Paraphrasing with Search Engine Query Logs

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1317–1325, Beijing, August 2010 Paraphrasing with Search Engine Query Logs Shiqi Zhao†‡, Haifeng Wang†, and Ting Liu‡ †Baidu Inc. ‡HIT Center for Information Retrieval, Harbin Institute of Technology {zhaoshiqi, wanghaifeng}@baidu.com, tliu@ir.hit.edu.cn Abstract This paper proposes a method that extracts paraphrases from search engine query logs. The method first extracts paraphrase query-title pairs based on an assumption that a search query and its correspond- ing clicked document titles may mean the same thing. It then extracts paraphrase query-query and title-title pairs from the query-title paraphrases with a pivot ap- proach. Paraphrases extracted in each step are validated with a binary classifier. We evaluate the method using a query log from Baidu1, a Chinese search engine. Experimental results show that the pro- posed method is effective, which extracts more than 3.5 million pairs of paraphrases with a precision of over 70%. The results also show that the extracted paraphrases can be used to generate high-quality para- phrase patterns. 1 Introduction The use of paraphrases is ubiquitous in hu- man languages, which also presents a challenge for natural language processing (NLP). Previous studies have shown that paraphrasing can play im- portant roles in plenty of areas, such as machine translation (MT) (Callison-Burch et al., 2006; Kauchak and Barzilay, 2006), question answer- ing (QA) (Duboue and Chu-Carroll, 2006; Riezler et al., 2007), natural language generation (NLG) (Iordanskaja et al., 1991), and so on. As a result, the research on paraphrasing and its applications have attracted significant interest. 1www.baidu.com This paper proposes a method that uses search engine query logs for extracting paraphrases, which is illustrated in Figure 1. Specifically, three kinds of paraphrases can be extracted with our method, which include (1) query-title (Q-T): a query and a document title that users clicked on; (2) query-query (Q-Q): two queries, for which users clicked on the same document title; (3) title- title (T-T): two titles that users clicked on for the same query. We train a classifier for each kind to filter incorrect pairs and refine the paraphrases. Extracting paraphrases using query logs has many advantages. First, query logs keep growing, which have no scale limitation. Second, query logs reflect web users’ real needs, hence the ex- tracted paraphrases may be more useful than that from other kinds of corpora. Third, paraphrases extracted from query logs can be directed applied in search engines for query suggestion and doc- ument reranking. In addition, we find that both queries and titles contain a good many question sentences, which can be useful in developing QA systems. We conduct experiments using a query log of a commercial Chinese search engine Baidu, from which we extracted about 2.7 million pairs of paraphrase Q-T, 0.4 million pairs of paraphrase Q- Q, and 0.4 million pairs of paraphrase T-T. The precision of the paraphrases is above 70%. In addition, we generate paraphrase patterns using the extracted paraphrases. The results show that 73,484 pairs of paraphrase patterns have been gen- erated, …

Content truncated to 500 words