Computing Paraphrasability of Syntactic Variants Using Web Snippets

Computing Paraphrasability of Syntactic Variants Using Web Snippets Atsushi Fujita Satoshi Sato Graduate School of Engineering, Nagoya University {fujita,ssato}@nuee.nagoya-u.ac.jp Abstract In a broad range of natural language pro- cessing tasks, large-scale knowledge-base of paraphrases is anticipated to improve their performance. The key issue in creating such a resource is to establish a practical method of computing semantic equivalence and syn- tactic substitutability, i.e., paraphrasability, between given pair of expressions. This paper addresses the issues of computing paraphrasability, focusing on syntactic vari- ants of predicate phrases. Our model esti- mates paraphrasability based on traditional distributional similarity measures, where the Web snippets are used to overcome the data sparseness problem in handling predicate phrases. Several feature sets are evaluated through empirical experiments. 1 Introduction One of the common characteristics of human lan- guages is that the same concept can be expressed by various linguistic expressions. Such linguistic vari- ations are called paraphrases. Handling paraphrases is one of the key issues in a broad range of natu- ral language processing (NLP) tasks. In informa- tion retrieval, information extraction, and question answering, technology of recognizing if or not the given pair of expressions are paraphrases is desired to gain a higher coverage. On the other hand, a sys- tem which generates paraphrases for given expres- sions is useful for text-transcoding tasks, such as machine translation and summarization, as well as beneficial to human, for instance, in text-to-speech, text simplification, and writing assistance. Paraphrase phenomena can roughly be divided into two groups according to their compositionality. Examples in (1) exhibit a degree of compositional- ity, while each example in (2) is composed of totally different lexical items. (1) a. be in our favor ⇔ be favorable for us b. show a sharp decrease ⇔ decrease sharply (Fujita et al., 2007) (2) a. burst into tears ⇔ cried b. comfort ⇔ console (Barzilay and McKeown, 2001) A number of studies have been carried out on both compositional (morpho-syntactic) and non- compositional (lexical and idiomatic) paraphrases (see Section 2). In most research, paraphrases have been represented with the similar templates, such as shown in (3) and (4). (3) a. N1 V N2 ⇔ N1’s V -ing of N2 b. N1 V N2 ⇔ N2 be V -en by N1 (Harris, 1957) (4) a. X wrote Y ⇔ X is the author of Y b. X solves Y ⇔ X deals with Y (Lin and Pantel, 2001) The weakness of these templates is that they should be applied only in some contexts. In other words, the lack of applicability conditions for slot fillers may lead incorrect paraphrases. One way to specify the applicability condition is to enumer- ate correct slot fillers. For example, Pantel et al. (2007) have harvested instances for the given para- phrase templates based on the co-occurrence statis- tics of slot fillers and lexicalized part of templates (e.g. “deal with” in (4b)). Yet, there is no method which assesses semantic equivalence and syntactic substitutability of resultant pairs of expressions. 537 In this paper, we propose …

Content truncated to 500 words