Minimally Supervised Method for Multilingual Paraphrase Extraction from Definition Sentences on the Web

Proceedings of NAACL-HLT 2013, pages 63–73, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Minimally Supervised Method for Multilingual Paraphrase Extraction from Definition Sentences on the Web Yulan Yan∗ Chikara Hashimoto‡ Kentaro Torisawa§ Takao Kawai¶ Jun’ichi Kazama∥ Stijn De Saeger∗∗ ∗ ‡ § ¶∥ ∗∗ Information Analysis Laboratory Universal Communication Research Institute National Institute of Information and Communications Technology (NICT) {∗ yulan, ‡ ch, § torisawa, ∗∗stijn}@nict.go.jp Abstract We propose a minimally supervised method for multilingual paraphrase extraction from definition sentences on the Web. Hashimoto et al. (2011) extracted paraphrases from Japanese definition sentences on the Web, as- suming that definition sentences defining the same concept tend to contain paraphrases. However, their method requires manually an- notated data and is language dependent. We extend their framework and develop a mini- mally supervised method applicable to multi- ple languages. Our experiments show that our method is comparable to Hashimoto et al.’s for Japanese and outperforms previous unsu- pervised methods for English, Japanese, and Chinese, and that our method extracts 10,000 paraphrases with 92% precision for English, 82.5% precision for Japanese, and 82% preci- sion for Chinese. 1 Introduction Automatic paraphrasing has been recognized as an important component for NLP systems, and many methods have been proposed to acquire paraphrase knowledge (Lin and Pantel, 2001; Barzilay and McKeown, 2001; Shinyama et al., 2002; Barzilay and Lee, 2003; Dolan et al., 2004; Callison-Burch, 2008; Hashimoto et al., 2011; Fujita et al., 2012). We propose a minimally supervised method for multilingual paraphrase extraction. Hashimoto et al. (2011) developed a method to extract paraphrases from definition sentences on the Web, based on their observation that definition sentences defining the same concept tend to contain many paraphrases. Their method consists of two steps; they extract def- inition sentences from the Web, and extract phrasal (1) a. Paraphrasing is the use of your own words to express the au- thor’s ideas without changing the meaning. b. Paraphrasing is defined as a process of transforming an expres- sion into another while keeping its meaning intact. (2) a. 言い換えとは、ある表現をその意味内容を変えずに別の 表現に置き換えることを言います。 (Paraphrasing refers to the replacement of an expression into another without changing the semantic content.) b. 言 い 換 え と は 、 あ る 言 語 表 現 を で き る だ け意味や内容を保ったまま同 一 言 語 の 別 の 表 現 に 変 換する処理である。 (Paraphrasing is a process of trans- forming an expression into another of the same language while preserving the meaning and content as much as possible.) (3) a. 意译是指译者在不改变原文意思的前提下，完全改变原 文的句子结构。 (Paraphrasing refers to the transformation of sentence structure by the translator without changing the meaning of original text.) b. 意译是指只保持原文内容，不保持原文形式的翻译方法。 (Paraphrasing is a translation method of keeping the content of original text but not keeping the expression.) Figure 1: Multilingual definition pairs on “paraphrasing.” paraphrases from the definition sentences. Both steps require supervised classifiers trained by manu- ally annotated data, and heavily depend on their tar- get language. However, the basic idea is actually language-independent. Figure 1 gives examples of definition sentences …

Content truncated to 500 words