id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_14700 | In recent work, some research explore "deep" expressions such as discourse commitments or strict logic for representing the text. | these expressions suffer from the limitation of inference inconvenience or translation loss. | contrasting |
train_14701 | Through logic inference, some implicit knowledge behind the text can be mined. | it is not easy to translate the natural language text into formal logic expressions and the translation process inevitably suffer from great information loss. | contrasting |
train_14702 | With the result in Table 2, it is true that these pre-trained word embeddings have a good effect on our performance (we use word2vec 3 on Chinese Gigaword Corpus for word pre-training). | as shown in Table 2, compared to standard pre-training, the influence of heterogenous data is more evident. | contrasting |
train_14703 | More recently, Zhou and Xu (2015) proposed L-STM RNN approach for English Semantic Role Labeling, which shared similar idea with our model. | the features used and the network architecture were different from ours. | contrasting |
train_14704 | Due to the commonality in natural language, negation focus plays a critical role in deep understanding of context. | existing studies for negation focus identification major on supervised learning which is timeconsuming and expensive due to manual preparation of annotated corpus. | contrasting |
train_14705 | In this paper, we thus focus on graph-based ranking methods (Mihalcea and Tarau, 2004) which first build a word graph according to word co-occurrences within document, and then use random walk algorithms (e.g., PageRank) to measure word importance. | for negation focus identification, the graph-based methods may suffer from the following two problems: (a) the words in graphbased methods are strongly connected by cooccurrence rather than semantic content, which do not necessarily guarantee that they are relevant to the negation focus in context; and (b) identifying a negation focus may be affected by not only the relatedness of surrounding words but also its importance in current document which is not considered in standard random walk algorithms. | contrasting |
train_14706 | Vector space models of language like the ones presented in (Collobert et al., 2011b;Mikolov et al., 2013;Pennington et al., 2014) create good representations for the individual words of a language. | the words in a language can be combined into infinitely many distinct, wellformed phrases and sentences. | contrasting |
train_14707 | Interpretation of events-determining what the author claims did or did not happen-is important for many NLP applications, such as news article summarization or biomedical information extraction. | detecting events and assessing their factuality is challenging. | contrasting |
train_14708 | (2012), who advocated for representing factuality from the reader's perspective as a distribution of categories, but their annotation process requires manual normalization of the text. | we model factuality from the author's perspective with scalar values, and we have an endto-end crowdsourced annotation pipeline. | contrasting |
train_14709 | A prerequisite relation describes a basic relation among concepts in cognition, education and other areas. | as a semantic relation, it has not been well studied in computational linguistics. | contrasting |
train_14710 | adverse drug reactions, infectious diseases) in particular communities. | in order for a machine to understand and make inferences on these health conditions, the ability to recognise when laymen's terms refer to a particular medical concept (i.e. | contrasting |
train_14711 | We firstly observe that for the vSim baseline, excepting for word vector representation with vector size 50 learned using GloVe from the Twitter collection, word vector representations learned using either CBOW or GloVe are more effective than the one-hot representation. | the difference between the MRR-5 performance is not statistically significant (p > 0.05, paired t-test). | contrasting |
train_14712 | We argue that some types are relatively simple to answer, partly due to the limited vocabulary used, which explains why simple lexical matching methods can perform well. | some questions require understanding of higher level concepts such as those of the story and its characters, and/or require inference. | contrasting |
train_14713 | Then we attempted to account for them by searching the text for the sentence indicating that she had gone home and reducing the weight for all subsequent sentences. | since the improvements due to these rules were negligible, we did not include them in our final system. | contrasting |
train_14714 | (2013) demonstrate that the MC160 and MC500 have similar ratings for clarity and grammar, and that humans perform equally well on both. | in many cases MC500 appears to be designed in such a way to confuse lexical algorithms and encourage the use of more sophisticated techniques necessary to deal with phenomena such as elimination questions, negation, and common knowledge not explicitly written in the story. | contrasting |
train_14715 | Existing learning algorithms have primarily focused on building actionable meaning representations which can, for example, directly query a database (Liang et al., 2011;Kwiatkowski et al., 2013) or instruct a robotic agent (Chen, 2012;Artzi and Zettlemoyer, 2013b). | due to their end-to-end nature, such models must be relearned for each new target application and have only been used to parse restricted styles of text, such as questions and imperatives. | contrasting |
train_14716 | AMR meaning bank provides a large new corpus that, for the first time, enables us to study the problem of grammar induction for broad-coverage semantic parsing. | it also presents significant challenges for existing algorithms, including much longer sentences, more complex syntactic phenomena and increased use of noncompositional semantics, such as within-sentence coreference. | contrasting |
train_14717 | Given the partial derivations, our gradient computation is identical to Equation 2. | in contrast to Collins and Roark (2004) our data does not include gold derivations. | contrasting |
train_14718 | To minimise handcrafting, Stent and Molina (2009) proposed learning sentence planning rules directly from a corpus of utterances labelled with Rhetorical Structure Theory (RST) discourse relations (Mann and Thompson, 1988). | the required corpus labelling is expensive and additional handcrafting is still needed to map the sentence plan to a valid syntactic form. | contrasting |
train_14719 | Paliwal, 1997) have been shown to be effective for sequential problems (Graves et al., 2013a;Sundermeyer et al., 2014). | applying a bidirectional network directly in the SC-LSTM generator is not straightforward since the generation process is sequential in time. | contrasting |
train_14720 | In addition, the slot error rate (ERR) as described in Section 3.5 was computed as an auxiliary metric alongside the BLEU score. | for the experiments it is computed at the corpus level, by averaging slot errors over each of the top 5 realisations in the entire corpus. | contrasting |
train_14721 | For every task, the +Expectation method has performances that often seem to be higher than the simple baseline (both for the 50d case or the 100d case). | only some of these differences are significant. | contrasting |
train_14722 | The system described in (Hosseini et al., 2014) handles only addition and subtraction problems, and requires additional annotated data for verb categories. | our system does not require any additional annotations and can handle a more general category of problems. | contrasting |
train_14723 | It tries to map numbers from the problem text to predefined equation templates. | they implicitly assume that similar equation forms have been seen in the training data. | contrasting |
train_14724 | However, they implicitly assume that similar equation forms have been seen in the training data. | our system can perform competitively, even when it has never seen similar expressions in training. | contrasting |
train_14725 | We believe that the improvement was mainly due to the dependence of the system of (Roy et al., 2015) on lexical and neighborhood of quantity features. | features from quantity schemas help us generalize across problem types. | contrasting |
train_14726 | First, in this method, an already existing knowledge base is heuristically aligned to texts, and the alignment results are treated as labeled data. | the heuristic alignment can fail, resulting in wrong label problem. | contrasting |
train_14727 | These methods have been shown to be effective for relation extraction. | their performance depends strongly on the quality of the designed features. | contrasting |
train_14728 | successfully incorporated multiinstance learning into traditional Backpropagation (BP) and Radial Basis Function (RBF) networks and optimized these networks by minimizing a sum-of-squares error function. | to their method, we define the objective function based on the cross-entropy principle. | contrasting |
train_14729 | The idea is to capture the most significant features (with the highest values) in each feature map. | despite the widespread use of single max pooling, this approach is insufficient for relation extraction. | contrasting |
train_14730 | Here we considered a fact or tuple as linked if both of its entities were linked to Freebase, as partiallylinked if only one of its entities was linked, and as non-linked otherwise. | to previous work (Riedel et al., 2013;, we retain partially-linked and non-linked facts in our dataset. | contrasting |
train_14731 | Other words with embeddings similar to "driving" that appear on the dependency path between the mentions will similarly receive high weight for the ART label. | if the embedding is similar but is not on the dependency path, it will have 0 weight. | contrasting |
train_14732 | Roth and Woodsend (2014) considered features similar to ours for semantic role labeling. | in prior work both of above approaches are only able to utilize limited information, usually one property for each word. | contrasting |
train_14733 | As we can see, POS tags, grammatical relations, and WordNet hypernyms are also discrete (like words per se). | no prevailing embedding learning method exists for POS tags, say. | contrasting |
train_14734 | The Bayesian model of X&T is important in providing insight into how people might reason about samples of data that exemplify categories. | it relies on having complete, built-in knowledge about the taxonomic hierarchy, including both the detailed composition of categories and the values for between-object similarities, drawn from adult similarity judgments. | contrasting |
train_14735 | In the FAS model, all the features for a word are dependent: increasing the probability of any feature results in decreasing the probability of others. | this interaction is not always desirable, as many features regularly co-occur in the world. | contrasting |
train_14736 | conditions), the model, like children, generalizes to the lowest level category in the hierarchy that is consistent with the training items, roughly equally preferring items from that category or lower, with slight preference for the lower categories. | after seeing a single training example (the (a) Adult data: 3 subord. | contrasting |
train_14737 | (2013) present latent variable models for unsupervised learning of latent character types in movie plot summaries and in English novels, taking authorial style into account. | even the state-of-the-art NLP work rather describes personas of fictional characters by their role in the story -e.g., action hero, valley girl, best friend, villain etc. | contrasting |
train_14738 | This poses a challenge for fictional characters. | strong correlations have been found between the self-reported and perceived personality traits (Mehl et al., 2006). | contrasting |
train_14739 | We have shown that exploiting the links between lexical resources to leverage more accurate semantic information can be beneficial for this type of tasks, oriented to actions performed by the entity. | the human annotator agreement in our task stays high above the performance achieved. | contrasting |
train_14740 | Expectation-maximization algorithms, such as those implemented in GIZA++ pervade the field of unsupervised word alignment. | these algorithms have a problem of over-fitting, leading to "garbage collector effects," where rare words tend to be erroneously aligned to untranslated words. | contrasting |
train_14741 | They associate two models via the agreement constraint and show that agreement-based joint training improves alignment accuracy significantly. | enforcing agreement in joint training faces a major problem: the two models are restricted to one-to-one alignments (Liang et al., 2006). | contrasting |
train_14742 | propose a joint model to process word segmentation and informal word detection. | text normalization is split to another task ). | contrasting |
train_14743 | Their model is trained on a partially annotated microblog corpus. | our model can be trained on existing annotated corpora in standard text. | contrasting |
train_14744 | release a Chinese microblog corpus for word segmentation and informal word detection. | there are no microblog corpora annotated Chinese word segmentation, POS tags, and normalized sentences. | contrasting |
train_14745 | A Naive h function is straightforwardly defined as the annotation of each local rule part of the tree in a bottom-up fashion: 1 Base: When ∃a ∈ A such that h(a) = ⊥ we say that the annotation has failed. | the naive procedure fails in a large number of cases. | contrasting |
train_14746 | Given a grammar in CNF, we can prove that for a sentence of length n, the number of derivation steps for a shift reduce parser is 3n − 1. | our tagset-preserving transformation also introduces rules of the form (b), which explains why the number of derivation steps may vary from 2n − 1 to 3n − 1. | contrasting |
train_14747 | This approach induces a compact feature representation by combining atomic features. | unlike traditional tensor models, it enables us to incorporate prior knowledge about desired feature interactions, eliminating invalid feature combinations. | contrasting |
train_14748 | These gains were obtained by a careful construction of features templates that combine standard dependency parsing features and typological features. | we propose an automated, tensor-based approach that can effectively capture the interaction between these features, yielding a richer representation for crosslingual transfer. | contrasting |
train_14749 | While these methods can automatically combine atomic features into a compact composite representation, they cannot take into account constraints on feature combination. | our method can capture features at different composition levels, and more generally can incorporate structural constraints based on prior knowledge. | contrasting |
train_14750 | In addition, we consider the semi-supervised transfer scenario, in which we assume 50 sentences in the target language are available with annotation. | we observe that random sentence selection of the supervised sample results in a big performance variance. | contrasting |
train_14751 | We believe that the reason is two-fold: (a) k-means finds a local maximum during clustering; (b) we do hard clustering instead of soft clustering. | we detected that the clustering algorithm gives a more diverse set of solutions, when the features are perturbed. | contrasting |
train_14752 | 2013; Choi and McCallum, 2013; Ballesteros and Bohnet, 2014). | these methods are based on discrete features and suffer from the problems of data sparsity and feature engineering (Chen and Manning, 2014). | contrasting |
train_14753 | Le and Zuidema (2014) proposed a generative re-ranking model with Inside-Outside Recursive Neural Network (IORNN), which can process trees both bottom-up and top-down. | iORNN works in generative way and just estimates the probability of a given tree, so iORNN cannot fully utilize the incorrect trees in k-best candidate results. | contrasting |
train_14754 | This is the case of many application like message dictation for example. | in the case of service at hand, an initial concept must be detected (the action), therefore, FAIL RAW improves the performance. | contrasting |
train_14755 | Moreover, continuous representation of words in neural network LMs can also support robust modeling (Bengio et al., 2003;Mikolov et al., 2010). | previous works are focused on maximizing performance in the same domain as that of the training data. | contrasting |
train_14756 | In other words, it is uncertain that these technologies robustly support out-of domain tasks. | latent words LMs (LWLMs) (Deschacht et al., 2012) are clearly effective for outof domain tasks. | contrasting |
train_14757 | LW is comparable to MKN and HPY, and inferior to RNN in terms of PPL. | in test sets (out-of domain tasks), PPL improved with the increase in the number of layers in LW. | contrasting |
train_14758 | We use the following features in our implementation of this model. | any relevant ASR and SMT feature may be readily added to this model. | contrasting |
train_14759 | As this problem is NP-hard, pruning low-weight concepts is required for the ILP solver to find optimal solutions efficiently Li et al., 2013). | reducing the number of concepts in the model has two undesirable consequences. | contrasting |
train_14760 | fluency, should be conducted using manual evaluation. | conducting formal human evaluation is somewhat problematic. | contrasting |
train_14761 | Initially, manual evaluation was carried out, where human judges were tasked to assess the quality of automatically generated summaries. | in an effort to make evaluation more scaleable, the automatic ROUGE 1 measure (Lin, 2004b) was introduced in DUC-2004. | contrasting |
train_14762 | In this variant, word embeddings are used, as we are proposing in this paper, to map text content within generated summaries to SCUs. | the SCUs still need to be manually identified, limiting this variant's scalability and applicability. | contrasting |
train_14763 | The MDL principle is widely useful in compression techniques of non-textual data, such as summarization of query results for OLAP applications. | (Lakshmanan et al., 2002;Bu et al., 2005) only a few works on text summarization using MDL can be found in the literature. | contrasting |
train_14764 | It is noteworthy that, in addition to the greedy approach, we also evaluated the global optimization with maximizing coverage and minimizing redundancy using Linear Programming (LP). | experimental results did not provide any improvement over the greedy approach. | contrasting |
train_14765 | Our work is based on the bipartite entity graph introduced by Guinaudeau and Strube (2013). | in their graph one set of nodes corresponds to entities whereas in our graph it corresponds to topics. | contrasting |
train_14766 | We use the DUC 2002 dataset to compare our results with state-of-the-art techniques. | to the PLOS Medicine data the DUC 2002 dataset contains very small articles. | contrasting |
train_14767 | However the difference between the results of Tgraph and Egraph are not significant. | to the entity graph based system, the coherence measure in our system is calculated by using a topic-based weighted projection graph, which is denser and hence more informative. | contrasting |
train_14768 | The results are comparable with the entity graph. | the entity graph is less informative and very sparse as compared to the topical graph. | contrasting |
train_14769 | Instead of performing SVD, we can also take s i ∈ R B as our sentence representation, which makes our method resemble the bigram coverage-based summarization approach. | this makes s i a very sparse vector. | contrasting |
train_14770 | Colmenares et.al construct a 1.3 million financial news headline dataset written in English for headline generation (Colmenares et al., 2015). | the data set is not publicly available. | contrasting |
train_14771 | Recently, recurrent neural network (RNN) have shown powerful abilities on speech recognition (Graves et al., 2013), machine translation and automatic dialog response (Shang et al., 2015). | there is rare research on the automatic text summarization by using deep models. | contrasting |
train_14772 | The first and second tier cities are facing growth difficulties. | o2o market in the third and fourth tier cities contains opportunities. | contrasting |
train_14773 | To some extent, this hope was validated through a number of works at the time, mostly involving machine translation applications, and constraining in more or less explicit ways the specification of r (van Noord, 1990). | for the non-statistical approaches to parsing then strongly dominant, robustness was an issue: a parser had to either accept or reject a given input x, with no intermediary options, and in order to be able to parse actual utterances, with all their empirical diversity, parsers had to be rather tolerant. | contrasting |
train_14774 | Modern QA systems rely on an independent component to pre-select candidate answer sentences, which utilizes various signals such as lexical matching and user behaviors. | the candidate sentences Table 5: Evaluation of answer triggering on the WIKIQA dataset. | contrasting |
train_14775 | Most work leverages multiple sources of information, such as search query history, Twitter feeds, Facebook likes, social network links, and user profiles. | in many situations, little of this information is available. | contrasting |
train_14776 | We term this procedure as On-Demand Augmentation (ODA), because the search can be performed during test time in an on-demand manner. | the previous approaches of adding edges or embeddings to the KB (Gardner et al., 2013), and vector space random walk PRA (Gardner et al., 2014) are batch procedures. | contrasting |
train_14777 | Using surface level relations and noun phrases for extracting meaningful relational facts is not a new idea (Hearst, 1992), (Brin, 1999), (Etzioni et al., 2004). | none of them make use of Knowledge Bases for improving information extraction. | contrasting |
train_14778 | It is then reasonable to expect that a state-of-the-art formal semantics provides an accurate computational basis of natural language inferences. | there are still obstacles in the way of achieving this goal. | contrasting |
train_14779 | Thus, E2 and E4 are relatively difficult to be recognized as events by themselves. | event coreference E1-E2, which is supported primarily by E2's participants Barclays and Zaragozano shared with E1, helps determine that E2 is an event. | contrasting |
train_14780 | That is, one can train a structured learning model to globally capture the interactions between two relevant tasks via a certain kind of structure, while making predictions specifically for these respective tasks. | no prior work has studied the interactions between event trigger identification and event coreference resolution. | contrasting |
train_14781 | If one uses this approach, a beam state may represent a partial assignment of an event trigger. | event coreference can be explored only from complete assignments of an event trigger. | contrasting |
train_14782 | It is possible for the model to predict an invalid encoded sequence that does not correspond to any word in the original vocabulary. | in our experiments, we did not observe any such sequences in the decoding of the test set. | contrasting |
train_14783 | the whole neural network (not just the output layer like the NNJM) for each noise sample and thus noise computation is more expensive. | for different epochs, we resampled the negative example for each positive example, so the BNNJM can make use of different negative examples. | contrasting |
train_14784 | For the number of top configurations k used to initialize each following stage, we know the larger k is, the better results in the next stage since Bayesian Optimization relies on good initial knowledge to fit good regression models (Feurer et al., 2015). | larger k value also leads to high computation cost at the next stage, since these initial settings will have to be evaluated first. | contrasting |
train_14785 | Notice that the weakly supervised classifier trained with 68M words obtains 68% accuracy on the AOC test set and 48% on the FB test set (row 1), which is not much higher. | considering this classifier is trained without any human labeled dialect data, the performance is expected and can be improved with better training data and models. | contrasting |
train_14786 | Precisions increase from the weakly supervised to the strongly supervised to the semi-supervised classifier, and the combined classifier generally outperforms all three classifiers, except for the Gulf dialect. | considering the smaller percentage of the Gulf dialect, we still observe significant improvement overall. | contrasting |
train_14787 | Recognizing claims exhibits similar behavior. | recognizing non-argumentative text performs better in the opposite direction. | contrasting |
train_14788 | These papers draw similar conclusions, showing that the the distribution of geotagged tweets over the US population is not random, and that higher usage is correlated with urban areas, high income, more ethnic minorities, and more young people. | this prior work did not consider the biases introduced by relying on geotagged messages, nor the consequences for geo-linguistic analysis. | contrasting |
train_14789 | As described in Section 3.2, the distant supervision labels were based on a linear combination of three heuristics that achieved at best an RMSE of 5.1 sentences. | with self-training, we can exploit the noisy heuristic labels by using only those labels that agree with the seed-trained model, thus reducing the amount of noise. | contrasting |
train_14790 | In this work, we also consider sentences as units of suggestion. | we observe that sentences might miss the context, or refer to something mentioned in the previous sentence. | contrasting |
train_14791 | And reposting messages, namely reposts, can provide valuable context information to the previous posts including their background, development, public opinions and so on. | a popular post usually attracts a large number of reposts. | contrasting |
train_14792 | A simple way to detect leaders on repost tree is to directly apply a binary classifier like SVM on each individual message. | these models assume reposts are independent without effectively leveraging abundant context along the repost tree paths, such as the reposting relations among different reposts on a path. | contrasting |
train_14793 | Therefore, this makes it possible to include true leaders misclassified as followers by leader detection module into summary. | allowing all messages to participate in ranking also increases the risk of selecting real followers. | contrasting |
train_14794 | Traditional approaches to zero anaphora resolution are based on manually created heuristic rules (Kameyama, 1986;Walker et al., 1994;Okumura and Tamura, 1996;Nakaiwa and Shirai, 1996), which are mainly motivated by the rules and preferences introduced in Centering Theory (Grosz et al., 1995). | the research trend of zero anaphora resolution has shifted from such rule-based approaches to machine learningbased approaches because in machine learning we can easily integrate many different types of information, such as morpho-syntactic, semantic and discourse-related information. | contrasting |
train_14795 | In these methods, the semantic compatibility between the contexts surrounding an anaphor and its antecedent (e.g., the compatibility of verbs kidnap and release given some arguments) was automatically extracted from raw texts in an unsupervised manner and used as features in a machine learning-based approach. | because the automatically acquired semantic compatibility is not always true or applicable in the context of any pair of an anaphor and its antecedent, the effectiveness of the compatibility features might be weakened. | contrasting |
train_14796 | Another important point is that in the NAIST Text Corpus, if the antecedent of a zero anaphor is not explicitly written in the corpus, it is simply annotated as 'exophoric', and the subject sharing relations between two predicates whose subject was annotated as exophoric cannot be captured. | in our cleaning procedure, the annotators additionally annotated such 'exophoric' subject sharing relations to take into account all subject sharing relations in the corpus. | contrasting |
train_14797 | When combining more than one subject sharing recognizer in our method, we construct the SSPN using the subject sharing relations recognized by at least one of those recognizers for transitive subject propagation. | in the baseline method, the SSPN was not constructed and zero anaphoric relations were identified using only the outputs of our subject detector and one of those recognizers. | contrasting |
train_14798 | The results are shown in Table 5 and demonstrate that the performance of all the methods without Step 5 does not reach that of the baseline method in F-score. | they retain high precision that ranges from 60% to 75%, preserving more than 10% of the recall on the DEP, DEP+ADJ, DEP+PNP and DEP+ADJ+PNP methods. | contrasting |
train_14799 | Given labeled data, supervised learning can be applied to obtain sentiment weights for each word. | the effectiveness of supervised sentiment analysis depends on having training data in the same domain as the target, and this is not always possible. | contrasting |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.