A Probabilistic Model for Measuring Grammaticality and Similarity of Automatically Generated Paraphrases of Predicate Phrases

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 225–232 Manchester, August 2008 A Probabilistic Model for Measuring Grammaticality and Similarity of Automatically Generated Paraphrases of Predicate Phrases Atsushi Fujita Satoshi Sato Graduate School of Engineering, Nagoya University {fujita,ssato}@nuee.nagoya-u.ac.jp Abstract The most critical issue in generating and recognizing paraphrases is development of wide-coverage paraphrase knowledge. Previous work on paraphrase acquisition has collected lexicalized pairs of expres- sions; however, the results do not ensure full coverage of the various paraphrase phenomena. This paper focuses on pro- ductive paraphrases realized by general transformation patterns, and addresses the issues in generating instances of phrasal paraphrases with those patterns. Our prob- abilistic model computes how two phrases are likely to be correct paraphrases. The model consists of two components: (i) a structured N-gram language model that ensures grammaticality and (ii) a distribu- tional similarity measure for estimating se- mantic equivalence and substitutability. 1 Introduction In many languages, a concept can be expressed with several different linguistic expressions. Han- dling such synonymous expressions in a given lan- guage, i.e., paraphrases, is one of the key issues in a broad range of natural language processing tasks. For example, the technology for identifying para- phrases would play an important role in aggregat- ing the wealth of uninhibited opinions about prod- ucts and services that are available on the Web, from both the consumers and producers viewpoint. On the other hand, whenever we draw up a docu- ment, we always seek the most appropriate expres- sion for conveying our ideas. In such a situation, a system that generates and proposes alternative ex- pressions would be extremely beneficial. c⃝ Atsushi Fujita and Satoshi Sato, 2008. Licensed under the Creative Commons Attribution-Noncommercial- Share Alike 3.0 Unported license. Some rights reserved. http://creativecommons.org/licenses/by-nc-sa/3.0/ Most of previous work on generating and recog- nizing paraphrases has been dedicated to develop- ing context-free paraphrase knowledge. It is typi- cally represented with pairs of fragmentary expres- sions that satisfy the following conditions: Condition 1. Semantically equivalent Condition 2. Substitutable in some context The most critical issue in developing such knowledge is ensuring the coverage of the para- phrase phenomena. To attain this coverage, we have proposed a strategy for dividing paraphrase phenomena into the following two classes (Fujita et al., 2007): (1) Non-productive (idiosyncratic) paraphrases a. burst into tears ⇔ cried b. comfort ⇔ console (Barzilay and McKeown, 2001) (2) Productive paraphrases a. be in our favor ⇔ be favorable to us b. show a sharp decrease ⇔ decrease sharply (Fujita et al., 2007) Typical examples of non-productive paraphrases are lexical paraphrases such as those shown in (1) and idiomatic paraphrases of literal phrases (e.g., “kick the bucket” ⇔ “die”). Knowledge of this class of paraphrases should be stored statically, because they cannot be represented with abstract patterns. On the other hand, a productive para- phrase is one having a degree of regularity, as exhibited by the examples in (2). It is therefore reasonable to represent them with a set of general patterns …

Content truncated to 500 words