Expanding Paraphrase Lexicons by Exploiting Lexical Variants

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 630–640, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Expanding Paraphrase Lexicons by Exploiting Lexical Variants Atsushi Fujita† Pierre Isabelle‡ †National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Souraku-gun, Kyoto, 619-0289, Japan atsushi.fujita@nict.go.jp ‡National Research Council Canada 1200 Montreal Road, Ottawa, Ontario, K1A 0R6, Canada Pierre.Isabelle@nrc.ca Abstract This study tackles the problem of paraphrase acquisition: achieving high coverage as well as accuracy. Our method first induces para- phrase patterns from given seed paraphrases, exploiting the generality of paraphrases exhib- ited by pairs of lexical variants, e.g., “amend- ment” and “amending,” in a fully empiri- cal way. It then searches monolingual cor- pora for new paraphrases that match the pat- terns. This can extract paraphrases compris- ing words that are completely different from those of the given seeds. In experiments, our method expanded seed sets by factors of 42 to 206, gaining 84% to 208% more cover- age than a previous method that generalizes only identical word forms. Human evaluation through a paraphrase substitution test demon- strated that the newly acquired paraphrases re- tained reasonable quality, given substantially high-quality seeds. 1 Introduction One of the characteristics of human languages is that the same semantic content can be expressed using several different linguistic expressions, i.e., para- phrases. Dealing with paraphrases is an important issue in a broad range of natural language process- ing (NLP) tasks (Madnani and Dorr, 2010; Androut- sopoulos and Malakasiotis, 2010). To adequately and robustly deal with paraphrases, a large-scale knowledge base containing words and phrases having approximately the same meaning is indispensable. Thus, the task of automatically creating such large-scale paraphrase lexicons has been drawing the attention of many researchers (see Section 2 for details). The challenge is to en- sure substantial coverage along with high accuracy despite the natural tension between these factors. Among the different types of language resources, monolingual corpora1 offer the largest coverage, but the quality of the extracted candidates is generally rather low. The difficulty lies in the manner of dis- tinguishing paraphrases from expressions that stand in different semantic relations, e.g., antonyms and sibling words, using only the statistics estimated from such corpora. In contrast, highly accurate para- phrases can be extracted from parallel or compara- ble corpora, but their coverage is limited owing to the limited availability of such corpora for most lan- guages. This study aims to improve coverage while main- taining accuracy. To that end, we propose a method that exploits the generality exhibited by pairs of lex- ical variants. Given a seed set of paraphrase pairs, our method first induces paraphrase patterns by gen- eralizing not only identical word forms (Fujita et al., 2012) but also pairs of lexical variants. For in- stance, from a seed pair (1a), a pattern (1b) is ac- quired, where the pair of lexical variants (“amend- ment”, “amending”) and the shared word form “reg- ulation” are generalized. (1) a. amendment …

Content truncated to 500 words