A Translation Model for Sentence Retrieval

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 684–691, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics A Translation Model for Sentence Retrieval Vanessa Murdock and W. Bruce Croft Center for Intelligent Information Retrieval Computer Science Department University of Massachusetts Amherst, MA 01003 {vanessa,croft}@cs.umass.edu Abstract In this work we propose a transla- tion model for monolingual sentence retrieval. We propose four methods for constructing a parallel corpus. Of the four methods proposed, a lexi- con learned from a bilingual Arabic- English corpus aligned at the sentence level performs best, significantly im- proving results over the query likeli- hood baseline. Further, we demon- strate that smoothing from the local context of the sentence improves re- trieval over the query likelihood base- line. 1 Introduction Sentence retrieval is the task of retrieving a rel- evant sentence in response to a user’s query. Tasks such as question answering, novelty de- tection and summarization often incorporate a sentence retrieval module. In previous work we examined sentence retrieval for question answer- ing (Murdock and Croft, 2004). This involves the comparison of two well-formed sentences, one a question, one a statement. In this work we compare well-formed sentences to queries, which can be typical keyword queries of 1 to 3 terms, or a set of sentences or sentence fragments. The TREC Novelty Track provides this type of data in the form of topic titles and descriptions, and sentence-level relevance judgments for a small subset of the collection. We present a translation model specifically for monolingual data, and show that it signif- icantly improves sentence retrieval over query- likelihood. Translation models train on a paral- lel corpus and in previous work we used a cor- pus of question/answer pairs. No such corpus is available for the novelty data, so in this pa- per we present four ways to construct a parallel corpus, to estimate a translation model. Many systems treat sentence retrieval as a type of document or passage retrieval. In our data a sentence is an average of 18 words, most of which occur once. A document is an average of 700 words, many of which are multiples of the same term. It is much less likely for a word and its synonym terms to appear in the same sentence than in the same document. Passages may be any length, either fixed or variable, but are somewhat arbitrarily desig- nated. Many systems that have a passage re- trieval module, on closer inspection have de- fined the passage to be a sentence. What is needed is a sentence retrieval mechanism that retains the benefits of passage retrieval, where a passage is longer than a sentence. We pro- pose that smoothing from the local context of the sentence improves retrieval over the query likelihood baseline, and the larger the context, the greater the improvement. We describe our translation model in sec- tion 2, along with our smoothing approach. In section 3 we discuss previous work in sentence retrieval for …

Content truncated to 500 words