Comparing Information Extraction Pattern Models

Proceedings of the Workshop on Information Extraction Beyond The Document, pages 12–19, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Comparing Information Extraction Pattern Models Mark Stevenson and Mark A. Greenwood Department of Computer Science University of Sheffield Sheffield, S1 4DP, UK {marks,m.greenwood}@dcs.shef.ac.uk Abstract Several recently reported techniques for the automatic acquisition of Information Extraction (IE) systems have used depen- dency trees as the basis of their extrac- tion pattern representation. These ap- proaches have used a variety of pattern models (schemes for representing IE pat- terns based on particular parts of the de- pendency analysis). An appropriate model should be expressive enough to represent the information which is to be extracted from text without being overly compli- cated. Four previously reported pattern models are evaluated using existing IE evaluation corpora and three dependency parsers. It was found that one model, linked chains, could represent around 95% of the information of interest without gen- erating an unwieldy number of possible patterns. 1 Introduction A common approach to Information Extraction (IE) is to use patterns which match against text and identify items of interest. Patterns are applied to text which has undergone various levels of lin- guistic analysis, such as phrase chunking (Soder- land, 1999) and full syntactic parsing (Gaizauskas et al., 1996). The approaches use different defini- tions of what constitutes a valid pattern. For exam- ple, the AutoSlog system (Riloff, 1993) uses pat- terns which match certain grammatical categories, mainly nouns and verbs, in phrase chunked text while Yangarber et al. (2000) use subject-verb- object tuples derived from a dependency parse. An appropriate pattern language must encode enough information about the text to be able to accurately identify the items of interest. However, it should not contain so much information as to be complex and impractical to apply. Several recent approaches to IE have used pat- terns based on a dependency analysis of the input text (Yangarber, 2003; Sudo et al., 2001; Sudo et al., 2003; Bunescu and Mooney, 2005; Stevenson and Greenwood, 2005). These approaches have used a variety of pattern models (schemes for rep- resenting IE patterns based on particular parts of the dependency tree). For example, Yangarber (2003) uses just subject-verb-object tuples while Sudo et al. (2003) allow any subpart of the tree to act as an extraction pattern. The set of patterns al- lowed by the first model is a proper subset of the second and therefore captures less of the informa- tion contained in the dependency tree. Little anal- ysis has been carried out into the appropriateness of each model. Sudo et al. (2003) compared three models in terms of their ability to identify event participants. The choice of pattern model has an effect on the number of potential patterns. This has impli- cations on the practical application for each ap- proach, particularly when used for automatic ac- quisition of IE systems using learning methods (Yangarber et al., 2000; Sudo et al., 2003; Bunescu and Mooney, 2005). This paper evaluates the ap- propriateness of …

Content truncated to 500 words