id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_15300 | E.g., the AI2 Elementary School Science Questions dataset (Khashabi et al., 2016) contains 1080 questions for students in elementary schools; NTCIR QA Lab (Shibuki et al., 2014) evaluates systems by the task of solving real-world university entrance exam questions; The Entrance Exams task at CLEF QA Track (Peñas et al., 2014;Rodrigo et al., 2015) evaluates the system's reading comprehension ability. | data provided in these existing tasks are far from sufficient for the training of advanced data-driven machine reading models, partially due to the expensive data generation process by human experts. | contrasting |
train_15301 | This encourages us to build a semantic parser that produces an explicit representation of what the question is asking, if we want to make quantitative progress on the question set. | while it is not hard to formalize the semantics: it is not clear how to devise a compositional transformation from the original question to the formal semantics, since the meaning is dispersed throughout the discourse, such that neither the maximization nor the minimization can be locally derived from some subtree of the syntactic structure. | contrasting |
train_15302 | Because of this, it is feasible to use discriminative approaches Hosseini et al., 2014;Roy and Roth, 2015;Zhou et al., 2015;Koncel-Kedziorski et al., 2015;Mitra and Baral, 2016) that extract the quantities, featurize the question, and then perform a weighted search over the space of instantiated operator trees or equation templates. | it is not clear how one can extend these discriminative techniques to handle the complex semantics found in Examples 1 and 2. | contrasting |
train_15303 | Template-based approaches Zhou et al., 2015;Upadhyay et al., 2016), on the other hand, leverage the built-in composition structure of equation system templates to formulate all types of math concepts seen in training data, such as (1 − n 1 ) * n 2 = x in Figure 1. | they suffer from two major shortcomings. | contrasting |
train_15304 | In general the results illustrate that performances of the models increase with the length of the window. | we found that for SECT model, its mean performance reached the peak while standard deviations narrowed when window size reaches 10. | contrasting |
train_15305 | Using the dependency tree as illustrated in Figure 2, SEDT is able to identify the subject phrase correctly, namely "The Annual Conference". | biDAF failed to anwer the question correctly and selected a noun phrase as the answer. | contrasting |
train_15306 | (2017) investigated a similar approach to generate embeddings for out-of-vocabulary words from their definitions and applied it to a number of different tasks. | their method mainly focuses on modeling generic concepts and is evaluated on tasks that do not require the understanding of world knowledge specifically. | contrasting |
train_15307 | The double encoder architecture mentioned above considers each context independently. | since each document consists of a sequence of contexts, the knowledge carried by other contexts in C could also provide useful information for the decision process of C i . | contrasting |
train_15308 | Note that alternatively, one could aggregate information about the past predictions through other means like policies or soft attention. | this would introduce extra complexities to the learning process. | contrasting |
train_15309 | We experimented with different configurations of defining contexts and entity definitions, such as expanding the context windows by including sentences that come before and after the one where blank is found, as well as taking more than one sentence out of the entity description. | results on validation set show that increasing the context window size and the definition size had very little impact on accuracies, but drastically increased the training time of all models. | contrasting |
train_15310 | Moreover, these datasets have the attractive quality that the answer is a short snippet of text within the paragraph, which narrows the search space of possible answer spans. | many of these models rely on large amounts of human-labeled data for training. | contrasting |
train_15311 | There has also been work on generating high-quality questions (Yuan et al., 2017;Serban et al., 2016;Labutov et al., 2015), but not how to best use them to train a model. | we use the two-stage SynNet to generate data tuples to directly boost performance of a model on a domain with no annotations. | contrasting |
train_15312 | We then transfer the MC model originally learned on the source domain to the target domain t using SGD on the synthetic data. | since the synthetic data is usually noisy, we alternatively train the MC model with mini-batches from x s and x t , which we call data-regularization. | contrasting |
train_15313 | An exception is the Dolphin18K dataset (Huang et al., 2016) which contains 18,000+ problems. | this dataset has not been made publicly available so far. | contrasting |
train_15314 | Ideally, we would generate all the paraphrases of q. | since this set could quickly become intractable, we restrict the number of candidate paraphrases to a manageable size. | contrasting |
train_15315 | These are then used by the other layers of the network to measure Q/AP relatedness. | convolution Tree Kernels (cTK) can be applied to relational structures built on top of syntactic/semantic structures derived from Q/AP text (Tymoshenko et al., 2016a). | contrasting |
train_15316 | Additionally, to have more reliable results, it is standard practice to apply n-fold cross-validation. | we cannot do this on the training (TRAIN) sets, since the embeddings learned in Sec. | contrasting |
train_15317 | It is possible to argue that part of the improvements of refinement models over STAGG in Table 3 may be due to model ensembling. | the performance gap between QUESREV and the alternative solutions enables us to isolate this effect for query revision approach. | contrasting |
train_15318 | As previously discussed, we can define the relevancy of a question in terms of the validity of its premises for an image, so we extract premises from each question Q and must find a suitable irrelevant image I − . | there are certainly many images for which one or more of Q's premises are false and an important design decision is then how to select I − from this set. | contrasting |
train_15319 | This reduces the size of the caption training set from 82,783 images to 70,194 images. | the complete caption training set is tokenized as a bag of words per image, and made available as image tag training data. | contrasting |
train_15320 | In addition, they have distinct visual patterns which are robust for current vision systems to recognize. | activity attributes are more difficult to conceptualize as they involve varying levels of abstractness, which are also more challenging for computer vision as they have less distinct visual patterns. | contrasting |
train_15321 | In addition to the attributes presented here, we also crowdsourced attributes for the emotion content of each verb (e.g., happiness, sadness, anger, and surprise). | we found these annotations to be skewed towards "no emotion", since most verbs do not strongly associate with a specific emotion. | contrasting |
train_15322 | From the numbers given in Section 3, variation in the number of chats in NALCS was much higher than LMS, which one may expect to have a critical effect in the language model. | our results seem to suggest that the L-Char-LSTM model can pickup other factors of the chat data (e.g. | contrasting |
train_15323 | Current reinforcement-based text generation works use traditional phrase-matching metrics (e.g., CIDEr, BLEU) as their reward function. | these metrics use undirected ngram matching of the machine-generated caption with the ground-truth caption, and hence fail to capture its directed logical correctness. | contrasting |
train_15324 | VerbNet (Kipper et al., 2008), the largest English verb argument structure resource, 1 organizes verbs and classes into a shallow hierarchy, but its structure has been handcrafted incrementally over time (starting with seminal work by Levin, 1993). | recently-developed, state-of-the-art machine learning methods offer a unique alternative approach to constructing such a hierarchy. | contrasting |
train_15325 | These results provide additional psychological evidence for the effects associated with Verb-Net's coarse distinctions: for unattested verbframe pairs, participants tend to assign a higher compatibility rating when the verb has sibling VerbNet classes that can take the frame. | the range of compatibility judgments is highly variable across all three categories, and BHC's finer-grained predictions fail to account for much of this variability. | contrasting |
train_15326 | In this example, the visual term "having fun" is also generated by the baseline NMT model, making it clear that at times what seems like a translation extracted exclusively from the image may have been learnt from the training text data. | none of the two baselines translated "Mexikanischen Setting" as "Mexican restaurant", but four out of the five multi-modal models did. | contrasting |
train_15327 | The summation expresses the goal of learning a behavior parame-terized by natural language instructions. | contextual Bandit Setting to most policy gradient approaches, we apply the objective to a contextual bandit setting where immediate reward is optimized rather than total expected reward. | contrasting |
train_15328 | While a simpler setup, for example decomposing the problem to source and target prediction and using a planner, is likely to perform better, we aim to minimize task-specific assumptions and engineering of separate modules. | to better understand the problem, we also report results for the decomposed task with a planner. | contrasting |
train_15329 | Persons with MCI show symptoms across several cognitive domains, where global cognitive ability, episodic memory, perceptual speed, and executive functioning are most clearly affected. | the performance of persons with MCI over-lap greatly with the performance of healthy controls, which highlights the complexity and heterogeneity of the diagnosis (Bäckman et al., 2005). | contrasting |
train_15330 | Clearly, we expect these new features to be somewhat correlated with each other, since function words tend also be high-frequency words. | many content words are also labeled as high-frequency in our methodology, such as bil (English: car) and potatis (English: potato). | contrasting |
train_15331 | In many systems, vague or none is also included as another relation type when a TLINK is not clear or missing. | current systems usually use a reduced set of relation types, mainly due to the following reasons. | contrasting |
train_15332 | We could of course sidestep the problem by exhaustively annotating the entire document corpus, by annotating all mentions of entities and checking relations between all pairs of mentions. | that would be a laborious and prohibitively expensive task: using the interfaces we've developed (Section 6), it costs about $15 to annotate a single document by non-expert crowdworkers, resulting in an estimated cost of at least $1,350,000 for a reasonably large corpus of 90,000 documents (Dang, 2016). | contrasting |
train_15333 | Even with large beam sizes, the locally normalized model underperforms these approaches. | by increasing model flexibility and performing search during training, the globally normalized model is able to recover from search errors and achieve much of the benefits of scoring all possible spans. | contrasting |
train_15334 | Second, its sequence-to-sequence neural architecture allows it to handle either one-hot symbolic input or dense vectors of acoustic features. | existing models are typically designed for "clean" symbolic input, then retrofitted with additional mechanisms to cope with acoustics. | contrasting |
train_15335 | Sparse text-based representations with linear context (DISTRIB and DEPS) significantly outperform some dense semantic representations. | no dense semantic models significantly outperform DISTRIB and DEPS. | contrasting |
train_15336 | Most of the above studies focus on synchronous multi-modal content, i.e., in which images are paired with text descriptions and videos are paired with subtitles. | we perform summarization from asynchronous (i.e., there is no given description for images and no subtitle for videos) multi-modal information about news topics, including multiple documents, images and videos, to generate a fixed length textual summary. | contrasting |
train_15337 | Specifically, when speech transcriptions are not considered, the informativeness of the summary is the worst. | adding speech transcriptions without guidance strategies decreases readability to a large extent, which indicates that guidance strategies are necessary for MMS. | contrasting |
train_15338 | The utterance "This movie is sick" can be ambiguous (either positive or negative) by itself, but if the speaker is also smiling at the same time, then it will be perceived as positive. | the same utterance with a frown would be perceived negatively. | contrasting |
train_15339 | When annotators are asked for objective judgments about a text (e.g., POS tags), the broader context in which the text is situated is often irrelevant. | many NLP tasks focus on inference of factors beyond words and syntax. | contrasting |
train_15340 | We ask an-notators to determine whether a given Twitter user supports Donald Trump or Hillary Clinton. | inferring something about a user from a single tweet that she writes may prove difficult. | contrasting |
train_15341 | 2010have been proposed (Guan et al., 2017;Tian and Zhu, 2012;Wauthier and Jordan, 2011;Passonneau and Carpenter, 2014). | our work is most similar to efforts outside the domain of NLP, where Dai et al. | contrasting |
train_15342 | While some annotation tasks only require information from short texts, in many others, we can elicit higher-quality labels by providing annotators with additional contextual information. | asking annotators to consider too much information would make their task slow and burdensome. | contrasting |
train_15343 | According to the Law of Innovation, polysemy was claimed to correlate positively with meaning change. | our analysis showed that polysemy is highly collinear with frequency, and as such, did not demonstrate independent contribution to semantic change. | contrasting |
train_15344 | Including user factors as direct features is beneficial when there is a linear relationship with the class label, such as with gender and sarcasm use. | user-factor adaptation can capture more complex relationships between user groups and their language expression. | contrasting |
train_15345 | Modeling users has a long history of successful applications in providing personalized information access (Dou et al., 2007;Teevan et al., 2005) and recommendations (Guy et al., 2009;Li et al., 2010;Morales et al., 2012). | this work models users to better understand their content via language processing tasks following ideas from demographics-aware and domain adaptation. | contrasting |
train_15346 | For example, in the sentence: (a) The two justices have been attending federalist society events for years, our model correctly disambiguated justices with the WordNet sense justice 3 n (public official) rather than justice 1 n (the quality of being just), and the corresponding softmax distribution was heavily biased towards words and senses related to persons or groups (commissioners, defendants, jury, cabinet, directors). | in the sentence: (b) Xavi Hernandez, the player of Barcelona, has 106 matches, the same model disambiguated matches with the wrong WordNet sense match 1 n (tool for starting a fire). | contrasting |
train_15347 | Search systems are often focused on providing relevant results for the "now", assuming both corpora and user needs that focus on the present. | many corpora today reflect significant longitudinal collections ranging from 20 years of the Web to hundreds of years of digitized newspapers and books. | contrasting |
train_15348 | In some sense, one can see the relation -both are war movies, and Lucas and Spielberg have worked together. | inglourious Basterds was created at 2009 and was directed by Quentin Tarantino. | contrasting |
train_15349 | From this point, we propose a novel inter-weighted layer to measure the importance of each word. | the more similar two sentences are, the more probably we can align each word of sentence S with several words of sentence T , and vice versa. | contrasting |
train_15350 | As sentence length grows, RNN will suffer from gradient vanishing problem. | gated mechanism, such as Long Short Term Memory(LSTM) (Hochreiter and Schmidhuber, 1997) is introduced to address it. | contrasting |
train_15351 | The PWIM is still competitive on WikiQA but gets an inferior performance on TrecQA. | our models both have state-of-the-art performances on three datasets which demonstrates our models have excellent generalization ability in different datasets. | contrasting |
train_15352 | Recently, several methods have been developed to induce taxonomies from text corpora (Wu et al., 2012;Yang, 2012;Luu et al., 2014). | this task is far from being solved for three reasons: i) Text corpora may vary in size, topic and quality. | contrasting |
train_15353 | (2014); Wang and He (2016), they suffer from extremely low recall for the Chinese language. | distributional approaches use word representations derived from contexts, independent of its hyponym or hypernym. | contrasting |
train_15354 | The ambiguity issue has been addressed in a few systems (Anke et al., 2016b;Wu et al., 2012;Ponzetto and Navigli, 2009) by word sense disambiguation. | it is not fully solved. | contrasting |
train_15355 | TAXI crawls a domainspecific corpora and employs lexico-syntactic patterns and substrings for domain is-a relation extraction. | the potential of distributional methods is not fully exploited as only one system uses such techniques. | contrasting |
train_15356 | The form of compositional function involves many kinds of neural networks, such as recurrent neural networks (Hochreiter and Schmidhuber, 1997;Chung et al., 2014), convolutional neural networks (Collobert et al., 2011;Kalchbrenner et al., 2014), and recursive neural networks (Socher et al., 2013;Tai et al., 2015;Zhu et al., 2015). | these methods show an obvious defect in representing idiomatic phrases, whose semantics are not literal compositions of the individual words. | contrasting |
train_15357 | Due to its importance, some previous work focuses on automatic identification of idioms (Katz and Giesbrecht, 2006;Li and Sporleder, 2009;Fazly et al., 2009;Peng et al., 2014;Salton et al., 2016). | challenge remains to take idioms into account to improve neural based semantic representations of phrases or sentences. | contrasting |
train_15358 | Differing from previous work, which evaluating idiom detection as a standalone task, we want to integrate idiom understanding into sentiment classification task. | most of existing sentiment datasets do not cover enough idioms or related linguistic phenomenon. | contrasting |
train_15359 | The result is a lexicon where each entry contains a logical form template and a set of possible phrases for triggering the template. | we have avoided binding grammar rules to particular phrases in order to handle lexical variations. | contrasting |
train_15360 | These generative approaches aim to solve the modeling problem of assigning higher probability mass to outputs that use reoccurring parts. | our learning algorithm uses caching as a way to constrain the search space for computational efficiency; the probabilities of the candidate outputs are assigned by a separate discriminative model. | contrasting |
train_15361 | This is indeed a pretty simple trick, and per-feature standardization (with zero mean) is also a standard data preprocessing method. | it is not self-evident that this kind of standardization shall be applied on pre-trained word embeddings before using them in deep neural networks, especially with the obvious downside of rendering the word embedding algorithm's loss function sub-optimal. | contrasting |
train_15362 | For example, since we generalize all the predicates to a compact form <pred-xx>, for irregular verbs like "became" ⇒ become-01, simply stemming the inflected verb form will not give us the correct concept even if the sense is predicted correctly. | since CAMR uses the alignment table to store all possible concept candidates for a word, adding our predicated label as a feature could potentially help the parser to choose the correct concept. | contrasting |
train_15363 | If yes, it then outputs a correct one. | to existing pipeline methods which first consider individual candidate answers separately and then make a prediction based on a threshold, we propose an end-to-end deep neural network framework, which is trained by a novel group-level objective function that directly optimizes the answer triggering performance. | contrasting |
train_15364 | An alternative pipeline approach is to first solve P 2 and then P 1 , i.e., first determine whether there's a correct answer in the candidate set and then rank all candidates to find a correct one. | as we will show using state-of-theart Multiple Instance Learning (MIL) algorithms in Section 4, P 2 by itself is currently a very challenging task, partly because of the difficulty of extracting features from a set of candidate answers that are effective for answer triggering. | contrasting |
train_15365 | In such cases, perhaps a multi-word embedding like Doc2Vec (Le and Mikolov, 2014) would be more appropriate. | if this were indeed the issue, one would expect vector offsets to perform equally poorly. | contrasting |
train_15366 | Given the asymmetric nature of word associations, KL-divergence seemed to be a natural fit. | it is vastly outperformed by even cosine similarity on the same set of embeddings. | contrasting |
train_15367 | Therefore, in order to learn the relationship representation between two arguments, we propose an attention mechanism that can select out the most important part from two arguments and perform the information interaction between two arguments. | one common issue involved in implicit discourse relationship identification is the lack of labeled data. | contrasting |
train_15368 | Therefore, the proper representations of ZPs are required so as to take advantage of semantic information when resolving ZPs. | representing ZPs is challenging because they are merely gaps that convey no actual content. | contrasting |
train_15369 | Meanwhile, considering that the antecedents of a ZP provide the necessary information for interpreting the gap (ZP), it is a natural way to express a ZP by its potential antecedents. | only some subsets of candidate antecedents are needed to represent a ZP 1 . | contrasting |
train_15370 | The semantic nature of discourse relations makes discourse parsing a difficult task. | the recent introduction of distributed representations of discourse units has seemingly led to significant improvements, with a claimed relative error reduction of 51% on fully labelled structures. | contrasting |
train_15371 | This avoids introducing discourse units that were not part of the original annotation, which a preliminary binarization of trees would have induced. | rST-Parseval considers approximately twice as many nodes as the original Parseval would on binarized trees (at most 2n − 2 nodes for n EDUs, compared to n − 1 attachments in a binary tree), and the relation labels of most nuclei are redundant with the nuclearity of a node and its sister (SPAN for a nucleus whose sisters are satellites, and the same label as its sisters for a nucleus whose sisters are nuclei). | contrasting |
train_15372 | Both aspects artificially raise the level of agreement between RST trees, especially when using manual EDU segmentation. | all the parsers in our sample except (Sagae, 2009;Heilman and Sagae, 2015) predict binary trees over manually segmented EDUs and evaluate them against right-heavy binarized reference trees. | contrasting |
train_15373 | Since they are trained on the same gold-standard data, one would expect RNN-GOLD to perform similarly to MAXENT. | in the case of the RNN-gold, the 50 tokens window may actually not have enough words to be filled with, because the gold-standard data is composed of the sentence with the it-pronoun and the three previous sentences, which in addition tend to be short. | contrasting |
train_15374 | Selectional preferences have long been claimed to be essential for coreference resolution. | they are mainly modeled only implicitly by current coreference resolvers. | contrasting |
train_15375 | Work on the automatic acquisition of selectional preferences has shown considerable progress (Dagan and Itai, 1990;Resnik, 1993;Agirre and Martinez, 2001;Pantel et al., 2007;Erk, 2007;Ritter et al., 2010;Van de Cruys, 2014). | today's coreference resolvers (Martschat and Strube, 2015;Wiseman et al., 2016;Clark and Manning, 2016a, i.a.) | contrasting |
train_15376 | This also holds for pilot and airplane. | captain and pilot, as well as ship and airplane have high paradigmatic similarity, i.e., they are seman-tically similar and occur in similar contexts. | contrasting |
train_15377 | In this way, the coreference resolver could learn to use selectional preferences mainly for mentions that are more likely to be anaphoric. | given that the F 1 score of current anaphoricity determiners or singleton detectors is only around 85 percent (Moosavi and Strube, 2016a, 2017), the effect of using system anaphoricity scores might be small. | contrasting |
train_15378 | RAPIER (Califf and Mooney, 2003) constraints are similar to our attributes, but are basic (surface form, POS tag, and hypernyms only), and expanding them will exponentially increase its complexity. | adding attributes to GrASP only increment runtime linearly (see Section 3.2). | contrasting |
train_15379 | Error Analysis The classifier has a small tendency towards labeling sentences with the majority class other. | sampling the training set yielded worse results for all classes. | contrasting |
train_15380 | To be more specific, in terms of accuracy, precision, recall, F1 and AUC, the average improvement for the baseline features are 4.33%, 10.30%, 4.32%, 11.01% and 10.40%, respectively. | we observe that the precision of U-GR+AF, although gives the second highest score among all feature combinations, is lower than that of UGR; we leave it for future work. | contrasting |
train_15381 | Up to this point, we focused on the task of extracting links between ACs. | recent work has shown that joint models that simultaneously try to complete multiple aspects of the subtask pipeline outperform models that focus on a single subtask (Persing and Ng, 2016;Stab and Gurevych, 2014b;Peldszus and Stede, 2015). | contrasting |
train_15382 | * indicates that this bin does not contain any major claim labels, and this average only applies to claim and premise classes. | we do not disable the model from predicting this class: the model was able to avoid predicting this class on its own. | contrasting |
train_15383 | One thing to notice here is that lattice nodes can have multiple predecessor states. | hidden states in LSTMs and other sequential RNNs are conditioned on only one predecessor state (h j in left column of Table 1), rendering standard RNNs unsuitable for the modeling of lattices. | contrasting |
train_15384 | This is not surprising, given that the lattice training data includes lattices of varying density, including lattices with very few paths or even only one path. | without fine-tuning on lattices, using lattices as input performs poorly (lattice/R and lattice/R+1). | contrasting |
train_15385 | More seriously, because the translation function is smooth, infrequent pairs tend to be wrongly seen as noise in the training process and so are largely ignored by the model. | to this, the conventional SMT approach is based on statistics of words and/or phrases, which, in principle, is a symbolic method that uses a discrete model and involves little parameter sharing. | contrasting |
train_15386 | This idea has been adopted in early research into neural-based MT methods, where neural models were utilized to improve SMT performance (Zhang et al., 2015). | this seems to be counterintuitive, as intuitively learning general rules should be the first step, rather than first memorizing special cases and then learning general rules. | contrasting |
train_15387 | Secondly, the results show that with both datasets, the lexical approach (NMT-L) can improve NMT performance, showing that using SMT knowledge helps NMT. | the improvement seems less significant than reported in (Arthur et al., 2016). | contrasting |
train_15388 | Gradual fine-tuning The second dynamic data selection technique, see Figure 1b, is inspired by the success of domain-specific fine-tuning (Luong and Manning, 2015;Zoph et al., 2016;Sennrich et al., 2016a;Freitag and Al-Onaizan, 2016), in which a model trained on a large general-domain bitext is trained for a few additional epochs only on small in-domain data. | rather than training a full model on the complete bitext G, we gradually decrease the training data size, starting from G and keeping only the top n sentence pairs for the duration of η epochs, where the top n pairs are defined by their CED s scores. | contrasting |
train_15389 | The results in Table 4 show that the bitext ranking plays a crucial role in the success of data selection. | the results also show that even in the absence of an appropriate bitext ranking, dynamic data selection-and in particular gradual fine-tuning-is still superior to static data selection. | contrasting |
train_15390 | We explain this result as follows: Compared to static selection, both sampling and gradual fine-tuning have better coverage due to their improved exploration of the data. | sampling also suffers from a surprise effect of observing new data in every epoch. | contrasting |
train_15391 | Overall, the fluency of the MT output improves when NMT is used, and the number of lexical choice errors is also reduced. | state-of-the-art NMT approaches based on an encoder-decoder architecture with an attention mechanism as introduced by exhibit weaknesses that sometimes lead to MT errors which a phrase-based MT system does not make. | contrasting |
train_15392 | In (Tang et al., 2016), the NMT decoder is modified to switch between using externally de-fined phrases and standard NMT word hypotheses. | only one target phrase per source phrase is considered, and the reported improvements are significant only when manually selected phrase pairs (mostly for rare named entities) are used. | contrasting |
train_15393 | When we use the LM trained on extra monolingual data, we get total improvements of 1.0% BLEU and 2.3% BLEU with the hybrid approach. | when we add this language model and a word penalty on top of the pure NMT system and tune scaling factors with MERT, we get small improvements (last row of Table 2) only on product descriptions. | contrasting |
train_15394 | Phrases play an important role in natural language understanding and machine translation (Sag et al., 2002; Villavicencio et al., 2005). | it is difficult to integrate them into current neural machine translation (NMT) which reads and generates sentences word by word. | contrasting |
train_15395 | Machine translation (MT) models typically require large, sentence-aligned bilingual texts to learn good translation models (Wu et al., 2016;Sennrich et al., 2016a;Koehn et al., 2003). | for many language pairs, such parallel texts may only be available in limited quantities, which is problematic. | contrasting |
train_15396 | BPR+MNN uses the same MNN seed translations as MNN, obtained from unannotated monolingual data of English and the foreign language, to learn the linear mapping between their embedding spaces. | unlike MNN, BPR+MNN uses the mapped word vectors to predict ranking in a supervised manner with BPR objective. | contrasting |
train_15397 | Instance weighting has been widely applied to phrase-based machine translation domain adaptation. | it is challenging to be applied to Neural Machine Translation (NMT) directly, because NMT is not a linear model. | contrasting |
train_15398 | In recent years, neural encoder-decoder models (Kalchbrenner and Blunsom, 2013; have significantly advanced the state of the art in NMT, and now consistently outperform Statistical Machine Translation (SMT) (Bojar et al., 2016). | their success hinges on the availability of sufficient amounts of parallel data, and contrary to the long line of research in SMT, there has only been a limited amount of work on how to effectively and efficiently make use of monolingual data which is typically amply available. | contrasting |
train_15399 | One downside of this approach is the significantly increased training time, due to training of a model in the reverse direction and translation of monolingual data. | we propose to train NMT models from scratch on both bilingual and target-side monolingual data in a multi-task setting. | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.