Paraphrase-Driven Learning for Open Question Answering

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1608–1618, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Paraphrase-Driven Learning for Open Question Answering Anthony Fader Luke Zettlemoyer Oren Etzioni Computer Science & Engineering University of Washington Seattle, WA 98195 {afader, lsz, etzioni}@cs.washington.edu Abstract We study question answering as a ma- chine learning problem, and induce a func- tion that maps open-domain questions to queries over a database of web extrac- tions. Given a large, community-authored, question-paraphrase corpus, we demon- strate that it is possible to learn a se- mantic lexicon and linear ranking func- tion without manually annotating ques- tions. Our approach automatically gener- alizes a seed lexicon and includes a scal- able, parallelized perceptron parameter es- timation scheme. Experiments show that our approach more than quadruples the re- call of the seed lexicon, with only an 8% loss in precision. 1 Introduction Open-domain question answering (QA) is a long- standing, unsolved problem. The central challenge is to automate every step of QA system construc- tion, including gathering large databases and an- swering questions against these databases. While there has been significant work on large-scale in- formation extraction (IE) from unstructured text (Banko et al., 2007; Hoffmann et al., 2010; Riedel et al., 2010), the problem of answering questions with the noisy knowledge bases that IE systems produce has received less attention. In this paper, we present an approach for learning to map ques- tions to formal queries over a large, open-domain database of extracted facts (Fader et al., 2011). Our system learns from a large, noisy, question- paraphrase corpus, where question clusters have a common but unknown query, and can span a diverse set of topics. Table 1 shows exam- ple paraphrase clusters for a set of factual ques- tions. Such data provides strong signal for learn- ing about lexical variation, but there are a number Who wrote the Winnie the Pooh books? Who is the author of winnie the pooh? What was the name of the authur of winnie the pooh? Who wrote the series of books for Winnie the poo? Who wrote the children’s storybook ‘Winnie the Pooh’? Who is poohs creator? What relieves a hangover? What is the best cure for a hangover? The best way to recover from a hangover? Best remedy for a hangover? What takes away a hangover? How do you lose a hangover? What helps hangover symptoms? What are social networking sites used for? Why do people use social networking sites worldwide? Advantages of using social network sites? Why do people use social networks a lot? Why do people communicate on social networking sites? What are the pros and cons of social networking sites? How do you say Santa Claus in Sweden? Say santa clause in sweden? How do you say santa clause in swedish? How do they say santa in Sweden? In Sweden what is santa called? Who is sweden santa? Table 1: Examples of paraphrase clusters from the WikiAnswers corpus. Within each cluster, …

Content truncated to 500 words