id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_15500 | NN-ML-CL extracted sentences that mention the good revenue performance of air-conditioners in Asia and Europe, which is the same as that in the gold summary. | nn-SE extracts sentences mentioning the bad revenue performance of fluorine resin and medical equipment, which are not described in the gold summary. | contrasting |
train_15501 | To overcome this, several studies have used artificial reference summaries (Sun et al., 2005;Svore et al., 2007;Woodsend and Lapata, 2010;Cheng and Lapata, 2016) compiled by collecting documents and corresponding highlights from other sources. | preparing such a parallel corpus often requires domain-specific or expert knowledge depending on the domain (Filippova et al., 2009;Parveen et al., 2016). | contrasting |
train_15502 | We first crawled 18121 English Wikinews and their source news articles via the associated URLs. | many Wikinews articles have very few source news articles and they are very short, and moreover, the URLs for many of the source news are out of date. | contrasting |
train_15503 | One common practice (Lee et al., 2012) to approach CD coreference task is to resolve event coreference in a megadocument created by concatenating topic-relevant documents, which essentially does not distinguish WD and CD event links. | intuitively, recognizing CD coreferent event pairs requires stricter evidence compared to WD event linking because it is riskier to link two event mentions from two distinct documents rather than the same document. | contrasting |
train_15504 | Like this work, several studies have considered both WD and CD event coreference resolution task together. | to simplify the problem, they (Lee et al., 2012;Harabagiu, 2010, 2014) created a meta-document by concatenating topic-relevant documents and treated both as an identical task. | contrasting |
train_15505 | However, only very small weights were assigned to the similarity and distance scores calculated using context embeddings. | in the classifier trained with cross-doc coreferent event mention pairs, the highest weight was assigned to the cosine similarity score calculated using context embeddings of two event mentions. | contrasting |
train_15506 | In neural text generation such as neural machine translation, summarization, and image captioning, beam search is widely used to improve the output text quality. | in the neural generation setting, hypotheses can finish in different steps, which makes it difficult to decide when to end beam search to ensure optimality. | contrasting |
train_15507 | We therefore propose a novel and simple beam search variant that will always return the optimalscore complete hypothesis (modulo beam size), and finish as soon as the optimality is established. | another well-known problem remains, that the generated sentences are often too short, compared to previous paradigms such as SMT (Shen et al., 2016). | contrasting |
train_15508 | Combining these two Theorems, it is interesting to note that our method is not just optimal but also faster. | optimal-score hypothesis, though satisfying in theory, is not ideal in practice, since neural models are notoriously bad in producing very short sentences, as opposed to older paradigms such as SMT (Shen et al., 2016). | contrasting |
train_15509 | In an extreme case, an agent may be micro-managed by a human user who uses the neural model to enforce grammar and style (e.g., a level of politeness, or a type of humor), while driving the content directly (e.g., by expressing In this manner, the neural model becomes an authoring tool, rather than an independent chat-bot. | in fully automated agent systems, the agent may be influenced by a knowledge database, or some other artificial information system, while running in a pre-set style or a style deemed best based on the course of the conversation. | contrasting |
train_15510 | Here, we follow a similar process as in (Wen et al., 2015) which generates multiple target hypotheses with stochastic sampling based on p(T |S), and then ranks them with the objective function 2 above. | as also observed by (Shao et al., 2017), step-by-step naive sampling can accumulate errors as the sequence gets longer. | contrasting |
train_15511 | This makes the model gradually adapt to the noisiness of the predicted label. | this method still relies upon a single current label, and, by omitting the distribution over the possible labels, this model loses information about the current stage. | contrasting |
train_15512 | However, this method still relies upon a single current label, and, by omitting the distribution over the possible labels, this model loses information about the current stage. | we propose to condition the next label on a predicted distribution of the current label. | contrasting |
train_15513 | (2016b) use keyword retrieval confidence as a reward. | it is widely acknowledged that manually defined reward functions can't possibly cover all crucial aspects and can lead to suboptimal generated utterances. | contrasting |
train_15514 | Training is formalized as a game in which the generative model is trained to generate outputs to fool the discriminator; the technique has been successfully applied to image generation. | to the best of our knowledge, this idea has not achieved comparable success in NLP. | contrasting |
train_15515 | This is similar to our work in that we also evaluate using the previous utterances, and the predicted DAs for them. | our modeling approaches are all based on DNNs, as de-scribed in more details in Section 3, and the interaction between utterances and DA labels is modeled in the hierarchical models in a more principled way. | contrasting |
train_15516 | (2011)'s model, individual reply structure paths from the first utterance to terminal utterances are teased apart into separate sequential conversations by duplicating utterances. | this method counts the same utterance multiple times and requires an aggregation method for making a final decision of the DA for each utterance. | contrasting |
train_15517 | The baseline models show a large variance in performance depending on the characteristics of the corpus. | our model has a low variance between the corpora, because the content word filtering, distinction between utterance-level and sentence-level DAs, and speaker preferences complement one another to adapt to different corpora. | contrasting |
train_15518 | One of the prevailing methods to build the system is using the generative Sequence-to-Sequence (Seq2Seq) model through neural networks. | the standard Seq2Seq model is prone to generate trivial responses. | contrasting |
train_15519 | (2017) developed several teaching strategies answering "how" the human teacher guide the learning process. | those previous teaching methods exclude "when" to teach from concern. | contrasting |
train_15520 | By investigating the real work mode in a call center, this framework makes a reasonable assumption that human teacher has access to the extracted dialogue states from the dialogue state tracker as well as the system's dialogue act, and can also reply in the same format. | there are two major problems in the previous framework. | contrasting |
train_15521 | In addition to early teaching, the teaching heuristics can be broadly divided into two categories: teacher-initiated heuristics and studentinitiated heuristics (Amir et al., 2016). | the teacher-initiated approaches require the constant long-term attention of the teacher to monitor the dialogue process (Torrey and Taylor, 2013;Amir et al., 2016), which is costly and impractical for real applications. | contrasting |
train_15522 | A straight way is to set a ratio threshold α, and consider it to be failure prognosis when T (s t ) < αR succ . | this assumes that the numerical scale of Q succ is consistent through the training period, which is not always the case. | contrasting |
train_15523 | Therefore, its RI should equal to zero. | if the success rate over a training process rises and falls and sometimes is below the threshold, it is risky. | contrasting |
train_15524 | In the seq2seq approach, the decoder network therefore only has to keep track of where it is in the output, and the content to generate can be transformed from the relevant parts in the source via the attention mechanism . | in conversation response generation, the prompt turn may be short and general (e.g., "what do you have planned tonight"), while an appropriate response may be long and informative. | contrasting |
train_15525 | This algorithm behaves similarly to standard beam search when the categorical distribution used during the process is sharp ('peaked'), since the samples are likely to be the top categories (words) . | when the distribution is smooth, many of the choices are likely. | contrasting |
train_15526 | In the case of reranking whole target sequences y, this becomes the marginal P (y), which corresponds to the same diversitypromoting objective used in (Li et al., 2015). | we found that our approximation works better in terms of N-choose-1 accuracy (see Section 5.2), which suggests that its value may be closer to the true conditional probability. | contrasting |
train_15527 | Our evaluation set is therefore not from the same distribution as our training set. | since our goal is to produce good general conversation responses, we found it to be a good general purpose evaluation set. | contrasting |
train_15528 | It is clear enough that this model progresses much slower, so we terminated it early. | it is surprising that the glimpse model with K = 10 progresses faster than the baseline model with only source-side attention, because the model is trained on examples with decoder-length fixed at 10, while the average response length is 38 in our data set. | contrasting |
train_15529 | There is a growing demand for intelligent personal assistants, mainly in the form of dialogue agents, that can help users accomplish tasks ranging from meeting scheduling to vacation planning. | most of the popular agents in today's market, such as Amazon Echo, Apple Siri, Google Home and Microsoft Cortana, can only handle very simple tasks, such as reporting weather and requesting songs. | contrasting |
train_15530 | A related but different extension to singledomain dialogues is multi-domain dialogues, where each domain is handled by a separate agent (Lison, 2011;Gasic et al., 2015a,b;Cuayáhuitl et al., 2016). | to compositedomain dialogues studied in this paper, a conversation in a multi-domain dialogue normally involves one domain, so completion of a task does not require solving sub-tasks in different domains. | contrasting |
train_15531 | We consider this paper to be the most complete study to date, across metrics, systems, datasets and domains, focusing on recent advances in data-driven NLG. | to previous work, we are the first to: • Target end-to-end data-driven NLG, where we compare 3 different approaches. | contrasting |
train_15532 | ness for the TGEN system on BAGEL (ρ = 0.33, p < 0.01, see Figure 1). | the wps metric (amongst most others) is not robust across systems and datasets: Its correlation on other datasets is very weak, (ρ ≤ .18) and its correlation with in-formativeness ratings of LOLS outputs is insignificant. | contrasting |
train_15533 | These results are quite promising, and suggest that neural models are a good fit for text generation. | the statistics of these datasets, shown in Table 1, indicate that these datasets use relatively simple language and record structure. | contrasting |
train_15534 | (2016) introduced the WIKIBIO dataset, which is at least an order of magnitude larger in terms of number of tokens and record types. | as shown in Table 1, this dataset too only contains short (single-sentence) generations, and relatively few records per generation. | contrasting |
train_15535 | (2015) to parameterize this probability; full details are given in the Appendix. | importantly, we note that the (s, y 1:T ) pairs typically used for training data-to-document systems are also sufficient for training the information extraction model presented above, since we can obtain (partial) supervision by simply checking whether a candidate record lexically matches a record in s. 1 since there may be multiple records r ∈ s with the same e and m but with different types r.t, we will not always be able to determine the type of a given entity-value pair found in the text. | contrasting |
train_15536 | (Hidayat, 2012) noted that in Facebook, users mostly preferred inter-sentential mixing and showed that 45% of the mixing originated from real lexical needs, 40% was used for conversations on a particular topic and the rest 5% for content clarification. | (Das and Gambäck, 2014) showed that in case of Facebook messages, intra-sentential mixing accounted for more than half of the cases while inter-sentential mixing accounted only for about one-third of the cases. | contrasting |
train_15537 | We also have a significant amount of unsupervised status update data (100k+ users). | the supervised datasets which have the SUD ground truth are pretty small, ranging from 896 for the intersection of the likes, status updates and SUD (LikeStatusSUD in Table 1) to 3508, which is the intersection of the likes and SUD (LikesSUD in Table 1). | contrasting |
train_15538 | 0.69 0.64 0.67 Table 8: Results obtained on the test set for the source identification task Most of the errors of the algorithm are due to information sources not recognized as NEs (in particular, when the source is a Twitter user), or NEs that are linked to the wrong DBpedia page. | in order to draw more interesting conclusions on the most suitable methods to address this task, we would need the increase the size of the dataset. | contrasting |
train_15539 | They first detect argumentative sentences, and second identify premises and claims. | none of them is neither interested in distinguishing facts from opinions nor to identify the arguments' sources. | contrasting |
train_15540 | In recent studies, embeddings v IN are usually used for measuring the similarity between words. | given the characteristics described in the previous paragraph and SGNS's equivalence with shifted positive pointwise mutual information (Levy and Goldberg, 2014), if we want to measure to what extent word w t tends to co-occur with w k in the training data, then we should use the similarity of In this study, we show the importance of using v OU T in a task where we need to see if a word matches its context. | contrasting |
train_15541 | Hence, this is a decaying weighting. | with uniform weights, we set α w j to be 1 for all w j in the context. | contrasting |
train_15542 | Close in spirit to our investigation, Schofield and Mehr (2016) train a number of classifiers over movie scripts for determining the gender of individual (and pairs) of speakers as well as the expected length of their relationships. | we focus on understanding how the gender of a given character implicitly relates to features that track their control over their own path (agency) and the world around them (power). | contrasting |
train_15543 | Advances in Natural Language Processing are leading to a point when text generation methods are deployed at scale. | in the quest to make these applications more likable, effective and hence more usable, these methods should consider a way to adapt themselves to the person or type of persons they are interacting with (Bates, 1994;Loyall and Bates, 1997) e.g., a student may learn better from a tutoring agent that expresses similar traits to himself (Baylor and Kim, 2004). | contrasting |
train_15544 | For instance in (Quinn et al., 2010), a topic model for legislative speech is defined. | those works study topics one at a time whereas a set of co-referenced topics is more relevant since it constitutes the core of a candidate's political program. | contrasting |
train_15545 | Following the tradition of sentiment analysis, many have proposed methods to automatically assess the quality of recommendations or comments based on subjective ratings of their usefulness (Liu et al., 2007;Siersdorfer et al., 2010;Becker et al., 2012;Momeni et al., 2013) or of persuasiveness (Wei et al., 2016). | information thought to be useful does not always prove so, and subjective ratings may be driven by biases. | contrasting |
train_15546 | MOOF forecasters were also linked directly to the form to update their forecast from the comment to minimize outside influences between the reading of the comment and prediction update. | as the forecasters were not in a laboratory, other Web browsing behavior in other tabs could not be controlled. | contrasting |
train_15547 | This model is then used to predict the quality of the comments in the 1/10th sample and compared to the true quality for those comments (using Pearson correlation in this case). | many of our scores for change in forecaster accuracy (benefit) are based simply on one change and thus quite unreliable. | contrasting |
train_15548 | Most studies listed thus far found length of comment to be the dominant predictor, with other features providing minimal benefit. | a few studies (including our own) have found this baseline can be overcome. | contrasting |
train_15549 | Extracting irregularly-formed bursty phrases as described in the previous paragraph is difficult since no restriction can be used anymore to filter out incomplete N-grams. | they are rare and little influence the overall accuracy even if they are correctly extracted. | contrasting |
train_15550 | In the field of trend detection on microblogs, this ignoring-minority strategy has become a de facto standard. | it always fails to extract irregularly-formed bursty phrases. | contrasting |
train_15551 | Owing to a good segmentation algorithm, it can potentially detect bursty phrases other than noun phrases, uni-grams, bi-grams, and tri-grams with high accuracy. | it is still possible to miss irregularly-formed bursty phrases because they are likely to be segmented incorrectly. | contrasting |
train_15552 | Among comparative methods, the segmentation-based method best achieved the min-z-score since it did not restrict the form of phrases. | it was still possible to miss very irregular phrases due to segmentation mistakes and the min-z-score was less than that of the proposed method. | contrasting |
train_15553 | Neural networks have achieved state-ofthe-art performance on several structuredoutput prediction tasks, trained in a fully supervised fashion. | annotated examples in structured domains are often costly to obtain, which thus limits the applications of neural networks. | contrasting |
train_15554 | The explicit supervision signals can be viewed as a source of immediate rewards, as we can often instantly know the correctness of the current action. | the implicit supervision can be viewed as a source of delayed rewards, where the reward of the actions can only be revealed later. | contrasting |
train_15555 | In , the authors demonstrated that labeling semantic parses is possible and often more effective with a sophisticated labeling interface. | collecting answers may still be easier or faster for certain problems or annotators. | contrasting |
train_15556 | Natural language processing research has largely adopted the outlined hierarchical models for mining arguments from text (Stab and Gurevych, 2014;Habernal and Gurevych, 2015;Peldszus and Stede, 2016). | the adequacy of the resulting overall structure for downstream analysis tasks of computational argumentation has rarely been evaluated (see Section 2 for details). | contrasting |
train_15557 | Recently, we generalized the model in order to make flows applicable to any type of information relevant for argumentation-related analysis tasks . | flows capture only sequential structure, whereas here we also model the hierarchical structure of overall argumentation. | contrasting |
train_15558 | To this end, we represent labels and positions as follows: Labels The tree kernel approaches in natural language processing discussed in Section 2 include text (usually words) in the leaf nodes. | we label each node v ∈ V with the type of the associated argument unit only. | contrasting |
train_15559 | In fact, we found that larger ηs do not delay most training instances in the first few iterations. | once the network obtains a reasonably high performance, schedulers start delaying instances for longer durations. | contrasting |
train_15560 | Most of them are based on the co-occurrence information of words and their contexts. | it is still an open question what is the best definition of context. | contrasting |
train_15561 | Furthermore, we systematically evaluate three word embedding models: CSG, CBOW and GLoVe. | c is thus odel; cond the numally larger generalize ords conpendencyntexts cap--word conence "Ausscope". | contrasting |
train_15562 | For example, in Figure 1, the context "scientist/nsubj" can only be predicted by word "discovers". | most of the word is connected to several contextual words. | contrasting |
train_15563 | In contrast, linear context type captures topical similarity with the "help" of unbound representation. | the above findings come with a major caveat: a lot seems to depend on the particular dataset, in addition to the model and context type. | contrasting |
train_15564 | The performance of unbound linear context and unbound DEPS context is similar. | for most models and categories, bound representation seems to outperform unbound representation. | contrasting |
train_15565 | Their system thus requires chunker or constituent parser. | we investigate the usefulness of syntactic information derived from dependency parses, and we extend their work in also comparing our results to the use of only POS tags and words. | contrasting |
train_15566 | As done in that study, we do sequence prediction using a neural network. | we extend their work significantly in reporting results for intra-sentential segmentation, in comparing more settings concerning the availability of information (tokenisation, POS tags), and in including syntactic information into our systems. | contrasting |
train_15567 | First, all previous systems were evaluated on the same set of 38 documents that initially contains 991 sentences -and more precisely on each sentence of this set for intra-sentential results. | soricut and Marcu (2003) do not consider sentences that are not exactly spanned by a discourse subtree (keeping only 941 sentences in the test set), and sporleder and Lapata (2005) only keep the sentences that contain intra-sentential EDUs (608 sentences). | contrasting |
train_15568 | The example actions not only make the learning safer but also can be directly used by the training of the student policy. | there are costs to the teaching of a human teacher. | contrasting |
train_15569 | dropout can be observed as claimed in (Gal and Ghahramani, 2016). | dropout dQN 1 seems to suffer premature and sub-optimal convergence, while our proposed dropout dQN 32, whose decision is based on multi votes (algorithm 1), can result in improvement of efficiency and better final performance. | contrasting |
train_15570 | (2016), which also makes predictions on a dataset derived from the IQ2 debates. | their work analyses speech signals, as opposed to textual data. | contrasting |
train_15571 | Considering applause as a sign of endorsement is not controversial, but laughter could be viewed as more ambiguous. | consider the audience of the debates: the debates air on the Bloomberg network and National Public Radio, suggesting a higher level of maturity of the audience, which is less likely to laugh at the participants, rather than at their jokes. | contrasting |
train_15572 | 3 Although the level of statistical significance is reported to be 99%, CIs in Figure 1(a) show that the proportion of each bootstrap distribution was substantially underestimated leading to overly narrow CI limits for both SAME and DIFF. | figure 1(b) shows CIs resulting from an accurately computed proportion of 95% of the same bootstrap distribution, where even at the lower level of 95% significance (as opposed to 99%) CIs for SAME and DIff now overlap, reversing the conclusion of strong reference bias. | contrasting |
train_15573 | These sentences are designed to dive deep into linguistic phenomena of interest, and to provide a much finer-grained analysis of the strengths and weaknesses of existing technologies, including NMT systems. | this strategy also necessitates that we work on fewer sentences. | contrasting |
train_15574 | Leveraging zero-shot learning to learn mapping functions between vector spaces of different languages is a promising approach to bilingual dictionary induction. | methods using this approach have not yet achieved high accuracy on the task. | contrasting |
train_15575 | to advances in methods for extracting general purpose knowledge Nakashole et al., 2013;Wijaya et al., 2014), the use of semantic knowledge has been explored for several natural language tasks (Nakashole and Mitchell, 2015;Yang and Mitchell, 2017). | for bilingual dictionary induction, and more generally, machine translation, the role of semantic knowledge has not been fully explored. | contrasting |
train_15576 | This is partly due to a better prediction of non-TQs, represented by the high corresponding F-scores for CL-SEQ for all three language pairs. | it is also linked to a better labelling of grammatical and lexical TQs, which can be seen by the high labelling precision (P*) in the context of high recall (R). | contrasting |
train_15577 | The matrix reveals that for predicted lexical tags and non-TQs, the majority were correctly classed into these coarse-grained classes. | grammatical tags proved more difficult to predict, the majority being classed as non-TQs, most likely a result of the fact that no such tag question was present on the German source side. | contrasting |
train_15578 | (2008) and Naim and Gildea (2015) propose models that can use orthographic similarities. | the model proposed by (Naim and Gildea, 2015) is only capable of producing a parallel lexicon and not translation. | contrasting |
train_15579 | This shows that our system is able to exploit regular sound correspondences to filter out a substantial number of false cognates, such as lexical borrowings or chance resemblances. | the overall contribution of the specific models is relatively small. | contrasting |
train_15580 | For example, "sickness" is difficult for our general model to associate with "bitterness, pain." | there are many instances where our system is successful in identifying non-obvious semantic similarity, often thanks to the word vector features of our model. | contrasting |
train_15581 | There are aspects of cognate identification that can only be detected by human experts, such as cognates that have undergone extensive phonetic and semantic changes, or large-scale lexical borrowing between languages. | we believe that our system represents a step towards automated cognate identification, and will prove a useful tool for historical linguists. | contrasting |
train_15582 | Uyghur is not supported by Google Translate, and the Uyghur Wikipedia has less than 3,000 articles. | the smallest Wikipedia size language in our test set is Yoruba, with 30K articles. | contrasting |
train_15583 | This benefit for English NLP has motivated the development of VerbNets for languages such as Spanish and Catalan (Aparicio et al., 2008), Czech (Pala and Horák, 2008), and Mandarin (Liu and Chiang, 2008). | end-to-end manual resource development using Levin's methodology is extremely time consuming, even when supported by translations of English VerbNet classes to other languages (Sun et al., 2010;Scarton et al., 2014). | contrasting |
train_15584 | Approaches which aim to learn verb classes automatically offer an attractive alternative. | existing methods rely on carefully engineered features that are extracted using sophisticated language-specific resources (Joanis et al., 2008;Sun et al., 2010;Falk et al., 2012, i.a. | contrasting |
train_15585 | K-Means is outperformed for each target language, confirming the superiority of spectral clustering established in prior work, e.g., (Scarton et al., 2014). | we find results with another clustering algorithm, hierarchical agglomerative clustering with Ward's linkage (Ward, 1963), on par with spectral clustering (1.4 points on average in favour of spectral, which is better on 4 out of 6 languages). | contrasting |
train_15586 | They clearly indicate that cross-lingual synonymy constraints are useful for both relationship types (compare the scores with XLing), with strong gains over the nonspecialised distributional space. | the inclusion of VerbNet information, while boosting classification scores for target languages and (trivially) for EN, deteriorates EN similarity scores across the board (compare XLing+VN against XLing in Tab. | contrasting |
train_15587 | This work has proven the potential of transferring lexical resources from resource-rich to resourcepoor languages using general-purpose cross-lingual dictionaries and bilingual vector spaces as means of transfer within a semantic specialisation frame-work. | we believe that the proposed basic framework may be upgraded and extended across several research paths in future work. | contrasting |
train_15588 | "Swimming in the lake" has no natural endpoint, as also shown by the linguistic test presented in (2). | "swimming across the lake" will necessarily be finished once the other side is reached. | contrasting |
train_15589 | Variance reduction by additive control variates has implicitly been used in doubly robust techniques (Dudik et al., 2011;Jiang and Li, 2016). | the connection to Monte Carlo techniques has not been made explicit until Thomas and Brunskill (2016), nor has the control variate technique of optimizing the variance reduction by adjusting a linear interpolation scalar (Ross, 2013) been applied in off-policy learning. | contrasting |
train_15590 | (2014), containing 1,391 positive pairs and 4,294 negative pairs. | the number of positive pairs is not sufficient for our propose. | contrasting |
train_15591 | As Figure 1 shows, in ACE data, for 70% of relations, two mentions are embedded in each other or separated by at most one word. | in SF, more than 46% of query, filler entity pairs are separated by at least 7 words. | contrasting |
train_15592 | One domain for which automated analysis is particularly useful is Internet security: researchers obtain large amounts of text data pertinent to active threats or ongoing cybercriminal activity, for which the ability to rapidly characterize that text and draw conclusions can reap major benefits (Krebs, 2013a,b). | conducting automatic analysis is difficult because this data is outof-domain for conventional NLP models, which harms the performance of both discrete models (McClosky et al., 2010) and deep models (Zhang et al., 2017). | contrasting |
train_15593 | Roughly 60% of posts in the two forums contain multiple annotated tokens that are distinct beyond stemming and lowercasing. | we analyzed 100 of these multiple product posts across Darkode and Hack Forums, and found that only 6 of them were actually selling multiple products, indicating that posts selling multiple types of products are actually quite rare (roughly 3% of cases overall). | contrasting |
train_15594 | And as expected, the LCRF baselines yields relatively lower results compared to the other models, since it cannot predict overlapping mentions. | 11 such results give us some idea on how much performance increase we can gain by properly recognizing overlapping mentions by looking at the results of LCRF (single), which in this case can be up to 9.7 points in F 1 -score in ACE-2004. | contrasting |
train_15595 | We believe it is possible that this slower convergence is due to the spurious structures issue in mention hypergraphs, which causes the objective function to be more complex to optimize. | some further analyses on the convergence issue and the impact of different ways of exploiting features (over different hyperedges) for the hypergraph-based models are needed. | contrasting |
train_15596 | In recent years, many text and language understanding tasks have been advanced by neural network architectures. | despite recent work, competitive ED systems still largely employ manually designed features. | contrasting |
train_15597 | MinIE's dictionary mode also makes use of the corpus frequency of constituents. | to ReVerb, MinIE uses frequency to inform minimization (instead of to prune) and applies it to subjects and arguments as well. | contrasting |
train_15598 | Past research has addressed citation sentiment (Athar and Teufel, 2012b,a), citation networks (Kas, 2011;Gabor et al., 2016;Sim et al., 2012;Do et al., 2013;Jaidka et al., 2014), summarization (Abu-Jbara and Radev, 2011) and some analysis of research community (Vogel and Jurafsky, 2012;Anderson et al., 2012;Luan et al., 2012Luan et al., , 2014bLevow et al., 2014). | due to scarce hand-annotated data resources, previous work on information extraction (IE) for scientific literature is very limited. | contrasting |
train_15599 | An analysis of confusion patterns show that the most frequent type confusions are between PROCESS and MATERIAL. | we observe that ULM+GRAPHFEAT* can greatly reduce the confusion, with 3.5% relative improvement of PRO-CESS and 3.6% relative improvement of PROCESS over ULM+GRAPHINTERP on token level. | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.