Adapting a Lexicalized-Grammar Parser to Contrasting Domains

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 475–484, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Adapting a Lexicalized-Grammar Parser to Contrasting Domains Laura Rimell and Stephen Clark Oxford University Computing Laboratory Wolfson Building, Parks Road Oxford, OX1 3QD, UK {laura.rimell,stephen.clark}@comlab.ox.ac.uk Abstract Most state-of-the-art wide-coverage parsers are trained on newspaper text and suffer a loss of accuracy in other domains, making parser adaptation a pressing issue. In this paper we demonstrate that a CCG parser can be adapted to two new domains, biomedical text and questions for a QA system, by us- ing manually-annotated training data at the POS and lexical category levels only. This ap- proach achieves parser accuracy comparable to that on newspaper data without the need for annotated parse trees in the new domain. We find that retraining at the lexical category level yields a larger performance increase for questions than for biomedical text and analyze the two datasets to investigate why different domains might behave differently for parser adaptation. 1 Introduction Most state-of-the-art wide-coverage parsers are based on the Penn Treebank (Marcus et al., 1993), making such parsers highly tuned to newspaper text. A pressing question facing the parsing community is how to adapt these parsers to other domains, such as biomedical research papers and web pages. A re- lated question is how to improve the performance of these parsers on constructions that are rare in the Penn Treebank, such as questions. Questions are particularly important since a question parser is a component in most Question Answering (QA) sys- tems (Harabagiu et al., 2001). In this paper we investigate parser adaptation in the context of lexicalized grammars, by using a parser based on Combinatory Categorial Grammar (CCG) (Steedman, 2000). A key property of CCG is that it is lexicalized, meaning that each word in a sentence is associated with an elementary syntactic structure. In the case of CCG this is a lexical cate- gory expressing subcategorization information. We exploit this property of CCG by performing manual annotation in the new domain, but only up to this level of representation, where the annotation can be carried out relatively quickly. Since CCG lexical cat- egories are so expressive, many of the syntactic char- acteristics of a domain are captured at this level. The two domains we consider are the biomedical domain and questions for a QA system. We use the term “domain” somewhat loosely here, since ques- tions are best described as a particular set of syn- tactic constructions, rather than a set of documents about a particular topic. However, we consider ques- tion data to be interesting in the context of domain adaptation for the following reasons: 1) there are few examples in the Penn Treebank (PTB) and so PTB parsers typically perform poorly on them; 2) questions form a fairly homogeneous set with re- spect to the syntactic constructions employed, and it is an interesting question how easy it is to adapt a parser to such data; and 3) QA …

Content truncated to 500 words