A Latent Discriminative Model for Compositional Entailment Relation Recognition using Natural Logic

Proceedings of COLING 2012: Technical Papers, pages 2805–2820, COLING 2012, Mumbai, December 2012. A Latent Discriminative Model for Compositional Entailment Relation Recognition Using Natural Logic Yotaro Watanabe1 Junta Mizuno1 Eric Nichols1 Naoaki Okazaki1,2 Kentaro Inui1 (1) Graduate School of Information Sciences, Tohoku University (2) Precursory Research for Embryonic Science and Technology, Japan Science and Technology Agency {yotaro-w,junta-m,eric,okazaki,inui}@ecei.tohoku.ac.jp ABSTRACT Recognizing semantic relations between sentences, such as entailment and contradiction, is a challenging task that requires detailed analysis of the interaction between diverse linguistic phenomena. In this paper, we propose a latent discriminative model that unifies a statistical framework and a theory of Natural Logic to capture complex interactions between linguistic phenomena. The proposed approach jointly models alignments, their local semantic relations, and a sentence-level semantic relation, and has hidden variables including alignment edits between sentences and their semantic relations, only requires sentences pairs annotated with sentence-level semantic relations as training data to learn appropriate alignments. In evalu- ation on a dataset including diverse linguistic phenomena, our proposed method achieved a competitive results on alignment prediction, and significant improvements on a sentence-level semantic relation recognition task compared to an alignment supervised model. Our analysis did not provide evidence that directly learning alignments and their labels using gold standard alignments contributed to semantic relation recognition performance and instead suggests that they can be detrimental to performance if used in a manner that prevents the learning of globally optimal alignments. KEYWORDS: Recognizing Textual Entailment, Natural Logic, Latent Variable Model. KEYWORDS IN L2: . 2805 1 Introduction Recognizing Textual Entailment (RTE) (Dagan et al., 2005) is the task of recognizing entailment relations between a given text pair, Text T and Hypothesis H. RTE is useful for many information access tasks that depend on natural language processing technologies, and a breakthrough would lead to significant progress in information retrieval, document summarization, and question answering, among other tasks. The majority of approaches proposed in previous work recognize entailment relations between a pair of texts by capturing lexical or structural correspondences. Methods include simple word overlap-based measures (Jijkoun and de Rijke, 2005) as well as alignment of syntactic and semantic dependencies (Sammons et al., 2009; Wang and Zhang, 2009). However, sentence- level semantic relations are affected by various linguistic phenomena: not only lexical semantic relations (synonyms, antonyms) but also monotonicity (e.g. downward-monotone caused by scope of negation), implicative/factive expressions, quantifiers, etc. Thus similarity measures are insufficient to capture these phenomena and their interactions. Transformation-based approaches are one way to capture the affects of diverse linguistic phenomena and their interactions, where a set of linguistic phenomena are decomposed into units. By doing so it becomes possible to consider their effects on entailment independently. A number of previous works explores transformation-based entailment relation recognition. The approach of Stern et al. (2011) recognizes a sentence-level semantic relation through a proof which represents a sequence of edits from T to H produced by applying various entailment rules and the operations such as insertion, deletion, moving subtrees, etc. In addition, Heilman and Smith …

Content truncated to 500 words