Enlarging Paraphrase Collections through Generalization and Instantiation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 631–642, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Enlarging Paraphrase Collections through Generalization and Instantiation Atsushi Fujita Future University Hakodate 116-2 Kameda-nakano-cho, Hakodate, Hokkaido, 041-8655, Japan fujita@fun.ac.jp Pierre Isabelle Roland Kuhn National Research Council Canada 283 Alexandre-Tach´e Boulevard, Gatineau, QC, J8X 3X7, Canada {Pierre.Isabelle, Roland.Kuhn}@nrc.ca Abstract This paper presents a paraphrase acquisition method that uncovers and exploits generali- ties underlying paraphrases: paraphrase pat- terns are first induced and then used to col- lect novel instances. Unlike existing methods, ours uses both bilingual parallel and monolin- gual corpora. While the former are regarded as a source of high-quality seed paraphrases, the latter are searched for paraphrases that match patterns learned from the seed paraphrases. We show how one can use monolingual cor- pora, which are far more numerous and larger than bilingual corpora, to obtain paraphrases that rival in quality those derived directly from bilingual corpora. In our experiments, the number of paraphrase pairs obtained in this way from monolingual corpora was a large multiple of the number of seed paraphrases. Human evaluation through a paraphrase sub- stitution test demonstrated that the newly ac- quired paraphrase pairs are of reasonable qual- ity. Remaining noise can be further reduced by filtering seed paraphrases. 1 Introduction Paraphrases are semantically equivalent expressions in the same language. Because “equivalence” is the most fundamental semantic relationship, techniques for generating and recognizing paraphrases play an important role in a wide range of natural language processing tasks (Madnani and Dorr, 2010). In the last decade, automatic acquisition of knowl- edge about paraphrases from corpora has been draw- ing the attention of many researchers. Typically, the acquired knowledge is simply represented as pairs of semantically equivalent sub-sentential expressions as in (1). (1) a. look like ⇔ resemble b. control system ⇔ controller The challenge in acquiring paraphrases is to ensure good coverage of the targeted classes of paraphrases along with a low proportion of incorrect pairs. How- ever, no matter what type of resource has been used, it has proven difficult to acquire paraphrase pairs with both high recall and high precision. Among various types of corpora, monolingual corpora can be considered the best source for high- coverage paraphrase acquisition, because there is far more monolingual than bilingual text avail- able. Most methods that exploit monolingual cor- pora rely on the Distributional Hypothesis (Harris, 1968): expressions that appear in similar contexts are expected to have similar meaning. However, if one uses purely distributional criteria, it is dif- ficult to distinguish real paraphrases from pairs of expressions that are related in other ways, such as antonyms and cousin words. In contrast, since the work in (Bannard and Callison-Burch, 2005), bilingual parallel corpora have been acknowledged as a good source of high- quality paraphrases: paraphrases are obtained by putting together expressions that receive the same translation in the other language (pivot language). Because translation expresses a specific meaning …

Content truncated to 500 words