Visual Denotations for Recognizing Textual Entailment

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2853–2859 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Visual Denotations for Recognizing Textual Entailment Dan Han1 dan.han@aist.go.jp Pascual Mart´ınez-G´omez1 pascual.mg@aist.go.jp Koji Mineshima2 mineshima.koji@ocha.ac.jp 1Artificial Intelligence Research Center, AIST 2Ochanomizu University Tokyo, Japan Abstract In the logic approach to Recognizing Textual Entailment, identifying phrase-to- phrase semantic relations is still an un- solved problem. Resources such as the Paraphrase Database offer limited cover- age despite their large size whereas unsu- pervised distributional models of meaning often fail to recognize phrasal entailments. We propose to map phrases to their visual denotations and compare their meaning in terms of their images. We show that our approach is effective in the task of Recog- nizing Textual Entailment when combined with specific linguistic and logic features. 1 Introduction and Related Work Recognizing Textual Entailment (RTE) is a chal- lenging task that was described as the best way of testing an NLP system’s semantic capacity (Cooper et al., 1994). In this task, given a text T and a hypothesis H, the objective is to recognize whether T implies H (yes), whether T contradicts H (no) or otherwise (unk). For example, given: (T) Some men walk in the tall and green grass. (H) Some people walk in the field. the system needs to recognize that T implies H (yes). Although humans can easily solve these problems, machines face great difficulties (Dagan et al., 2013). RTE has been approached from dif- ferent perspectives, ranging from purely statistical systems (Lai and Hockenmaier, 2014; Zhao et al., 2014) to purely logical (Bos et al., 2004; Abzian- idze, 2015; Mineshima et al., 2015) and hybrid systems (Beltagy et al., 2013). We evaluate our idea on top of a logic system since they generally offer a high precision and in- terpretability, which is useful to our purposes. In this approach, there are two main challenges. The first challenge is to model the logical semantic composition of sentences guided by the syntax and logical words (e.g. most, not, some, every). The second challenge is to introduce lexical knowledge that describes the relationship between words or phrases (e.g. men → people, tall and green grass → field). Whereas the relationship men → people can be found in high precision ontological resources such as WordNet (Miller, 1995), phrasal relations such as tall and green grass → field are not avail- able in databases such as the Paraphrase Database (PPDB) (Ganitkevitch et al., 2013) despite their large size. Moreover, although unsupervised dis- tributional similarity models have an infinite do- main (given a compositional function on words), they often fail to identify entailments (e.g. guitar has a high similarity to piano but they do not en- tail each other). To address these issues, Roller et al. (2014) investigated supervised methods to identify word-to-word hypernym relations given word vectors whereas Beltagy et al. (2016) pro- posed a mechanism to extract phrase pairs from T and H and train a classifier to identify para- …

Content truncated to 500 words