Application-driven Statistical Paraphrase Generation

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 834–842, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Application-driven Statistical Paraphrase Generation Shiqi Zhao, Xiang Lan, Ting Liu, Sheng Li Information Retrieval Lab, Harbin Institute of Technology 6F Aoxiao Building, No.27 Jiaohua Street, Nangang District Harbin, 150001, China {zhaosq,xlan,tliu,lisheng}@ir.hit.edu.cn Abstract Paraphrase generation (PG) is important in plenty of NLP applications. However, the research of PG is far from enough. In this paper, we propose a novel method for statistical paraphrase generation (SPG), which can (1) achieve various applications based on a uniform statistical model, and (2) naturally combine multiple resources to enhance the PG performance. In our experiments, we use the proposed method to generate paraphrases for three differ- ent applications. The results show that the method can be easily transformed from one application to another and generate valuable and interesting paraphrases. 1 Introduction Paraphrases are alternative ways that convey the same meaning. There are two main threads in the research of paraphrasing, i.e., paraphrase recogni- tion and paraphrase generation (PG). Paraphrase generation aims to generate a paraphrase for a source sentence in a certain application. PG shows its importance in many areas, such as question expansion in question answering (QA) (Duboue and Chu-Carroll, 2006), text polishing in natu- ral language generation (NLG) (Iordanskaja et al., 1991), text simplification in computer-aided read- ing (Carroll et al., 1999), and sentence similarity computation in the automatic evaluation of ma- chine translation (MT) (Kauchak and Barzilay, 2006) and summarization (Zhou et al., 2006). This paper presents a method for statistical paraphrase generation (SPG). As far as we know, this is the first statistical model specially designed for paraphrase generation. It’s distinguishing fea- ture is that it achieves various applications with a uniform model. In addition, it exploits multiple resources, including paraphrase phrases, patterns, and collocations, to resolve the data shortage prob- lem and generate more varied paraphrases. We consider three paraphrase applications in our experiments, including sentence compression, sentence simplification, and sentence similarity computation. The proposed method generates paraphrases for the input sentences in each appli- cation. The generated paraphrases are then man- ually scored based on adequacy, fluency, and us- ability. The results show that the proposed method is promising, which generates useful paraphrases for the given applications. In addition, comparison experiments show that our method outperforms a conventional SMT-based PG method. 2 Related Work Conventional methods for paraphrase generation can be classified as follows: Rule-based methods: Rule-based PG methods build on a set of paraphrase rules or patterns, which are either hand crafted or automatically collected. In the early rule-based PG research, the paraphrase rules are generally manually writ- ten (McKeown, 1979; Zong et al., 2001), which is expensive and arduous. Some researchers then tried to automatically extract paraphrase rules (Lin and Pantel, 2001; Barzilay and Lee, 2003; Zhao et al., 2008b), which facilitates the rule-based PG methods. However, it has been shown that the coverage of the paraphrase patterns is not …

Content truncated to 500 words