id
stringlengths 7
12
| sentence1
stringlengths 6
1.27k
| sentence2
stringlengths 6
926
| label
stringclasses 4
values |
|---|---|---|---|
train_3100 | Prior work has generated sentences that describe 2D images (Farhadi et al., 2010;Kulkarni et al., 2011;Karpathy et al., 2014) and referring expressions for specific objects in images (FitzGerald et al., 2013;Kazemzadeh et al., 2014). | generating scenes is currently out of reach for purely image-based approaches. | contrasting |
train_3101 | We see that for the simple seed sentences both the rule-based and combined model approach the quality of human-created scenes. | all methods have significantly lower ratings for the more complex MTurk sentences. | contrasting |
train_3102 | Thus, we expect that our method for lexical grounding can facilitate development of text-to-scene systems in other languages. | additional data collection and experiments are necessary to confirm this and identify challenges specific to other languages. | contrasting |
train_3103 | Instead of computing the similarity directly as we do for DIRECTSIM, we first transform the representation of the g-phrase in one TEXTCHUNK using a transformation matrix M ∈ R d×d , then compute the match score by inner product and sigmoid activation: Our motivation is that for a TEXTCHUNK relation like clause coherence, the two TEXTCHUNKS need not have any direct similarity. | if we map the representations of TEXTCHUNK S 1 into an appropriate space then we can hope that similarity between these transformed representations of S 1 and the representations of TEXTCHUNK S 2 do yield useful features. | contrasting |
train_3104 | Titov and McDonald (2008) emphasize the importance of an unsupervised approach for aspect detection. | the authors also indicate that standard LDA (Blei et al., 2003) methods capture global topics and not necessarily pertinent aspects -a challenge that we address in this work. | contrasting |
train_3105 | We found that in 63% of these missed predictions, Seed-edLDA predicts LECTURE-CONTENT, instead of QUIZ-CONTENT, and vice versa. | pSL-Joint uses both coarse and fine SeededLDA scores and captures the dependency between a coarse aspect and its corresponding fine aspect. | contrasting |
train_3106 | The only requirement is that the learned embeddings should be compatible within each individual fact. | they fail to discover the intrinsic geometric structure of the embedding space. | contrasting |
train_3107 | With enough GRE data, the prediction error may be backpropagated to directly adjust or learn embedding in the look-up tables. | given the limited size of the GRE data, we only employed the top hidden layers to non-linearly merge the distances between a word pair that are obtained within each of the modules in the Contrast Inference Layer. | contrasting |
train_3108 | Without the sampling, the F-score achieved on the test data is 83%. | the findings presented above should not be simply taken as that distributional hypothesis is not useful for learning lexical contrast. | contrasting |
train_3109 | Our regularizer framework is able to propagate weights from the conjunctive features seen in training to unseen conjunctive features that are close to each other in the projected space (these are the yellow and red cells in the matrix). | 1 and 2 regularization techniques can not put weight on unseen conjunctions. | contrasting |
train_3110 | Most similarly to our work, Weiss and Taskar (2013) improve performance for several structured vision tasks by dynamically selecting features at runtime. | they use a reinforcement learning approach whose computational tradeoffs are better suited to vision problems with expensive features. | contrasting |
train_3111 | Searn (Daumé III et al., 2009) and DAgger (Ross et al., 2011) are two popular principled frameworks for reducing sequential prediction to classification by learning a classifier on additional synthetic training data. | as we do in our experiments, practitioners often see good results by training on the gold standard labels with an off-the-shelf classification algorithm, as though classifying IID data (Bengtson and Roth, 2008;Choi and Palmer, 2012). | contrasting |
train_3112 | (These two requirements are both incorporated into our novel parameter estimation algorithm.) | other work (Weiss and Taskar, 2013;He et al., 2013) learns a separate classifier to determine when to add features. | contrasting |
train_3113 | (2012) greatly increase available raw material for paths by augmenting KB-schema relations with relations defined by the text connecting mentions of entities in a large corpus (also known as OpenIE relations (Banko et al., 2007)). | these symbolic methods can produce many millions of distinct paths, each of which is categorically distinct, treated by PRA as a dis-tinct feature. | contrasting |
train_3114 | We introduce a word-representation model to capture meaningful semantic regularities for words and adopt a framework based on a convolutional neural network (CNN) to capture sentence-level clues. | cNN can only capture the most important information in a sentence and may miss valuable facts when considering multiple-event sentences. | contrasting |
train_3115 | In S1, beats is a trigger of type Elect. | in S2, beats is a trigger of type Attack, which is more common than type Elect. | contrasting |
train_3116 | e i is another embedding for e i , see Morin and Bengio (2005) The CNN, with max-pooling layers, is a good choice to capture the semantics of long-distance words within a sentence (Collobert et al., 2011). | as noted in the section 1, traditional CNN is incapable of addressing the event extraction problem. | contrasting |
train_3117 | To extract the most important features (max value) within each feature map, traditional CNNs (Collobert et al., 2011;Kim, 2014;Zeng et al., 2014) take one feature map as a pool and only get one max value for each feature map. | single max-pooling is not sufficient for event extraction. | contrasting |
train_3118 | (2013) as the traditional method, which uses traditional lexical features, such as n-grams, POS tags and some entity information. | we only use word embedding as our lexical feature. | contrasting |
train_3119 | The RPI BLENDER KBP system (Yu et al., 2014) casts SFV in this framework, using a graph propagation method that modeled the credibility of systems, sources, and response values. | they only report scores on the 2013 SFV data which contain less complicated and easier queries compared to the 2014 data. | contrasting |
train_3120 | Another direction is to use more complex translation models such as hierarchical models (Chiang, 2007). | these approaches suffer from the long-distance reordering issue and computational complexity. | contrasting |
train_3121 | Statistical syntactic parsers are usually trained on tree-annotated corpora. | corpora annotated with BTG parse trees are unavailable, and only the gold standard permutation y is available. | contrasting |
train_3122 | Experimental results suggest the effectiveness of online learning as a way to exploit user feedback to tailor QE predictions to their quality standards and to cope with the heterogeneity of data coming from different domains. | though robust to user and domain changes, the method is solely driven by the distance computed between predicted and true labels, and it does not exploit any notion of similarity between tasks (e.g. | contrasting |
train_3123 | The methods in these two strands have shown their effectiveness on lexical selection. | correlations between sentence-and document-level contexts have never been explored before. | contrasting |
train_3124 | The politics topic can be further used to enable the decoder to select a correct translation "issue" for another source word "K /wèntǐ", which is consistent with this topic. | if we know that this document mainly focuses on the politics topic, the candiate translation "stance" will be more compatible with the context of "á|/lìchǎng" than the candiate translation "attitude". | contrasting |
train_3125 | QA has had a long history of using pipeline models that extract a limited number of high-level features from induced representations of questionanswer pairs, and then built a classifier using some labelled corpora. | we learnt these structures and performed machine comprehension jointly through a unified max-margin framework. | contrasting |
train_3126 | Previous work in the literature has demonstrated the effectiveness of the category information for question retrieval (Cao et al., 2010;Zhou et al., 2013). | we argue that the category information benefits the word embedding learning in this work. | contrasting |
train_3127 | (2014) used convolutional neural networks to answer single-relation questions on REVERB (Fader et al., 2011). | the system worked on relation-entity triples instead of more structured knowledge bases. | contrasting |
train_3128 | Its correct response should contain only one entity (vivian liberto). | our system addi-tionally outputs june carter cash who is his second wife, because both the candidate answers are connected to johnny cash by the relation people.person.spouse s. In order to solve this issue, we need to define some ad-hoc operators used for comparisons or develop more advanced semantic representations. | contrasting |
train_3129 | The first row of Table 5 confirms that the hubs found in the neighbourhoods of ridgemapped query terms are items that tend to be closer to the search space mean vector, and that this effect is radically reduced with max-margin estimation. | the second row of the table shows another factor at play, that has a major role in the cross-modal setting, and it is only partially addressed by max-margin estimation: Namely, in vision-to-language mapping, there is a strong tendency for hubs (that, recall, have an important effect on performance, as they enter many nearest neighbour lists) to be close to a training data point. | contrasting |
train_3130 | However, it is not trivial to combine the algebraic objects created by distributional semantics to get a sensible distributional representation for more complex expressions, consisting of several words. | the formalism of the λ -calculus provides us with general, advanced and efficient methods for composition that can model meaning composition not only of simple phrases, but also more complex phenomena such as coercion or composition with fine-grained types (Asher, 2011;Luo, 2010;Bassac et al., 2010). | contrasting |
train_3131 | • Previous research has indicated that the lexical function approach is able to achieve better results using a reduced space with svd. | the negative values that result from svd are detrimental for the multiplicative approach. | contrasting |
train_3132 | The authors evaluated their models using a similarity task that is similar to the one used by Mitchell & Lapata. | they use more complex compositional expressions: rather than using compositions of two words (such as a verb and an object), they use simple transitive phrases (subject-verbobject). | contrasting |
train_3133 | Because the number of rule indicators n o is fairly large (approximately 4000 in the Penn Treebank), the multiplication by W in the model is also expensive. | because only a small number of rules can apply to a given span and split point, f o is sparse and we can selectively compute the terms necessary for the final bilinear product. | contrasting |
train_3134 | Given a sentence x, graph-based models formulates the parsing process as a searching problem: where y * (x) is tree with highest score, Y (x) is the set of all trees compatible with x, θ are model parameters and Score(x,ŷ(x); θ) represents how likely that a particular treeŷ(x) is the correct analysis for x. | the size of Y (x) is exponential large, which makes it impractical to solve equation (1) directly. | contrasting |
train_3135 | The simplest subgraph uses a first-order factorization (McDonald et al., 2005) which decomposes a dependency tree into single dependency arcs ( The most common choice for ScoreF (x, c; θ), which is the score function for subgraph c in the tree, is a simple linear function: where f (x, c) is the feature representation of subgraph c and w is the corresponding weight vector. | the effectiveness of this function relies heavily on the design of feature vector f (x, c). | contrasting |
train_3136 | Le and Zuidema (2014) proprosed an infinite-order model based on recursive neural network. | their model can only be used as an reranking model since decoding is intractable. | contrasting |
train_3137 | Furthermore, because the neural network uses a distributed representation, it is able to model lexical, part-of-speech (POS) tag, and arc label similarities in a continuous space. | although their model outperforms its greedy hand-engineered counterparts, it is not competitive with state-of-the-art dependency parsers that are trained for structured search. | contrasting |
train_3138 | As in prior work, we train the neural network to model the probability of individual parse actions. | we do not use these probabilities directly for prediction. | contrasting |
train_3139 | The word embeddings capture similar distributional information as word clusters and give consistent improvements by providing a good initialization and information about words not seen in the treebank data. | obtaining more training data is even more important than a good initialization. | contrasting |
train_3140 | These sentences are of course easier to parse, having an average length of 15 words, compared to 24 words for the tune set overall. | because we only use these sentences to extract individual transition decisions, the shorter length does not seem to hurt their utility. | contrasting |
train_3141 | (1992), which added an external stack memory to an RNN. | our architecture provides an embedding of the complete contents of the stack, whereas theirs made only the top of the stack visible to the RNN. | contrasting |
train_3142 | Here, we learned from observable stack manipulation operations (i.e., supervision from a treebank), and the computed embeddings of final parser states were not used for any further prediction. | this could be reversed, giving a device that learns to construct context-free programs (e.g., expression trees) given only observed outputs; one application would be unsupervised parsing. | contrasting |
train_3143 | However, removing the nsubj edge in For most dependencies, this semantics can be hard-coded with high accuracy. | there are at least two cases where more attention is warranted. | contrasting |
train_3144 | A standard pipeline for statistical relational learning involves two steps: one first constructs the knowledge base (KB) from text, and then performs the learning and reasoning tasks using probabilistic first-order logics. | a key issue is that information extraction (IE) errors from text affect the quality of the KB, and propagate to the reasoning task. | contrasting |
train_3145 | For example, prepositional phrases (PPs) express crucial information that IE methods need to extract. | pps are a major source of syntactic ambiguity. | contrasting |
train_3146 | The Collins baseline achieves 84% accuracy on the benchmark Wall Street Journal PP dataset. | drawing a distinction in the precision of different prepositions provides useful insights on its performance. | contrasting |
train_3147 | If "with" is used in the sense of "contains" , then the PP is a likely noun attachment, as in n1 contains n2 in the quad ate, cookies, with, cranberries. | if "with" is used in the sense of "accompanied by", then the PP is a likely verb attachment, as in the quad visted, P aris, with, Sue. | contrasting |
train_3148 | To estimate these parameters, we could use the labeled data as training data and use standard gradient descent to minimize the logistic regression cost function. | we also leverage the unlabeled data. | contrasting |
train_3149 | Both support comparisons, and ideally we can detect some level of similarity. | if we consider only sub-trees, the two dependency trees share in common only two fragments: [#camera] and [is]. | contrasting |
train_3150 | PT kernel would find that the parse in Figure 2(a) with all its subgraphs can be matched as a whole within the parse in Figure 2(b), identifying a close match. | pT kernel is prone to two drawbacks. | contrasting |
train_3151 | usually results in an SRL that generates densely labeled sentences, resulting in better recall but poorer precision. | training data that is sparsely labeled results in an SRL that weighs the option of not assigning a label with higher probability, resulting in better precision and poorer recall. | contrasting |
train_3152 | (2013) applied projection learning for opinion mining in Chinese. | this work only addresses agent detection and requires translating the MPQA corpus. | contrasting |
train_3153 | We ran experiments using the same data and MPQA partitions as Johansson and Moschitti (2013). | since our system is designed for predicting opinion, agents and targets together, we removed the documents that were not annotated with targets. | contrasting |
train_3154 | The work in (Gui et al., 2014) follows the same line although it considers knowledge transferring between two languages. | the main focus of our work is to filter out the noisy knowledge having sentiment changes by wrong translations. | contrasting |
train_3155 | It is reasonable that a Chinese classifier performs better on Chinese text than an English classifier performs on the translated English text due to the different language distributions and MT errors. | as shown in Tables 3 and 4, the better performance of our proposed method compared with that of the self-boosting method further suggests the effectiveness of our proposed knowledge validation model. | contrasting |
train_3156 | Directly employing the translated resources for sentiment classification in the target language is simple and could get acceptable results. | the gap between the source language and target language inevitably impacts the performance of sentiment classification. | contrasting |
train_3157 | Specifically, for a given document pair (x E , x C ), we can learn bilingual embeddings to reconstruct The sum of 5 losses is used as the loss function of bilingual embeddings: In the unsupervised phase, we have learned the bilingual embeddings, which could capture the semantic information within and across languages. | the sentiment polarities of text are ignored in the unsupervised phase. | contrasting |
train_3158 | In this paper, we only evaluate BSWE when dimension d varies from 50 to 500. | there is still space for further improvement if d continues to increase. | contrasting |
train_3159 | We can see that the distribution of factoids relating to venues, methodology and applications is similar for the two datasets. | factoids related to definitional sentences are almost completely missing in the citing sentences data. | contrasting |
train_3160 | Scientific summarization: Early work in scientific summarization used abstracts of scientific articles to produce summaries of specific scientific papers (Kupiec et al., 1995). | later work (Elkiss et al., 2008) showed that citation sentences are as important in understanding the main contributions of a paper. | contrasting |
train_3161 | In the basic version described above, the scorer is trained to score full sentence plan trees. | it is also used to score incomplete sentence plans during the decoding. | contrasting |
train_3162 | Extractive models can generate more readable headlines, because the final title is derived by tailoring human-written sentences. | extractive models give less informative titles (Alfonseca et al., 2013), because sentences are very sparse, making high-recall candidate extraction difficult. | contrasting |
train_3163 | They leverage titles of sub-documents for supervised training. | we generate a title for a single document using an unsupervised model. | contrasting |
train_3164 | The proposed techniques in Section 4 and 5 are generic methodologies and not tied to any particular models such as any sequence models and instanced based models. | because of superior performance over CRF, we use a hidden unit CRF (HUCRF) of Maaten et al. | contrasting |
train_3165 | For instance, one could simply use the sparse feature vector representation of each label. | ccA's low-dimensional projection is computationally more convenient and arguably more generalizable. | contrasting |
train_3166 | In particular, proposed a frame-semantics based framework for automatically inducing semantic slots given raw audios. | these approaches generally do not explicitly learn the latent factor representations to model the measurement errors (Skrondal and Rabe-Hesketh, 2004), nor do they jointly consider the complex lexical, syntactic, and semantic relations among words, slots, and utterances. | contrasting |
train_3167 | Following their success of unsupervised SLU, recent studies have also obtained interesting results on the tasks of relation detection Chen et al., 2014a), entity extraction (Wang et al., 2014), and extending domain coverage (El-Kahky et al., 2014;Chen and Rudnicky, 2014). | most of the studies above do not explicitly learn latent factor representations from the data-while we hypothesize that the better robustness in noisy data can be achieved by explicitly modeling the measurement errors (usually produced by automatic speech recognizers (ASR)) using latent variable models and taking additional local and global semantic constraints into account. | contrasting |
train_3168 | fluency features can help capture long-range dependencies for disfluency detection. | the UT model does not perform as well as BCT. | contrasting |
train_3169 | But this model suffers from the problem that the number of transition actions is not identical for different hypotheses in decoding, leading to the failure of performing optimal search. | our novel right-to-left transition-based joint method caters to the disfluency constraint which can not only overcome the decoding deficiency in previous work but also achieve significantly higher performance on disfluency detection as well as dependency parsing. | contrasting |
train_3170 | On one hand, an entity is modeled to have multiple internal representations, each regarding one or more closely related facts. | an entity query is decomposed into one or more subqueries, each describing a fact about target entities. | contrasting |
train_3171 | Structure penalty The depth of a factoid node indicates the level of information granularity. | we also need to control the depth of the factoid hierarchy. | contrasting |
train_3172 | One possible reason is that Wikipedia articles contain much irrelevant information regarding entities, and these two existing models cannot easily make use of additional information. | with Wikipedia full-text available, both of our proposed models achieve obviously better performances. | contrasting |
train_3173 | Typical models include voting model (Macdonald and Ounis, 2006), graph model (Serdyukov et al., 2008), etc. | it is not easy to generalize these models for open domain entity retrieval. | contrasting |
train_3174 | Our work is also related to the information extraction and knowledge representation field since our framework involves extraction and aggregation of knowledge from free texts. | most existing approaches takes two extreme ways: either extract relations based on pre-defined ontology, such as DBpedia (Lehmann et al., 2014); or cluster relation without referring to some ontology, such as OpenIE (Etzioni et al., 2011). | contrasting |
train_3175 | Another line of work tries to improve the accuracy by introducing ontologies (Fodeh et al., 2011;Kumar and Salim, 2012) and structured knowledge bases such as WordNet (Nastase et al., 2010), which provide semantic information about words such as synonym and antonym (Sansonnet and Bouchet, 2010). | these methods primarily rely on special resources constructed with supervision or even manually, which are difficult to expand and in turn limit their applications in practice. | contrasting |
train_3176 | As we will see below, S1, S2, and S3 are error metrics, so lower scores on them imply better performance. | p C is a correlation metric, so higher correlation implies better performance. | contrasting |
train_3177 | For example, one of the groups on the "Matrix" disambiguation page has a name "Business and government" and there is no Wikipedia category by that name. | the group names generated by our (and baseline) method are from the Wikipedia categories (which forms our topic DAG). | contrasting |
train_3178 | 8 Our method is robust across different relationships in terms of NDCG. | we observe some variation in ERR as this metric is more sensitive to the distribution of relevant items than NDCG-the distribution over relevance grades varies per relationship type. | contrasting |
train_3179 | Our approach consists of the following steps: (1) generating high quality embedding for each training instance; (2) pre-training with the stacked denoising auto-encoder (Bengio et al., 2003) for feature dimension reduction; and (3) supervised fine-tuning to optimize the neural networks towards a similarity measure (e.g., dot product). | morph resolution is significantly different from the traditional entity linking task since the latter mainly focuses on formal and explicit entities (e.g., "薄 熙 来 (Bo Xilai)") which tend to have stable referents in Wikipedia. | contrasting |
train_3180 | We found that (Huang et al., 2013) failed to resolve many unpopular morphs (e.g., "小马 (Little Ma)" is a morph referring to Ma Yingjiu, and it only appeared once in the data), because it heavily relies on aggregating contextual and temporal information from multiple instances of each morph. | our unsupervised resolution approach only leverages the pre-trained word embeddings to capture the semantics of morph mentions and entities. | contrasting |
train_3181 | (2011) recognize the importance of combining both local and global context for robust disambiguation. | their approach is limited to EL and optimization is performed in a discrete setting. | contrasting |
train_3182 | Note that our formulation could easily be extended to using additional coefficients for each ob-jective. | these hyper-parameters would have to be estimated on development data and therefore, this method could hurt generalization. | contrasting |
train_3183 | They should not be applied during disambiguation because these priors can be very strong and are not domain independent. | they provide a good initialization which is important for successful continuous optimization. | contrasting |
train_3184 | On the other hand, our system is slightly worse on the KORE dataset compared to Babelfy (6 errors more in total), which might be due to the strong priors and the small context. | the dataset is rather small, containing only 50 sentences, and has been artificially tailored to the use of highly ambiguous entity mentions. | contrasting |
train_3185 | Because the corpora contain more text we find significantly higher overlap with the different keyphrase corpora. | this comes at the cost of not being able to isolate the domainspecific keyphrases. | contrasting |
train_3186 | This is an advantage over softmax classifiers. | sampling informative negative classes/examples can have a significant impact in the effectiveness of the learned model. | contrasting |
train_3187 | This effect of WPEs is reported by (Zeng et al., 2014). | when using only the text span between the target nouns, the impact of WPE is much smaller. | contrasting |
train_3188 | follow the supervised DA paradigm and assume some labeled data in the target domains. | plank and Moschitti (2013) and Nguyen and Grishman (2014) work on the unsupervised DA. | contrasting |
train_3189 | (2014; study the factor-based compositional embedding models. | none of this work examines word embeddings for tree kernels as well as domain adaptation as we do. | contrasting |
train_3190 | The research on negation and speculation identification in English has received a noticeable boost. | in contrast to the significant achievements concerning English, the research progress in Chinese language is quite limited. | contrasting |
train_3191 | Currently, mainstream approaches formulated the scope resolution as a chunking problem, which classifies every word of a sentence as being inside or outside the scope of a cue. | unlike in English, we found that plenty of errors occurred in Chinese scope resolution by using words as the basic identifying candidate. | contrasting |
train_3192 | In this paper, we are concerned with manipulation actions, that is actions performed by agents (humans or robots) on objects, resulting in some physical change of the object. | most of the current AI systems require manually defined semantic rules. | contrasting |
train_3193 | In total, we are able to collect 90 testing detections and they serve as the testing corpus. | since many of the objects used in the testing data are not present in the training set, an object model-free approach is adopted and thus "subject" and "patient" fields are filled with segment IDs instead of a specific object name. | contrasting |
train_3194 | This by itself is a very interesting research topic and it is beyond this paper's scope. | by applying a couple of common sense Axioms on the testing data, we can provide some flavor of this idea. | contrasting |
train_3195 | Here we want to mention a caveat: there are definitely other common sense Axioms that we are not able to address in the current implementation. | from the case studies presented, we can see that using the presented formal framework, our system is able to reason about manipulation action goals instead of just observing what is happening visually. | contrasting |
train_3196 | Both in TransE and TransH, the embeddings of entities and relations are in the same space. | entities and relations are different types objects, it is insufficient to model them in the same space. | contrasting |
train_3197 | In both TransH and TransR/CTransR, all types of entities share the same mapping vectors/matrices. | different types of entities have different attributes and functions, it is insufficient to let them share the same transform parameters of a relation. | contrasting |
train_3198 | The most likely reason for this is that, as the correction of these error types typically only involves the addition or removal of an -s suffix, i.e., a minor change in number morphology, there is very little room for annotators to disagree. | the next highest category, article and determiner errors (ArtOrDet), has a slightly larger confusion set, {the, a/an, }, which may account for the slightly lower score. | contrasting |
train_3199 | Traditional approaches to word sense disambiguation (WSD) rest on the assumption that there exists a single, unambiguous communicative intention underlying every word in a document. | writers sometimes intend for a word to be interpreted as simultaneously carrying multiple distinct meanings. | contrasting |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.