highly-relevant knowledge. Specifically, we denote ˆk = (ˆp1, . . . , ˆpm) as the selected knowledge entries, and feed the input sequence x = {[CLS], ˆp1, [SEP], . . . , [SEP], ˆpm, [SEP], u1, [SEP], . . . , [SEP], ulh, [SEP], q, [SEP], r, [SEP]} to BERT. The final matching score g(ˆk, h, q, r) can be computed based on [CLS] representation. 3.3 Pre-training Strategies On the basis of BERT, we further jointly train it with three tasks including 1) query-passage matching task; 2) query-dialogue history matching task; 3) multi-turn response matching task. The former two tasks could help the model in knowledge selection and knowledge (and dialogue history) comprehension, while the last task is designed for matching the proper response with the given query and background knowledge (dialogue 1The scale of the knowledge referenced by each dialogue usually exceeds the limitation of input length in PLMs. history). By this means, the model can be learned to select relevant knowledge and distinguish the proper response, with the help of a large number of ungrounded dialogues and ad-hoc retrieval corpora. 3.3.1 Query-Passage Matching Although there exist a huge amount of conversation data on social media, it is hard to collect sufficient dialogues that are naturally grounded on knowledge documents. Existing studies (Dinan et al., 2019) usually extract the relevant knowledge before the response matching or jointly train the knowledge retrieval and response selection in a multi-task manner. However, both methods need in-domain knowledge-grounded dialogue data (with gold knowledge label) to train, making the model hard to generalize to a new domain. Fortunately, the ad-hoc retrieval task (Harman, 2005; Khattab and Zaharia, 2020) in the information retrieval area provides a potential solution to simulate the process of knowledge seeking. To take advantage of the parallel data in the ad-hoc retrieval task, we consider incorporating the query-passage matching task, so as to help the knowledge selection and knowledge comprehension for our task. Given a query-passage pair (q, p), we first concatenate the query q and the passage p as a single consecutive token sequence with special tokens separating them, which is formulated as: Sqp = {[CLS], wp 1, . . . , wp np, [SEP], wq 1, . . . , wq nq} (2) where wp i , wq j denotes the i-th and j-th token of knowledge entry p and query q respectively. For each token in Sqp i , token, segment and position 4450 embeddings are summated and fed into BERT. It is worth noting that here we set the segment embedding of the knowledge to be the same as the dialogue history. Finally, we feed the output representation of [CLS] Eqp [CLS] into a MLP to obtain the final query-passage matching score g(q, p). The loss function of each training sample for query-passage matching task is defined by Lp(q, p+, p− 1 , . . . , p− np) = −log( eg(q,p+) eg(q,p+) + Pδp j=1 eg(q,p− j ) ) (3) where p+ stands for the positive passage for q, p− j is the j-th negative passage and δp is the number of negative passage. 3.3.2 Query-Dialogue History Matching In multi-turn dialogues, the conversation history (excluding the latest query) is a piece of supplementary information for the current query and can be regarded as another format of background knowledge during the response matching. Besides, due to the natural sequential relationship between dialogue turns, the dialogue query usually shows a strong semantic relevance with the previous turns in the dialogue history. Inspired by such characteristics, we design a query-dialogue history matching task with the multi-turn dialogue context, so as to enhance the capability of the model to comprehend the dialogue history with the given dialogue query and to rank relevant passages with these pseudo query-passage pairs. Specifically, we first concatenate the dialogue history into a long sequence. The task requires the model to predict whether a query q = {wq 1, . . . , wq nq} and a dialogue history sequence h = {wh 1, . . . , wh nh} are consecutive and relevant. We concatenate two sequences into a single consecutive sequence with [SEP] tokens, Sqh = {[CLS], wh 1 , . . . , wh nh, [SEP], wq 1, . . . , wq nq} (4) For each word in Sqh, token, segment and position embeddings are summated and fed into BERT. Finally, we feed Eqh [CLS] into a MLP to obtain the final query-history matching score g(q, h). The loss function of each training sample for queryhistory matching task is defined by Lh(q, h+, h− 1 , . . . , h− nh) = −log( eg(q,h+) eg(q,h+) + Pδh j=1 eg(q,h− j ) ) (5) where h+ stands for the true dialogue history for q, h− j is the j-th negative dialogue history randomly sampled from the training set and δh is the number of sampled dialogue history. 3.3.3 Multi-turn Response Matching The above two tasks are designed for empowering the model to knowledge or history comprehension and knowledge selection. In this task, we aim at training the model to match reasonable responses based on dialogue history and query. Since we treat the dialogue history as a special form of background knowledge and they share the same segment embeddings in the PLMs, our model can acquire the ability to identify the proper response with either dialogue history or the background knowledge through the multi-turn response matching task. Specifically, we format the multi-turn dialogues as query-history-response triples and requires the model to predict whether a response candidate r = {wr 1, . . . , wr nr} is appropriate for a given query q = {wq 1, . . . , wq nq} and a concatenated dialogue history sequence h = {wh 1, . . . , wh nh}. Concretely, we concatenate three input sequences into a single consecutive tokens sequence with [SEP] tokens, Shqr = {[CLS], wh 1 , . . . , wh nh, [SEP], wq 1, . . . , wq nq, [SEP], wr 1, . . . , wr nr} (6) Similarly, we feed an embedding sequence of which each entry is a summation of token, segment and position embeddings into BERT. Finally, we feed Ehqr [CLS] into a MLP to obtain the final response matching score g(h, q, r). The loss function of each training sample for multi-turn response matching task is defined by Lr(h, q, r+, r− 1 , . . . , r− δr) = −log( eg(h,q,r+) eg(h,q,r+) + Pnr i=j eg(h,q,r− j ) ) (7) where r+ is the true response for a given q and h, r− j is the j-th negative response candidate randomly sampled from the training set and δr is the number of negative response candidate. 3.3.4 Joint Learning We adopt a multi-task learning manner and define the final objective function as: Lfinal = Lp + Lh + Lr (8) In this way, all tasks are jointly learned so that the model can effectively leverage two training 4451 corpus and learn to select relevant knowledge and distinguish the proper response. 3.4 Calculating Matching Score After learning model from Dc and Dp, we first rank {pi}nk i=1 according to g(q, ki) and then select top m knowledge entries {p1, . . . , pm} for the subsequent response matching process. Here we design two strategies to compute the final matching score g(k, h, q, r). In the first strategy, we directly concatenate the selected knowledge and dialogue history as a long sequence of background knowledge and feed into the model to obtain the final matching score, which is formulated as, g(k, h, q, r) = g(p1 ⊕. . . ⊕pm ⊕c, q, r) (9) where ⊕denotes the concatenation operation. In the second strategy, we treat each selected knowledge entry and the dialogue history equally as the background knowledge, and compute the matching degree between each query, background knowledge, and the response candidates with the trained model. Consequently, the matching score is defined as an integration of a set of knowledgegrounded response matching scores, formulated as, g(k, h, q, r) = g(h, q, r)+ max i∈(0,m) g(pi, q, r) (10) where m is the number of selected knowledge entries. We name our model with the two strategies as PTKGCcat and PTKGCsep respectively. We compare the two learning strategies through empirical studies, as will be reported in the next section. 4 Experiments 4.1 Datasets and Evaluation Metrics Training Set. We adopt MS MARCO passage ranking dataset (Nguyen et al., 2016) built on Bing’s search for query-passage matching task. The dataset contains 8.8M passages from Web pages gathered from Bing’s results to real-world queries and each passage contains an average of 55 words. Each query is associated with sparse relevance judgments of one (or very few) passage marked as relevant. The training set contains about 500k pairs of query and relevant passage, and another 400M pairs of query and passages that have not been marked as relevant, from which the negatives are sampled in our task. For the query-dialogue history matching task and multi-turn response matching task, we use the multi-turn dialogue corpus constructed from the Reddit (Dziri et al., 2018). The dataset contains more than 15 million dialogues and each dialogue has at least 3 utterances. After the pre-processing, we randomly sample 2.28M/20K dialogues as the training/validation set. For each dialogue session, we regard the last turn as the response, the last but one as the query, and the rest as the positive dialogue history. The negative dialogue histories are randomly sampled from the whole dialogue set. On average, each dialogue contains 4.3 utterances, and the average length of the utterances is 42.5. Test Set. We tested our proposed method on the Wizard-of-Wikipedia (WoW) (Dinan et al., 2019) and CMU DoG (Zhou et al., 2018a). Both datasets contain multi-turn dialogues grounded on a set of background knowledge and are built with crowd-sourcing on Amazon Mechanical Turk. In WoW, the given knowledge collection is obtained from Wikipedia and covers a wide range of topics or domains, while in CMU DoG, the underlying knowledge focuses on the movie domain. Unlike CMU DoG where the golden knowledge index for each turn is unknown, the golden knowledge index for each turn is provided in WoW. Two configurations (e.g., test-seen and test-unseen) are provided in WoW. Following existing works (Dinan et al., 2019; Zhao et al., 2019), positive responses are true responses from humans and negative ones are randomly sampled. The ratio between positive and negative responses is 1 : 99 for WoW and 1 : 19 for CMU DoG. More details of the two benchmarks are shown in Appendix A.1. Evaluation Metrics. Following previous works on knowledge-grounded response selection (Gu et al., 2020b; Zhao et al., 2019), we also employ recall n at k Rn@k (where n = 100 for WoW and n = 20 for CMU DoG and k = {1, 2, 5}) as the evaluation metrics. 4.2 Implementation Details Our model is implemented by PyTorch (Paszke et al., 2019). Without loss of generality, we select English uncased BERTbase (110M) as the matching model. During the training, the maximum lengths of the knowledge (a.k.a., passage), the dialogue history, the query, and the response candidate were set to 128, 120 60, and 40. Intuitively, the last tokens in the dialogue history and the previous 4452 Models Test Seen Test Unseen R@1 R@2 R@5 R@1 R@2 R@5 IR Baseline 17.8 14.2 BoW MemNet 71.3 33.1 Two-stage Transformer 84.2 63.1 Transformer MemNet 87.4 69.8 DIM (Gu et al., 2019) 83.1 91.1 95.7 60.3 77.8 92.3 FIRE (Gu et al., 2020b) 88.3 95.3 97.7 68.3 84.5 95.1 PTKGCcat 85.7 94.6 98.2 65.5 82.0 94.7 PTKGCsep 89.5 96.7 98.9 69.6 85.8 96.3 Table 1: Evaluation results on the test set of WoW. tokens in the query and response candidate are more important, so we cut off the previous tokens for the context but do the cut-off in the reverse direction for the query and response candidate if the sequences are longer than the maximum length. We set a batch size of 32 for multi-turn response matching and query-dialogue history matching, and 8 for query-document matching in order to train these tasks jointly under the circumstance of training examples inequality. We set δp = 6, δh = 1 and δr = 12 for the query-passage matching, the query-dialogue history matching and the multiturn response matching respectively. Particularly, the negative dialogue histories are sampled from other training instances in a batch. The model is optimized using Adam optimizer with a learning rate set as 5e −6. The learning rate is scheduled by warmup and linear decay. A dropout rate of 0.1 is applied for all linear transformation layers. The gradient clipping threshold is set as 10.0. Early stopping on the corresponding validation data is adopted as a regularization strategy. During the testing, we vary the number of selected knowledgeentries m ∈{1, . . . , 15} and set m = 2 for PTKGCcat and set m = 14 for PTKGCsep because they achieve the best performance. 4.3 Baselines Since the characteristics of the two data sets are different (only WoW provides the golden knowledge label), we compare the proposed model with the baselines on both data sets individually. Baselines on WoW. 1) IR Baseline (Dinan et al., 2019) uses simple word overlap for response selection; 2) BoW MemNet (Dinan et al., 2019) is a memory network where knowledge entries are embedded via bag-of-words representation, and the model learns the knowledge selection and response matching jointly; 3) Transformer MemNet (Dinan et al., 2019) is an extension of BoW MemNet, Models R@1 R@2 R@5 Starspace (Wu et al., 2018) 50.7 64.5 80.3 BoW MemNet (Zhang et al., 2018) 51.6 65.8 81.4 KV Profile Memory (Zhang et al., 2018) 56.1 69.9 82.4 Transformer MemNet (Mazar´e et al., 2018) 60.3 74.4 87.4 DGMN (Zhao et al., 2019) 65.6 78.3 91.2 DIM (Gu et al., 2019) 78.7 89.0 97.1 FIRE (Gu et al., 2020b) 81.8 90.8 97.4 PTKGCcat 61.6 73.5 86.1 PTKGCsep 66.1 77.8 88.7 Table 2: Evaluation results on the test set of CMU DoG. and the dialogue history, response candidate and knowledge entries are encoded with Transformer encoder (Vaswani et al., 2017) pre-trained on a large data set. 4) Two-stage Transformer (Dinan et al., 2019) trains two separately models for knowledge selection and response retrieval respectively. A best-performing model on the knowledge selection task is used for the dialogue retrieval task. Baselines on CMU DoG 1) Starspace (Wu et al., 2018) selects the response by the cosine similarity between a concatenated sequence of dialogue context, knowledge, and the response candidate represented by StarSpace (Wu et al., 2018); 2) BoW MemNet (Zhang et al., 2018) is a memory network with the bag-of-words representation of knowledge entries as the memory items; 3) KV Profile Memory (Zhang et al., 2018) is a key-value memory network grounded on knowledge profiles; 4) Transformer MemNet (Mazar´e et al., 2018) is similar to BoW MemNet and all utterances are encoded with a pre-trained Transformer; 5) DGMN (Zhao et al., 2019) lets the dialogue context and all knowledge entries interact with the response candidate respectively via the cross-attention; 6) DIM (Gu et al., 2019) is similar to DGMN and all utterance are encoded with BiLSTMs; 7) FIRE (Gu et al., 2020b) first filters the context and knowledge and then use the filtered context and knowledge to perform the iterative response matching process. 4.4 Evaluation Results Performance of Response Selection. Table 1 and Table 2 report the evaluation results of response selection on WoW and CMU DoG where PTKGCcat and PTKGCsep represent the final matching score computed with the first strategy (Equation 9) and the second strategy (Equation 10) respectively. We can see that PTKGCsep is 4453 Models Wizard of Wikipedia CMU DoG Test Seen Test Unseen R@1 R@2 R@5 R@1 R@2 R@5 R@1 R@2 R@5 PTKGCsep 89.5 96.7 98.9 69.6 85.8 96.3 66.1 77.8 88.7 PTKGCsep (q) 70.6 79.7 86.8 55.9 70.8 83.4 47.3 58.8 75.0 PTKGCsep (q+h) 84.9 93.9 97.8 64.9 81.7 94.3 59.5 72.3 86.1 PTKGCsep (q+k) 89.5 96.4 98.6 67.0 84.0 96.0 62.7 73.8 84.8 PTKGCsep,m=1 85.6 94.4 97.9 66.7 82.8 94.3 60.4 72.5 86.0 PTKGCsep,m=1 - Lp 84.7 93.5 97.5 63.4 80.5 94.0 58.7 70.8 85.6 PTKGCsep,m=1 - Lh 84.9 93.7 97.6 65.5 81.7 94.1 59.4 71.4 85.3 Table 3: Ablation study. Models Wizard Seen Wizard Unseen R@1 R@2 R@5 R@1 R@2 R@5 Random 2.7 2.3 IR Baseline 5.8 7.6 BoW MemNet 23.0 8.9 Transformer 22.5 12.2 Transformer (w/ pretrain) 25.5 22.9 Our Model 22.0 31.2 48.8 23.1 32.1 50.7 Our Model - Lp 12.8 22.6 45.2 13.3 23.3 45.5 Our Model - Lh 21.2 29.9 47.6 22.7 31.2 49.2 Table 4: The performance of knowledge selection on the test sets of WoW data. All baselines come from Dinan et al. (2019). The details for all baselines are shown in Appendix A.2. consistently better than PTKGCcat over all metrics on two data sets, demonstrating that individually representing each knowledge-query-response triple with BERT can lead to a more optimal matching signal than representing a single long sequence. Our explanation to the phenomenon is that there is information loss when a long sequence composed of the knowledge and dialogue history passes through the deep architecture of BERT. Thus, the earlier different knowledge entries and dialogue history are fused together, the more information of dialogue history or background knowledge will be lost in matching. Particularly, on the WoW, in terms of R@1, our PTKGCsep achieves a comparable performance with the existing stateof-the-art models that are learned from the crowdsourced training set, indicating that the model can effectively learn how to leverage external knowledge feed for response selection through the proposed pre-training approach. Notably, we can observe that our PTKGCsep performs worse than DIM and FIRE on the CMU DoG. Our explanation to the phenomenon is that the dialogue and knowledge in CMU DoG focus on the movie domain while our train data including ad-hoc retrieval corpora and multi-turn dialogues come from the open domain. Thus, our model may not select proper knowledge entries and can not well recognize the semantics clues for response matching due to the domain shift. Despite this, PTKGCsep can still show better performance than several existing models, such as Transformer MemNet and DGMN, though PTKGCsep does not access any training examples in the benchmarks. Performance of Knowledge Selection. We also assess the ability of models to predict the knowledge selected by human wizards in WoW data. The results are shown in Table 4. We can find that the performance of our method is comparable with various supervised methods trained on the gold knowledge index. In particular, on the testseen, our model is slightly worse than Transformer (w/ pretrain), while on the test-unseen, our model achieves slightly better results. The results demonstrate the advantages of our pretraining tasks and the good generalization ability of our model. 4.5 Discussions Ablation Study. We conduct a comprehensive ablation study to investigate the impact of different inputs and different tasks. First, we remove the dialogue history, knowledge, and both of them from the model, which is denoted as PTKGCsep(q+k), PTKGCsep(q+h) and PTKGCsep(q) respectively. According to the results of the first four rows in Table 3, we can find that both the dialogue history and knowledge are crucial for response selection as removing anyone will generally cause a performance drop on the two data. Besides, the background knowledge is more critical for response selection as removing the background knowledge causes more significant performance degradation than removing the dialogue history. Then, we remove each training task individually from PTKGCsep, and denote the models 4454 Models Wizard Seen Wizard Unseen R@1 R@2 R@5 R@1 R@2 R@5 PTKGCsep (q+h) 84.9 93.9 97.8 64.9 81.7 94.3 PTKGCsep (q+h) -Lh 84.1 93.7 97.7 64.3 81.9 93.8 PTKGCsep (q+h) -Lp 83.4 93.5 97.9 60.9 80.2 93.5 PTKGCsep (q+h) -Lh-Lp 83.2 93.8 97.6 60.9 80.1 93.8 Table 5: Ablation study of our model without considering the grounded knowledge. as PTKGCsep-X, where X ∈{Lp, Lh} meaning query-passage matching task and query-dialogue history matching task respectively. Table 4 shows the ablation results of knowledge selection. We can find that both tasks are useful in the learning of knowledge selection, and query-passage matching plays a dominant role since the performance of knowledge selection drops dramatically when the task is removed from the pre-training process. The last two rows in Table 3 show the ablation results of response selection. We report the ablation results when only 1 knowledge is provided since the knowledge recalls for different ablated models and the full model are very close when m is large (m = 14). We can see that both tasks are helpful and the performance of response selection drops more when removing the query-passage matching task. Particularly, Lp plays a more important role and the performance on test-unseen of WoW drops more obvious when removing each training task. To further investigate the impact of our pretraining tasks on the performance of the multiturn response selection (without considering the grounded knowledge), we conduct an ablation study and the results are shown in Table 5. We can observe that the performance of the response matching model (no grounded knowledge) drops obviously when removing one of the pretraining tasks or both tasks. Particularly, the query-passage matching task contributes more to the response selection. The impact of the number of selected knowledge. We further study how the number of selected knowledge (m) influences the performance of PTKGCsep. Figure 2 shows how the performance of our model changes with respect to different numbers of selected knowledge. We observe that the performance increases monotonically until the knowledge number reaches a certain value, and then stable when the number keeps increasing. The results are rational because more knowledge entries can provide more useful 0.85 0.86 0.87 0.88 0.89 0.90 0.856 0.864 0.869 0.875 0.877 0.882 0.885 0.887 0.889 0.891 0.892 0.893 0.894 0.895 0.895 Seen Unseen 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 The number of selected knowledge (m) 0.65 0.66 0.67 0.68 0.69 0.70 0.667 0.672 0.675 0.682 0.682 0.681 0.682 0.682 0.685 0.687 0.688 0.690 0.692 0.696 0.696 R100@1 Figure 2: The performance of response selection across different number of selected knowledge. information for response matching, but when the knowledge becomes enough, the noise will be brought to matching. 5 Conclusion In this paper, we study response matching in knowledge-grounded conversations under a zeroresource setting. In particular, we propose decomposing the training of the knowledge-grounded response selection into three tasks and joint train all tasks in a unified pre-trained language model. Our model can be learned to select relevant knowledge and distinguish proper response, with the help of ad-hoc retrieval corpora and amount of multiturn dialogues. Experimental results on two benchmarks indicate that our model achieves a comparable performance with several existing methods trained on crowd-sourced data. In the future, we would like to explore the ability of our proposed method in retrieval-augmented dialogues. Acknowledgement We would like to thank the anonymous reviewers for their constructive comments. This work was supported by the National Key Research and Development Program of China (No. 2020YFB1406702), the National Science Foundation of China (NSFC No. 61876196) and Beijing Outstanding Young Scientist Program (No. BJJWZYJH012019100020098). Rui Yan is the corresponding author, and is supported as a young fellow at Beijing Academy of Artificial Intelligence (BAAI). 4455 References Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4171–4186. Association for Computational Linguistics. Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2019. Wizard of wikipedia: Knowledge-powered conversational agents. In International Conference on Learning Representations. Nouha Dziri, Ehsan Kamalloo, Kory W Mathewson, and Osmar R Zaiane. 2018. Augmenting neural response generation with context-aware topical attention. arXiv preprint arXiv:1811.01063. Marjan Ghazvininejad, Chris Brockett, Ming-Wei Chang, Bill Dolan, Jianfeng Gao, Wen-tau Yih, and Michel Galley. 2018. A knowledge-grounded neural conversation model. In The Thirty-Second AAAI Conference on Artificial Intelligence, pages 5110– 5117. Jia-Chen Gu, Tianda Li, Quan Liu, Zhen-Hua Ling, Zhiming Su, Si Wei, and Xiaodan Zhu. 2020a. Speaker-aware bert for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 29th ACM International Conference on Information and Knowledge Management, CIKM ’20, pages 2041–2044. ACM. Jia-Chen Gu, Zhen-Hua Ling, Xiaodan Zhu, and Quan Liu. 2019. Dually interactive matching network for personalized response selection in retrieval-based chatbots. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1845–1854, Hong Kong, China. Jia-Chen Gu, Zhenhua Ling, Quan Liu, Zhigang Chen, and Xiaodan Zhu. 2020b. Filtering before iteratively referring for knowledge-grounded response selection in retrieval-based chatbots. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1412–1422, Online. Association for Computational Linguistics. Donna K Harman. 2005. The trec ad hoc experiments. Matthew Henderson, Paweł Budzianowski, I˜nigo Casanueva, Sam Coope, Daniela Gerz, Girish Kumar, Nikola Mrkˇsi´c, Georgios Spithourakis, Pei-Hao Su, Ivan Vuli´c, and Tsung-Hsien Wen. 2019. A repository of conversational datasets. In Proceedings of the First Workshop on NLP for Conversational AI, pages 1–10, Florence, Italy. Zongcheng Ji, Zhengdong Lu, and Hang Li. 2014. An information retrieval approach to short text conversation. arXiv preprint arXiv:1408.6988. Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 39–48. Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. A diversity-promoting objective function for neural conversation models. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 110–119, San Diego, California. Association for Computational Linguistics. Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2020. Zero-resource knowledge-grounded dialogue generation. In Proceedings of the 34th Conference on Neural Information Processing Systems. Ryan Lowe, Nissan Pow, Iulian Serban, and Joelle Pineau. 2015. The Ubuntu dialogue corpus: A large dataset for research in unstructured multiturn dialogue systems. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 285–294, Prague, Czech Republic. Association for Computational Linguistics. Pierre-Emmanuel Mazar´e, Samuel Humeau, Martin Raison, and Antoine Bordes. 2018. Training millions of personalized dialogue agents. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2775–2779, Brussels, Belgium. Association for Computational Linguistics. Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. Ms marco: A human generated machine reading comprehension dataset. In CoCo@ NIPS. Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. Iulian Vlad Serban, Alessandro Sordoni, Yoshua Bengio, Aaron C Courville, and Joelle Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, volume 16, pages 3776–3784. Chongyang Tao, Wei Wu, Can Xu, Wenpeng Hu, Dongyan Zhao, and Rui Yan. 2019. One time of interaction may not be enough: Go deep with an interaction-over-interaction network for response selection in dialogues. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 1–11. 4456 Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc. Jesse Vig and Kalai Ramea. 2019. Comparison of transfer-learning approaches for response selection in multi-turn conversations. In Workshop on DSTC7. Hao Wang, Zhengdong Lu, Hang Li, and Enhong Chen. 2013. A dataset for research on shorttext conversations. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 935–945. Association for Computational Linguistics. Mingxuan Wang, Zhengdong Lu, Hang Li, and Qun Liu. 2015. Syntax-based deep matching of short texts. In IJCAI, pages 1354–1361. Tsung-Hsien Wen, David Vandyke, Nikola Mrkˇsi´c, Milica Gaˇsi´c, Lina M. Rojas-Barahona, Pei-Hao Su, Stefan Ultes, and Steve Young. 2017. A networkbased end-to-end trainable task-oriented dialogue system. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 438–449. Association for Computational Linguistics. Taesun Whang, Dongyub Lee, Chanhee Lee, Kisu Yang, Dongsuk Oh, and HeuiSeok Lim. 2020. An effective domain adaptive post-training method for bert in response selection. In Proceedings of INTERSPEECH 2020, pages 1585–1589. Ledell Yu Wu, Adam Fisch, Sumit Chopra, Keith Adams, Antoine Bordes, and Jason Weston. 2018. Starspace: Embed all the things! In Thirty-Second AAAI Conference on Artificial Intelligence, pages 5569–5577. Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner, et al. 2016. Google’s neural machine translation system: Bridging the gap between human and machine translation. CoRR, abs/1609.08144. Yu Wu, Wei Wu, Chen Xing, Ming Zhou, and Zhoujun Li. 2017. Sequential matching network: A new architecture for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 496–505. Association for Computational Linguistics. Ruijian Xu, Chongyang Tao, Daxin Jiang, Xueliang Zhao, Dongyan Zhao, and Rui Yan. 2020. Learning an effective context-response matching model with self-supervised tasks for retrieval-based dialogues. In Proceedings of the Thirty-Fifth AAAI Conference on Artificial Intelligence. Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc. Chunyuan Yuan, Wei Zhou, Mingming Li, Shangwen Lv, Fuqing Zhu, Jizhong Han, and Songlin Hu. 2019. Multi-hop selector network for multi-turn response selection in retrieval-based chatbots. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 111–120. Association for Computational Linguistics. Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2204–2213. Association for Computational Linguistics. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT : Largescale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270–278, Online. Association for Computational Linguistics. Xueliang Zhao, Chongyang Tao, Wei Wu, Can Xu, Dongyan Zhao, and Rui Yan. 2019. A documentgrounded matching network for response selection in retrieval-based chatbots. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, pages 5443–5449. Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledgegrounded dialogue generation with pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3377–3390, Online. Association for Computational Linguistics. Kangyan Zhou, Shrimai Prabhumoye, and Alan W Black. 2018a. A dataset for document grounded conversations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 708–713, Brussels, Belgium. Association for Computational Linguistics. Xiangyang Zhou, Lu Li, Daxiang Dong, Yi Liu, Ying Chen, Wayne Xin Zhao, Dianhai Yu, and Hua Wu. 2018b. Multi-turn response selection for chatbots with deep attention matching network. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 1118–1127. Association for Computational Linguistics. 4457 A Appendices A.1 Details of Test Sets Statistics Wizard of Wikipedia CMU DoG Test Seen Test Unseen Test Avg. # turns 9.0 9.1 12.4 Avg, # words per turn 16.4 16.1 18.1 Avg. # knowledge entries 60.8 61.0 31.8 Avg. # words per knowledge 36.9 37.0 27.0 Table 6: The statistics of test sets of two benchmarks. We tested our proposed method on the Wizardof-Wikipedia (WoW) (Dinan et al., 2019) and CMU DoG (Zhou et al., 2018a). Both datasets contain multi-turn dialogues grounded on a set of background knowledge and are built with crowdsourcing on Amazon Mechanical Turk. In the WoW dataset, one of the paired speakers is asked to play the role of a knowledgeable expert with access to the given knowledge collection obtained from Wikipedia, while the other of a curious learner. The dataset consists of 968 complete knowledge-grounded dialogues for testing. It is worth noting that the golden knowledge index for each turn is available in the dataset. Response selection is performed at every turn of a complete dialogue, which results in 7512 for testing in total. Following the setting of the original paper, positive responses are true responses from humans and negative ones are randomly sampled. The ratio between positive and negative responses is 1 : 99 in testing sets. Besides, the test set is divided into two subsets: Test Seen and Test Unseen. The former shares 533 common topics with the training set, while the latter contains 58 new topics uncovered by the training or validation set. The CMU DoG data contains knowledgegrounded human-human conversations where the underlying knowledge comes from wiki articles and focuses on the movie domain. Similar to Dinan et al. (2019), the dataset was also built in two scenarios. In the first scenario, only one worker can access the provided knowledge collections, and he/she is responsible for introducing the movie to the other worker; while in the second scenario, both workers know the knowledge and they are asked to discuss the content. Different from WoW, the golden knowledge index for each turn is unknown for both scenarios. Since the data size for an individual scenario is small, we merge the data of the two scenarios following the setting with Zhao et al. (2019). Finally, there are 537 dialogues for testing. We evaluate the performance of the response selection at every turn of a dialogue, which results in 6637 samples for testing. We adopted the version shared in Zhao et al. (2019), where 19 negative candidates were randomly sampled for each utterance from the same set. More details about the two benchmarks can be seen in Table 6. A.2 Baselines for Knowledge Selection To compare the performance of knowledge selection, we choose the following baselines from Dinan et al. (2019) including (1) Random: the model randomly selects a knowledge entry from a set of knowledge entries; (2) IR Baseline: the model uses simple word overlap between the dialogue context and the knowledge entry to select the relevant knowledge; (3) BoW MemNet: the model is based on memory network where each memory item is a bag-of-words representation of a knowledge entry, and the gold knowledge labels for each turn are used to train the model; (4) Transformer: the model trains a context-knowledge matching network based on Transformer architecture; (5) Transformer (w/ pretrain): the model is similar to the former model, but the transformer is pre-trained on Reddit data and fine-tuned for the knowledge selection task. A.3 Results of Low-Resource Setting Ration (t) Wizard Seen Wizard Unseen R@1 R@2 R@5 R@1 R@2 R@5 0% 89.5 96.7 98.9 69.6 85.8 96.3 10% 90.8 97.1 99.4 73.2 86.9 96.8 50% 91.5 97.1 99.3 73.9 87.9 96.9 100% 92.2 97.6 99.4 74.3 88.1 97.1 Table 7: Evaluation results of our model in the lowresource setting on the Wizard of Wikipedia data. As an additional experiment, we also evaluate the proposed model for a low-resource setting. We randomly sample t ∈{10%, 50%, 100%} portion of training data from WoW, and use the data to finetune our model. The results are shown in Table 7. We can find that with only 10% training data, our model can significantly outperform existing models, indicating the advantages of our pretraining tasks. With 100% training data, our model can achieve 2.7% improvement in terms of R@1 on the test-seen and 4.7% improvement on the testunseen.
|
2021
|
343
|
Dependency-driven Relation Extraction with Attentive Graph Convolutional Networks Yuanhe Tian♥⇤, Guimin Chen}⇤, Yan Song♠~†, Xiang Wan~ ♥University of Washington }QTrade ♠The Chinese University of Hong Kong (Shenzhen) ~Shenzhen Research Institute of Big Data ♥[email protected] }[email protected] ♠[email protected] [email protected] Abstract Syntactic information, especially dependency trees, has been widely used by existing studies to improve relation extraction with better semantic guidance for analyzing the context information associated with the given entities. However, most existing studies suffer from the noise in the dependency trees, especially when they are automatically generated, so that intensively leveraging dependency information may introduce confusions to relation classification and necessary pruning is of great importance in this task. In this paper, we propose a dependency-driven approach for relation extraction with attentive graph convolutional networks (A-GCN). In this approach, an attention mechanism upon graph convolutional networks is applied to different contextual words in the dependency tree obtained from an offthe-shelf dependency parser, to distinguish the importance of different word dependencies. Consider that dependency types among words also contain important contextual guidance, which is potentially helpful for relation extraction, we also include the type information in A-GCN modeling. Experimental results on two English benchmark datasets demonstrate the effectiveness of our A-GCN, which outperforms previous studies and achieves state-ofthe-art performance on both datasets.1 1 Introduction Relation extraction (RE), which aims to detect the relationship between entity mentions from raw text, is one of the most important tasks in information extraction and retrieval, and plays a crucial role in supporting many downstream natural language processing (NLP) applications such as text mining (Distiawan et al., 2019), sentiment analysis (Sun *Equal contribution. †Corresponding author. 1The code and models involved in this paper are released at https://github.com/cuhksz-nlp/RE-AGCN. Figure 1: An illustration of noises in the dependency tree that can hurt relation extraction, where the word dependency connected in between “pumpkin mixture” and “bowl” (whose relation is content-container) may introduce confusion to the system when the object is to predict the relation between “milk” and “pumpkin mixture” (whose relation is entity-destination). et al., 2019), question answering (Xu et al., 2016a), and summarization (Wang and Cardie, 2012). Recently, neural RE methods (Zeng et al., 2014; Zhang and Wang, 2015; Xu et al., 2015; dos Santos et al., 2015; Zhang et al., 2015; Wang et al., 2016; Zhou et al., 2016; Zhang et al., 2017) with powerful encoders (such as CNN, RNN, and Transformers) have significantly improved model performance for RE without requiring any elaborately designed systems or manually constructed features. These methods are superior in capturing contextual information and thus enable RE systems to better understand the text and identify relations between entities in the given text. Adopting neural models to help RE is not only straightforward and effective, but is also expected to incorporate more diverse and informative knowledge into RE systems. Among all different knowledge sources, syntactic information, especially the dependency trees, have been demonstrated to be beneficial in many studies (Miwa and Bansal, 2016; Zhang et al., 2018; Sun et al., 2020; Chen et al., 2021) because they provide long-distance word connections between useful words and thus accordingly guide the system to better extract relations between entity pairs. However, intensively leveraging dependency information may not always lead to good RE performance, because the noise in the dependency tree can potentially introduce confusions to relation classification (Xu et al., 2015; Yu et al., 2020), especially when those trees are automatically generated. For example, Figure 1 shows an example sentence with its dependency tree, where the dependency connection between “pumpkin mixture” and “bowl” may introduce noise when the object is to predict the relation between “milk” and “pumpkin mixture”. Therefore, previous studies have always required necessary pruning strategies before encoding the dependency information through a particular model such as LSTM (Xu et al., 2015) or graph convolutional networks (GCN) (Zhang et al., 2018). Because fixed pruning strategies are not guaranteed to result in a sub-tree with all important contextual information included and with all noise filtered out, it is necessary to design an appropriate way for distinguishing the noise in the dependency tree and modelling them accordingly. In this paper, we propose a dependency-driven neural approach for RE, where attentive graph neural network (A-GCN) is proposed to distinguish the important contextual information for this task. Furthermore, given that the dependency types (e.g., nominal subject) that associate with dependency connections are also potentially useful for RE since they contain the syntactic instruction among connected words, we further improve A-GCN by introducing type information into it. Specifically, we first obtain the dependency tree of an input sentence from an off-the-shelf toolkit, then build the graph over the dependency tree, and assign different weights to different labeled dependency connections between any two words, with the weights computed based on the connections and their dependency types, lastly predict relations by the AGCN according to the learned weights. In doing so, not only is A-GCN able to distinguish important contextual information from dependency trees and leverage them accordingly, such that reliance on pruning strategies is unnecessary, but A-GCN can also leverage the dependency type information that is omitted by most previous studies (in particular, the studies that also use attention mechanism (Guo et al., 2019)). Experimental results on two English benchmark datasets, i.e., ACE2005EN and SemEval 2010 Task 8, demonstrate the effectiveness of our approach to RE through A-GCN equipped with dependency type information. State-of-the-art performance is observed on both datasets. 2 The Proposed Approach RE is conventionally performed as a typical classification task. Our approach follows this paradigm by using A-GCN and incorporates dependency information to improve model performance, where the overall architecture of our model is illustrated in Figure 2. Specifically, given an unstructured input sentence X = x1, · · · , xn with n words and let E1 and E2 denote two entities in X, our approach predicts the relation br between E1 and E2 by br = arg max r2R p (r|A-GCN (X, TX )) (1) where TX is the dependency tree of X obtained from an off-the-shelf toolkit, R is the relation type set; p computes the probability of a particular relation r 2 R given the two entities and br the output of A-GCN, which takes X and TX as the input. Following texts start with a brief introduction of the standard GCN model, then elaborate our proposed A-GCN equipped with dependency type information, and lastly illustrate the process of applying A-GCN to the classification paradigm for RE. 2.1 Standard Graph Convolutional Networks Generally, a good text representation is a prerequisite to achieve outstanding model performance (Song et al., 2017; Bojanowski et al., 2017; Song et al., 2018; Song and Shi, 2018; Hajdik et al., 2019). To enhance the text representation and thus obtain a good understanding of the running text, many studies (Song et al., 2009, 2012; Song and Xia, 2013; Xu et al., 2015; Miwa and Bansal, 2016; Zhang et al., 2019; Mandya et al., 2020; Nie et al., 2020) tried to leverage contextual features, such as n-grams and syntactic information, through different model architectures. Among all these architecture choices, graph convolutional networks (GCN) is a widely used architecture to encode the information in a graph, where in each GCN layer, information in each node communicates to its neighbors through the connections between them. The effectiveness of GCN models to encode the contextual information over a graph of an input sentence has been demonstrated by many previous studies (Zhang et al., 2018; Guo et al., 2019; Sun et al., 2020; Chen et al., 2020; Yu et al., 2020; Mandya et al., 2020; Tian et al., 2020c, 2021a). Normally, the graph in the standard GCN model is built from word dependencies and is represented by an adjacency matrix A = (ai,j)n⇥n where ai,j = 1 if i = j or there is a dependency connection2 (arc) between two words xi and xj in the dependency tree TX and ai,j = 0 otherwise. Based on A, for 2Normally the direction of the connection is ignored. Figure 2: The overall architecture of the proposed A-GCN for RE illustrated with an example input sentence (the two entities “defamation” and “bishop” are highlighted in blue and red colors, respectively) and its dependency tree. The left part shows our A-GCN model where the attention weights are applied to different connections to model the dependency type-aware contextual information. The right part illustrates the adjacency matrix A for the dependency graph and the process to compute the attention weights (i.e., p(l) i,j) for different connections. each word xi 2 X, the l-th GCN layer gathers the information carried by its context words in TX and computes the output representation h(l)i for xi by: h(l) i = σ n X j=1 ai,j ⇣ W(l) · h(l−1) j +b(l)⌘! (2) where h(l−1) j denotes the output representation of xj from the (l-1)-th GCN layer3, W(l) and b(l) are trainable matrices and the bias for the l-th GCN layer, respectively, and σ is the ReLU activation. 2.2 A-GCN with Dependency Type It is noted that in standard GCN (e.g., Eq. (2)), the connections among words are treated equally (i.e., ai,j is either 0 or 1). Therefore, GCN-based models for RE are not able to distinguish the importance of different connections and thus pruning on them is of great importance for RE. Therefore, we propose A-GCN for this task, which uses an attention mechanism to compute the weights for different connections so that the model is able to 3h(0) j is the output of the encoder for xj. leverage different dependency connections accordingly. In addition, the standard GCN and most previous studies omit the dependency types associated with the dependency connections, where those types contain highly useful information for RE and are introduced into A-GCN in this work. Specifically, we firstly represent dependency types in TX by a type matrix T = (ti,j)n⇥n, where ti,j is the dependency type (e.g., nsubj) associated with the directed dependency connection4 between xi and xj. Next, we map each type ti,j to its embedding et i,j. Then, at the l-th GCN layer, the weight for the connection between xi and xj is computed by p(l) i,j = ai,j · exp ⇣ s(l) i · s(l) j ⌘ Pn j=1 ai,j · exp ⇣ s(l) i · s(l) j ⌘ (3) where ai,j 2 A, “·” denotes inner production, and s(l) i and s(l) i are two intermediate vectors for xi and 4It means ti,j and tj,i are represented in different dependency types to model directions of connections between xi and xj. For example, if ti,j is nsubj, then tj,i is #nsubj. xj, respectively, which are computed by s(l) i = h(l−1) i ⊕et i,j (4) and s(l) j = h(l−1) j ⊕et i,j (5) with “⊕” denoting the vector concatenation operation. Afterwards, we apply the weight p(l) i,j to the associated dependency connection between xi and xj and obtain the output representation of xi by h(l) i = σ n X j=1 p(l) i,j ⇣ W(l) · eh(l−1) j + b(l)⌘! (6) with σ, W(l), and b(l) following the same notations in Eq. (2) for standard GCN, and eh(l−1) j (a typeenhanced representation for xj) computed by eh(l−1) j = h(l−1) j + W(l) T · et i,j (7) where W(l) T maps the dependency type embedding et i,j to the same dimension as h(l−1) j . Compared with standard GCN (i.e., Eq. (2)), our approach uses numerical weighting (i.e., p(l) i,j 2 [0, 1]) rather than a binary choice for ai,j, to distinguish the importance of different connections so as to leverage them accordingly. In addition, we integrate the dependency type information into both the computed weight (i.e., p(l) i,j) and the output representation of xi (i.e., h(l) i ), which is not considered in most previous studies. 2.3 Relation Extraction with A-GCN Before applying A-GCN for RE, we firstly encode the input X into hidden vectors by BERT (Devlin et al., 2019) with h(0) i denoting the hidden vector for xi, where the hidden vector (denoted as hX ) for the special sentence initial token “[CLS]” is used as the representation for the entire sentence. Next, we feed h(0) i to our proposed A-GCN model with L layers and obtain the corresponding output h(L) i . Then, we apply the max pooling mechanism to the output hidden vectors of the words that belongs to an entity mention (i.e., Ek, k = 1, 2) to compute the representation for entity (denoted as hEk) by hEk = MaxPooling({h(L) i |xi 2 Ek}) (8) Afterwards, we concatenate the representations of the sentence (i.e., hX ) and two entities (i.e., hE1 and hE2) and apply a trainable matrix WR to the computed vector to map it to the output space by o = WR · (hX ⊕hE1 ⊕hE2) (9) ACE05 SEMEVAL # INSTANCES TRAIN 48,198 8,000 DEV 11,854 TEST 10,097 2,717 Table 1: The number of unique instances (i.e., entity pairs) of ACE05 and SemEval benchmark datasets. where o is a |R|-dimensional vector with each of its value referring to a relation type in the relation type set R. Finally, we apply a softmax function of o to predict the relation br between E1 and E2 by br = arg max exp (ou) P|R| u=1 exp (ou) (10) with ou representing the value at dimension u in o. 3 Experimental Settings 3.1 Datasets In the experiments, we use two English benchmark datasets for RE, namely, ACE2005EN (ACE05)5 and SemEval 2010 Task 8 (SemEval)6 (Hendrickx et al., 2010). For ACE05, we use its English section and follow previous studies (Miwa and Bansal, 2016; Christopoulou et al., 2018; Ye et al., 2019) to pre-process it (two small subsets cts and un are removed) and split the documents into training, development, and test sets7. For SemEval, we use its official train/test split8. The numbers of unique relation types in ACE05 and SemEval are 7 and 19, respectively. We report the number of instances (i.e., entity pairs), for train/dev/test sets of ACE05 and SemEval benchmark datasets in Table 1. 3.2 Dependency Graph Construction To construct graphs for A-GCN, we use Standard CoreNLP Toolkits (SCT)9 to obtain the dependency tree TX for each input sentence X. Although our approach is able to distinguish the importance of different dependency connections through the attention mechanism, it is still beneficial if we can filter out those dependency connections that bring confusions to RE through particular pruning strategies. Motivated by previous studies (Xu et al., 2015; 5We obtain the official data (LDC2006T06) from https: //catalog.ldc.upenn.edu/LDC2006T06. 6The data is downloaded from http://docs.google. com/View?docid=dfvxd49s_36c28v9pmw. 7We follow the train/dev/test splits specified by Miwa and Bansal (2016) at https://github.com/tticoin/ LSTM-ER/tree/master/data/ace2005/split 8SemEval only includes the training and test sets. 9We download the version 3.9.2 from https:// stanfordnlp.github.io/CoreNLP/. Figure 3: An illustration on the two (i.e., local and global) groups of dependency connections for an example sentence (entities are highlighted in red color) with an adjacency matrix (on the right) built upon all connections from the two groups. Local and global connections are represented in orange and blue colors, respectively, Zhang et al., 2018; Yu et al., 2020), in this paper, we construct the graph for A-GCN by including two groups of dependency connections, namely, the local connections and the global connections. In detail, local connections include all dependencies that directly connect to the heads of two entities and global connections include all dependencies along the shortest dependency path (SDP) between the head of two entities, where in many cases words that do not directly connected to the two entities are also involved. With an example sentence including two entities (i.e., “company” and benchmarking), Figure 3 illustrates the two groups of dependency connections and the resulted adjacency matrix, which is built with the connections from the two groups10. It is worth noting that, when the SDP is short, there might be more connections in the local group than that in the global one. 3.3 Implementation Following Soares et al. (2019), we insert four special tokens (i.e., “<e1>”, “</e1>”, “<e2>”, and “</e2>”) into the input sentence to mark the boundary11 of the two entities to be investigated, which allows the encoder to distinguish the position of entities during encoding and thus improves model performance. For the encoder, we try BERT (Devlin et al., 2019), because it is a powerful pre-trained language model which and whose variants have achieved state-of-the-art performance in many NLP tasks (Wu and He, 2019; Soares et al., 2019; Wu et al., 2019; Diao et al., 2020; Song et al., 2020; Antoun et al., 2020; Tian et al., 2020a,b,d, 2021b; Qin et al., 2021; Song et al., 2021). Specifically, we use the uncased version of BERT-base and 10We do not distinguish the two groups of connections in A-GCN once they are represented by the adjacency matrix. 11For example, “<e1>” and “</e1>” are respectively inserted right before and after the entity E1 in the input X. BERT-large12 following the default settings (e.g., for BERT-base, we use 12 layers of multi-head attentions with 768-dimensional hidden vectors; for BERT-large, we use 24 layers of multi-head attentions with 1024-dimensional hidden vectors). For A-GCN, we randomly initialize all trainable parameters and the dependency type embeddings. For evaluation, we follow previous studies to use the standard micro-F1 scores13 for ACE05 and use the macro-averaged F1 scores14 for SemEval. In our experiments, we try different combinations of hyper-parameters, and tune them on the dev set, then evaluate on the test set by the model that achieves the highest F1 score on the dev set.15 4 Results 4.1 Overall Results In the experiments, we run our A-GCN models using BERT-base and BERT-large encoder on graphs with and without applying dependency pruning strategies, which correspond to the graph built upon the combined local and global connections (“L + G”), as well as the one constructed by the full dependency graph (“Full”), respectively. We also run baselines with standard GCN and standard graph attentive networks (GAT) (Veliˇckovi´c et al., 2017) with the same graph. For both standard GCN and AGCN, we try different numbers of layers (i.e. 1 to 3 layers). In addition, we try BERT-base and BERTlarge baselines without using any dependency information. Table 2 shows the F1 scores of our A-GCN 12We download different BERT models from https:// github.com/huggingface/transformers. 13We use the evaluation script from sklearn framework. 14We use the official evaluation script downloaded from http://semeval2.fbk.eu/scorers/task08/ SemEval2010_task8_scorer-v1.2.zip. 15We report the hyper-parameter settings of different models with their size and running speed in Appendix A and B. ID MODELS ACE05 SEMEVAL 1 BERT-BASE 75.31 87.87 2 + GAT (FULL) 76.16 88.39 3 + GAT (L + G) 75.79 88.53 4 + 1 GCN LAYER (FULL) 74.91 87.58 5 + 1 A-GCN LAYER (FULL) 76.63 88.34 6 + 1 GCN LAYER (L + G) 75.51 88.64 7 + 1 A-GCN LAYER (L + G) 77.10 89.03 8 + 2 GCN LAYERS (FULL) 75.09 88.66 9 + 2 A-GCN LAYERS (FULL) 77.25 88.70 10 + 2 GCN LAYERS (L + G) 76.11 88.62 11 + 2 A-GCN LAYERS (L + G) 77.30 89.16 12 + 3 GCN LAYERS (FULL) 75.69 88.54 13 + 3 A-GCN LAYERS (FULL) 76.26 88.63 14 + 3 GCN LAYERS (L + G) 76.85 88.33 15 + 3 A-GCN LAYERS (L + G) 76.36 88.70 (a) BERT-base ID MODELS ACE05 SEMEVAL 1 BERT-LARGE 76.79 89.02 2 + GAT (FULL) 78.25 89.39 3 + GAT (L + G) 78.71 89.44 4 + 1 GCN LAYER (FULL) 77.63 88.98 5 + 1 A-GCN LAYER (FULL) 78.53 89.54 6 + 1 GCN LAYER (L + G) 77.49 89.11 7 + 1 A-GCN LAYER (L + G) 78.48 89.69 8 + 2 GCN LAYERS (FULL) 78.67 89.43 9 + 2 A-GCN LAYERS (FULL) 78.91 89.70 10 + 2 GCN LAYERS (L + G) 78.82 89.42 11 + 2 A-GCN LAYERS (L + G) 79.05 89.85 12 + 3 GCN LAYERS (FULL) 78.08 89.62 13 + 3 A-GCN LAYERS (FULL) 78.45 89.46 14 + 3 GCN LAYERS (L + G) 78.64 89.19 15 + 3 A-GCN LAYERS (L + G) 78.83 89.56 (b) BERT-large Table 2: F1 scores of our A-GCN models and the baselines (i.e., BERT-only, standard GAT, and standard GCN) under different settings with BERT-base (a) and BERT-large (b) used. All graph-based models (i.e., GAT, GCN, and A-GCN) are tested with two settings: the first is using the full graph (FULL) with all dependency connections involved and the second is using the combination of local and global connections (L + G). We also run GCN and A-GCN with different numbers of layers (i.e., 1 to 3 layers) for fair comparisons. models and all the aforementioned baselines on the test set of ACE05 and SemEval.16 There are several observations. First, A-GCN functions well when using BERT-base or BERTlarge as encoder, where the consistent improvement is observed over the BERT-only baselines (ID: 1) across two benchmark datasets, even though the BERT baselines have already achieve good performance. Second, for both datasets, A-GCN outperforms GAT (ID: 2, 3) and standard GCN baselines (ID: 4, 6, 8, 10, 12, 14) with the same graph (i.e., either “L + G” or “Full”) and equal number of layers. Particularly, when full dependency graph is used, it is noted that, in some cases (e.g., ID: 8 for BERT-base on ACE05), standard GCN obtains very limited improvements (or even worse results) over the BERT-only baseline (ID: 1), whereas our A-GCN models (e.g., ID: 9 for BERT-base) is able to consistently outperform the BERT-only baseline and achieve higher performance. We attribute this observation to the attention mechanism used to weigh different dependency connections, which allows A-GCN to distinguish the noise in the graph and thus leverage useful dependency information accordingly. Third, among the models with different numbers of A-GCN layers, the ones (e.g., ID: 11 for BERT-base and ID: 11 for BERT-large) with two A-GCN layers achieves the highest scores, where similar tread is observed from the standard GCN baselines. Besides, we find that our A-GCN 16For the same group of models, we report the F1 scores on the development sets in Appendix C and the mean and standard deviation of their test set results in Appendix D. models (as well as the standard GCN baselines) with the local and global connections (i.e., “L + G”) consistently outperform the ones with full dependency graph (i.e., “Full”). These observations are relatively intuitive since the dependency information may introduce more noise to RE when it is leveraged in an intensive way (e.g., by using more layers or the full dependency tree without pruning). 4.2 Comparison with Previous Studies In addition, we compare our best models (with “L + G” or “Full” graphs) using BERT-large encoder and two A-GCN layers (ID: 9 and 11) with previous studies. The test results (F1 scores) are reported in Table 3, where our model with both local and global connections (i.e., “L + G”) outperforms all previous studies and achieves state-ofthe-art performance on the two benchmark datasets. Specifically, compared with Guo et al. (2019) who proposed an graph-based approach with attentions to leverage dependency connections, our approach leverages both dependency connections and dependency types among all input words and thus provides a better way to comprehensively leverage the dependency information. In addition, although Mandya et al. (2020) proposed an approach to leverage both dependency connections and dependency types through attentions, they added the dependency type directly to the input word embeddings along with POS embeddings, and the attention in their approach is a separate stand-alone module which is added on the top of the GCN layer. On the contrary, in our approach, the dependency type MODELS ACE05 SEMEVAL XU ET AL. (2015) 83.7 WANG ET AL. (2016) 88.0 ZHANG ET AL. (2018) 84.8 CHRISTOPOULOU ET AL. (2018) 64.2 YE ET AL. (2019) 68.9 WU AND HE (2019) (BERT) 89.2 SOARES ET AL. (2019) (BERT) 89.5 GUO ET AL. (2019) 85.4 SUN ET AL. (2020) 86.0 MANDYA ET AL. (2020) 85.9 YU ET AL. (2020) 86.4 A-GCN (BERT) (FULL) 78.91 89.70 A-GCN (BERT) (L + G) 79.05 89.85 Table 3: The comparison (F1 scores) between previous studies and our best models using two A-GCN layers and BERT-large encoder on ACE05 and SemEval. is added to each A-GCN layer and the attention mechanism is directly applied to each dependency connection in the A-GCN layer. Therefore, compared with Mandya et al. (2020), our A-GCN encodes the dependency connections and dependency types in a more intensive manner and thus can better leverage them to guide the process of predicting the relations between the given entities. 5 Analyses 5.1 The Effect of A-GCN Dependency information is supposed to be beneficial for RE because it contains long-distance wordword relations, which could be extremely useful when the given two entities are far away from each other in the input sentence. To explore the effect of A-GCN in capturing such long-distance wordword relations to help with RE, we split the test instances into different groups according to their entities’ distances (i.e., the number of words between the two entities) and run models on these groups to test their performance. Figure 4 shows the performance of our best performing A-GCN model with BERT-large (ID: 11 in Table 2) and its corresponding standard GCN and BERT-large baselines on the three groups of test instances from the test set of SemEval, where the category name indicates the range of the entity distance.17 It is observed that, A-GCN outperforms the two baselines on all groups of test instances and the improvement becomes larger when the entity distance increases. This observation confirms that our approach is able to leverage dependency information and capture long-distance word-word relations to improve RE. 17For example, a test sentence whose distance in between two entities is 7 will fall into the group (5, 10]. Figure 4: Performance (F1 scores) of different models (i.e., BERT-only, two layers of standard GCN, and two layers of A-GCN) with the BERT-large encoder on three groups of test instances from SemEval, where each group is generated based on the distance (i.e., number of words) between two entities in an instance. 5.2 The Effect of Graph Construction In the main experiments, we try A-GCN with the graph built upon the combined local and global connections (“L + G”). To explore the effect of the local connections and the global connections for AGCN, we run our approach using two A-GCN layers with the graph constructed by local connections (“L”) or global connections (“G”) alone. Table 4 presents the experimental results (F1 scores) of different models with BERT-base and BERT-large encoders, where the results from BERT-only baselines, A-GCN (L + G), and A-GCN (Full) are also copied from Table 2 for reference. Compared to A-GCN (L + G), models with the graph constructed by either local connections (i.e., A-GCN (L)) or global connections (i.e., A-GCN (G)) achieve lower performance, which complies with our intuition because both groups of connections contain important contextual features for RE. Interestingly, it is found that A-GCN (L) outperforms A-GCN (G) with both BERT-base and BERT-large encoders. A possible explanation could be the following. There are overlaps between local and global connections (e.g., the connection between “range” and “restrictions” in Figure 3). Therefore, A-GCN (L) can not only leverage the contextual information associated with the entities themselves, but is also partially18 benefited from the overlapping connections on the SDP between the two entities, which leads A-GCN (L) to achieve a higher performance than A-GCN (G). 5.3 Ablation Study Compared with the standard GCN, A-GCN enhances it from two aspects: (1) using an attention 18When there is only one word on the shortest dependency path between two entities, all global connections are included in local ones, e.g., “defamation” and “bishop” in Figure 2. ID MODELS ACE2005 SEMEVAL 1 BERT-BASE 75.31 87.87 2 + A-GCN (L) 76.92 88.89 3 + A-GCN (G) 76.72 88.89 4 + A-GCN (L + G) 77.30 89.16 5 + A-GCN (FULL) 77.25 88.70 6 BERT-LARGE 76.79 89.02 7 + A-GCN (L) 78.61 89.70 8 + A-GCN (G) 78.40 89.38 9 + A-GCN (L + G) 79.05 89.85 10 + A-GCN (FULL) 78.91 89.70 Table 4: Performance of our models with two A-GCN layers using the graphs built upon (1) only local connections (L), (2) only global connections (G), (3) the combination of local and global connections (G + L) , and (4) full dependency graph (FULL). The performance of BERT-only baseline is also reported for reference. mechanism to weigh different dependency connections and (2) introducing dependency types to the process to encode more detailed dependency information. To better investigate the effect of each individual enhancement (i.e., the attention mechanism or the dependency type information), we conduct an ablation study on our best model, i.e., two layers of A-GCN (L + G) with BERT-base and BERTlarge encoder. Table 5 reports the experimental results of different models, where the performance of BERT-only baseline and the standard GCN baseline (i.e., the one uses neither the attention mechanism nor dependency types) are also reported for reference. The results clearly indicate that, the ablation of either enhancement (i.e., the attention mechanism or the dependency type information) could result in worse results (compared with full A-GCN). Between the two enhancements, the ablation of the attention mechanism hurts A-GCN more, which indicates the ability of distinguishing important connections and leveraging them accordingly plays a more important role in RE. 5.4 Case Study To explore in detail that how A-GCN leverages dependency connections and types to improve RE, we conduct a case study with our A-GCN models with different dependency graphs (i.e., two layers of A-GCN (Full) and A-GCN (L + G) with BERTlarge encoder) on an example sentence “A central vacuum is a vacuum motor and filtration system built inside a canister.”. Figure 5 shows the sentence where both the two models correctly predict the relation between “motor” (E1) and “canister” ATT. TYPE ACE2005 SEMEVAL BERT-BASE BASELINE 75.31 87.87 p p 77.30 89.16 ⇥ p 77.00 88.07 p ⇥ 76.27 88.50 GCN 76.11 88.62 BERT-LARGE BASELINE 76.79 89.02 p p 79.05 89.85 ⇥ p 78.92 89.26 p ⇥ 78.22 89.37 GCN 77.92 89.13 Table 5: The ablation study on the attention mechanism (ATT.) and dependency types (TYPE) in our best model, i.e., two layers of A-GCN (L + G). “p” and “⇥” stand for that whether a module is used. The F1 scores of BERT-only and the standard two layers of GCN (L + G) are also reported for references. (E2) (highlighted in the red color) to be “ContentContainer”, whereas the baseline GCN (Full) and GCN (L + G) models fail to do so. We also visualize the attention weights assigned to different dependency connections extracted from the last AGCN layer, with darker and thicker lines referring to higher weights. In this example, for A-GCN (Full), we observe that the connection between “built” and “canister” along SDP and the connection between “inside” and “canister” receive the highest weights, where this is valid because the dependency type, i.e., obl (oblique nominal), associated with the connection (between “built” and “canister”) reveals that “canister” could be the position where the action (i.e., build) takes place, and is further confirmed by another dependency connection and type (i.e., case) between “inside” and “canister”. Therefore, it is proved that our model learn from the contextual information carried by such important connections and results in correct RE prediction. Similarly, A-GCN (L + G) also correctly perform RE on this case by highlighting the same dependency connections as those from the A-GCN (Full) with much higher weights (because many dependency connections are filtered out). 6 Related Work Recently, neural networks with integrating external knowledge or resources play important roles in RE because of their superiority in better capturing contextual information (Shen and Huang, 2016; Soares et al., 2019). Particularly, as one kind of such knowledge, dependency parses show their effectiveness in supporting RE for its ability Figure 5: Visualizations of weights assigned to different dependency connections of A-GCN (Full) and AGCN (L + G) for an example input, where darker and thicker lines refer to connections with higher weights. in capturing long-distance word relations (Zhang et al., 2018; Guo et al., 2019). However, intensively leveraging dependency information could introduce confusions to RE (Xu et al., 2016b; Yu et al., 2020) so that necessary pruning is required to alleviate this problem. E.g., Xu et al. (2015) proposed to use the connections along the shortest dependency path between the two entities and apply LSTM to model them; Miwa and Bansal (2016) proposed to prune the original dependency tree into the lowest common ancestor subtree. However, these pruning strategies are either too aggressive or modest, so that the resulted graph might lose some important contexts or filled with more noise. Zhang et al. (2018) adopted GCN to model the dependencies and proposed a trade-off pruning strategy in between Xu et al. (2015) and Miwa and Bansal (2016). Besides, there are other graphbased models for RE that utilize layers of multihead attentions (Guo et al., 2019), dynamic pruning (Yu et al., 2020), and additional attention layers (Mandya et al., 2020) to encode dependency trees. Compared with the aforementioned methods, especially the graph-based ones, our approach offers an alternative to enhance RE with A-GCN by using attention mechanism and dependency type, which are effective and efficient improvement to standard GCN without requiring complicated model design. 7 Conclusion In this paper, we propose A-GCN to leverage dependency information for relation extraction, where an attention mechanism is applied to dependency connections to applying weighting on both connections and types so as to better distinguish the important dependency information and leverage them accordingly. In doing so, A-GCN is able to dynamically learn from different dependency connections so that less-informative dependencies are smartly pruned. Experimental results and analyses on two English benchmark datasets for relation extraction demonstrate the effectiveness of our approach, especially for entities with long word-sequence distances, where state-of-theart performance is obtained on both datasets. Acknowledgements This work is supported by Chinese Key-Area Research and Development Program of Guangdong Province (2020B0101350001) and NSFC under the project “The Essential Algorithms and Technologies for Standardized Analytics of Clinical Texts” (12026610). This work is also partially supported by Shenzhen Institute of Artificial Intelligence and Robotics for Society under the project “Automatic Knowledge Enhanced Natural Language Understanding and Its Applications” (AC01202101001). We also thank Mr. Peilin Zhou for providing the first version of the model architecture figure. References Wissam Antoun, Fady Baly, and Hazem Hajj. 2020. AraBERT: Transformer-based Model for Arabic Language Understanding. arXiv preprint arXiv:2003.00104. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching Word Vectors with Subword Information. Transactions of the Association for Computational Linguistics, 5:135–146. Guimin Chen, Yuanhe Tian, and Yan Song. 2020. Joint Aspect Extraction and Sentiment Analysis with Directional Graph Convolutional Networks. In Proceedings of the 28th International Conference on Computational Linguistics, pages 272–279. Guimin Chen, Yuanhe Tian, Yan Song, and Xiang Wan. 2021. Relation Extraction with Type-aware Map Memories of Word Dependencies. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021. Fenia Christopoulou, Makoto Miwa, and Sophia Ananiadou. 2018. A Walk-based Model on Entity Graphs for Relation Extraction. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 81–88. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186. Shizhe Diao, Jiaxin Bai, Yan Song, Tong Zhang, and Yonggang Wang. 2020. ZEN: Pre-training Chinese Text Encoder Enhanced by N-gram Representations. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4729–4740. Bayu Distiawan, Gerhard Weikum, Jianzhong Qi, and Rui Zhang. 2019. Neural Relation Extraction for Knowledge Base Enrichment. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 229–240. Zhijiang Guo, Yan Zhang, and Wei Lu. 2019. Attention Guided Graph Convolutional Networks for Relation Extraction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 241–251. Valerie Hajdik, Jan Buys, Michael Wayne Goodman, and Emily M. Bender. 2019. Neural Text Generation from Rich Semantic Representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2259–2266, Minneapolis, Minnesota. Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid ´O S´eaghdha, Sebastian Pad´o, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2010. SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations between Pairs of Nominals. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 33–38. Angrosh Mandya, Danushka Bollegala, and Frans Coenen. 2020. Graph Convolution over Multiple Dependency Sub-graphs for Relation Extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6424–6435. Makoto Miwa and Mohit Bansal. 2016. End-to-End Relation Extraction using LSTMs on Sequences and Tree Structures. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1105– 1116. Yuyang Nie, Yuanhe Tian, Yan Song, Xiang Ao, and Xiang Wan. 2020. Improving Named Entity Recognition with Attentive Ensemble of Syntactic Information. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4231–4245. Han Qin, Guimin Chen, Yuanhe Tian, and Yan Song. 2021. Improving Arabic Diacritization with Regularized Decoding and Adversarial Training. In Proceedings of the Joint Conference of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing. C´ıcero dos Santos, Bing Xiang, and Bowen Zhou. 2015. Classifying Relations by Ranking with Convolutional Neural Networks. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 626–634. Yatian Shen and Xuanjing Huang. 2016. Attentionbased Convolutional Neural Network for Semantic Relation Extraction. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2526– 2536, Osaka, Japan. Livio Baldini Soares, Nicholas FitzGerald, Jeffrey Ling, and Tom Kwiatkowski. 2019. Matching the Blanks: Distributional Similarity for Relation Learning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2895–2905. Yan Song, Chunyu Kit, and Xiao Chen. 2009. Transliteration of Name Entity via Improved Statistical Translation on Character Sequences. In Proceedings of the 2009 Named Entities Workshop: Shared Task on Transliteration (NEWS 2009), pages 57–60. Yan Song, Prescott Klassen, Fei Xia, and Chunyu Kit. 2012. Entropy-based Training Data Selection for Domain Adaptation. In Proceedings of COLING 2012: Posters, pages 1191–1200. Yan Song, Chia-Jung Lee, and Fei Xia. 2017. Learning Word Representations with Regularization from Prior Knowledge. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 143–152. Yan Song and Shuming Shi. 2018. Complementary Learning of Word Embeddings. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 4368– 4374. Yan Song, Shuming Shi, and Jing Li. 2018. Joint Learning Embeddings for Chinese Words and Their Components via Ladder Structured Networks. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4375–4381. Yan Song, Yuanhe Tian, Nan Wang, and Fei Xia. 2020. Summarizing Medical Conversations via Identifying Important Utterances. In Proceedings of the 28th International Conference on Computational Linguistics, pages 717–729. Yan Song and Fei Xia. 2013. A Common Case of Jekyll and Hyde: The Synergistic Effect of Using Divided Source Training Data for Feature Augmentation. In Proceedings of the Sixth International Joint Conference on Natural Language Processing, pages 623–631. Yan Song, Tong Zhang, Yonggang Wang, and Kai-Fu Lee. 2021. ZEN 2.0: Continue Training and Adaption for N-gram Enhanced Text Encoders. arXiv preprint arXiv:2105.01279. Kai Sun, Richong Zhang, Yongyi Mao, Samuel Mensah, and Xudong Liu. 2020. Relation Extraction with Convolutional Network over Learnable SyntaxTransport Graph. In AAAI, pages 8928–8935. Kai Sun, Richong Zhang, Samuel Mensah, Yongyi Mao, and Xudong Liu. 2019. Aspect-level Sentiment Analysis via Convolution over Dependency Tree. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5683–5692. Yuanhe Tian, Guimin Chen, and Yan Song. 2021a. Aspect-based Sentiment Analysis with Type-aware Graph Convolutional Networks and Layer Ensemble. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2910–2922, Online. Yuanhe Tian, Guimin Chen, and Yan Song. 2021b. Enhancing Aspect-level Sentiment Analysis with Word Dependencies. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3726–3739, Online. Yuanhe Tian, Wang Shen, Yan Song, Fei Xia, Min He, and Kenli Li. 2020a. Improving Biomedical Named Entity Recognition with Syntactic Information. BMC Bioinformatics, 21:1471–2105. Yuanhe Tian, Yan Song, and Fei Xia. 2020b. Joint Chinese Word Segmentation and Part-of-speech Tagging via Multi-channel Attention of Character Ngrams. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2073–2084. Yuanhe Tian, Yan Song, and Fei Xia. 2020c. Supertagging Combinatory Categorial Grammar with Attentive Graph Convolutional Networks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6037–6044. Yuanhe Tian, Yan Song, Fei Xia, and Tong Zhang. 2020d. Improving Constituency Parsing with Span Attention. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1691– 1703. Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2017. Graph Attention Networks. arXiv preprint arXiv:1710.10903. Linlin Wang, Zhu Cao, Gerard De Melo, and Zhiyuan Liu. 2016. Relation Classification via Multi-Level Attention CNNs. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1298– 1307. Lu Wang and Claire Cardie. 2012. Focused Meeting Summarization via Unsupervised Relation Extraction. In Proceedings of the 13th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 304–313. Shanchan Wu and Yifan He. 2019. Enriching Pretrained Language Model with Entity Information for Relation Classification. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pages 2361–2364. Zhaofeng Wu, Yan Song, Sicong Huang, Yuanhe Tian, and Fei Xia. 2019. WTMED at MEDIQA 2019: A Hybrid Approach to Biomedical Natural Language Inference. In Proceedings of the 18th BioNLP Workshop and Shared Task, pages 415–426, Florence, Italy. Kun Xu, Siva Reddy, Yansong Feng, Songfang Huang, and Dongyan Zhao. 2016a. Question Answering on Freebase via Relation Extraction and Textual Evidence. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2326–2336. Kun Xu, Siva Reddy, Yansong Feng, Songfang Huang, and Dongyan Zhao. 2016b. Question answering on Freebase via relation extraction and textual evidence. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Yan Xu, Lili Mou, Ge Li, Yunchuan Chen, Hao Peng, and Zhi Jin. 2015. Classifying Relations via Long Short Term Memory Networks Along Shortest Dependency Paths. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1785–1794. Wei Ye, Bo Li, Rui Xie, Zhonghao Sheng, Long Chen, and Shikun Zhang. 2019. Exploiting Entity BIO Tag Embeddings and Multi-task Learning for Relation Extraction with Imbalanced Data. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1351–1360. Bowen Yu, Mengge Xue, Zhenyu Zhang, Tingwen Liu, Wang Yubin, and Bin Wang. 2020. Learning to Prune Dependency Trees with Rethinking for Neural Relation Extraction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3842–3852, Barcelona, Spain (Online). Daojian Zeng, Kang Liu, Siwei Lai, Guangyou Zhou, and Jun Zhao. 2014. Relation Classification via Convolutional Deep Neural Network. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2335–2344. Dongxu Zhang and Dong Wang. 2015. Relation Classification via Recurrent Neural Network. arXiv preprint arXiv:1508.01006. Hongming Zhang, Yan Song, and Yangqiu Song. 2019. Incorporating Context and External Knowledge for Pronoun Coreference Resolution. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 872–881. Shu Zhang, Dequan Zheng, Xinchen Hu, and Ming Yang. 2015. Bidirectional Long Short-Term Memory Networks for Relation Classification. In Proceedings of the 29th Pacific Asia Conference on Language, Information and Computation, pages 73–78. Yuhao Zhang, Peng Qi, and Christopher D. Manning. 2018. Graph Convolution over Pruned Dependency Trees Improves Relation Extraction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2205–2215. Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D. Manning. 2017. Positionaware Attention and Supervised Data Improve Slot Filling. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 35–45. Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li, Hongwei Hao, and Bo Xu. 2016. Attention-Based Bidirectional Long Short-Term Memory Networks for Relation Classification. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 207–212. Appendix A. Hyper-parameter Settings Table 6 reports the hyper-parameters tested in training our models. We test all combinations of them for each model and use the one achieving the highest F1 score in our final experiments. The best hyper-parameter setting is highlighted in boldface. Hyper-parameters Values Learning Rate 5e −6, 1e −5, 2e −5, 3e −5 Warmup Rate 0.06, 0.1 Dropout Rate 0.1 Batch Size 16, 32, 64, 128 Table 6: The hyper-parameters tested in tuning our models. The best ones used in our final experiments are highlighted in boldface. Appendix B. Model Size and Running Speed Table 7 reports the number of trainable parameters and the inference speed (sentences per second) of the baseline (i.e., BERT, BERT + GAT and BERT + GCN) and our models (i.e., BERT + A-GCN) on ACE2005 and SemEval datasets. All models are performed on an NVIDIA Tesla V100 GPU. Appendix C. Experimental Results on the Development Set Table 8 reports the F1 scores of different models on the development set of ACE2005.19 Appendix D. Mean and Deviation of the Results In the experiments, we test models with different configurations. For each model, we train it with the best hyper-parameter setting using five different random seeds. We report the mean (µ) and standard deviation (σ) of the F1 scores on the test set of ACE2005 and SemEval in Table 9. 19SemEval does not have an official dev set. Models ACE2005 SemEval Para. Speed Para. Speed BERT-base 109M 27.7 109M 54.7 + GAT (Full) 110M 26.2 110M 51.8 + GAT (L + G) 110M 26.2 110M 51.8 + 1 GCN layer (Full) 110M 26.4 110M 52.2 + 1 A-GCN layer (Full) 110M 25.1 110M 50.4 + 1 GCN layer (L + G) 110M 26.4 110M 52.2 + 1 A-GCN layer (L + G) 110M 25.1 110M 50.4 + 2 GCN layers (Full) 111M 24.8 111M 49.9 + 2 A-GCN layers (Full) 111M 24.1 111M 48.7 + 2 GCN layers (L + G) 111M 24.8 111M 49.9 + 2 A-GCN layers (L + G) 111M 24.1 111M 48.7 + 3 GCN layers (Full) 112M 23.1 112M 47.9 + 3 A-GCN layers (Full) 112M 23.0 112M 47.2 + 3 GCN layers (L + G) 112M 23.1 112M 47.9 + 3 A-GCN layers (L + G) 112M 23.0 112M 47.2 (a) BERT-base Models ACE2005 SemEval Para. Speed Para. Speed BERT-large 335M 8.9 335M 17.1 + GAT (Full) 337M 8.4 337M 16.7 + GAT (L + G) 337M 8.4 337M 16.7 + 1 GCN layer (Full) 337M 8.6 337M 16.9 + 1 A-GCN layer (Full) 337M 8.1 337M 16.6 + 1 GCN layer (L + G) 337M 8.6 337M 16.9 + 1 A-GCN layer (L + G) 337M 8.1 337M 16.6 + 2 GCN layers (Full) 338M 8.0 338M 16.3 + 2 A-GCN layers (Full) 338M 7.8 338M 16.1 + 2 GCN layers (L + G) 338M 8.0 338M 16.3 + 2 A-GCN layers (L + G) 338M 7.8 338M 16.1 + 3 GCN layers (Full) 339M 7.4 339M 15.8 + 3 A-GCN layers (Full) 339M 7.2 339M 15.5 + 3 GCN layers (L + G) 339M 7.4 339M 15.8 + 3 A-GCN layers (L + G) 339M 7.2 339M 15.5 (b) BERT-large Table 7: Numbers of trainable parameters (Para.) in different models and the inference speed (sentences per second) of these models on the test sets of both datasets. Models BERT-base BERT-Large Baseline 75.03 76.11 GAT (Full) 75.33 76.87 GAT (L + G) 75.31 76.93 + 1 GCN layer (Full) 74.97 76.13 + 1 A-GCN layer (Full) 76.49 77.33 + 1 GCN layer (L + G) 75.80 77.19 + 1 A-GCN layer (L + G) 76.00 77.49 + 2 GCN layers (Full) 75.36 77.35 + 2 A-GCN layers (Full) 76.65 77.55 + 2 GCN layers (L + G) 76.59 77.48 + 2 A-GCN layers (L + G) 76.90 77.82 + 3 GCN layers (Full) 75.61 77.33 + 3 A-GCN layers (Full) 76.45 77.54 + 3 GCN layers (L + G) 76.48 77.36 + 3 A-GCN layers (L + G) 76.58 77.65 Table 8: F1 scores of our A-GCN models and the baselines (i.e., BERT-only, standard GAT, and standard GCN) under different settings with BERT-base and BERT-large on the development set of ACE2005. All graph-based models (i.e., GAT, GCN, and A-GCN) are tested with two settings: the first is using the full graph (FULL) with all dependency connections involved and the second is using the combination of local and global connections (L + G). We also run GCN and A-GCN with different numbers of layers (i.e., 1 to 3 layers) for fair comparisons. Models ACE2005 SemEval µ σ µ σ BERT-base 75.22 0.31 87.39 0.26 + GAT (Full) 75.87 0.23 88.16 0.44 + GAT (L + G) 75.47 0.27 88.15 0.28 + 1 GCN layer (Full) 74.51 0.13 87.34 0.29 + 1 A-GCN layer (Full) 74.39 0.21 88.02 0.30 + 1 GCN layer (L + G) 75.28 0.23 88.43 0.17 + 1 A-GCN layer (L + G) 76.70 0.37 88.69 0.28 + 2 GCN layers (Full) 74.73 0.24 88.13 0.31 + 2 A-GCN layers (Full) 76.95 0.21 88.35 0.34 + 2 GCN layers (L + G) 75.60 0.42 88.30 0.23 + 2 A-GCN layers (L + G) 77.06 0.13 88.81 0.28 + 3 GCN layers (Full) 75.37 0.15 88.26 0.21 + 3 A-GCN layers (Full) 75.94 0.28 88.29 0.26 + 3 GCN layers (L + G) 76.48 0.38 88.10 0.16 + 3 A-GCN layers (L + G) 75.87 0.45 88.46 0.25 (a) BERT-base Models ACE2005 SemEval µ σ µ σ BERT-large 76.55 0.17 88.63 0.26 + GAT (Full) 77.96 0.18 89.10 0.21 + GAT (L + G) 78.33 0.38 89.13 0.31 + 1 GCN layer (Full) 77.30 0.28 88.52 0.31 + 1 A-GCN layer (Full) 78.15 0.37 89.05 0.49 + 1 GCN layer (L + G) 76.98 0.49 88.80 0.28 + 1 A-GCN layer (L + G) 78.04 0.32 89.32 0.22 + 2 GCN layers (Full) 78.56 0.41 89.16 0.26 + 2 A-GCN layers (Full) 78.68 0.22 89.34 0.33 + 2 GCN layers (L + G) 78.40 0.33 89.22 0.17 + 2 A-GCN layers (L + G) 78.83 0.21 89.41 0.44 + 3 GCN layers (Full) 77.58 0.32 89.14 0.36 + 3 A-GCN layers (Full) 78.03 0.32 89.16 0.17 + 3 GCN layers (L + G) 78.64 0.27 88.93 0.26 + 3 A-GCN layers (L + G) 78.55 0.45 89.20 0.33 (b) BERT-large Table 9: The mean µ and standard deviation σ of F1 scores of our A-GCN model and baselines on the test set of ACE2005 and SemEval for relation extraction.
|
2021
|
344
|
Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4472–4485 August 1–6, 2021. ©2021 Association for Computational Linguistics 4472 Evaluating Entity Disambiguation and the Role of Popularity in Retrieval-Based NLP Anthony Chen∗ Pallavi Gudipati Shayne Longpre Xiao Ling Sameer Singh University of California, Irvine Apple {anthony.chen, sameer}@uci.edu {pgudipati, slongpre, xiaoling}@apple.com Abstract Retrieval is a core component for open-domain NLP tasks. In open-domain tasks, multiple entities can share a name, making disambiguation an inherent yet under-explored problem. We propose an evaluation benchmark for assessing the entity disambiguation capabilities of these retrievers, which we call Ambiguous Entity Retrieval (AmbER) sets. We define an AmbER set as a collection of entities that share a name along with queries about those entities. By covering the set of entities for polysemous names, AmbER sets act as a challenging test of entity disambiguation. We create AmbER sets for three popular open-domain tasks: fact checking, slot filling, and question answering, and evaluate a diverse set of retrievers. We find that the retrievers exhibit popularity bias, significantly under-performing on rarer entities that share a name, e.g., they are twice as likely to retrieve erroneous documents on queries for the less popular entity under the same name. These experiments on AmbER sets show their utility as an evaluation tool and highlight the weaknesses of popular retrieval systems.1 1 Introduction Substantial progress in NLP has been made on “closed” tasks, where queries are paired with relevant documents (Rajpurkar et al., 2016; Dua et al., 2019). However, there is growing interest in “opendomain” tasks, where relevant documents need to be retrieved from a knowledge source before an NLP system can perform reasoning and produce an answer (Chen et al., 2017; Petroni et al., 2021). The open-domain setting better reflects real-world usage for tasks where relevant information is generally not provided (e.g., fact checking). ∗Work started during an internship at Apple. 1The AmbER sets used in this paper and the code to generate them are available at https://github.com/ anthonywchen/AmbER-Sets. Q: Which battle did Abe Lincoln fight in? A: World War II Wikipedia Documents Ranked by BLINK: 1. Abraham Lincoln 2. Abraham Lincoln in the Black Hawk War 3. Abraham Lincoln (captain) 4. Benjamin Lincoln 5. Lincoln Nebraska 6. Lincoln England Q: What musical instrument does Abe Lincoln play? A: Trombone Wikipedia Documents Ranked by BLINK: 1. Abraham Lincoln 2. John Wilkes Booth 3. Abe (musical) 4. Nebraska 5. Lincoln Nebraska 6. Abe Lincoln (musician) Figure 1: Queries for two entities (president & musician) with the name “Abe Lincoln”. Retrieving the gold document involves disambiguating which “Abe Lincoln” each query is asking about. BLINK performs sub-optimally on the second query, as it ranks the document of the president over the gold document. Because success hinges on finding relevant documents, open-domain progress has been closely tied to improvements in retrieval systems2 (Lee et al., 2019; Karpukhin et al., 2020; Lewis et al., 2020b). A crucial challenge when interacting with a large knowledge source (e.g., Wikipedia) is entity ambiguity, the phenomenon where a single name can map to multiple entities. Resolving this ambiguity is referred to as entity disambiguation and is an important step for effective retrieval. For example, given the query “What musical instrument does Abe Lincoln play?”, documents about the musician should rank higher than other entities with the same name (Figure 1). Although entity disambiguation has been extensively studied in entity linking (Hoffart et al., 2011; Rao et al., 2013; Sevgili et al., 2For example, replacing the BM25 retriever with DPR on Natural Questions increases exact match by 15 points. 4473 2020) and search (Balog et al., 2010, 2011), in the context of open-domain NLP, it is unclear how good retrieval systems are when faced with queries with ambiguous entities. Evaluating entity ambiguity is challenging because the popularity of entities follows a long-tail (Figure 2) and rare entities are seldom covered in naturally-occurring datasets. In this paper we introduce AmbER sets, a benchmark for evaluating the entity disambiguation capabilities of retrievers across multiple NLP tasks. Each AmbER set is a collection of Wikidata entities that share a name, and their corresponding queries for specific NLP tasks. For each set, we define the head entity as the most popular entity and tail entities as the less popular ones. By creating queries for multiple entities that share a name, AmbER sets provide an accurate test of entity disambiguation capabilities of retrievers and help assess the role of entity popularity in disambiguation. We show examples of AmbER sets for the question answering task in Table 1. We automatically create AmbER sets by mining the Wikidata knowledge graph (Vrandecic and Kr¨otzsch, 2014) for relevant names and entities, and leveraging task-specific templates to generate inputs for three tasks: fact checking, slot filling, and question answering (Figure 3). In total, our AmbER sets contain 80k task-specific queries which we align to the Wikipedia snapshot from KILT (Petroni et al., 2021). We use AmbER sets to conduct a systematic study of various retrieval systems that operate under different principles, such as token overlap and dense embedding similarity. Retrievers perform very differently on AmbER sets in terms of absolute retrieval numbers, with Bootleg (Orr et al., 2020), an entity-linking-based retriever, performing best. Despite these differences, all retrievers exhibit a large degree of popularity bias, underperforming on inputs concerning tail entities. TFIDF, a token-based retriever, performs about four times worse on tail entity inputs compared to head entity inputs. Even with Bootleg, the best performing retriever, performance on tail entities is still 1.5 times lower than on head entities. Our results on AmbER sets demonstrate that there is significant work to be done on making retrievers robust in handling entity disambiguation. 2 AmbER Sets Retrieving relevant documents from large knowledge sources such as Wikipedia is an important Figure 2: The Long Tail of Entity Popularity: Graph of the Wikipedia pageviews (in October 2019) for each Wikidata entity, ranked by popularity. Gray are 100k randomly sampled entities, while red/blue are entities with the name “Abe Lincoln”. first step in the open-domain pipeline. An inherent problem in working with such sources is entity disambiguation: resolving a name (mention) to an entity in the knowledge source. Entity disambiguation can be challenging because many entities share a name, and the popularity of entities follows a long-tail distribution (Figure 2). Despite the importance of entity disambiguation, it remains an understudied problem for open-domain NLP. We introduce AmbER sets for evaluating entity disambiguation capabilities of retrievers and analyze the role of entity popularity in disambiguation. 2.1 What is an AmbER Set? We first provide an intuition for an AmbER set before concretely defining one. Consider two entities, a president and a musician, both of which have the name “Abe Lincoln” (Figure 1). Now, consider the query “Which battle did Abe Lincoln fight in?” and assume a retriever correctly returns the article about the president for this query. Simply because the correct document was retrieved does not mean a retriever has the ability to disambiguate between the president and the musician, as the president is much more popular. We should only be confident in its ability to disambiguate entities if we also pose a query about the less popular musician and the retriever again returns the correct document (as opposed to the document about the president). Based on this intuition, we define an AmbER set as a collection of queries that satisfy the following: • Criteria 1: Polysemous Name: The queries in an AmbER set are all about entities that share a common name (e.g., Abe Lincoln). 4474 QID Input Answer Gold Document AmbER-H Q517 What wars did Napoleon participate in? Napoleon Wars Napoleon Q3335909 What sport does Napoleon play? Rugby Napolioni Nalaga Q3335909 Which team does Napoleon play for? Fiji National Napolioni Nalaga Q117012 What movement did Yoko Ono participate in? Fluxus Yoko Ono Q16264827 Which sport does Yoko Ono participate in? Judo Yoko Ono (judoka) AmbER-N Q312 Which industry is Apple in? Electronics Apple Inc. Q532100 What is the record label of Apple? Page One Apple (band) Q7714007 Who acted in Apple? Ray Shell The Apple (1980 film) Q788822 Who is a cast member on Her? Steve Zissis Her (film) Q788822 Who is Her’s screenwriter? Spike Jonze Her (film) Q28441308 Who performed Her? Aaron Tippin Her (song) Table 1: Examples of QA AmbER sets. An AmbER set is a collection of entities that share a name, with instantiated queries for each entity. In this work, we use Wikidata to collect entities (QID). We also create queries for two more tasks, fact checking and slot filling (omitted from this table). • Criteria 2: Disparity in Popularity: An AmbER set contains queries about both the most popular entity for a name (the head entity), e.g., the president, and the less popular entities (the tail entities), e.g., the musician. • Criteria 3: Resolvable Ambiguity: The content of the query should be sufficient to resolve to the correct entity. The query “Which battle did Abe Lincoln fight in?” satisfies this criteria, because there is only one Abe Lincoln that fought in a war, while “Where was Abe Lincoln born?” does not since it applies to all Abe Lincolns. We provide examples of AmbER sets for the task of question answering in Table 1. 2.2 Open-Domain Tasks In this work, we create AmbER sets for three tasks: fact checking, slot filling, and question answering (Table 2). We consider these three tasks for three reasons. First, these three set of tasks are diverse in nature. In this work, slot filling is a generation task, question answering is a span selection task, and fact checking is a classification task. Second, the training sets available for each task are quite disparate. The largest fact checking training set, FEVER (Thorne et al., 2018), has 80k instances, while the slot filling dataset, T-REx (Elsahar et al., 2018), has over 2 million instances. The final reason we study these three tasks is that their inputs are short and easy to create. 3 Creating AmbER Sets While AmbER sets can be manually created, doing so can be time-consuming, requiring a human to manually scour a knowledge base for polysemous Task Input Instance Output FC John Mayer plays music. True SF Nike [SEP] country USA QA Whose face is on $100 bill? Benjamin Franklin Table 2: Examples for each open-domain NLP task. names and related entities before manually writing queries for those entities. Instead, we present a pipeline for automatically creating AmbER sets using the Wikidata knowledge graph (Vrandecic and Kr¨otzsch, 2014). In this section, we describe two different collections of AmbER sets, and discuss our automatic pipeline for creating AmbER sets. 3.1 Two Collections of AmbER Sets A natural question is “How do retrievers handle entity ambiguity when two entities have the same entity type as opposed when they have different types?”. To answer this question, we create two collections of AmbER sets. The first is AmbERH, a collection of AmbER sets where all entities are humans. The choice to restrict AmbER-H to humans is motivated by the fact that humans have properties that help distinguish themselves from other humans, generally based on occupation. The second is AmbER-N, a collection of AmbER sets where all entities contained are non-humans, and disambiguation of a name is between non-human entities with different entity types. This is because a non-human entity, like a movie, does not generally have a single distinguishing property to distinguish from other movies. This makes it natural to compare non-human entities to other non-human entities with different types. We specify the entity types in each collection in Table 3. 4475 “Davy Jones” Name David Bowie* Popularity: 4.09 Wikidata Entities Davy Jones (racing driver) Popularity: 2.49 Davy Jones (baseball) Popularity: 1.93 Gender: Male Birthplace: Brixton Gender: Male Sport: Baseball Gender: Male Sport: Auto Racing Movement: New Wave Wikidata Properties Sports Team: Chicago White Sox Task Specific Inputs QA: Which movement is Davy Jones associated with? SF: Davy Jones [SEP] movement FC: Davy Jones participated in the new wave movement. TRUE Davy Jones participated in the baroque music movement. FALSE QA: Which team does Davy Jones play for? SF: Davy Jones [SEP] member of sports team FC: Davy Jones plays for the Chicago White Sox. TRUE Davy Jones plays for the Philadelphia Phillies. FALSE *born Davy Jones Q5383 Q1178405 Q5242203 Figure 3: Automated creation of AmbER sets for three tasks. We collect sets of entities from Wikipedia that share a name, where the most popular entity is the head entity (in red) and others are tail entities (in blue), along with their properties and associated values. We filter out properties that do not help distinguish entities in the set (gray-ed out), and remove entities that do not have any properties remaining. From the remaining properties, we instantiate queries via templates for three tasks: question answering (QA), slot filling (SF), and fact checking (FC). 3.2 Automatic Creation of AmbER Sets We now describe a pipeline to automatically create AmbER sets for three tasks: fact checking, slot filling, and question answering. We provide a visualization of the pipeline in Figure 3. Collecting Names and Entities We begin by collecting all entity aliases3 in Wikidata. From these aliases, we filter for those that are shared by multiple Wikidata entities. Each entity in Wikidata is represented by a unique QID. The entities must have an entity type from Table 3 depending on the collection we are collecting AmbER sets for. Each alias and associated entities form the basis for an AmbER set. Within each set, we define the head and tail entities based on the number of Wikipedia page views for the month of October 2019. We filter out AmbER sets where the percentage gap in popularity between the head entity and the most popular tail entity is less than 10% to account for noise in the monthly page views. Collecting Distinguishing Properties We gather properties and associated values for each entity from Wikidata. We only retain properties that are in a specified list (Table 3), as they are useful for resolving ambiguity (Criteria 3). We also filter a property if two entities within an AmbER set have that property, ensuring that the remaining properties can be used to disambiguate between entities with the same name. These properties are used to instantiate the queries. Aligning Entities to Wikipedia We use the KILT Wikipedia snapshot (Petroni et al., 2021) as 3Aliases are all possible names for an entity. Entity Type Property (PID) Percent AmbER-H Human instrument (P1303) 17.01 movement (P135) 2.04 appears in (P1441) 0.08 killed by (P157) 0.19 PhD student (P185) 0.42 military branch (P241) 12.22 sports position (P413) 12.82 sports team (P54) 17.25 battles or wars (P607) 12.29 sport (P641) 25.68 AmbER-N Album performer (P175) 16.57 record label (P264) 7.11 tracklist (P658) 0.21 Business industry (P452) 0.65 City population (P1082) 0.24 Film cast member (P161) 27.14 screenwriter (P58) 18.28 Literary Work author (P50) 11.13 Musical Group record label (P264) 2.1 Song performer (P175) 4.42 record label (P264) 0.62 TV Series cast member (P161) 2.01 # seasons (P2437) 1.85 screenwriter (P58) 0.21 Written Work author (P50) 7.43 Table 3: Distinguishing Properties selected to create queries based on whether they are sufficient to resolve ambiguity. We provide the percent breakdown of how often each property occurs in each AmbER collection. the knowledge source for AmbER sets for better reproducibility. Each Wikipedia document in KILT has an associated QID. For each entity, we find all Wikipedia documents with that associated QID. After this alignment, we apply a round of filtering on the tuples. For each tuple, we check that the value of the tuple is within the first 350 tokens of the aligned Wikipedia article. If not, we remove 4476 AmbER-H AmbER-N # AmbER Sets 2,093 5,237 Averages per AmbER Set . . . # entities 2.98 2.42 . . . # entities w/ properties 2.03 2.06 . . . # properties 2.84 2.64 # Input Queries 23,768 55,216 . . . Question Answering (QA) 5,942 13,804 . . . Slot Filling (SF) 5,942 13,804 . . . Fact checking (FC) 11,884 27,608 Table 4: Statistics of AmbER collections. the tuple.4 Aligned Wikipedia articles that contain the tuple value serve as gold documents. Instantiating AmbER Instances Recall that our goal was to create AmbER sets for three tasks: fact checking, slot filling, and question answering. We are able to create queries for all three tasks simultaneously using the collected Wikidata tuples. For question answering and fact checking, we use templates based on properties to instantiate inputs. Three of the authors wrote a template each for each property for the two tasks. Duplicate templates are removed, resulting in an average of 3 question answering templates per property and 2.7 fact checking templates per property. See Appendix B for the complete list of templates. For slot filling, we create a single input from each Wikidata tuple by concatenating the AmbER set name with the property name, and using the value of the tuple as the answer. For question answering, we also create a single input for each tuple by filling in the template with the AmbER set name and using the value of the tuple as the answer. For fact checking, we create two inputs for each tuple, one claim that is true using the tuple value and one claim that is false. The false claim is created by finding the most popular value for the tuple property that does not match the tuple value5. 3.3 Dataset Statistics We provide statistics for AmbER sets in Table 4. On average, each AmbER set has about three entities that share the same name. Of these three entities, on average, only two have properties after filtering. In total, our AmbER sets contain about 80k task-specific input queries. 4This reduces the number of tuples for AmbER-H from 17,079 to 5,942 and for AmbER-N from 22,219 to 13,804. 5 The most popular instrument in Wikidata is piano. Therefore, given the true claim “Abe Lincoln played the trombone.”, the false claim would be “Abe Lincoln played the piano.”. 3.4 Limitations Since our pipeline is automated and relies on Wikipedia and Wikidata, there are a few limitations worth noting. AmbER sets will be affected by incompleteness of the knowledge source, sometimes resulting ambiguous queries if a property is missing from Wikidata, but answerable from Wikipedia text. For this reason, we only select a few properties for each type (Table 3). Second, even though we author multiple templates for each property, the reliance on these templates limits the syntactic diversity in the queries (not a critical concern, since we are only evaluating existing models). Also, we use Wikipedia page views as a proxy for real-world popularity of entities. Defining popularity in this way may be problematic, as page views for an entity can fluctuate, and may make our pipeline difficult to generalize to other knowledge sources, where this information may not be available. Several design choices in creating AmbER sets are worth further investigation. We limit AmbER sets to a pre-specified list of entity types and properties to ensure that entities in an AmbER set are distinguishable. This precludes other properties that may be useful in distinguishing entities, reducing the diversity in AmbER sets. Another design choice is we allow any alias in Wikidata to form an AmbER sets, however, not all aliases are canonical ways to refer to the entity. For instance, Shaquille O’Neal has the unusual alias “The Big Cactus”, potentially leading to a somewhat unrealistic query “What sport did The Big Cactus play?”. We plan to revisit the these design choices in future work. 4 Evaluation Setup Retrieval Systems The primary focus of this work is to evaluate entity ambiguity of retrieval systems. We consider four retrievers based on different retrieval paradigms. The first three are TF-IDF, a token-based retriever using sparse embeddings, DPR (Karpukhin et al., 2020), a dense embedding based retriever, and BLINK (Wu et al., 2020), a linker-based retriever which ranks documents based on input entities. These three retrievers have been thoroughly evaluated on a number of open-domain tasks in Petroni et al. (2021) with no obvious winner across tasks. Encouraged by the disambiguation success on rare entities by Orr et al. (2020), we also evaluate a retriever based on Bootleg, another entity linker. We provide additional details about these retrievers in Appendix D. 4477 Collection Retriever Fact Checking (FC) Slot Filling (SF) Question Answering (QA) All Head Tail ∀ All Head Tail ∀ All Head Tail ∀ AmbER-H TF-IDF 17.3 28.5 8.2 0.0 18.8 31.9 8.1 0.0 16.7 28.2 7.3 0.1 DPR 18.1 23.9 13.3 0.1 8.0 11.6 5.1 0.3 13.1 19.6 7.9 1.1 BLINK 55.9 64.4 49.0 5.6 38.2 57.0 22.9 11.5 31.7 40.5 24.6 6.6 Bootleg 34.8 43.0 28.2 0.7 56.5 63.9 50.6 25.3 67.2 77.1 59.1 36.1 AmbER-N TF-IDF 9.4 13.6 4.9 0.0 13.4 21.0 5.2 0.2 13.9 21.7 5.4 0.3 DPR 36.9 48.0 24.8 4.4 29.9 40.9 18.0 6.0 36.2 49.2 22.2 9.3 BLINK 11.7 13.9 9.4 0.0 5.7 7.3 3.9 0.7 35.2 44.7 24.9 10.1 Bootleg 3.5 4.6 2.4 0.0 52.3 61.3 42.5 22.4 59.8 69.5 49.3 29.0 Table 5: Top-1 retrieval results on each collection of AmbER sets. We report accuracy@1 results on all instances as well as results on instances about head entities and instances about tail entities. We also report a set-level metric, all correct (∀), the percentage of AmbER sets where all inputs had the correct document retrieved. FC SF QA Head Tail Head Tail Head Tail H* TF-IDF 19.5 67.5 28.2 75.7 27.9 76.1 DPR 1.2 10.0 2.3 23.8 2.6 27.0 BLINK 9.8 32.2 14.0 58.2 4.4 27.6 Bootleg 6.2 24.7 9.3 30.5 3.7 28.7 N* TF-IDF 10.1 49.9 22.0 76.9 23.0 76.8 DPR 6.2 32.2 9.1 48.3 8.7 44.0 BLINK 5.8 22.8 5.1 32.2 5.5 31.9 Bootleg 7.7 26.1 16.1 36.2 7.8 31.6 * H represents AmbER-H and N represents AmbER-N. Table 6: Entity confusion measures the % of queries the gold document ranks worse (lower) than a document for another entity with the same name (i.e., another entity in the AmbER set). Retrievers are four times as likely to exhibit this when dealing tail queries. Downstream Models The dominant approach to open-domain tasks is a two-stage process where a retriever first finds relevant documents, followed by a downstream model that processes these documents to produce an answer. We evaluate the end-to-end performance on AmbER sets by training downstream NLP models on our tasks of interest. For fact checking, we fine-tune a BERT classifier (Devlin et al., 2019) on FEVER (Thorne et al., 2018). For question answering, we fine-tune a RoBERTa model (Liu et al., 2019) on Natural Questions (Kwiatkowski et al., 2019). For slot filling, a generation task, we fine-tune a BART model (Lewis et al., 2020a) on T-Rex (Elsahar et al., 2018). We provide example training instances in Table 2 and additional details on the models in Appendix E. We use the AllenNLP and HuggingFace Transformers library to finetune our downstream models (Gardner et al., 2018; Wolf et al., 2020). 5 Results In this section, we evaluate existing open-domain NLP pipelines using AmbER sets. We also conduct Figure 4: Popularity Gap vs Retrieval Gap. We bin QA queries of pairs of head and tail entities based on the popularity gap between the entities. For each bin, we calculate the retrieval accuracy@1 difference on the head and tail queries. Larger popularity gaps tend to lead to a wider gaps in retrieval performance. The red line is retrievers’ performance gaps between head and tail queries on the entire collection. a user study to evaluate the quality of the queries in the AmbER sets. Top Document Retrieval We report retrieval performance in Table 5 in terms of retriever accuracy@1 (the % of instances where the first retrieved document is the gold document). For each task, we report values on the entire AmbER set (“All”), as well as instances corresponding only to “Head” entities or to “Tail” entities. We also report a metric we call all correct (∀), the fraction 4478 Task System Results All Head Tail H FC BERT (Oracle) 77.7 73.6 80.3 BERT + BLINK 59.8 60.1 57.7 SF BART (Oracle) 83.9 85.0 83.5 BART + BLINK 34.4 38.2 32.6 QA BERT (Oracle) 71.4 77.7 83.0 BERT + BLINK 27.5 33.8 22.3 N FC BERT (Oracle) 66.6 63.9 69.5 BERT + DPR 60.9 61.4 60.4 SF BART (Oracle) 82.1 80.1 84.3 BART + DPR 18.6 18.6 18.6 QA BERT (Oracle) 83.5 85.1 81.8 BERT + DPR 26.0 31.3 20.4 Table 7: End-to-end performance on AmbER sets. We evaluate systems in an oracle setting, where the gold document is provided, and a retrieval setting, where 20 documents are provided from a retriever. of AmbER sets in which all queries had the correct document retrieved. All retrievers do better on head entities compared to tail entities. Since BLINK, Bootleg, and DPR are initialized using pre-trained language models, they may have a predisposition towards being biased to more popular entities. However, we find TF-IDF also does better on head entities, perhaps because more popular entities have longer Wikipedia pages, possibly increasing term-frequency scores. Second, there are large discrepancies between a retriever’s performance on different tasks for an AmbER collection. For instance, DPR does substantially worse on slot filling compared to its performance on question answering. This is surprising since queries for all tasks are created from the same set of Wikidata tuples. Finally, we find that retrievers are mostly incorrect on getting all the queries in a set correct, with some receiving a ∀score of 0 on some tasks. Overall, we find that the Bootleg retriever on average does the best across tasks, however there is significant scope for improvement. Entity Confusion To explicitly evaluate whether retrievers get confused by entities in the same AmbER set, we compute entity confusion for retrievers defined as the percentage of queries where the retriever ranks a document for an incorrect entity from the same AmbER set over the gold document (Table 6). We find that across retrievers, tasks, and AmbER collections, entity confusion is twice as high for tail entity inputs. This result indicates that the popularity of an entity for a given name plays a significant role in retrieval performance. Effect of Popularity Gap Since the difference in popularity between the head and tail entities can vary considerably, these results obfuscate the effect of the size of the popularity gap. We explore how the gap in popularity between head and tail entities translates to the gaps in performance on their associated queries. For a head entity with popularity ph and a tail entity with popularity pt from the same AmbER set, we calculate popularity gap, ph−pt pt , and bin associated head/tail inputs based on the gap6. For each bin, we calculate the difference in accuracy@1 between the head and tail entity queries. Results for QA AmbER sets (Figure 4) show that there is a strong correlation between the popularity gap and the difference in performance. End to End Results We evaluate end to end performance in several evaluation settings with all results provided in Table 7. The metrics used are F1 for slot filling and question answering and accuracy for fact checking. In the “oracle” setting, we directly provide the downstream NLP model the gold document, and find that the gap between head entities and tail entities is fairly small. This suggests that in closed NLP settings, where the gold document is known, entity disambiguation is not a major concern. In the regular retrieval setting, we provide the model the top 20 documents as ranked by a retrieval system (BLINK and DPR), and find that retrievers still perform better on head entity queries (see Appendix A). The downstream systems that use retrieved documents display a noticeable gap in end-to-end performance between head and tail entity inputs. This is expected, as retrieval systems perform worse on tail entities. User Study AmbER sets are created in a largely automatic process, raising questions about data quality. To address these questions, we conduct a small user study on AmbER sets to evaluate whether the queries are resolvable by humans. We present a query from a QA AmbER set along with three documents for the entities from the same AmbER set, one of which is the gold document. We first ask the user to select the relevant document, then we ask the user to select an answer span from the selected document. In total, we asked 7 subjects to examine about 120 queries across AmbERH and AmbER-N, and computed their accuracy in 6Bin width of 20%. Queries with a popularity gap higher than 100% are binned into the highest bin. 4479 System AmbER-H AmbER-N Doc Acc. EM Doc Acc. EM TF-IDF 43.3 50.3 DPR 69.1 68.3 BLINK 69.1 74.1 Bootleg 79.6 73.1 BERT 71.8 75.5 Human 100 78.8 97.9 77.5 Table 8: User study on AmbER QA. Humans are nearly perfect in identifying the correct document for each query (Doc Acc), while existing retrievers frequently fail. When the gold document is provided to downstream NLP models (BERT), they do almost as well as humans in answering the question (EM). selecting the correct document and answer (Table 8). We also compare retrievers for this task, i.e. select from 3 documents for the same queries, and find that humans perform very well on the document selection task compared to retrievers on both sets. We also compare the accuracy of answer selection, and see that the closed domain NLP model (fine-tuned BERT) is as almost accurate as humans on the same set of queries7. This further confirms that closed NLP models are not the source of bias towards head entities, but the retrievers are. 6 Related Work Entity Ambiguity As previously mentioned, entity ambiguity is when a single name can match multiple entities in a knowledge source. Entity ambiguity has been most studied in the context of entity linking (Rao et al., 2013). To improve disambiguation, entity linkers have included auxiliary information such as entity types (Onoe and Durrett, 2020) and entity descriptions (Logeswaran et al., 2019). A recent thread of work aims to study how language models recall and leverage information about names and entities. Prabhakaran et al. (2019) shows that names can have a measurable effect on the prediction of sentiment analysis systems. Shwartz et al. (2020) demonstrates that pre-trained language models implicitly resolve entity ambiguity by grounding names to entities based on the pretraining corpus. The problem of entity ambiguity also appears implicitly in entity-centric tasks such as determining the semantic relatedness between entities (Hoffart et al., 2012) and entity-oriented 7The relatively low answer score is due to artifacts in using EM for QA evaluation, and is consistent with human performance on span selection (Rajpurkar et al., 2016)). search (Balog et al., 2010, 2011). We draw inspiration from these works by studying entity ambiguity in the context of open-domain NLP. Popularity Bias System’s that perform worse on the long-tail suffer from what is known as popularity bias. This problem has been studied extensively in the recommendation systems literature, where recommendation systems are known to often ignore the long-tail of products and instead recommend very popular items (Abdollahpouri et al., 2017; Chen et al., 2020). This has the effect of unfairly hurting users who would prefer these less-popular items (Abdollahpouri et al., 2019; Ciampaglia et al., 2018). We explore popularity bias from the angle of retrieval as opposed to recommendation, and find popularity bias exists in retrieval systems. Open-Domain Ambiguity Ambiguity is an inherent problem when it comes to open-domain reasoning. Min et al. (2020) showed that half of instances sampled from Natural Questions are ambiguous, with multiple correct answers. AmbER sets are similar in that the ambiguity is in terms of the entity in the query, however, in contrast to Natural Questions, AmbER set inputs have been constructed such that the ambiguity is resolvable. Challenge Sets There have been many evaluation sets specifically designed to assess a model’s ability to handle a specific phenomenon (Naik et al., 2018; Zhao et al., 2018; McCoy et al., 2019; Warstadt et al., 2020; Richardson et al., 2020; Jeretic et al., 2020; Ribeiro et al., 2019). Some of these challenge sets, similar to AmbER sets, use templates to generate a large amount of evaluation data quickly (Richardson et al., 2020; McCoy et al., 2019; Ribeiro et al., 2020). AmbER sets can be viewed as a challenge set for assessing opendomain systems’ ability to handle entity ambiguity. 7 Conclusion Entity ambiguity is an inherent problem in retrieval, as many entities can share a name. For evaluating disambiguation capabilities of retrievers, we introduce AmbER sets; an AmbER set is a collection of task-specific queries about entities that share a name, but the queries have sufficient content to resolve the correct entity. We create a broad range of AmbER sets, covering many entity types, with input queries for three open-domain NLP tasks: fact checking, slot filling, and question answering. Our experiments demonstrate the struggles of current 4480 retrievers in handling entity ambiguity. In particular, we find that the popularity of an entity in relation to other entities that share a name plays a significant role during disambiguation. For instance, we find that all tested retrievers are about twice as likely to retrieve erroneous documents when dealing with less popular entities than the most popular entity with the same name. Future goals include improving entity disambiguation capabilities of retrievers, perhaps more directly incorporating ideas from entity linking and coreference resolution. The AmbER sets and the code for the generation pipeline is available at https: //github.com/anthonywchen/AmbER-Sets. Acknowledgements We would like to thank Jo Daiber, Michael Tu, Russ Webb, Matt Gardner, Robert Logan, Sherry Tongshuang Wu, and the anonymous reviewers for providing valuable feedback for our work. This work is funded in part by the DARPA MCS program under Contract No. N660011924033 with the United States Office Of Naval Research. References Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher. 2017. Controlling popularity bias in learning-to-rank recommendation. In Proceedings of the Eleventh ACM Conference on Recommender Systems, RecSys 2017, Como, Italy, August 27-31, 2017, pages 42–46. ACM. Himan Abdollahpouri, Masoud Mansoury, Robin Burke, and Bamshad Mobasher. 2019. The unfairness of popularity bias in recommendation. arXiv preprint arXiv:1907.13286. K. Balog, Pavel Serdyukov, and Arjen P. de Vries. 2010. Overview of the trec 2010 entity track. In TREC. K. Balog, Pavel Serdyukov, and Arjen P. de Vries. 2011. Overview of the trec 2011 entity track. In TREC. Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer opendomain questions. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1870– 1879, Vancouver, Canada. Association for Computational Linguistics. J. Chen, Hande Dong, Xiao lei Wang, Fuli Feng, MingChieh Wang, and X. He. 2020. Bias and debias in recommender system: A survey and future directions. arXiv preprint arXiv:2010.03240. Giovanni Luca Ciampaglia, Azadeh Nematzadeh, Filippo Menczer, and Alessandro Flammini. 2018. How algorithmic popularity bias hinders or promotes quality. Scientific Reports, 8. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. Dheeru Dua, Yizhong Wang, Pradeep Dasigi, Gabriel Stanovsky, Sameer Singh, and Matt Gardner. 2019. DROP: A reading comprehension benchmark requiring discrete reasoning over paragraphs. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 2368–2378, Minneapolis, Minnesota. Association for Computational Linguistics. Hady Elsahar, Pavlos Vougiouklis, Arslen Remaci, Christophe Gravier, Jonathon Hare, Frederique Laforest, and Elena Simperl. 2018. T-REx: A large scale alignment of natural language with knowledge base triples. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke Zettlemoyer. 2018. AllenNLP: A deep semantic natural language processing platform. In Proceedings of Workshop for NLP Open Source Software (NLP-OSS), pages 1– 6, Melbourne, Australia. Association for Computational Linguistics. Johannes Hoffart, Stephan Seufert, Dat Ba Nguyen, Martin Theobald, and Gerhard Weikum. 2012. KORE: keyphrase overlap relatedness for entity disambiguation. In 21st ACM International Conference on Information and Knowledge Management, CIKM’12, Maui, HI, USA, October 29 - November 02, 2012, pages 545–554. ACM. Johannes Hoffart, Mohamed Amir Yosef, Ilaria Bordino, Hagen F¨urstenau, Manfred Pinkal, Marc Spaniol, Bilyana Taneva, Stefan Thater, and Gerhard Weikum. 2011. Robust disambiguation of named entities in text. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 782–792, Edinburgh, Scotland, UK. Association for Computational Linguistics. Paloma Jeretic, Alex Warstadt, Suvrat Bhooshan, and Adina Williams. 2020. Are natural language inference models IMPPRESsive? Learning IMPlicature 4481 and PRESupposition. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8690–8705, Online. Association for Computational Linguistics. Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769– 6781, Online. Association for Computational Linguistics. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452–466. Kenton Lee, Ming-Wei Chang, and Kristina Toutanova. 2019. Latent retrieval for weakly supervised open domain question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6086–6096, Florence, Italy. Association for Computational Linguistics. Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020a. BART: Denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871–7880, Online. Association for Computational Linguistics. Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, Sebastian Riedel, and Douwe Kiela. 2020b. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3449–3460, Florence, Italy. Association for Computational Linguistics. Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3428–3448, Florence, Italy. Association for Computational Linguistics. Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2020. AmbigQA: Answering ambiguous open-domain questions. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5783– 5797, Online. Association for Computational Linguistics. Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. Stress test evaluation for natural language inference. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2340–2353, Santa Fe, New Mexico, USA. Association for Computational Linguistics. Yasumasa Onoe and Greg Durrett. 2020. Fine-grained entity typing for domain independent entity linking. In AAAI. Laurel Orr, Megan Leszczynski, Simran Arora, Sen Wu, Neel Guha, Xiao Ling, and Christopher R´e. 2020. Bootleg: Chasing the tail with self-supervised named entity disambiguation. arXiv preprint arXiv:2010.10363. Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Rockt¨aschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523–2544, Online. Association for Computational Linguistics. Vinodkumar Prabhakaran, Ben Hutchinson, and Margaret Mitchell. 2019. Perturbation sensitivity analysis to detect unintended model biases. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 5740–5745, Hong Kong, China. Association for Computational Linguistics. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392, Austin, Texas. Association for Computational Linguistics. Delip Rao, Paul McNamee, and Mark Dredze. 2013. Entity linking: Finding extracted entities in a knowledge base. In Multi-source, Multilingual Information Extraction and Summarization. 4482 Marco Tulio Ribeiro, Carlos Guestrin, and Sameer Singh. 2019. Are red roses red? evaluating consistency of question-answering models. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6174–6184, Florence, Italy. Association for Computational Linguistics. Marco Tulio Ribeiro, Tongshuang Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond accuracy: Behavioral testing of NLP models with CheckList. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4902– 4912, Online. Association for Computational Linguistics. Kyle Richardson, H. Hu, L. Moss, and A. Sabharwal. 2020. Probing natural language inference models through semantic fragments. In AAAI. Ozge Sevgili, Artem Shelmanov, Mikhail V. Arkhipov, Alexander Panchenko, and Christian Biemann. 2020. Neural entity linking: A survey of models based on deep learning. arXiv preprint arXiv:2006.00575. Vered Shwartz, Rachel Rudinger, and Oyvind Tafjord. 2020. “you are grounded!”: Latent name artifacts in pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6850–6861, Online. Association for Computational Linguistics. James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819, New Orleans, Louisiana. Association for Computational Linguistics. Denny Vrandecic and M. Kr¨otzsch. 2014. Wikidata: a free collaborative knowledgebase. Commun. ACM, 57:78–85. Alex Warstadt, Alicia Parrish, Haokun Liu, Anhad Mohananey, Wei Peng, Sheng-Fu Wang, and Samuel R. Bowman. 2020. BLiMP: The benchmark of linguistic minimal pairs for English. Transactions of the Association for Computational Linguistics, 8:377–392. Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38–45. Association for Computational Linguistics. Ledell Yu Wu, F. Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Zero-shot entity linking with dense entity retrieval. In EMNLP. Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 15–20, New Orleans, Louisiana. Association for Computational Linguistics. 4483 Appendix A Top-20 Retrieval Results We provide results for top-20 retrieval in Table 9. Top-20 retrieval is used for providing documents in the end-to-end evaluation setting. In this setting, retrieval accuracy measures whether a gold document appears in one of the top-20 retrieved documents. Similar to top-1 retrieval, retrievers continue to perform better on head queries. B Task Specific Templates Table 10 contains the templates used to instantiate the task-specific inputs. Templates were written on a per-property basis. We note that many of the properties share templates that are very similar. C Computational Resources All experiments (e.g., training baselines, generating AmbER sets, etc.) were conducted on a machine with 500 GB of RAM, 64 CPUs, and using an NVIDIA TitanRTX with 24 GB of RAM. Retrieval on a collection of AmbER sets takes about 12 hours for the most time-consuming retriever, BLINK. Training a downstream model takes roughly 5 hours and inference on a collection of AmbER sets takes less than 30 minutes. D Retriever Details For BLINK, DPR, and TF-IDF, we use the retriever code in the KILT repository released by Facebook8. For Bootleg, we use the code provided by the Hazy Research group9. E Downstream Model Details For question answering, we train a RoBERTa-Large model on Natural Questions. We use the negative documents in Natural Questions to train a “noanswer” classifier using the [CLS] token. During inference, we take the highest-scoring span where the answer is not classified as “no-answer”. For slot filling, we train a BART-base model. For each slot filling instance, we train with the top non-gold document retrieved by TF-IDF as a negative document. For this negative document, we train the model to generate a “none” token, and during inference, we take the highest scoring answer that is 8https://github.com/facebookresearch/ KILT 9https://github.com/HazyResearch/ bootleg not “none”. For fact checking, we train a three-way (i.e., SUPPORTS, REFUTES, NEUTRAL) BERTbase classifier. Similar to slot filling, we train with the top non-gold document retrieved by TF-IDF as a negative document and train the model to classify this negative document as NEUTRAL. During inference, we take the highest scoring prediction that is not NEUTRAL. When training baselines models, we do not tune over hyperparameters and train with a batch size of 32 for 3 epochs. 4484 Collection Retriever Fact Checking Slot Filling Question Answering All Head Tail ∀ All Head Tail ∀ All Head Tail ∀ AmbER-H TF-IDF 65.8 78.5 55.4 26.7 72.0 83.5 62.5 55.6 72.6 82.0 64.8 55.9 DPR 39.8 51.0 30.6 4.1 26.6 37.0 18.1 6.8 36.1 49.3 25.3 9.6 BLINK 78.6 82.0 76.0 43.8 73.3 73.9 72.8 64.6 58.8 60.3 57.5 32.2 Bootleg 96.5 97.6 95.6 93.2 96.6 97.7 95.7 93.6 96.5 97.6 95.6 93.5 AmbER-N TF-IDF 50.8 57.0 44.1 12.0 46.8 53.4 39.7 35.3 52.0 59.1 44.4 40.7 DPR 62.3 75.8 47.7 27.8 57.3 71.4 42.0 29.4 63.4 77.9 47.8 37.2 BLINK 33.5 38.7 27.9 1.3 18.2 21.5 14.6 5.8 74.7 80.6 68.3 53.0 Bootleg 79.3 80.2 78.4 61.5 89.6 91.9 87.1 85.3 83.8 83.6 84.1 71.1 Table 9: Top-20 retrieval results measuring retrieval accuracy and ∀. 4485 Property Question Answering Template Fact Checking Template AmbER-H instrument Which musical instrument did $name play? What musical instrument does $name play? What instrument does $name play? $name plays the $object. $name plays the musical instrument $object. The $object is played by $name. movement What movement did $name participate in? Which movement is $name associated with? What movement is $name associated with? $name was a member of the $object movement. $name participated in the $object movement. $name was a part of the $object movement. appears in What works does the fictional entity $name appear in? What work is the character $name present in? Which work was the character $name in? $name is a character in $object. $name is a fictional character in $object. $object features the fictional character $name. doctoral student Who were the doctoral students of $name? Who are $name’s doctoral students? Who did $name advise? $name has a doctoral student named $object. $name’s doctoral student is $object. $name advised their student $object. military branch What branch of the military does $name belong to? Which military branch does $name belong to? What military branch is $name affiliated with? $name is a member of the $object. $name belongs to the military branch $object. $name belongs to the $object branch of the military. sports position What is the position that $name plays? What position does $name play? Which position does $name play? $name plays the $object position. $name plays as a $object. sports team $name plays for which team? What team does $name play for? Which team does $name play for? $name is a player on the $object. $name plays for the $object team. $name plays for the $object. battles or wars What were the wars that $name participated in? Which battle did $name fight in? Which war did $name fight? $name fought in the $object. $name fought in $object. sport Which sport does $name participate in? Which sport does $name play? What sport does $name play? $name plays $object. $name plays the sport $object. AmbER-N performer Who performs $name? Who is the performer of $name? Who performed $name? $object performs in $name. $object is the performer of $name . $name was performed by $object. record label What is the record label of $name.? What is the record label for $name? $name belongs to which record label? $object is the record label for $name. $name’s record label is $object. tracklist What song appears in the album $name? What song appears on $name? What are the tracks in $name? $name belongs to $object tracklist. $object is on the release of $name . $object is a song in the $name tracklist. industry Which industry is $name in? In what industry is $name? What is $name’s industry? $name is in the industry of $object. The company $name is in the $object industry. $name’s industry is $object. population What is the total population of $name? What is the population of $name? How many people live in $name? The population of $name is $object. $name’s population is $object. $name has a population of $object. cast member Who acted in $name? Who is a cast member on $name? Who starred in $name? $object was a cast member in $name. $object appeared in $name. $object acted in $name. screenwriter Who was the screenwriter for $name? Who was screenwriter for $name? Who is $name’s screenwriter? $name’s screenwriter is $object. $object wrote the screenplay of $name. $object screenwrote $name. # seasons How many seasons are there in $name? How many seasons does $name have? How many seasons were there in $name? There were $object seasons in $name. $name has $object seasons. author Who is the author of $name? Who wrote $name? Who authored $name? $name wrote $object. $name is written by $object. $object authored $name. Table 10: Templates used to instantiate the task-specific inputs.
|
2021
|
345
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.