id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_16000 | Our dataset contains five classes as mentioned in Section 6.1. | previous work only investigates binary relation detection. | contrasting |
train_16001 | Many efforts have been devoted to identifying valid instances from noisy data. | most existing methods handle each relation in isolation, regardless of rich semantic correlations located in relation hierarchies. | contrasting |
train_16002 | These methods focus on adding an extra model to reduce the noisy label. | stacking extra model does not fundamentally solve the problem of inadequate supervision signals of distant supervision, and will introduce expensive training costs. | contrasting |
train_16003 | Distant supervision (Mintz et al., 2009) aimed to obtain large-scale training data automatically, which becomes the most versatile supervision method. | it suffers from the noisy label problem. | contrasting |
train_16004 | Such relations require the model to predict both relation types and the entity order correctly. | for undirected relations, such as Synonym-of in ScienceIE and Associated in Phenebank, both directions can be accepted. | contrasting |
train_16005 | Schutte et al., 2011;Meena et al., 2012;Zarrieß et al., 2016;Shore and Skantze, 2017). | discerning REs from non-referring language in dialog is not trivial. | contrasting |
train_16006 | Given a literary model of reference, they should have supplied exactly enough information to identify the referent and no more (the big pinkish asteroid to the left), adhering to Grice's (1975) maxim of quantity. | rL can undergo not only expansion but also replacement: For color, speaker B proposes both purple and pinkish, but only pinkish is then accepted by A. | contrasting |
train_16007 | However, RL can undergo not only expansion but also replacement: For color, speaker B proposes both purple and pinkish, but only pinkish is then accepted by A. | to a literary model of reference, a collaborative model represents reference resolution as a process of iteratively presenting RL to the other participant(s) in a dialog, which is then either accepted as being sufficient to identify a referent or rejected as insufficient (Clark and Wilkes-Gibbs, 1986, 9). | contrasting |
train_16008 | (2011) algorithmically extracted RL as utterance(s) preceding a discrete event in a shared environment within a certain timeframe. | one drawback to this method is that "the references must be contained in instructions that cause events involving the referents" and "it must be possible to automatically detect these events" (Schutte et al., 2011, 189). | contrasting |
train_16009 | Other approaches use language structure to infer RL, namely in parsing said language using a combination of statistical or rule-based methods. | both entail that a solution be specialized for language specific to a given domain, such as for route-following instructions (Meena et al., 2012) or for a specific instructor-manipulator pair task (Shore and Skantze, 2017): Meena et al. | contrasting |
train_16010 | Deepak (2016) presents MiXKmeans, a variation of k-means algorithm, suited for clustering threads present on forums and Community Question Answering websites. | most techniques use unsupervised clustering to group similar questions/queries, without modeling intents. | contrasting |
train_16011 | Finally, our work is related to a large body of research on dialog acts (Stolcke et al., 2000;Kim et al., 2010;Chen et al., 2018): our low-level intent labels (Table 1) can be seen as very finegrained dialog acts (Core and Allen, 1997;Bunt et al., 2010;Oraby et al., 2017). | our paper's objective is different as our goal is not to rigidly define intents and then exploit them to derive a semantic interpretation. | contrasting |
train_16012 | In fact, w learns to score the edges since the structural feature vector decomposes over the edges. | imposing a structure onto the output is supposed to produce a better w, which we test in our experiments described in Section 5. | contrasting |
train_16013 | Here, we apply the same LSSVM, LSP, and the SVM classifier models trained in the experiments of the previous paragraph. | we recompute all the four unsupervised clustering baselines supplying them with the new k -the number of gold intent-based annotated clusters. | contrasting |
train_16014 | A typical conversation involves recalling important points from this background knowledge and producing them appropriately in the context of the conversation. | most existing large scale datasets (Lowe et al., 2015b;Ritter et al., 2010;Serban et al., 2016) simply contain a sequence of utterances and responses without any explicit background knowledge associated with them. | contrasting |
train_16015 | There has been considerable work on incorporating background knowledge in the context of goal-oriented dialog datasets even before the advent of large-scale datasets for deep learning (Raux et al., 2005;Seneff et al., 1991) as well as in recent times (Rojas-Barahona et al., 2017;Williams et al., 2016; where datasets include small sized knowledge graphs as background knowledge. | the conversations in these datasets are very templated and nowhere close to open conversations in specific domains such as the ones contained in our dataset. | contrasting |
train_16016 | Additionally, existing large-scale datasets are noisy as they are extracted from online forums which are inherently noisy. | since we use crowdsourcing, the extent of noise is reduced since there are humans in the loop who were explicitly instructed to use only clean sentences from the external knowledge sources. | contrasting |
train_16017 | While playing the role of Speaker 1, the worker was not restricted to copy/modify sentences from the background resources but was given the freedom to create (write) original sentences. | when playing the role of Speaker 2, the worker was strictly instructed to copy/modify sentences from the shown resources such that they were relevant in the current context of the conversation. | contrasting |
train_16018 | We would expect a poor performance for BiDAF (ms) as the resource has more noise because of the sentences from irrelevant resources. | we speculate the model learns to regard irrelevant sentences as noise and learns to focus on sentences corresponding to the correct resource resulting in improved performance (however, this is only a hypothesis and it needs to be verified). | contrasting |
train_16019 | Thanks to the automatic generation process, these datasets can be very large in size, leading to significant research progresses. | compared to how humans would create cloze questions and evaluate reading comprehension ability, the automatic generation process bears some inevitable issues. | contrasting |
train_16020 | Large-scale automatically-generated cloze tests (Hermann et al., 2015;Hill et al., 2016;Onishi et al., 2016) lead to significant research advancements. | generated questions do not consider language phenomenon to be tested and are relatively easy to solve. | contrasting |
train_16021 | Grammar questions are easily differentiated from other two categories. | the teachers themselves cannot specify a clear distinction between reasoning questions and vocabulary questions since all questions require comprehending the words within the context and conducting some level of reasoning by recognizing incomplete information or conceptual overlap. | contrasting |
train_16022 | The same conclusion can also be drawn from the success of a concurrent work ELMo which uses LM representations as word vectors and achieves state-ofthe-art results on six language tasks (Peters et al., 2018). | if we increase the context length to three sentences, the accuracy of 1B-LM only has a marginal improvement. | contrasting |
train_16023 | Additionally, the 1B-LM is trained on the sentence level, which might also result in the inability to track paragraph level information. | to investigate the differences between training on sentence level and on paragraph level, a prohibitive amount of computational resource is required to train a large model on the 1 Billion Word Corpus. | contrasting |
train_16024 | Automatic question answering (QA) has made big strides with several open-domain and machine comprehension systems built using large-scale annotated datasets (Voorhees et al., 1999;Ferrucci et al., 2010;Rajpurkar et al., 2016;Joshi et al., 2017). | in the clinical domain this problem remains relatively unexplored. | contrasting |
train_16025 | Recent advances in crowd-sourcing and search engines have resulted in an explosion of large-scale (100K) MC datasets for factoid QA, having ample redundant evidence in text (Rajpurkar et al., 2016;Trischler et al., 2016;Joshi et al., 2017;Dhingra et al., 2017). | complex domainspecific MC datasets such as MCTest (Richardson et al., 2013), biological process modeling (Berant et al., 2014), BioASQ (Tsatsaronis et al., 2015), InsuranceQA (Feng et al., 2015), etc have been limited in scale (500-10K) because of the complexity of the task or the need for expert annotations that cannot be crowd-sourced or gathered from the web. | contrasting |
train_16026 | The performance of the proposed models is summarized in Table 5. emrQL results are not directly comparable with GeoQuery and ATIS because of the differences in the lexicon and tools available for the domains. | it helps us establish that QL learning in emrQA is non-trivial and supports significant future work. | contrasting |
train_16027 | When the "oracle" science fact f used by the question author is provided to the knowledge-enhanced reader, it improves over the knowledge-less models by about 5%. | there is still a large gap, showing that the core fact is insufficient to answer the question. | contrasting |
train_16028 | In the second-order false-belief task, Sally observes the new location of the milk; thus, she has a true belief about the milk's location. | anne's belief about Sally's mental state does not match reality because anne does not know that Sally has observed the change in the environment. | contrasting |
train_16029 | Note that the median accuracy (the dark blue line in box) is close to 1 for these questions. | the model often fails (median accuracy around 0.5) given the first-order question ("where will Sally look for the milk") and the false-belief and second-order false-belief tasks. | contrasting |
train_16030 | The MemN2N model succeeds at both bAbi task 1 and the reality question given the true-belief task. | the correct answer to the memory question ("where was the milk at the beginning?") | contrasting |
train_16031 | Second, the syntactic parsers are unreliable on account of the risk of erroneous syntactic input, which may lead to error propagation and an unsatisfactory SRL performance. | syntactic information is considered closely related to semantic relation and plays an essential role in SRL task (Punyakanok et al., 2008). | contrasting |
train_16032 | Following Marcheggiani and Titov (2017), we focus on argument labeling and formulate SRL as sequence labeling problem. | we differ by (1) leveraging enhanced word representation, (2) applying recent advances in recurrent neural networks (RNNs), such as highway connections (Srivastava et al., 2015), (3) using deep encoder with residual connections (He et al., 2016), (4) further extending Syntax Aware Long Short-Term Memory (SA-LSTM) (Qian et al., 2017) for SRL, and 5introducing the Tree-Structured Long Short-Term Memory (Tree-LSTM) (Tai et al., 2015) to model syntactic information for SRL. | contrasting |
train_16033 | More specifically, Tree-LSTM only considers information from arbitrary child units so that each node lacks of the information from parent. | our Syn-GCN and SA-LSTM combine bidirectional information, both head-to-dependent and dependent-to-head. | contrasting |
train_16034 | The DFS approach is an applicable linearization of trees since a recursive traversal, which starts at the root and explores all outgoing edges, is guaranteed to visit all of the graph's nodes. | dFS linearization is not directly applicable to SdP, as its graphs often consist of several non-connected components. | contrasting |
train_16035 | In addition, an ablation study shows that multi-tasking the PRIMARY tasks is beneficial over a single task setting, which in turn is outperformed by the inclusion of the AUXILIARY tasks. | simulating disjoint annotations -with sDP's complete overlap of annotated sentences, multi-task learning often deals with disjoint training data. | contrasting |
train_16036 | We can see the question is asking "What state is San Antonio located in?". | the natural language word order in Indone- sian is different from English, where the phrase "berada di" that corresponds to m 2 (i.e., loc) appears between "San Antonio" (which corresponds to m 5 -san antonio ) and "what" (which corresponds to m 1 -answer). | contrasting |
train_16037 | Prior work, such as ✏-greedy exploration (Guu et al., 2017), has reduced the severity of this problem by introducing random noise in the search procedure to avoid saturating the search on high-scoring spurious programs. | random noise need not bias the search towards the correct program(s). | contrasting |
train_16038 | We therefore rewrite Equation 1as follows: As such we now define our model as: During training, the target length is set to T y = T y . | at test time, the target length generally varies according to the domain, genre, and language at hand. | contrasting |
train_16039 | In Section 3.1.1, we used the Euclidean distance of vector pairs to define M (G) and Sinkhorn distance d sh (G). | in our preliminary experiment, we found that Euclidean distance of unnormalized vectors gave poor performance. | contrasting |
train_16040 | When using our alignment process, our method is competitive with the TRANSLATE TRAIN baseline, suggesting that it might be possible to encode similarity between languages directly in the embedding spaces generated by the encoders. | these methods are still below the other machine translation baseline TRANSLATE TEST, which significantly outperforms the multilingual sentence encoder approach by up to 6% (Swahili). | contrasting |
train_16041 | A Burst Information Network (BINet) is a graphbased text stream representation and has proven effective for multiple text stream mining tasks (Ge et al., 2016a,b,c). | to many information networks (e.g., (Ji, 2009;Li et al., 2014)), BI-Nets are specially for text streams. | contrasting |
train_16042 | One reason is that the topic overlap of coordinated cross-lingual text streams is not so significant as the Wikipedia data used for their experiments, and the other reason is that their approaches focus on common fundamental words like "城市(city)" while our targets are OOVs like "东协(ASEAN)" which do not frequently appear in a corpus. | our approach is more practical: it not only works well in easily available and endless coordinated text streams without high content overlap requirement, but also can accurately mine translations of many OOVs which do not appear frequently and really need mining their translations. | contrasting |
train_16043 | In contrast to the word-level alignment methods, we attempt to mine burst-level alignment to largely narrow down candidates, and introduce powerful clues for improving accuracy and discovering various language knowledge. | to previous cross-lingual projection work like data transfer (Pado and Lapata, 2009) and model transfer (McDonald et al., 2011), we do not require any parallel data. | contrasting |
train_16044 | Therefore, the semantic meanings of words can reflect in the contexts according to these distributed representation models. | homographic puns always have multiple meanings. | contrasting |
train_16045 | In recent years, homographic puns have increasingly become a respectable research topic, which widely appears in rhetoric and literary criticism. | there were little related works in the fields of computational linguistics and natural language processing by (Miller, Tristan and Turković, Mladen, 2016). | contrasting |
train_16046 | The best perform of SemEval2017 task7 is Fermi. | fermi only evaluates on 675 of 2250 homographic contexts (Miller, Tristan and Hempelmann, Christian and Gurevych, Iryna, 2017) in SemEval2017 task7. | contrasting |
train_16047 | Chinese spelling check (CSC) is a challenging yet meaningful task, which not only serves as a preprocessing in many natural language processing (NLP) applications, but also facilitates reading and understanding of running texts in peoples' daily lives. | to utilize datadriven approaches for CSC, there is one major limitation that annotated corpora are not enough in applying algorithms and building models. | contrasting |
train_16048 | To build an annotated corpus of P-style errors, we follow the similar inspiration with those for V-style errors and OCR tools, and adopted a pipeline as shown in Figure 3. | given the availability of various speech recognition datasets, we employ a simpler approach. | contrasting |
train_16049 | easy to be considered as correct. | considering the preceding word 政企 (translation: politics and industry) and the subsequent words 是一 种痼疾 (translation: a kind of chronic disease), we can see that 部分 does not fit the current context and should be corrected as 不分. | contrasting |
train_16050 | An possible reason is that with more instances containing different spelling error including in the training dataset, the number of unseen spelling error in the testing dataset is reduced, thus facilitating the model to detect more spelling errors. | the improvement of the precision is not so obvious as that of the recall. | contrasting |
train_16051 | Grammatical error correction (GEC) systems deployed in language learning environments are expected to accurately correct errors in learners' writing. | in practice, they often produce spurious corrections and fail to correct many errors, thereby misleading learners. | contrasting |
train_16052 | This level of performance is impressive since GEC is a difficult task given the diversity and complexity of language errors. | in real-world use cases such as language learning, erroneous feedback from automatic GEC systems can potentially mislead language learners. | contrasting |
train_16053 | They use a dataset of learner sentences manually annotated with subjective scores of grammaticality. | their method was to assess learner writing and not for system evaluation. | contrasting |
train_16054 | number of edits number of reference tokens In MT, the reference translations for HTER are targeted, i.e., they are created by post-editing system translated sentences. | in GEC, highquality datasets annotated by experts with minimal edits are available (Dahlmeier et al., 2013;Yannakoudakis et al., 2011) and GEC systems are typically trained to make minimal changes to input sentences. | contrasting |
train_16055 | Because most POS tagging methods are based on supervised models, they usually require a large amount of labeled data for training. | the existing labeled datasets for Twitter are much smaller than those for newswire text. | contrasting |
train_16056 | Hence, to help POS tagging for Twitter, most domain adaptation methods try to leverage newswire datasets by learning the shared features between the two domains. | from a linguistic perspective, Twitter users not only tend to mimic the formal expressions of traditional media, like news, but they also appear to be developing linguistically informal styles. | contrasting |
train_16057 | It is a dataset of POS-tagged tweets consisting almost entirely of tweets sampled from one particular day (October 27, 2010). | the test set was introduced in (Owoputi et al., 2013), and contains 574 tweets (DAILY547). | contrasting |
train_16058 | We can see that our method could significantly outperform the TweetNLP Tagger. | our method was worse than the ARK tagger. | contrasting |
train_16059 | 's limited window-based representation (97.3%). | existing local models "regularly make egregious errors" (Manning, 2011), notably on imperative detection 1 . | contrasting |
train_16060 | The best performing tagger (Bohnet et al., 2018) was 3.9% above the next best model. | it still has an error rate of 25%. | contrasting |
train_16061 | The improvements from predicted case results are interesting, since in nonneural parsers, predicted case usually harms accuracy (Tsarfaty et al., 2010). | we note that our taggers use gold POS, which might help. | contrasting |
train_16062 | For most other dependencies (and all dependencies in German), Lemma is the most important feature, suggesting a strong reliance on lexical semantics of nouns and verbs. | we also notice that the model sometimes attends to features like Aspect, Polarity, and Verb-Form-since these features are present only on verbs, we suspect that the model may simply use them as convenient signals that a word is verb, and thus a likely head for a given noun. | contrasting |
train_16063 | Research in affective computing has mainly focused on understanding affect (emotions and sentiment) in monologues. | with increasing interactions of humans with machines, researchers now aim at building agents that can seamlessly analyze affective content in conversations. | contrasting |
train_16064 | The gold annotations are available for every 0.2 seconds in each video (Nicolle et al., 2012). | to align with our problem statement, we approximate the utterance-level annotation as the mean of the continuous values within the spoken utterance. | contrasting |
train_16065 | This suggests that multi-hop is more crucial than the latter. | best performance is achieved by variant 6 which contains all the proposed modules in its pipeline. | contrasting |
train_16066 | 3a), where all regions without a positive label can be used as negative examples. | query-Adaptive R-CNN is trained using the open-vocabulary phrases annotated to the regions (Fig. | contrasting |
train_16067 | Using hard negative examples has proven to be effective in the object detection to train a discriminative detector (Felzenszwalb et al., 2010;Shrivastava et al., 2016). | adding negative examples is usually not easy in the openvocabulary setting, because it is not guaranteed that a region without a positive label is negative. | contrasting |
train_16068 | 1) The first approach uses the WordNet hierarchy: if two categories have parentchild relationships in WordNet (Miller, 1995), they are not mutually exclusive. | the converse is not necessarily true; e.g., man and skier are not mutually exclusive but do not have the parent-child relationship in the WordNet hierarchy. | contrasting |
train_16069 | Using a softmax layer augmented with hidden state vectors, they predict the verb and the nominal fillers of its roles. | to above works on ImSitu, we do not link the roles of a verb to their lexical fillers. | contrasting |
train_16070 | Using the triplet loss with memory modules led to greater performance when compared to the attn 3 model, but the performance sits around the use of either triplet only or memory only. | when we increase the N diag to 16 or 32, we find a jump in performance. | contrasting |
train_16071 | As in this earlier work, we are also interested in navigation with a topological map of the environment. | we do not process symbolic phrases. | contrasting |
train_16072 | We also observe that inherent ambiguities in instruction following make exact goal identification difficult, as demonstrated by imperfect human performance. | the gap to human-level performance still remains large across both tasks. | contrasting |
train_16073 | On CHAI, CHAP-LOT18 and MISRA17 both fail to learn, while our approach shows an improvement on stop distance (SD). | all models perform poorly on CHAI, especially on manipulation (MA). | contrasting |
train_16074 | Furthermore, arbitrary IRF kernels are not guaranteed to afford analytical estimator functions or unique real-valued solutions. | recent advances in machine learning have led to libraries like Tensorflow (Abadi et al., 2015)which uses auto-differentiation to support optimization of arbitrary computation graphs -and Edward (Tran et al., 2016) -which enables black box variational inference (BBVI) on Tensorflow graphs. | contrasting |
train_16075 | They argue that this constitutes the first strong evidence of memory effects in broadcoverage sentence processing. | it turns out that when one baseline predictor -probabilistic context free grammar (PCFG) surprisal -is spilled over one position, the reported effects disappear: p = 0.816 for constituent wrap-up and p = 0.370 for dependency locality. | contrasting |
train_16076 | the use of the canonical HRF to convolve predictors in fMRI models prior to linear regression. | since DTSR is domain-general, it can be a valuable component in any analysis toolchain for time series. | contrasting |
train_16077 | (2016) has formalized the grammar of these queries and proposed semisupervised algorithms for the adaptation of parsers originally designed to parse according to the standard dependency grammar, so that they can account for the unique forest grammar of queries. | their algorithms rely on resources typically not available outside of big web corporates. | contrasting |
train_16078 | We observe a universal improvement across all POS tags for each of the three variations of the system compared to the baseline. | it is notable that the biggest gains in HDLAS are for open word classes: NOUNs, VERBs and ADJs. | contrasting |
train_16079 | These kinds of depth bounds in sentence processing have been used to explain the relative difficulty of center-embedded sentences compared to more right-branching paraphrases like It was awful for the plant's parts to fail. | depth-bounded grammar induction has never been compared against unbounded induction in the same system, in part because most previous depth-bounding models are built around sequence models, the complexity of which grows exponentially with the maximum allowed depth. | contrasting |
train_16080 | tion, we generate a series of tables of size O(|w| × |w|) with elements of size O(|Q P | × |Q P |). | since we use the sliding window technique, we only save O(1) of these tables, leading to an overall space complexity of O((|w||Q P |) 2 ). | contrasting |
train_16081 | They aim at maximizing the probability of generating a response given an input query, and generally use the maximum likelihood estimation (MLE) as their objective function. | various problems occur when Seq2Seq models are used for dialogue generation tasks. | contrasting |
train_16082 | It estimates the beliefs of possible user's goals at every dialogue turn. | for most current approaches, it's difficult to scale to large dialogue domains. | contrasting |
train_16083 | the slots and values are defined in advance, and can't change dynamically. | this is not flexible in practice (Xu and Hu, 2018). | contrasting |
train_16084 | For example, in Set an alarm at 8 am for Monday and Wednesday, 8 am needs to be associated with both Monday and Wednesday which would require graph-structured repre-sentations. | we found that just 0.3% of our dataset would require a more expressive representation to model adequately. | contrasting |
train_16085 | Many deep learning approaches have been explored to train a classifier with those datasets to develop an automatic abusive language detection system (Badjatiya et al., 2017;Park and Fung, 2017;Pavlopoulos et al., 2017). | these works do not explicitly address any model bias in their models. | contrasting |
train_16086 | This could be attributed to the difference in the source and target tasks (abusive & sexist). | the decrease was marginal (less than 4%), while the drop in bias was significant. | contrasting |
train_16087 | ;Zhang et al., 2018) employ adversarial training methods to make the classifiers unbiased toward certain variables. | those works do not deal with natural language where features like gender and race are latent variables inside the language. | contrasting |
train_16088 | Talk page data already underpins research on social phenomena such as conversational behavior (Danescu-Niculescu-Mizil et al., 2012, 2013, disputes (Wang and Cardie, 2014b), antisocial behavior (Wulczyn et al., 2017;Zhang et al., 2018) and collaboration (Kittur et al., 2007;Halfaker et al., 2009). | the scope of such studies has so far been limited by a view of the conversation that is incomplete in two crucial ways: first, it only captures a subset of all discussions; and second, it only accounts for the final form of each conversation, which frequently differs from the interlocutors experience as the conversation develops. | contrasting |
train_16089 | One way to tag multiple datasets is to concatenate all the datasets with all the output labels and train a single BiLSTM-CRF model. | this assumes that each text snippet is completely annotated across the label sets, which is not true. | contrasting |
train_16090 | They generate large quantities of parse trees by parsing unlabeled data with two existing parsers and selecting only the sentences for which the two parsers produced the same trees. | the trees produced this way have noise 1 and tend to be short sentences, since it is easier for different parsers to get consistent results. | contrasting |
train_16091 | 2018, we then combine the input vector ( .., L) by a Softmaxnormalized weight (W woc ) and a scalar parameter (γ): The parameters of word ordering model are fixed during the training of parsing model. | the weight and scalar parameters are tuned to better adapt to it. | contrasting |
train_16092 | This might be partly explained by the fact that learning morphological features at the word level is difficult due to data sparsity -indeed the rarest French words in our dataset are observed only 10 times in the training data. | additional experiments showed that our finding is consistent across different word frequency bins: that is, even the embeddings of frequent words do not encode morphological features better than the majority baseline. | contrasting |
train_16093 | This can be explained by the fact that number remains most often unchanged through translation, and is marked in all target languages -albeit to different extents. | tense is determined by the semantics but also by languagespecific usage, while gender has little semantic value and is mostly assigned to nouns arbitrarily. | contrasting |
train_16094 | Initially, we also experimented with only Levenshtein distance as loss, similar to previous work on character-level problems (Leblond et al., 2018;Bahdanau et al., 2017). | models did not learn much, which we attribute to sparse training signal as all action sequences producing the same y would incur the same sequence-level loss, including intuitively very wasteful ones, e.g. | contrasting |
train_16095 | In the image domain, these perturbations are often virtually indistinguishable to human perception, causing humans and state-of-the-art models to disagree. | in the natural language domain, small perturbations are clearly perceptible, and the replacement of a single word can drastically alter the semantics of the document. | contrasting |
train_16096 | In terms of training speed, we do not observe obvious decrease, which in turn demonstrates the advantage of our disagreement regularizations. | the combinations of different disagreement regularizations fail to further improve translation performance (Rows 5-7). | contrasting |
train_16097 | This is also consistent with the finding in (Tu et al., 2016), which indicates that neural networks can model linguistic information in their own way. | to attended positions, it seems that the multi-head attention prefer to encoding the differences among multiple heads in the learned representations. | contrasting |
train_16098 | Typical acquisition functions select examples for which the current predictor is most uncertain. | how precisely to quantify uncertainty, especially for neural networks, remains an open question. | contrasting |
train_16099 | For the input-to-hidden matrix W x the linear combination (W x x t ) is normally distributed. | sampling the same W x for all timesteps and sampling the same noise i for preactivations for all timesteps are not equivalent. | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.