On-demand Injection of Lexical Knowledge for Recognising Textual Entailment

Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 710–720, Valencia, Spain, April 3-7, 2017. c⃝2017 Association for Computational Linguistics On-demand Injection of Lexical Knowledge for Recognising Textual Entailment Pascual Mart´ınez-G´omez1 pascual.mg@aist.go.jp Koji Mineshima2 mineshima.koji@ocha.ac.jp Yusuke Miyao1,3,4 yusuke@nii.ac.jp Daisuke Bekki1,2,3 bekki@is.ocha.ac.jp 1Artificial Intelligence Research Center, AIST 2Ochanomizu University 3National Institute of Informatics and JST, PRESTO 4The Graduate University for Advanced Studies (SOKENDAI) Tokyo, Japan Abstract We approach the recognition of textual en- tailment using logical semantic represen- tations and a theorem prover. In this setup, lexical divergences that preserve semantic entailment between the source and target texts need to be explicitly stated. However, recognising subsentential semantic rela- tions is not trivial. We address this prob- lem by monitoring the proof of the the- orem and detecting unprovable sub-goals that share predicate arguments with logi- cal premises. If a linguistic relation exists, then an appropriate axiom is constructed on-demand and the theorem proving con- tinues. Experiments show that this ap- proach is effective and precise, produc- ing a system that outperforms other logic- based systems and is competitive with state-of-the-art statistical methods. 1 Introduction Recognising Textual Entailment (RTE) is a chal- lenging NLP application where the objective is to judge whether a text fragment H logically fol- lows from another text fragment T (Dagan et al., 2013). Advances in RTE have potentially positive implications in other areas such as fact checking, question-answering or information retrieval. So- lutions to the RTE problem span a wide array of methods. Some methods are purely statistical (Lai and Hockenmaier, 2014; Zhao et al., 2014), where a classifying function is estimated using lexical or syntactic features. Other methods are purely se- mantic (Bos et al., 2004), where logical formulas that represent the text fragments are constructed and used in a formal proof system. And yet others are hybrid systems (Beltagy et al., 2013), where a combination of statistical features and logical for- mulas are used to judge entailment relations. In this paper, we adopt a strategy based on logics, encouraged by the high-performance that these systems achieve in linguistically challeng- ing datasets (Abzianidze, 2015; Mineshima et al., 2015). An important advantage of these systems (including ours) is that they are unsupervised, thus no training data is necessary and no parameters need to be adjusted. Under the perspective of these logic-based sys- tems, there are mainly two associated challenges when solving RTE problems. The first challenge is to model the logics of the language with the pur- pose to represent the semantics of text fragments accurately. To this end, we follow the standard practice in formal semantics where the meaning of sentences is represented using logical formulas. The second challenge is to account for lexical re- lations between text fragments, typically between words or non-compositional phrases. We dedicate our efforts to the latter challenge, and assume that wide coverage linguistic resources are available to signal potential relations between lexical items in text fragments. The question is …

Content truncated to 500 words