Leveraging Multiple MT Engines for Paraphrase Generation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1326–1334, Beijing, August 2010 Leveraging Multiple MT Engines for Paraphrase Generation Shiqi Zhao†‡, Haifeng Wang†, Xiang Lan‡, and Ting Liu‡ †Baidu Inc. ‡HIT Center for Information Retrieval, Harbin Institute of Technology {zhaoshiqi, wanghaifeng}@baidu.com, {xlan, tliu}@ir.hit.edu.cn Abstract This paper proposes a method that lever- ages multiple machine translation (MT) engines for paraphrase generation (PG). The method includes two stages. Firstly, we use a multi-pivot approach to acquire a set of candidate paraphrases for a source sentence S. Then, we employ two kinds of techniques, namely the selection-based technique and the decoding-based tech- nique, to produce a best paraphrase T for S using the candidates acquired in the first stage. Experimental results show that: (1) The multi-pivot approach is effective for obtaining plenty of valuable candi- date paraphrases. (2) Both the selection- based and decoding-based techniques can make good use of the candidates and pro- duce high-quality paraphrases. Moreover, these two techniques are complementary. (3) The proposed method outperforms a state-of-the-art paraphrase generation ap- proach. 1 Introduction This paper addresses the problem of paraphrase generation (PG), which seeks to generate para- phrases for sentences. PG is important in many natural language processing (NLP) applications. For example, in machine translation (MT), a sentence can be paraphrased so as to make it more translatable (Zhang and Yamamoto, 2002; Callison-Burch et al., 2006). In question answer- ing (QA), a question can be paraphrased to im- prove the coverage of answer extraction (Duboue and Chu-Carroll, 2006; Riezler et al., 2007). In natural language generation (NLG), paraphrasing can help to increase the expressive power of the NLG systems (Iordanskaja et al., 1991). In this paper, we propose a novel PG method. For an English sentence S, the method first ac- quires a set of candidate paraphrases with a multi- pivot approach, which uses MT engines to auto- matically translate S into multiple pivot languages and then translate them back into English. Fur- thermore, the method employs two kinds of tech- niques to produce a best paraphrase T for S us- ing the candidates, i.e., the selection-based and decoding-based techniques. The former selects a best paraphrase from the candidates based on Minimum Bayes Risk (MBR), while the latter trains a MT model using the candidates and gen- erates paraphrases with a MT decoder. We evaluate our method on a set of 1182 En- glish sentences. The results show that: (1) al- though the candidate paraphrases acquired by MT engines are noisy, they provide good raw ma- terials for further paraphrase generation; (2) the selection-based technique is effective, which re- sults in the best performance; (3) the decoding- based technique is promising, which can generate paraphrases that are different from the candidates; (4) both the selection-based and decoding-based techniques outperform a state-of-the-art approach SPG (Zhao et al., 2009). 2 Related Work 2.1 Methods for Paraphrase Generation MT-based method is the mainstream method on PG. It regards PG as a monolingual machine trans- lation problem, i.e., “translating” a …

Content truncated to 500 words