Augmentable Paraphrase Extraction Framework

International Joint Conference on Natural Language Processing, pages 706–711, Nagoya, Japan, 14-18 October 2013. Augmentable Paraphrase Extraction Framework Mei-Hua Chen ac, Yi-Chun Chen b, Shih-Ting Huang b, Jason S. Chang b aInstitute of Information Systems and Applications and bDepartment of Computer Science, National Tsing Hua University, HsinChu, Taiwan, R.O.C. 30013 cDepartment of Foreign Languages and Literature, Hua Fan University, Taipei, Taiwan, R.O.C. 22301 {chen.meihua, pieyaaa, koromiko1104,jason.jschang}@gmail.com Abstract Paraphrase extraction relying on a single factor such as distribution similarity or translation similarity might lead to the loss of some linguistic properties. In this paper, we propose a paraphrase extraction framework, which accommodates various linguistically motivated factors to optimize the quality of paraphrase extraction. The major contributions of this study lie in the augmentable paraphrasing framework and the three kinds of factors conducive to both semantic and syntactic correctness. A manual evaluation showed that our model achieves more successful results than the state-of-the-art methods. 1. Introduction Paraphrasing provides an alternative way to express an idea using different words. Early work on paraphrase acquisition has been mainly based on either distributional similarity (e.g., Lin and Pantel, 2001) or the pivot-based approach (e.g., Bannard and Callison-Burch, 2005). Both methods have their strengths and limitations. Distributional similarity is capable of extracting syntactically correct paraphrases, but may risk including antonymous phrases as paraphrases. On the other hand, the pivot approach has the advantage of preserving semantic similarity among the generated paraphrases; however, the quality and quantity of the paraphrases closely correlates with the techniques of bilingual phrase alignment. Considering single factors, existing paraphrasing methods could lose some linguistic properties. In view of this, we attempt to differentiate the importance of the paraphrase candidates based on various factors. In this paper, we take a graphical view of the paraphrasing issue. To achieve the goal mentioned above, we adopt the Weighted PageRank Algorithm (Xing and Ghorbani, 2004). English phrases are treated as nodes. The edge weights are determined by various factors such as semantic similarity or syntactic similarity between nodes. It means that the performance of the ranked paraphrase candidates depends on the factors we selected and added. In other words, our framework is augmentable and is able to accommodate various factors to optimize the quality of paraphrase extraction. In this case, we propose three linguistically motivated factors to improve the performance of the paraphrase extraction. Lexical distributional similarity is used to ensure that the contexts in which the generated paraphrases appear are similar whereas syntactic distributional similarity is adopted for the purpose of maintaining the syntactic correctness. Translation similarity, one more factor, is capable of preserving semantic equivalence. These three selected factors adopted together effectively achieve better performance on paraphrase extraction. The evaluation shows that our model achieves more satisfactory results than the state-of-the-art pivot-based methods and graph-based methods. 2. Related Work Several approaches have been proposed to extract paraphrases. Earlier studies have focused on extracting paraphrases from monolingual corpora. Barzilay and Mckeown (2001) determine that the phrases in a monolingual parallel corpus are paraphrases of one …

Content truncated to 500 words